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)];
687 #ifdef MDB_USE_SYSV_SEM
688 #define mti_semid mt1.mtb.mtb_semid
692 char mt2_wmname[MNAME_LEN];
693 #define mti_wmname mt2.mt2_wmname
695 pthread_mutex_t mt2_wmutex;
696 #define mti_wmutex mt2.mt2_wmutex
698 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
701 MDB_reader mti_readers[1];
704 /** Lockfile format signature: version, features and field layout */
705 #define MDB_LOCK_FORMAT \
707 ((MDB_LOCK_VERSION) \
708 /* Flags which describe functionality */ \
709 + (((MNAME_LEN) == 0) << 18) /* MDB_USE_SYSV_SEM */ \
710 + (((MDB_PIDLOCK) != 0) << 16)))
713 /** Common header for all page types.
714 * Overflow records occupy a number of contiguous pages with no
715 * headers on any page after the first.
717 typedef struct MDB_page {
718 #define mp_pgno mp_p.p_pgno
719 #define mp_next mp_p.p_next
721 pgno_t p_pgno; /**< page number */
722 struct MDB_page *p_next; /**< for in-memory list of freed pages */
725 /** @defgroup mdb_page Page Flags
727 * Flags for the page headers.
730 #define P_BRANCH 0x01 /**< branch page */
731 #define P_LEAF 0x02 /**< leaf page */
732 #define P_OVERFLOW 0x04 /**< overflow page */
733 #define P_META 0x08 /**< meta page */
734 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
735 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
736 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
737 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
738 #define P_KEEP 0x8000 /**< leave this page alone during spill */
740 uint16_t mp_flags; /**< @ref mdb_page */
741 #define mp_lower mp_pb.pb.pb_lower
742 #define mp_upper mp_pb.pb.pb_upper
743 #define mp_pages mp_pb.pb_pages
746 indx_t pb_lower; /**< lower bound of free space */
747 indx_t pb_upper; /**< upper bound of free space */
749 uint32_t pb_pages; /**< number of overflow pages */
751 indx_t mp_ptrs[1]; /**< dynamic size */
754 /** Size of the page header, excluding dynamic data at the end */
755 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
757 /** Address of first usable data byte in a page, after the header */
758 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
760 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
761 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
763 /** Number of nodes on a page */
764 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
766 /** The amount of space remaining in the page */
767 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
769 /** The percentage of space used in the page, in tenths of a percent. */
770 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
771 ((env)->me_psize - PAGEHDRSZ))
772 /** The minimum page fill factor, in tenths of a percent.
773 * Pages emptier than this are candidates for merging.
775 #define FILL_THRESHOLD 250
777 /** Test if a page is a leaf page */
778 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
779 /** Test if a page is a LEAF2 page */
780 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
781 /** Test if a page is a branch page */
782 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
783 /** Test if a page is an overflow page */
784 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
785 /** Test if a page is a sub page */
786 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
788 /** The number of overflow pages needed to store the given size. */
789 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
791 /** Link in #MDB_txn.%mt_loose_pgs list */
792 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
794 /** Header for a single key/data pair within a page.
795 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
796 * We guarantee 2-byte alignment for 'MDB_node's.
798 typedef struct MDB_node {
799 /** lo and hi are used for data size on leaf nodes and for
800 * child pgno on branch nodes. On 64 bit platforms, flags
801 * is also used for pgno. (Branch nodes have no flags).
802 * They are in host byte order in case that lets some
803 * accesses be optimized into a 32-bit word access.
805 #if BYTE_ORDER == LITTLE_ENDIAN
806 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
808 unsigned short mn_hi, mn_lo;
810 /** @defgroup mdb_node Node Flags
812 * Flags for node headers.
815 #define F_BIGDATA 0x01 /**< data put on overflow page */
816 #define F_SUBDATA 0x02 /**< data is a sub-database */
817 #define F_DUPDATA 0x04 /**< data has duplicates */
819 /** valid flags for #mdb_node_add() */
820 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
823 unsigned short mn_flags; /**< @ref mdb_node */
824 unsigned short mn_ksize; /**< key size */
825 char mn_data[1]; /**< key and data are appended here */
828 /** Size of the node header, excluding dynamic data at the end */
829 #define NODESIZE offsetof(MDB_node, mn_data)
831 /** Bit position of top word in page number, for shifting mn_flags */
832 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
834 /** Size of a node in a branch page with a given key.
835 * This is just the node header plus the key, there is no data.
837 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
839 /** Size of a node in a leaf page with a given key and data.
840 * This is node header plus key plus data size.
842 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
844 /** Address of node \b i in page \b p */
845 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
847 /** Address of the key for the node */
848 #define NODEKEY(node) (void *)((node)->mn_data)
850 /** Address of the data for a node */
851 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
853 /** Get the page number pointed to by a branch node */
854 #define NODEPGNO(node) \
855 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
856 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
857 /** Set the page number in a branch node */
858 #define SETPGNO(node,pgno) do { \
859 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
860 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
862 /** Get the size of the data in a leaf node */
863 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
864 /** Set the size of the data for a leaf node */
865 #define SETDSZ(node,size) do { \
866 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
867 /** The size of a key in a node */
868 #define NODEKSZ(node) ((node)->mn_ksize)
870 /** Copy a page number from src to dst */
872 #define COPY_PGNO(dst,src) dst = src
874 #if SIZE_MAX > 4294967295UL
875 #define COPY_PGNO(dst,src) do { \
876 unsigned short *s, *d; \
877 s = (unsigned short *)&(src); \
878 d = (unsigned short *)&(dst); \
885 #define COPY_PGNO(dst,src) do { \
886 unsigned short *s, *d; \
887 s = (unsigned short *)&(src); \
888 d = (unsigned short *)&(dst); \
894 /** The address of a key in a LEAF2 page.
895 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
896 * There are no node headers, keys are stored contiguously.
898 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
900 /** Set the \b node's key into \b keyptr, if requested. */
901 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
902 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
904 /** Set the \b node's key into \b key. */
905 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
907 /** Information about a single database in the environment. */
908 typedef struct MDB_db {
909 uint32_t md_pad; /**< also ksize for LEAF2 pages */
910 uint16_t md_flags; /**< @ref mdb_dbi_open */
911 uint16_t md_depth; /**< depth of this tree */
912 pgno_t md_branch_pages; /**< number of internal pages */
913 pgno_t md_leaf_pages; /**< number of leaf pages */
914 pgno_t md_overflow_pages; /**< number of overflow pages */
915 size_t md_entries; /**< number of data items */
916 pgno_t md_root; /**< the root page of this tree */
919 /** mdb_dbi_open flags */
920 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
921 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
922 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
923 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
925 /** Handle for the DB used to track free pages. */
927 /** Handle for the default DB. */
930 /** Meta page content.
931 * A meta page is the start point for accessing a database snapshot.
932 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
934 typedef struct MDB_meta {
935 /** Stamp identifying this as an LMDB file. It must be set
938 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
940 void *mm_address; /**< address for fixed mapping */
941 size_t mm_mapsize; /**< size of mmap region */
942 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
943 /** The size of pages used in this DB */
944 #define mm_psize mm_dbs[0].md_pad
945 /** Any persistent environment flags. @ref mdb_env */
946 #define mm_flags mm_dbs[0].md_flags
947 pgno_t mm_last_pg; /**< last used page in file */
948 volatile txnid_t mm_txnid; /**< txnid that committed this page */
951 /** Buffer for a stack-allocated meta page.
952 * The members define size and alignment, and silence type
953 * aliasing warnings. They are not used directly; that could
954 * mean incorrectly using several union members in parallel.
956 typedef union MDB_metabuf {
959 char mm_pad[PAGEHDRSZ];
964 /** Auxiliary DB info.
965 * The information here is mostly static/read-only. There is
966 * only a single copy of this record in the environment.
968 typedef struct MDB_dbx {
969 MDB_val md_name; /**< name of the database */
970 MDB_cmp_func *md_cmp; /**< function for comparing keys */
971 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
972 MDB_rel_func *md_rel; /**< user relocate function */
973 void *md_relctx; /**< user-provided context for md_rel */
976 /** A database transaction.
977 * Every operation requires a transaction handle.
980 MDB_txn *mt_parent; /**< parent of a nested txn */
981 MDB_txn *mt_child; /**< nested txn under this txn */
982 pgno_t mt_next_pgno; /**< next unallocated page */
983 /** The ID of this transaction. IDs are integers incrementing from 1.
984 * Only committed write transactions increment the ID. If a transaction
985 * aborts, the ID may be re-used by the next writer.
988 MDB_env *mt_env; /**< the DB environment */
989 /** The list of pages that became unused during this transaction.
992 /** The list of loose pages that became unused and may be reused
993 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
995 MDB_page *mt_loose_pgs;
996 /* #Number of loose pages (#mt_loose_pgs) */
998 /** The sorted list of dirty pages we temporarily wrote to disk
999 * because the dirty list was full. page numbers in here are
1000 * shifted left by 1, deleted slots have the LSB set.
1002 MDB_IDL mt_spill_pgs;
1004 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1005 MDB_ID2L dirty_list;
1006 /** For read txns: This thread/txn's reader table slot, or NULL. */
1009 /** Array of records for each DB known in the environment. */
1011 /** Array of MDB_db records for each known DB */
1013 /** Array of sequence numbers for each DB handle */
1014 unsigned int *mt_dbiseqs;
1015 /** @defgroup mt_dbflag Transaction DB Flags
1019 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1020 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1021 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1022 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1024 /** In write txns, array of cursors for each DB */
1025 MDB_cursor **mt_cursors;
1026 /** Array of flags for each DB */
1027 unsigned char *mt_dbflags;
1028 /** Number of DB records in use. This number only ever increments;
1029 * we don't decrement it when individual DB handles are closed.
1033 /** @defgroup mdb_txn Transaction Flags
1037 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1038 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1039 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1040 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1042 unsigned int mt_flags; /**< @ref mdb_txn */
1043 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1044 * Includes ancestor txns' dirty pages not hidden by other txns'
1045 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1046 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1048 unsigned int mt_dirty_room;
1051 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1052 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1053 * raise this on a 64 bit machine.
1055 #define CURSOR_STACK 32
1059 /** Cursors are used for all DB operations.
1060 * A cursor holds a path of (page pointer, key index) from the DB
1061 * root to a position in the DB, plus other state. #MDB_DUPSORT
1062 * cursors include an xcursor to the current data item. Write txns
1063 * track their cursors and keep them up to date when data moves.
1064 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1065 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1068 /** Next cursor on this DB in this txn */
1069 MDB_cursor *mc_next;
1070 /** Backup of the original cursor if this cursor is a shadow */
1071 MDB_cursor *mc_backup;
1072 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1073 struct MDB_xcursor *mc_xcursor;
1074 /** The transaction that owns this cursor */
1076 /** The database handle this cursor operates on */
1078 /** The database record for this cursor */
1080 /** The database auxiliary record for this cursor */
1082 /** The @ref mt_dbflag for this database */
1083 unsigned char *mc_dbflag;
1084 unsigned short mc_snum; /**< number of pushed pages */
1085 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1086 /** @defgroup mdb_cursor Cursor Flags
1088 * Cursor state flags.
1091 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1092 #define C_EOF 0x02 /**< No more data */
1093 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1094 #define C_DEL 0x08 /**< last op was a cursor_del */
1095 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1096 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1098 unsigned int mc_flags; /**< @ref mdb_cursor */
1099 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1100 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1103 /** Context for sorted-dup records.
1104 * We could have gone to a fully recursive design, with arbitrarily
1105 * deep nesting of sub-databases. But for now we only handle these
1106 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1108 typedef struct MDB_xcursor {
1109 /** A sub-cursor for traversing the Dup DB */
1110 MDB_cursor mx_cursor;
1111 /** The database record for this Dup DB */
1113 /** The auxiliary DB record for this Dup DB */
1115 /** The @ref mt_dbflag for this Dup DB */
1116 unsigned char mx_dbflag;
1119 /** State of FreeDB old pages, stored in the MDB_env */
1120 typedef struct MDB_pgstate {
1121 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1122 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1125 /** The database environment. */
1127 HANDLE me_fd; /**< The main data file */
1128 HANDLE me_lfd; /**< The lock file */
1129 HANDLE me_mfd; /**< just for writing the meta pages */
1130 /** Failed to update the meta page. Probably an I/O error. */
1131 #define MDB_FATAL_ERROR 0x80000000U
1132 /** Some fields are initialized. */
1133 #define MDB_ENV_ACTIVE 0x20000000U
1134 /** me_txkey is set */
1135 #define MDB_ENV_TXKEY 0x10000000U
1136 uint32_t me_flags; /**< @ref mdb_env */
1137 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1138 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1139 unsigned int me_maxreaders; /**< size of the reader table */
1140 unsigned int me_numreaders; /**< max numreaders set by this env */
1141 MDB_dbi me_numdbs; /**< number of DBs opened */
1142 MDB_dbi me_maxdbs; /**< size of the DB table */
1143 MDB_PID_T me_pid; /**< process ID of this env */
1144 char *me_path; /**< path to the DB files */
1145 char *me_map; /**< the memory map of the data file */
1146 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1147 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1148 void *me_pbuf; /**< scratch area for DUPSORT put() */
1149 MDB_txn *me_txn; /**< current write transaction */
1150 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1151 size_t me_mapsize; /**< size of the data memory map */
1152 off_t me_size; /**< current file size */
1153 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1154 MDB_dbx *me_dbxs; /**< array of static DB info */
1155 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1156 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1157 pthread_key_t me_txkey; /**< thread-key for readers */
1158 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1159 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1160 # define me_pglast me_pgstate.mf_pglast
1161 # define me_pghead me_pgstate.mf_pghead
1162 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1163 /** IDL of pages that became unused in a write txn */
1164 MDB_IDL me_free_pgs;
1165 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1166 MDB_ID2L me_dirty_list;
1167 /** Max number of freelist items that can fit in a single overflow page */
1169 /** Max size of a node on a page */
1170 unsigned int me_nodemax;
1171 #if !(MDB_MAXKEYSIZE)
1172 unsigned int me_maxkey; /**< max size of a key */
1174 int me_live_reader; /**< have liveness lock in reader table */
1176 int me_pidquery; /**< Used in OpenProcess */
1178 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM)
1179 /* Windows mutexes/SysV semaphores do not reside in shared mem */
1180 mdb_mutex_t me_rmutex;
1181 mdb_mutex_t me_wmutex;
1183 void *me_userctx; /**< User-settable context */
1184 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1187 /** Nested transaction */
1188 typedef struct MDB_ntxn {
1189 MDB_txn mnt_txn; /**< the transaction */
1190 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1193 /** max number of pages to commit in one writev() call */
1194 #define MDB_COMMIT_PAGES 64
1195 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1196 #undef MDB_COMMIT_PAGES
1197 #define MDB_COMMIT_PAGES IOV_MAX
1200 /** max bytes to write in one call */
1201 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1203 /** Check \b txn and \b dbi arguments to a function */
1204 #define TXN_DBI_EXIST(txn, dbi) \
1205 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1207 /** Check for misused \b dbi handles */
1208 #define TXN_DBI_CHANGED(txn, dbi) \
1209 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1211 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1212 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1213 static int mdb_page_touch(MDB_cursor *mc);
1215 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1216 static int mdb_page_search_root(MDB_cursor *mc,
1217 MDB_val *key, int modify);
1218 #define MDB_PS_MODIFY 1
1219 #define MDB_PS_ROOTONLY 2
1220 #define MDB_PS_FIRST 4
1221 #define MDB_PS_LAST 8
1222 static int mdb_page_search(MDB_cursor *mc,
1223 MDB_val *key, int flags);
1224 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1226 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1227 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1228 pgno_t newpgno, unsigned int nflags);
1230 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1231 static int mdb_env_pick_meta(const MDB_env *env);
1232 static int mdb_env_write_meta(MDB_txn *txn);
1233 #if !(defined(_WIN32) || defined(MDB_USE_SYSV_SEM)) /* Drop unused excl arg */
1234 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1236 static void mdb_env_close0(MDB_env *env, int excl);
1238 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1239 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1240 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1241 static void mdb_node_del(MDB_cursor *mc, int ksize);
1242 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1243 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1244 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1245 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1246 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1248 static int mdb_rebalance(MDB_cursor *mc);
1249 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1251 static void mdb_cursor_pop(MDB_cursor *mc);
1252 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1254 static int mdb_cursor_del0(MDB_cursor *mc);
1255 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1256 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1257 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1258 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1259 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1261 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1262 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1264 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1265 static void mdb_xcursor_init0(MDB_cursor *mc);
1266 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1268 static int mdb_drop0(MDB_cursor *mc, int subs);
1269 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1270 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1273 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1277 static SECURITY_DESCRIPTOR mdb_null_sd;
1278 static SECURITY_ATTRIBUTES mdb_all_sa;
1279 static int mdb_sec_inited;
1282 /** Return the library version info. */
1284 mdb_version(int *major, int *minor, int *patch)
1286 if (major) *major = MDB_VERSION_MAJOR;
1287 if (minor) *minor = MDB_VERSION_MINOR;
1288 if (patch) *patch = MDB_VERSION_PATCH;
1289 return MDB_VERSION_STRING;
1292 /** Table of descriptions for LMDB @ref errors */
1293 static char *const mdb_errstr[] = {
1294 "MDB_KEYEXIST: Key/data pair already exists",
1295 "MDB_NOTFOUND: No matching key/data pair found",
1296 "MDB_PAGE_NOTFOUND: Requested page not found",
1297 "MDB_CORRUPTED: Located page was wrong type",
1298 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1299 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1300 "MDB_INVALID: File is not an LMDB file",
1301 "MDB_MAP_FULL: Environment mapsize limit reached",
1302 "MDB_DBS_FULL: Environment maxdbs limit reached",
1303 "MDB_READERS_FULL: Environment maxreaders limit reached",
1304 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1305 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1306 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1307 "MDB_PAGE_FULL: Internal error - page has no more space",
1308 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1309 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1310 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1311 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1312 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1313 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1317 mdb_strerror(int err)
1320 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1321 * This works as long as no function between the call to mdb_strerror
1322 * and the actual use of the message uses more than 4K of stack.
1325 char buf[1024], *ptr = buf;
1329 return ("Successful return: 0");
1331 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1332 i = err - MDB_KEYEXIST;
1333 return mdb_errstr[i];
1337 /* These are the C-runtime error codes we use. The comment indicates
1338 * their numeric value, and the Win32 error they would correspond to
1339 * if the error actually came from a Win32 API. A major mess, we should
1340 * have used LMDB-specific error codes for everything.
1343 case ENOENT: /* 2, FILE_NOT_FOUND */
1344 case EIO: /* 5, ACCESS_DENIED */
1345 case ENOMEM: /* 12, INVALID_ACCESS */
1346 case EACCES: /* 13, INVALID_DATA */
1347 case EBUSY: /* 16, CURRENT_DIRECTORY */
1348 case EINVAL: /* 22, BAD_COMMAND */
1349 case ENOSPC: /* 28, OUT_OF_PAPER */
1350 return strerror(err);
1355 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1356 FORMAT_MESSAGE_IGNORE_INSERTS,
1357 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1360 return strerror(err);
1364 /** assert(3) variant in cursor context */
1365 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1366 /** assert(3) variant in transaction context */
1367 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1368 /** assert(3) variant in environment context */
1369 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1372 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1373 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1376 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1377 const char *func, const char *file, int line)
1380 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1381 file, line, expr_txt, func);
1382 if (env->me_assert_func)
1383 env->me_assert_func(env, buf);
1384 fprintf(stderr, "%s\n", buf);
1388 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1392 /** Return the page number of \b mp which may be sub-page, for debug output */
1394 mdb_dbg_pgno(MDB_page *mp)
1397 COPY_PGNO(ret, mp->mp_pgno);
1401 /** Display a key in hexadecimal and return the address of the result.
1402 * @param[in] key the key to display
1403 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1404 * @return The key in hexadecimal form.
1407 mdb_dkey(MDB_val *key, char *buf)
1410 unsigned char *c = key->mv_data;
1416 if (key->mv_size > DKBUF_MAXKEYSIZE)
1417 return "MDB_MAXKEYSIZE";
1418 /* may want to make this a dynamic check: if the key is mostly
1419 * printable characters, print it as-is instead of converting to hex.
1423 for (i=0; i<key->mv_size; i++)
1424 ptr += sprintf(ptr, "%02x", *c++);
1426 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1432 mdb_leafnode_type(MDB_node *n)
1434 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1435 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1436 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1439 /** Display all the keys in the page. */
1441 mdb_page_list(MDB_page *mp)
1443 pgno_t pgno = mdb_dbg_pgno(mp);
1444 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1446 unsigned int i, nkeys, nsize, total = 0;
1450 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1451 case P_BRANCH: type = "Branch page"; break;
1452 case P_LEAF: type = "Leaf page"; break;
1453 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1454 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1455 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1457 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1458 pgno, mp->mp_pages, state);
1461 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1462 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1465 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1469 nkeys = NUMKEYS(mp);
1470 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1472 for (i=0; i<nkeys; i++) {
1473 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1474 key.mv_size = nsize = mp->mp_pad;
1475 key.mv_data = LEAF2KEY(mp, i, nsize);
1477 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1480 node = NODEPTR(mp, i);
1481 key.mv_size = node->mn_ksize;
1482 key.mv_data = node->mn_data;
1483 nsize = NODESIZE + key.mv_size;
1484 if (IS_BRANCH(mp)) {
1485 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1489 if (F_ISSET(node->mn_flags, F_BIGDATA))
1490 nsize += sizeof(pgno_t);
1492 nsize += NODEDSZ(node);
1494 nsize += sizeof(indx_t);
1495 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1496 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1498 total = EVEN(total);
1500 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1501 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1505 mdb_cursor_chk(MDB_cursor *mc)
1511 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1512 for (i=0; i<mc->mc_top; i++) {
1514 node = NODEPTR(mp, mc->mc_ki[i]);
1515 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1518 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1524 /** Count all the pages in each DB and in the freelist
1525 * and make sure it matches the actual number of pages
1527 * All named DBs must be open for a correct count.
1529 static void mdb_audit(MDB_txn *txn)
1533 MDB_ID freecount, count;
1538 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1539 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1540 freecount += *(MDB_ID *)data.mv_data;
1541 mdb_tassert(txn, rc == MDB_NOTFOUND);
1544 for (i = 0; i<txn->mt_numdbs; i++) {
1546 if (!(txn->mt_dbflags[i] & DB_VALID))
1548 mdb_cursor_init(&mc, txn, i, &mx);
1549 if (txn->mt_dbs[i].md_root == P_INVALID)
1551 count += txn->mt_dbs[i].md_branch_pages +
1552 txn->mt_dbs[i].md_leaf_pages +
1553 txn->mt_dbs[i].md_overflow_pages;
1554 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1555 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1556 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1559 mp = mc.mc_pg[mc.mc_top];
1560 for (j=0; j<NUMKEYS(mp); j++) {
1561 MDB_node *leaf = NODEPTR(mp, j);
1562 if (leaf->mn_flags & F_SUBDATA) {
1564 memcpy(&db, NODEDATA(leaf), sizeof(db));
1565 count += db.md_branch_pages + db.md_leaf_pages +
1566 db.md_overflow_pages;
1570 mdb_tassert(txn, rc == MDB_NOTFOUND);
1573 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1574 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1575 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1581 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1583 return txn->mt_dbxs[dbi].md_cmp(a, b);
1587 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1589 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1592 /** Allocate memory for a page.
1593 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1596 mdb_page_malloc(MDB_txn *txn, unsigned num)
1598 MDB_env *env = txn->mt_env;
1599 MDB_page *ret = env->me_dpages;
1600 size_t psize = env->me_psize, sz = psize, off;
1601 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1602 * For a single page alloc, we init everything after the page header.
1603 * For multi-page, we init the final page; if the caller needed that
1604 * many pages they will be filling in at least up to the last page.
1608 VGMEMP_ALLOC(env, ret, sz);
1609 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1610 env->me_dpages = ret->mp_next;
1613 psize -= off = PAGEHDRSZ;
1618 if ((ret = malloc(sz)) != NULL) {
1619 VGMEMP_ALLOC(env, ret, sz);
1620 if (!(env->me_flags & MDB_NOMEMINIT)) {
1621 memset((char *)ret + off, 0, psize);
1625 txn->mt_flags |= MDB_TXN_ERROR;
1629 /** Free a single page.
1630 * Saves single pages to a list, for future reuse.
1631 * (This is not used for multi-page overflow pages.)
1634 mdb_page_free(MDB_env *env, MDB_page *mp)
1636 mp->mp_next = env->me_dpages;
1637 VGMEMP_FREE(env, mp);
1638 env->me_dpages = mp;
1641 /** Free a dirty page */
1643 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1645 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1646 mdb_page_free(env, dp);
1648 /* large pages just get freed directly */
1649 VGMEMP_FREE(env, dp);
1654 /** Return all dirty pages to dpage list */
1656 mdb_dlist_free(MDB_txn *txn)
1658 MDB_env *env = txn->mt_env;
1659 MDB_ID2L dl = txn->mt_u.dirty_list;
1660 unsigned i, n = dl[0].mid;
1662 for (i = 1; i <= n; i++) {
1663 mdb_dpage_free(env, dl[i].mptr);
1668 /** Loosen or free a single page.
1669 * Saves single pages to a list for future reuse
1670 * in this same txn. It has been pulled from the freeDB
1671 * and already resides on the dirty list, but has been
1672 * deleted. Use these pages first before pulling again
1675 * If the page wasn't dirtied in this txn, just add it
1676 * to this txn's free list.
1679 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1682 pgno_t pgno = mp->mp_pgno;
1683 MDB_txn *txn = mc->mc_txn;
1685 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1686 if (txn->mt_parent) {
1687 MDB_ID2 *dl = txn->mt_u.dirty_list;
1688 /* If txn has a parent, make sure the page is in our
1692 unsigned x = mdb_mid2l_search(dl, pgno);
1693 if (x <= dl[0].mid && dl[x].mid == pgno) {
1694 if (mp != dl[x].mptr) { /* bad cursor? */
1695 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1696 txn->mt_flags |= MDB_TXN_ERROR;
1697 return MDB_CORRUPTED;
1704 /* no parent txn, so it's just ours */
1709 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1711 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1712 txn->mt_loose_pgs = mp;
1713 txn->mt_loose_count++;
1714 mp->mp_flags |= P_LOOSE;
1716 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1724 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1725 * @param[in] mc A cursor handle for the current operation.
1726 * @param[in] pflags Flags of the pages to update:
1727 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1728 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1729 * @return 0 on success, non-zero on failure.
1732 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1734 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1735 MDB_txn *txn = mc->mc_txn;
1741 int rc = MDB_SUCCESS, level;
1743 /* Mark pages seen by cursors */
1744 if (mc->mc_flags & C_UNTRACK)
1745 mc = NULL; /* will find mc in mt_cursors */
1746 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1747 for (; mc; mc=mc->mc_next) {
1748 if (!(mc->mc_flags & C_INITIALIZED))
1750 for (m3 = mc;; m3 = &mx->mx_cursor) {
1752 for (j=0; j<m3->mc_snum; j++) {
1754 if ((mp->mp_flags & Mask) == pflags)
1755 mp->mp_flags ^= P_KEEP;
1757 mx = m3->mc_xcursor;
1758 /* Proceed to mx if it is at a sub-database */
1759 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1761 if (! (mp && (mp->mp_flags & P_LEAF)))
1763 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1764 if (!(leaf->mn_flags & F_SUBDATA))
1773 /* Mark dirty root pages */
1774 for (i=0; i<txn->mt_numdbs; i++) {
1775 if (txn->mt_dbflags[i] & DB_DIRTY) {
1776 pgno_t pgno = txn->mt_dbs[i].md_root;
1777 if (pgno == P_INVALID)
1779 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1781 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1782 dp->mp_flags ^= P_KEEP;
1790 static int mdb_page_flush(MDB_txn *txn, int keep);
1792 /** Spill pages from the dirty list back to disk.
1793 * This is intended to prevent running into #MDB_TXN_FULL situations,
1794 * but note that they may still occur in a few cases:
1795 * 1) our estimate of the txn size could be too small. Currently this
1796 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1797 * 2) child txns may run out of space if their parents dirtied a
1798 * lot of pages and never spilled them. TODO: we probably should do
1799 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1800 * the parent's dirty_room is below a given threshold.
1802 * Otherwise, if not using nested txns, it is expected that apps will
1803 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1804 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1805 * If the txn never references them again, they can be left alone.
1806 * If the txn only reads them, they can be used without any fuss.
1807 * If the txn writes them again, they can be dirtied immediately without
1808 * going thru all of the work of #mdb_page_touch(). Such references are
1809 * handled by #mdb_page_unspill().
1811 * Also note, we never spill DB root pages, nor pages of active cursors,
1812 * because we'll need these back again soon anyway. And in nested txns,
1813 * we can't spill a page in a child txn if it was already spilled in a
1814 * parent txn. That would alter the parent txns' data even though
1815 * the child hasn't committed yet, and we'd have no way to undo it if
1816 * the child aborted.
1818 * @param[in] m0 cursor A cursor handle identifying the transaction and
1819 * database for which we are checking space.
1820 * @param[in] key For a put operation, the key being stored.
1821 * @param[in] data For a put operation, the data being stored.
1822 * @return 0 on success, non-zero on failure.
1825 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1827 MDB_txn *txn = m0->mc_txn;
1829 MDB_ID2L dl = txn->mt_u.dirty_list;
1830 unsigned int i, j, need;
1833 if (m0->mc_flags & C_SUB)
1836 /* Estimate how much space this op will take */
1837 i = m0->mc_db->md_depth;
1838 /* Named DBs also dirty the main DB */
1839 if (m0->mc_dbi > MAIN_DBI)
1840 i += txn->mt_dbs[MAIN_DBI].md_depth;
1841 /* For puts, roughly factor in the key+data size */
1843 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1844 i += i; /* double it for good measure */
1847 if (txn->mt_dirty_room > i)
1850 if (!txn->mt_spill_pgs) {
1851 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1852 if (!txn->mt_spill_pgs)
1855 /* purge deleted slots */
1856 MDB_IDL sl = txn->mt_spill_pgs;
1857 unsigned int num = sl[0];
1859 for (i=1; i<=num; i++) {
1866 /* Preserve pages which may soon be dirtied again */
1867 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1870 /* Less aggressive spill - we originally spilled the entire dirty list,
1871 * with a few exceptions for cursor pages and DB root pages. But this
1872 * turns out to be a lot of wasted effort because in a large txn many
1873 * of those pages will need to be used again. So now we spill only 1/8th
1874 * of the dirty pages. Testing revealed this to be a good tradeoff,
1875 * better than 1/2, 1/4, or 1/10.
1877 if (need < MDB_IDL_UM_MAX / 8)
1878 need = MDB_IDL_UM_MAX / 8;
1880 /* Save the page IDs of all the pages we're flushing */
1881 /* flush from the tail forward, this saves a lot of shifting later on. */
1882 for (i=dl[0].mid; i && need; i--) {
1883 MDB_ID pn = dl[i].mid << 1;
1885 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1887 /* Can't spill twice, make sure it's not already in a parent's
1890 if (txn->mt_parent) {
1892 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1893 if (tx2->mt_spill_pgs) {
1894 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1895 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1896 dp->mp_flags |= P_KEEP;
1904 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1908 mdb_midl_sort(txn->mt_spill_pgs);
1910 /* Flush the spilled part of dirty list */
1911 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1914 /* Reset any dirty pages we kept that page_flush didn't see */
1915 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1918 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1922 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1924 mdb_find_oldest(MDB_txn *txn)
1927 txnid_t mr, oldest = txn->mt_txnid - 1;
1928 if (txn->mt_env->me_txns) {
1929 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1930 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1941 /** Add a page to the txn's dirty list */
1943 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1946 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1948 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1949 insert = mdb_mid2l_append;
1951 insert = mdb_mid2l_insert;
1953 mid.mid = mp->mp_pgno;
1955 rc = insert(txn->mt_u.dirty_list, &mid);
1956 mdb_tassert(txn, rc == 0);
1957 txn->mt_dirty_room--;
1960 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1961 * me_pghead and mt_next_pgno.
1963 * If there are free pages available from older transactions, they
1964 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1965 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1966 * and move me_pglast to say which records were consumed. Only this
1967 * function can create me_pghead and move me_pglast/mt_next_pgno.
1968 * @param[in] mc cursor A cursor handle identifying the transaction and
1969 * database for which we are allocating.
1970 * @param[in] num the number of pages to allocate.
1971 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1972 * will always be satisfied by a single contiguous chunk of memory.
1973 * @return 0 on success, non-zero on failure.
1976 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1978 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1979 /* Get at most <Max_retries> more freeDB records once me_pghead
1980 * has enough pages. If not enough, use new pages from the map.
1981 * If <Paranoid> and mc is updating the freeDB, only get new
1982 * records if me_pghead is empty. Then the freelist cannot play
1983 * catch-up with itself by growing while trying to save it.
1985 enum { Paranoid = 1, Max_retries = 500 };
1987 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1989 int rc, retry = num * 60;
1990 MDB_txn *txn = mc->mc_txn;
1991 MDB_env *env = txn->mt_env;
1992 pgno_t pgno, *mop = env->me_pghead;
1993 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1995 txnid_t oldest = 0, last;
2000 /* If there are any loose pages, just use them */
2001 if (num == 1 && txn->mt_loose_pgs) {
2002 np = txn->mt_loose_pgs;
2003 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2004 txn->mt_loose_count--;
2005 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2013 /* If our dirty list is already full, we can't do anything */
2014 if (txn->mt_dirty_room == 0) {
2019 for (op = MDB_FIRST;; op = MDB_NEXT) {
2024 /* Seek a big enough contiguous page range. Prefer
2025 * pages at the tail, just truncating the list.
2031 if (mop[i-n2] == pgno+n2)
2038 if (op == MDB_FIRST) { /* 1st iteration */
2039 /* Prepare to fetch more and coalesce */
2040 last = env->me_pglast;
2041 oldest = env->me_pgoldest;
2042 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2045 key.mv_data = &last; /* will look up last+1 */
2046 key.mv_size = sizeof(last);
2048 if (Paranoid && mc->mc_dbi == FREE_DBI)
2051 if (Paranoid && retry < 0 && mop_len)
2055 /* Do not fetch more if the record will be too recent */
2056 if (oldest <= last) {
2058 oldest = mdb_find_oldest(txn);
2059 env->me_pgoldest = oldest;
2065 rc = mdb_cursor_get(&m2, &key, NULL, op);
2067 if (rc == MDB_NOTFOUND)
2071 last = *(txnid_t*)key.mv_data;
2072 if (oldest <= last) {
2074 oldest = mdb_find_oldest(txn);
2075 env->me_pgoldest = oldest;
2081 np = m2.mc_pg[m2.mc_top];
2082 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2083 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2086 idl = (MDB_ID *) data.mv_data;
2089 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2094 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2096 mop = env->me_pghead;
2098 env->me_pglast = last;
2100 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2101 last, txn->mt_dbs[FREE_DBI].md_root, i));
2103 DPRINTF(("IDL %"Z"u", idl[j]));
2105 /* Merge in descending sorted order */
2106 mdb_midl_xmerge(mop, idl);
2110 /* Use new pages from the map when nothing suitable in the freeDB */
2112 pgno = txn->mt_next_pgno;
2113 if (pgno + num >= env->me_maxpg) {
2114 DPUTS("DB size maxed out");
2120 if (env->me_flags & MDB_WRITEMAP) {
2121 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2123 if (!(np = mdb_page_malloc(txn, num))) {
2129 mop[0] = mop_len -= num;
2130 /* Move any stragglers down */
2131 for (j = i-num; j < mop_len; )
2132 mop[++j] = mop[++i];
2134 txn->mt_next_pgno = pgno + num;
2137 mdb_page_dirty(txn, np);
2143 txn->mt_flags |= MDB_TXN_ERROR;
2147 /** Copy the used portions of a non-overflow page.
2148 * @param[in] dst page to copy into
2149 * @param[in] src page to copy from
2150 * @param[in] psize size of a page
2153 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2155 enum { Align = sizeof(pgno_t) };
2156 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2158 /* If page isn't full, just copy the used portion. Adjust
2159 * alignment so memcpy may copy words instead of bytes.
2161 if ((unused &= -Align) && !IS_LEAF2(src)) {
2162 upper = (upper + PAGEBASE) & -Align;
2163 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2164 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2167 memcpy(dst, src, psize - unused);
2171 /** Pull a page off the txn's spill list, if present.
2172 * If a page being referenced was spilled to disk in this txn, bring
2173 * it back and make it dirty/writable again.
2174 * @param[in] txn the transaction handle.
2175 * @param[in] mp the page being referenced. It must not be dirty.
2176 * @param[out] ret the writable page, if any. ret is unchanged if
2177 * mp wasn't spilled.
2180 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2182 MDB_env *env = txn->mt_env;
2185 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2187 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2188 if (!tx2->mt_spill_pgs)
2190 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2191 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2194 if (txn->mt_dirty_room == 0)
2195 return MDB_TXN_FULL;
2196 if (IS_OVERFLOW(mp))
2200 if (env->me_flags & MDB_WRITEMAP) {
2203 np = mdb_page_malloc(txn, num);
2207 memcpy(np, mp, num * env->me_psize);
2209 mdb_page_copy(np, mp, env->me_psize);
2212 /* If in current txn, this page is no longer spilled.
2213 * If it happens to be the last page, truncate the spill list.
2214 * Otherwise mark it as deleted by setting the LSB.
2216 if (x == txn->mt_spill_pgs[0])
2217 txn->mt_spill_pgs[0]--;
2219 txn->mt_spill_pgs[x] |= 1;
2220 } /* otherwise, if belonging to a parent txn, the
2221 * page remains spilled until child commits
2224 mdb_page_dirty(txn, np);
2225 np->mp_flags |= P_DIRTY;
2233 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2234 * @param[in] mc cursor pointing to the page to be touched
2235 * @return 0 on success, non-zero on failure.
2238 mdb_page_touch(MDB_cursor *mc)
2240 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2241 MDB_txn *txn = mc->mc_txn;
2242 MDB_cursor *m2, *m3;
2246 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2247 if (txn->mt_flags & MDB_TXN_SPILLS) {
2249 rc = mdb_page_unspill(txn, mp, &np);
2255 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2256 (rc = mdb_page_alloc(mc, 1, &np)))
2259 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2260 mp->mp_pgno, pgno));
2261 mdb_cassert(mc, mp->mp_pgno != pgno);
2262 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2263 /* Update the parent page, if any, to point to the new page */
2265 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2266 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2267 SETPGNO(node, pgno);
2269 mc->mc_db->md_root = pgno;
2271 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2272 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2274 /* If txn has a parent, make sure the page is in our
2278 unsigned x = mdb_mid2l_search(dl, pgno);
2279 if (x <= dl[0].mid && dl[x].mid == pgno) {
2280 if (mp != dl[x].mptr) { /* bad cursor? */
2281 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2282 txn->mt_flags |= MDB_TXN_ERROR;
2283 return MDB_CORRUPTED;
2288 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2290 np = mdb_page_malloc(txn, 1);
2295 rc = mdb_mid2l_insert(dl, &mid);
2296 mdb_cassert(mc, rc == 0);
2301 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2303 np->mp_flags |= P_DIRTY;
2306 /* Adjust cursors pointing to mp */
2307 mc->mc_pg[mc->mc_top] = np;
2308 m2 = txn->mt_cursors[mc->mc_dbi];
2309 if (mc->mc_flags & C_SUB) {
2310 for (; m2; m2=m2->mc_next) {
2311 m3 = &m2->mc_xcursor->mx_cursor;
2312 if (m3->mc_snum < mc->mc_snum) continue;
2313 if (m3->mc_pg[mc->mc_top] == mp)
2314 m3->mc_pg[mc->mc_top] = np;
2317 for (; m2; m2=m2->mc_next) {
2318 if (m2->mc_snum < mc->mc_snum) continue;
2319 if (m2->mc_pg[mc->mc_top] == mp) {
2320 m2->mc_pg[mc->mc_top] = np;
2321 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2323 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2325 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2326 if (!(leaf->mn_flags & F_SUBDATA))
2327 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2335 txn->mt_flags |= MDB_TXN_ERROR;
2340 mdb_env_sync(MDB_env *env, int force)
2343 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2344 if (env->me_flags & MDB_WRITEMAP) {
2345 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2346 ? MS_ASYNC : MS_SYNC;
2347 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2350 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2354 if (MDB_FDATASYNC(env->me_fd))
2361 /** Back up parent txn's cursors, then grab the originals for tracking */
2363 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2365 MDB_cursor *mc, *bk;
2370 for (i = src->mt_numdbs; --i >= 0; ) {
2371 if ((mc = src->mt_cursors[i]) != NULL) {
2372 size = sizeof(MDB_cursor);
2374 size += sizeof(MDB_xcursor);
2375 for (; mc; mc = bk->mc_next) {
2381 mc->mc_db = &dst->mt_dbs[i];
2382 /* Kill pointers into src - and dst to reduce abuse: The
2383 * user may not use mc until dst ends. Otherwise we'd...
2385 mc->mc_txn = NULL; /* ...set this to dst */
2386 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2387 if ((mx = mc->mc_xcursor) != NULL) {
2388 *(MDB_xcursor *)(bk+1) = *mx;
2389 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2391 mc->mc_next = dst->mt_cursors[i];
2392 dst->mt_cursors[i] = mc;
2399 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2400 * @param[in] txn the transaction handle.
2401 * @param[in] merge true to keep changes to parent cursors, false to revert.
2402 * @return 0 on success, non-zero on failure.
2405 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2407 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2411 for (i = txn->mt_numdbs; --i >= 0; ) {
2412 for (mc = cursors[i]; mc; mc = next) {
2414 if ((bk = mc->mc_backup) != NULL) {
2416 /* Commit changes to parent txn */
2417 mc->mc_next = bk->mc_next;
2418 mc->mc_backup = bk->mc_backup;
2419 mc->mc_txn = bk->mc_txn;
2420 mc->mc_db = bk->mc_db;
2421 mc->mc_dbflag = bk->mc_dbflag;
2422 if ((mx = mc->mc_xcursor) != NULL)
2423 mx->mx_cursor.mc_txn = bk->mc_txn;
2425 /* Abort nested txn */
2427 if ((mx = mc->mc_xcursor) != NULL)
2428 *mx = *(MDB_xcursor *)(bk+1);
2432 /* Only malloced cursors are permanently tracked. */
2440 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2443 mdb_txn_reset0(MDB_txn *txn, const char *act);
2445 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2451 Pidset = F_SETLK, Pidcheck = F_GETLK
2455 /** Set or check a pid lock. Set returns 0 on success.
2456 * Check returns 0 if the process is certainly dead, nonzero if it may
2457 * be alive (the lock exists or an error happened so we do not know).
2459 * On Windows Pidset is a no-op, we merely check for the existence
2460 * of the process with the given pid. On POSIX we use a single byte
2461 * lock on the lockfile, set at an offset equal to the pid.
2464 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2466 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2469 if (op == Pidcheck) {
2470 h = OpenProcess(env->me_pidquery, FALSE, pid);
2471 /* No documented "no such process" code, but other program use this: */
2473 return ErrCode() != ERROR_INVALID_PARAMETER;
2474 /* A process exists until all handles to it close. Has it exited? */
2475 ret = WaitForSingleObject(h, 0) != 0;
2482 struct flock lock_info;
2483 memset(&lock_info, 0, sizeof(lock_info));
2484 lock_info.l_type = F_WRLCK;
2485 lock_info.l_whence = SEEK_SET;
2486 lock_info.l_start = pid;
2487 lock_info.l_len = 1;
2488 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2489 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2491 } else if ((rc = ErrCode()) == EINTR) {
2499 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2500 * @param[in] txn the transaction handle to initialize
2501 * @return 0 on success, non-zero on failure.
2504 mdb_txn_renew0(MDB_txn *txn)
2506 MDB_env *env = txn->mt_env;
2507 MDB_txninfo *ti = env->me_txns;
2511 int rc, new_notls = 0;
2513 if (txn->mt_flags & MDB_TXN_RDONLY) {
2515 txn->mt_numdbs = env->me_numdbs;
2516 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2518 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2519 txn->mt_txnid = meta->mm_txnid;
2520 txn->mt_u.reader = NULL;
2522 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2523 pthread_getspecific(env->me_txkey);
2525 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2526 return MDB_BAD_RSLOT;
2528 MDB_PID_T pid = env->me_pid;
2529 MDB_THR_T tid = pthread_self();
2530 mdb_mutex_t *rmutex = MDB_MUTEX(env, r);
2532 if (!env->me_live_reader) {
2533 rc = mdb_reader_pid(env, Pidset, pid);
2536 env->me_live_reader = 1;
2539 if (LOCK_MUTEX(rc, env, rmutex))
2541 nr = ti->mti_numreaders;
2542 for (i=0; i<nr; i++)
2543 if (ti->mti_readers[i].mr_pid == 0)
2545 if (i == env->me_maxreaders) {
2546 UNLOCK_MUTEX(rmutex);
2547 return MDB_READERS_FULL;
2549 r = &ti->mti_readers[i];
2550 r->mr_txnid = (txnid_t)-1;
2552 r->mr_pid = pid; /* should be written last, see ITS#7971. */
2554 ti->mti_numreaders = ++nr;
2555 /* Save numreaders for un-mutexed mdb_env_close() */
2556 env->me_numreaders = nr;
2557 UNLOCK_MUTEX(rmutex);
2559 new_notls = (env->me_flags & MDB_NOTLS);
2560 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2565 do /* LY: Retry on a race, ITS#7970. */
2566 r->mr_txnid = ti->mti_txnid;
2567 while(r->mr_txnid != ti->mti_txnid);
2568 txn->mt_txnid = r->mr_txnid;
2569 txn->mt_u.reader = r;
2570 meta = env->me_metas[txn->mt_txnid & 1];
2574 if (LOCK_MUTEX(rc, env, MDB_MUTEX(env, w)))
2576 #ifdef MDB_USE_SYSV_SEM
2577 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2578 txn->mt_txnid = meta->mm_txnid;
2579 /* Update mti_txnid like mdb_mutex_failed() would,
2580 * in case last writer crashed before updating it.
2582 ti->mti_txnid = txn->mt_txnid;
2584 txn->mt_txnid = ti->mti_txnid;
2585 meta = env->me_metas[txn->mt_txnid & 1];
2588 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2589 txn->mt_txnid = meta->mm_txnid;
2592 txn->mt_numdbs = env->me_numdbs;
2595 if (txn->mt_txnid == mdb_debug_start)
2599 txn->mt_child = NULL;
2600 txn->mt_loose_pgs = NULL;
2601 txn->mt_loose_count = 0;
2602 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2603 txn->mt_u.dirty_list = env->me_dirty_list;
2604 txn->mt_u.dirty_list[0].mid = 0;
2605 txn->mt_free_pgs = env->me_free_pgs;
2606 txn->mt_free_pgs[0] = 0;
2607 txn->mt_spill_pgs = NULL;
2609 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2612 /* Copy the DB info and flags */
2613 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2615 /* Moved to here to avoid a data race in read TXNs */
2616 txn->mt_next_pgno = meta->mm_last_pg+1;
2618 for (i=2; i<txn->mt_numdbs; i++) {
2619 x = env->me_dbflags[i];
2620 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2621 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2623 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2625 if (env->me_maxpg < txn->mt_next_pgno) {
2626 mdb_txn_reset0(txn, "renew0-mapfail");
2628 txn->mt_u.reader->mr_pid = 0;
2629 txn->mt_u.reader = NULL;
2631 return MDB_MAP_RESIZED;
2638 mdb_txn_renew(MDB_txn *txn)
2642 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2645 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2646 DPUTS("environment had fatal error, must shutdown!");
2650 rc = mdb_txn_renew0(txn);
2651 if (rc == MDB_SUCCESS) {
2652 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2653 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2654 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2660 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2664 int rc, size, tsize = sizeof(MDB_txn);
2666 if (env->me_flags & MDB_FATAL_ERROR) {
2667 DPUTS("environment had fatal error, must shutdown!");
2670 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2673 /* Nested transactions: Max 1 child, write txns only, no writemap */
2674 if (parent->mt_child ||
2675 (flags & MDB_RDONLY) ||
2676 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2677 (env->me_flags & MDB_WRITEMAP))
2679 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2681 tsize = sizeof(MDB_ntxn);
2684 if (!(flags & MDB_RDONLY)) {
2686 txn = env->me_txn0; /* just reuse preallocated write txn */
2689 /* child txns use own copy of cursors */
2690 size += env->me_maxdbs * sizeof(MDB_cursor *);
2692 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2694 if ((txn = calloc(1, size)) == NULL) {
2695 DPRINTF(("calloc: %s", strerror(errno)));
2698 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2699 if (flags & MDB_RDONLY) {
2700 txn->mt_flags |= MDB_TXN_RDONLY;
2701 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2702 txn->mt_dbiseqs = env->me_dbiseqs;
2704 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2706 txn->mt_dbiseqs = parent->mt_dbiseqs;
2707 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2709 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2710 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2718 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2719 if (!txn->mt_u.dirty_list ||
2720 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2722 free(txn->mt_u.dirty_list);
2726 txn->mt_txnid = parent->mt_txnid;
2727 txn->mt_dirty_room = parent->mt_dirty_room;
2728 txn->mt_u.dirty_list[0].mid = 0;
2729 txn->mt_spill_pgs = NULL;
2730 txn->mt_next_pgno = parent->mt_next_pgno;
2731 parent->mt_child = txn;
2732 txn->mt_parent = parent;
2733 txn->mt_numdbs = parent->mt_numdbs;
2734 txn->mt_flags = parent->mt_flags;
2735 txn->mt_dbxs = parent->mt_dbxs;
2736 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2737 /* Copy parent's mt_dbflags, but clear DB_NEW */
2738 for (i=0; i<txn->mt_numdbs; i++)
2739 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2741 ntxn = (MDB_ntxn *)txn;
2742 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2743 if (env->me_pghead) {
2744 size = MDB_IDL_SIZEOF(env->me_pghead);
2745 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2747 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2752 rc = mdb_cursor_shadow(parent, txn);
2754 mdb_txn_reset0(txn, "beginchild-fail");
2756 rc = mdb_txn_renew0(txn);
2759 if (txn != env->me_txn0)
2763 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2764 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2765 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2772 mdb_txn_env(MDB_txn *txn)
2774 if(!txn) return NULL;
2779 mdb_txn_id(MDB_txn *txn)
2781 if(!txn) return (txnid_t)-1;
2782 return txn->mt_txnid;
2785 /** Export or close DBI handles opened in this txn. */
2787 mdb_dbis_update(MDB_txn *txn, int keep)
2790 MDB_dbi n = txn->mt_numdbs;
2791 MDB_env *env = txn->mt_env;
2792 unsigned char *tdbflags = txn->mt_dbflags;
2794 for (i = n; --i >= 2;) {
2795 if (tdbflags[i] & DB_NEW) {
2797 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2799 char *ptr = env->me_dbxs[i].md_name.mv_data;
2801 env->me_dbxs[i].md_name.mv_data = NULL;
2802 env->me_dbxs[i].md_name.mv_size = 0;
2803 env->me_dbflags[i] = 0;
2804 env->me_dbiseqs[i]++;
2810 if (keep && env->me_numdbs < n)
2814 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2815 * May be called twice for readonly txns: First reset it, then abort.
2816 * @param[in] txn the transaction handle to reset
2817 * @param[in] act why the transaction is being reset
2820 mdb_txn_reset0(MDB_txn *txn, const char *act)
2822 MDB_env *env = txn->mt_env;
2824 /* Close any DBI handles opened in this txn */
2825 mdb_dbis_update(txn, 0);
2827 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2828 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2829 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2831 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2832 if (txn->mt_u.reader) {
2833 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2834 if (!(env->me_flags & MDB_NOTLS))
2835 txn->mt_u.reader = NULL; /* txn does not own reader */
2837 txn->mt_numdbs = 0; /* close nothing if called again */
2838 txn->mt_dbxs = NULL; /* mark txn as reset */
2840 pgno_t *pghead = env->me_pghead;
2842 mdb_cursors_close(txn, 0);
2843 if (!(env->me_flags & MDB_WRITEMAP)) {
2844 mdb_dlist_free(txn);
2847 if (!txn->mt_parent) {
2848 if (mdb_midl_shrink(&txn->mt_free_pgs))
2849 env->me_free_pgs = txn->mt_free_pgs;
2851 env->me_pghead = NULL;
2855 /* The writer mutex was locked in mdb_txn_begin. */
2857 UNLOCK_MUTEX(MDB_MUTEX(env, w));
2859 txn->mt_parent->mt_child = NULL;
2860 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2861 mdb_midl_free(txn->mt_free_pgs);
2862 mdb_midl_free(txn->mt_spill_pgs);
2863 free(txn->mt_u.dirty_list);
2866 mdb_midl_free(pghead);
2871 mdb_txn_reset(MDB_txn *txn)
2876 /* This call is only valid for read-only txns */
2877 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2880 mdb_txn_reset0(txn, "reset");
2884 mdb_txn_abort(MDB_txn *txn)
2890 mdb_txn_abort(txn->mt_child);
2892 mdb_txn_reset0(txn, "abort");
2893 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2894 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2895 txn->mt_u.reader->mr_pid = 0;
2897 if (txn != txn->mt_env->me_txn0)
2901 /** Save the freelist as of this transaction to the freeDB.
2902 * This changes the freelist. Keep trying until it stabilizes.
2905 mdb_freelist_save(MDB_txn *txn)
2907 /* env->me_pghead[] can grow and shrink during this call.
2908 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2909 * Page numbers cannot disappear from txn->mt_free_pgs[].
2912 MDB_env *env = txn->mt_env;
2913 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2914 txnid_t pglast = 0, head_id = 0;
2915 pgno_t freecnt = 0, *free_pgs, *mop;
2916 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2918 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2920 if (env->me_pghead) {
2921 /* Make sure first page of freeDB is touched and on freelist */
2922 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2923 if (rc && rc != MDB_NOTFOUND)
2927 if (!env->me_pghead && txn->mt_loose_pgs) {
2928 /* Put loose page numbers in mt_free_pgs, since
2929 * we may be unable to return them to me_pghead.
2931 MDB_page *mp = txn->mt_loose_pgs;
2932 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2934 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2935 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2936 txn->mt_loose_pgs = NULL;
2937 txn->mt_loose_count = 0;
2940 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2941 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2942 ? SSIZE_MAX : maxfree_1pg;
2945 /* Come back here after each Put() in case freelist changed */
2950 /* If using records from freeDB which we have not yet
2951 * deleted, delete them and any we reserved for me_pghead.
2953 while (pglast < env->me_pglast) {
2954 rc = mdb_cursor_first(&mc, &key, NULL);
2957 pglast = head_id = *(txnid_t *)key.mv_data;
2958 total_room = head_room = 0;
2959 mdb_tassert(txn, pglast <= env->me_pglast);
2960 rc = mdb_cursor_del(&mc, 0);
2965 /* Save the IDL of pages freed by this txn, to a single record */
2966 if (freecnt < txn->mt_free_pgs[0]) {
2968 /* Make sure last page of freeDB is touched and on freelist */
2969 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2970 if (rc && rc != MDB_NOTFOUND)
2973 free_pgs = txn->mt_free_pgs;
2974 /* Write to last page of freeDB */
2975 key.mv_size = sizeof(txn->mt_txnid);
2976 key.mv_data = &txn->mt_txnid;
2978 freecnt = free_pgs[0];
2979 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2980 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2983 /* Retry if mt_free_pgs[] grew during the Put() */
2984 free_pgs = txn->mt_free_pgs;
2985 } while (freecnt < free_pgs[0]);
2986 mdb_midl_sort(free_pgs);
2987 memcpy(data.mv_data, free_pgs, data.mv_size);
2990 unsigned int i = free_pgs[0];
2991 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2992 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2994 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3000 mop = env->me_pghead;
3001 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3003 /* Reserve records for me_pghead[]. Split it if multi-page,
3004 * to avoid searching freeDB for a page range. Use keys in
3005 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3007 if (total_room >= mop_len) {
3008 if (total_room == mop_len || --more < 0)
3010 } else if (head_room >= maxfree_1pg && head_id > 1) {
3011 /* Keep current record (overflow page), add a new one */
3015 /* (Re)write {key = head_id, IDL length = head_room} */
3016 total_room -= head_room;
3017 head_room = mop_len - total_room;
3018 if (head_room > maxfree_1pg && head_id > 1) {
3019 /* Overflow multi-page for part of me_pghead */
3020 head_room /= head_id; /* amortize page sizes */
3021 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3022 } else if (head_room < 0) {
3023 /* Rare case, not bothering to delete this record */
3026 key.mv_size = sizeof(head_id);
3027 key.mv_data = &head_id;
3028 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3029 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3032 /* IDL is initially empty, zero out at least the length */
3033 pgs = (pgno_t *)data.mv_data;
3034 j = head_room > clean_limit ? head_room : 0;
3038 total_room += head_room;
3041 /* Return loose page numbers to me_pghead, though usually none are
3042 * left at this point. The pages themselves remain in dirty_list.
3044 if (txn->mt_loose_pgs) {
3045 MDB_page *mp = txn->mt_loose_pgs;
3046 unsigned count = txn->mt_loose_count;
3048 /* Room for loose pages + temp IDL with same */
3049 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3051 mop = env->me_pghead;
3052 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3053 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3054 loose[ ++count ] = mp->mp_pgno;
3056 mdb_midl_sort(loose);
3057 mdb_midl_xmerge(mop, loose);
3058 txn->mt_loose_pgs = NULL;
3059 txn->mt_loose_count = 0;
3063 /* Fill in the reserved me_pghead records */
3069 rc = mdb_cursor_first(&mc, &key, &data);
3070 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3071 txnid_t id = *(txnid_t *)key.mv_data;
3072 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3075 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3077 if (len > mop_len) {
3079 data.mv_size = (len + 1) * sizeof(MDB_ID);
3081 data.mv_data = mop -= len;
3084 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3086 if (rc || !(mop_len -= len))
3093 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3094 * @param[in] txn the transaction that's being committed
3095 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3096 * @return 0 on success, non-zero on failure.
3099 mdb_page_flush(MDB_txn *txn, int keep)
3101 MDB_env *env = txn->mt_env;
3102 MDB_ID2L dl = txn->mt_u.dirty_list;
3103 unsigned psize = env->me_psize, j;
3104 int i, pagecount = dl[0].mid, rc;
3105 size_t size = 0, pos = 0;
3107 MDB_page *dp = NULL;
3111 struct iovec iov[MDB_COMMIT_PAGES];
3112 ssize_t wpos = 0, wsize = 0, wres;
3113 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3119 if (env->me_flags & MDB_WRITEMAP) {
3120 /* Clear dirty flags */
3121 while (++i <= pagecount) {
3123 /* Don't flush this page yet */
3124 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3125 dp->mp_flags &= ~P_KEEP;
3129 dp->mp_flags &= ~P_DIRTY;
3134 /* Write the pages */
3136 if (++i <= pagecount) {
3138 /* Don't flush this page yet */
3139 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3140 dp->mp_flags &= ~P_KEEP;
3145 /* clear dirty flag */
3146 dp->mp_flags &= ~P_DIRTY;
3149 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3154 /* Windows actually supports scatter/gather I/O, but only on
3155 * unbuffered file handles. Since we're relying on the OS page
3156 * cache for all our data, that's self-defeating. So we just
3157 * write pages one at a time. We use the ov structure to set
3158 * the write offset, to at least save the overhead of a Seek
3161 DPRINTF(("committing page %"Z"u", pgno));
3162 memset(&ov, 0, sizeof(ov));
3163 ov.Offset = pos & 0xffffffff;
3164 ov.OffsetHigh = pos >> 16 >> 16;
3165 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3167 DPRINTF(("WriteFile: %d", rc));
3171 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3172 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3174 /* Write previous page(s) */
3175 #ifdef MDB_USE_PWRITEV
3176 wres = pwritev(env->me_fd, iov, n, wpos);
3179 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3181 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3183 DPRINTF(("lseek: %s", strerror(rc)));
3186 wres = writev(env->me_fd, iov, n);
3189 if (wres != wsize) {
3192 DPRINTF(("Write error: %s", strerror(rc)));
3194 rc = EIO; /* TODO: Use which error code? */
3195 DPUTS("short write, filesystem full?");
3206 DPRINTF(("committing page %"Z"u", pgno));
3207 next_pos = pos + size;
3208 iov[n].iov_len = size;
3209 iov[n].iov_base = (char *)dp;
3215 /* MIPS has cache coherency issues, this is a no-op everywhere else
3216 * Note: for any size >= on-chip cache size, entire on-chip cache is
3219 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3221 for (i = keep; ++i <= pagecount; ) {
3223 /* This is a page we skipped above */
3226 dl[j].mid = dp->mp_pgno;
3229 mdb_dpage_free(env, dp);
3234 txn->mt_dirty_room += i - j;
3240 mdb_txn_commit(MDB_txn *txn)
3246 if (txn == NULL || txn->mt_env == NULL)
3249 if (txn->mt_child) {
3250 rc = mdb_txn_commit(txn->mt_child);
3251 txn->mt_child = NULL;
3258 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3259 mdb_dbis_update(txn, 1);
3260 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3265 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3266 DPUTS("error flag is set, can't commit");
3268 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3273 if (txn->mt_parent) {
3274 MDB_txn *parent = txn->mt_parent;
3278 unsigned x, y, len, ps_len;
3280 /* Append our free list to parent's */
3281 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3284 mdb_midl_free(txn->mt_free_pgs);
3285 /* Failures after this must either undo the changes
3286 * to the parent or set MDB_TXN_ERROR in the parent.
3289 parent->mt_next_pgno = txn->mt_next_pgno;
3290 parent->mt_flags = txn->mt_flags;
3292 /* Merge our cursors into parent's and close them */
3293 mdb_cursors_close(txn, 1);
3295 /* Update parent's DB table. */
3296 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3297 parent->mt_numdbs = txn->mt_numdbs;
3298 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3299 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3300 for (i=2; i<txn->mt_numdbs; i++) {
3301 /* preserve parent's DB_NEW status */
3302 x = parent->mt_dbflags[i] & DB_NEW;
3303 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3306 dst = parent->mt_u.dirty_list;
3307 src = txn->mt_u.dirty_list;
3308 /* Remove anything in our dirty list from parent's spill list */
3309 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3311 pspill[0] = (pgno_t)-1;
3312 /* Mark our dirty pages as deleted in parent spill list */
3313 for (i=0, len=src[0].mid; ++i <= len; ) {
3314 MDB_ID pn = src[i].mid << 1;
3315 while (pn > pspill[x])
3317 if (pn == pspill[x]) {
3322 /* Squash deleted pagenums if we deleted any */
3323 for (x=y; ++x <= ps_len; )
3324 if (!(pspill[x] & 1))
3325 pspill[++y] = pspill[x];
3329 /* Find len = length of merging our dirty list with parent's */
3331 dst[0].mid = 0; /* simplify loops */
3332 if (parent->mt_parent) {
3333 len = x + src[0].mid;
3334 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3335 for (i = x; y && i; y--) {
3336 pgno_t yp = src[y].mid;
3337 while (yp < dst[i].mid)
3339 if (yp == dst[i].mid) {
3344 } else { /* Simplify the above for single-ancestor case */
3345 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3347 /* Merge our dirty list with parent's */
3349 for (i = len; y; dst[i--] = src[y--]) {
3350 pgno_t yp = src[y].mid;
3351 while (yp < dst[x].mid)
3352 dst[i--] = dst[x--];
3353 if (yp == dst[x].mid)
3354 free(dst[x--].mptr);
3356 mdb_tassert(txn, i == x);
3358 free(txn->mt_u.dirty_list);
3359 parent->mt_dirty_room = txn->mt_dirty_room;
3360 if (txn->mt_spill_pgs) {
3361 if (parent->mt_spill_pgs) {
3362 /* TODO: Prevent failure here, so parent does not fail */
3363 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3365 parent->mt_flags |= MDB_TXN_ERROR;
3366 mdb_midl_free(txn->mt_spill_pgs);
3367 mdb_midl_sort(parent->mt_spill_pgs);
3369 parent->mt_spill_pgs = txn->mt_spill_pgs;
3373 /* Append our loose page list to parent's */
3374 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3376 *lp = txn->mt_loose_pgs;
3377 parent->mt_loose_count += txn->mt_loose_count;
3379 parent->mt_child = NULL;
3380 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3385 if (txn != env->me_txn) {
3386 DPUTS("attempt to commit unknown transaction");
3391 mdb_cursors_close(txn, 0);
3393 if (!txn->mt_u.dirty_list[0].mid &&
3394 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3397 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3398 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3400 /* Update DB root pointers */
3401 if (txn->mt_numdbs > 2) {
3405 data.mv_size = sizeof(MDB_db);
3407 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3408 for (i = 2; i < txn->mt_numdbs; i++) {
3409 if (txn->mt_dbflags[i] & DB_DIRTY) {
3410 if (TXN_DBI_CHANGED(txn, i)) {
3414 data.mv_data = &txn->mt_dbs[i];
3415 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3422 rc = mdb_freelist_save(txn);
3426 mdb_midl_free(env->me_pghead);
3427 env->me_pghead = NULL;
3428 if (mdb_midl_shrink(&txn->mt_free_pgs))
3429 env->me_free_pgs = txn->mt_free_pgs;
3435 if ((rc = mdb_page_flush(txn, 0)) ||
3436 (rc = mdb_env_sync(env, 0)) ||
3437 (rc = mdb_env_write_meta(txn)))
3440 /* Free P_LOOSE pages left behind in dirty_list */
3441 if (!(env->me_flags & MDB_WRITEMAP))
3442 mdb_dlist_free(txn);
3447 mdb_dbis_update(txn, 1);
3450 UNLOCK_MUTEX(MDB_MUTEX(env, w));
3451 if (txn != env->me_txn0)
3461 /** Read the environment parameters of a DB environment before
3462 * mapping it into memory.
3463 * @param[in] env the environment handle
3464 * @param[out] meta address of where to store the meta information
3465 * @return 0 on success, non-zero on failure.
3468 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3474 enum { Size = sizeof(pbuf) };
3476 /* We don't know the page size yet, so use a minimum value.
3477 * Read both meta pages so we can use the latest one.
3480 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3484 memset(&ov, 0, sizeof(ov));
3486 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3487 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3490 rc = pread(env->me_fd, &pbuf, Size, off);
3493 if (rc == 0 && off == 0)
3495 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3496 DPRINTF(("read: %s", mdb_strerror(rc)));
3500 p = (MDB_page *)&pbuf;
3502 if (!F_ISSET(p->mp_flags, P_META)) {
3503 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3508 if (m->mm_magic != MDB_MAGIC) {
3509 DPUTS("meta has invalid magic");
3513 if (m->mm_version != MDB_DATA_VERSION) {
3514 DPRINTF(("database is version %u, expected version %u",
3515 m->mm_version, MDB_DATA_VERSION));
3516 return MDB_VERSION_MISMATCH;
3519 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3525 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3527 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3529 meta->mm_magic = MDB_MAGIC;
3530 meta->mm_version = MDB_DATA_VERSION;
3531 meta->mm_mapsize = env->me_mapsize;
3532 meta->mm_psize = env->me_psize;
3533 meta->mm_last_pg = 1;
3534 meta->mm_flags = env->me_flags & 0xffff;
3535 meta->mm_flags |= MDB_INTEGERKEY;
3536 meta->mm_dbs[0].md_root = P_INVALID;
3537 meta->mm_dbs[1].md_root = P_INVALID;
3540 /** Write the environment parameters of a freshly created DB environment.
3541 * @param[in] env the environment handle
3542 * @param[in] meta the #MDB_meta to write
3543 * @return 0 on success, non-zero on failure.
3546 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3554 memset(&ov, 0, sizeof(ov));
3555 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3557 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3560 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3561 len = pwrite(fd, ptr, size, pos); \
3562 rc = (len >= 0); } while(0)
3565 DPUTS("writing new meta page");
3567 psize = env->me_psize;
3569 p = calloc(2, psize);
3571 p->mp_flags = P_META;
3572 *(MDB_meta *)METADATA(p) = *meta;
3574 q = (MDB_page *)((char *)p + psize);
3576 q->mp_flags = P_META;
3577 *(MDB_meta *)METADATA(q) = *meta;
3579 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3582 else if ((unsigned) len == psize * 2)
3590 /** Update the environment info to commit a transaction.
3591 * @param[in] txn the transaction that's being committed
3592 * @return 0 on success, non-zero on failure.
3595 mdb_env_write_meta(MDB_txn *txn)
3598 MDB_meta meta, metab, *mp;
3601 int rc, len, toggle;
3610 toggle = txn->mt_txnid & 1;
3611 DPRINTF(("writing meta page %d for root page %"Z"u",
3612 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3615 mp = env->me_metas[toggle];
3616 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3617 /* Persist any increases of mapsize config */
3618 if (mapsize < env->me_mapsize)
3619 mapsize = env->me_mapsize;
3621 if (env->me_flags & MDB_WRITEMAP) {
3622 mp->mm_mapsize = mapsize;
3623 mp->mm_dbs[0] = txn->mt_dbs[0];
3624 mp->mm_dbs[1] = txn->mt_dbs[1];
3625 mp->mm_last_pg = txn->mt_next_pgno - 1;
3626 #if !(defined(_MSC_VER) || defined(__i386__) || defined(__x86_64__))
3627 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3628 __sync_synchronize();
3630 mp->mm_txnid = txn->mt_txnid;
3631 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3632 unsigned meta_size = env->me_psize;
3633 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3636 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3637 if (meta_size < env->me_os_psize)
3638 meta_size += meta_size;
3643 if (MDB_MSYNC(ptr, meta_size, rc)) {
3650 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3651 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3653 meta.mm_mapsize = mapsize;
3654 meta.mm_dbs[0] = txn->mt_dbs[0];
3655 meta.mm_dbs[1] = txn->mt_dbs[1];
3656 meta.mm_last_pg = txn->mt_next_pgno - 1;
3657 meta.mm_txnid = txn->mt_txnid;
3659 off = offsetof(MDB_meta, mm_mapsize);
3660 ptr = (char *)&meta + off;
3661 len = sizeof(MDB_meta) - off;
3663 off += env->me_psize;
3666 /* Write to the SYNC fd */
3667 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3668 env->me_fd : env->me_mfd;
3671 memset(&ov, 0, sizeof(ov));
3673 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3677 rc = pwrite(mfd, ptr, len, off);
3680 rc = rc < 0 ? ErrCode() : EIO;
3681 DPUTS("write failed, disk error?");
3682 /* On a failure, the pagecache still contains the new data.
3683 * Write some old data back, to prevent it from being used.
3684 * Use the non-SYNC fd; we know it will fail anyway.
3686 meta.mm_last_pg = metab.mm_last_pg;
3687 meta.mm_txnid = metab.mm_txnid;
3689 memset(&ov, 0, sizeof(ov));
3691 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3693 r2 = pwrite(env->me_fd, ptr, len, off);
3694 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3697 env->me_flags |= MDB_FATAL_ERROR;
3700 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3701 CACHEFLUSH(env->me_map + off, len, DCACHE);
3703 /* Memory ordering issues are irrelevant; since the entire writer
3704 * is wrapped by wmutex, all of these changes will become visible
3705 * after the wmutex is unlocked. Since the DB is multi-version,
3706 * readers will get consistent data regardless of how fresh or
3707 * how stale their view of these values is.
3710 env->me_txns->mti_txnid = txn->mt_txnid;
3715 /** Check both meta pages to see which one is newer.
3716 * @param[in] env the environment handle
3717 * @return meta toggle (0 or 1).
3720 mdb_env_pick_meta(const MDB_env *env)
3722 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3726 mdb_env_create(MDB_env **env)
3730 e = calloc(1, sizeof(MDB_env));
3734 e->me_maxreaders = DEFAULT_READERS;
3735 e->me_maxdbs = e->me_numdbs = 2;
3736 e->me_fd = INVALID_HANDLE_VALUE;
3737 e->me_lfd = INVALID_HANDLE_VALUE;
3738 e->me_mfd = INVALID_HANDLE_VALUE;
3739 #ifdef MDB_USE_SYSV_SEM
3740 e->me_rmutex.semid = -1;
3741 e->me_wmutex.semid = -1;
3743 e->me_pid = getpid();
3744 GET_PAGESIZE(e->me_os_psize);
3745 VGMEMP_CREATE(e,0,0);
3751 mdb_env_map(MDB_env *env, void *addr)
3754 unsigned int flags = env->me_flags;
3758 LONG sizelo, sizehi;
3761 if (flags & MDB_RDONLY) {
3762 /* Don't set explicit map size, use whatever exists */
3767 msize = env->me_mapsize;
3768 sizelo = msize & 0xffffffff;
3769 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3771 /* Windows won't create mappings for zero length files.
3772 * and won't map more than the file size.
3773 * Just set the maxsize right now.
3775 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3776 || !SetEndOfFile(env->me_fd)
3777 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3781 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3782 PAGE_READWRITE : PAGE_READONLY,
3783 sizehi, sizelo, NULL);
3786 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3787 FILE_MAP_WRITE : FILE_MAP_READ,
3789 rc = env->me_map ? 0 : ErrCode();
3794 int prot = PROT_READ;
3795 if (flags & MDB_WRITEMAP) {
3797 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3800 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3802 if (env->me_map == MAP_FAILED) {
3807 if (flags & MDB_NORDAHEAD) {
3808 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3810 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3812 #ifdef POSIX_MADV_RANDOM
3813 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3814 #endif /* POSIX_MADV_RANDOM */
3815 #endif /* MADV_RANDOM */
3819 /* Can happen because the address argument to mmap() is just a
3820 * hint. mmap() can pick another, e.g. if the range is in use.
3821 * The MAP_FIXED flag would prevent that, but then mmap could
3822 * instead unmap existing pages to make room for the new map.
3824 if (addr && env->me_map != addr)
3825 return EBUSY; /* TODO: Make a new MDB_* error code? */
3827 p = (MDB_page *)env->me_map;
3828 env->me_metas[0] = METADATA(p);
3829 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3835 mdb_env_set_mapsize(MDB_env *env, size_t size)
3837 /* If env is already open, caller is responsible for making
3838 * sure there are no active txns.
3846 meta = env->me_metas[mdb_env_pick_meta(env)];
3848 size = meta->mm_mapsize;
3850 /* Silently round up to minimum if the size is too small */
3851 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3855 munmap(env->me_map, env->me_mapsize);
3856 env->me_mapsize = size;
3857 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3858 rc = mdb_env_map(env, old);
3862 env->me_mapsize = size;
3864 env->me_maxpg = env->me_mapsize / env->me_psize;
3869 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3873 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3878 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3880 if (env->me_map || readers < 1)
3882 env->me_maxreaders = readers;
3887 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3889 if (!env || !readers)
3891 *readers = env->me_maxreaders;
3896 mdb_fsize(HANDLE fd, size_t *size)
3899 LARGE_INTEGER fsize;
3901 if (!GetFileSizeEx(fd, &fsize))
3904 *size = fsize.QuadPart;
3916 /** Further setup required for opening an LMDB environment
3919 mdb_env_open2(MDB_env *env)
3921 unsigned int flags = env->me_flags;
3922 int i, newenv = 0, rc;
3926 /* See if we should use QueryLimited */
3928 if ((rc & 0xff) > 5)
3929 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3931 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3934 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3937 DPUTS("new mdbenv");
3939 env->me_psize = env->me_os_psize;
3940 if (env->me_psize > MAX_PAGESIZE)
3941 env->me_psize = MAX_PAGESIZE;
3942 memset(&meta, 0, sizeof(meta));
3943 mdb_env_init_meta0(env, &meta);
3944 meta.mm_mapsize = DEFAULT_MAPSIZE;
3946 env->me_psize = meta.mm_psize;
3949 /* Was a mapsize configured? */
3950 if (!env->me_mapsize) {
3951 env->me_mapsize = meta.mm_mapsize;
3954 /* Make sure mapsize >= committed data size. Even when using
3955 * mm_mapsize, which could be broken in old files (ITS#7789).
3957 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3958 if (env->me_mapsize < minsize)
3959 env->me_mapsize = minsize;
3961 meta.mm_mapsize = env->me_mapsize;
3963 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3968 if (flags & MDB_FIXEDMAP)
3969 meta.mm_address = env->me_map;
3970 i = mdb_env_init_meta(env, &meta);
3971 if (i != MDB_SUCCESS) {
3976 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3977 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3979 #if !(MDB_MAXKEYSIZE)
3980 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3982 env->me_maxpg = env->me_mapsize / env->me_psize;
3986 int toggle = mdb_env_pick_meta(env);
3987 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3989 DPRINTF(("opened database version %u, pagesize %u",
3990 env->me_metas[0]->mm_version, env->me_psize));
3991 DPRINTF(("using meta page %d", toggle));
3992 DPRINTF(("depth: %u", db->md_depth));
3993 DPRINTF(("entries: %"Z"u", db->md_entries));
3994 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3995 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3996 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3997 DPRINTF(("root: %"Z"u", db->md_root));
4005 /** Release a reader thread's slot in the reader lock table.
4006 * This function is called automatically when a thread exits.
4007 * @param[in] ptr This points to the slot in the reader lock table.
4010 mdb_env_reader_dest(void *ptr)
4012 MDB_reader *reader = ptr;
4018 /** Junk for arranging thread-specific callbacks on Windows. This is
4019 * necessarily platform and compiler-specific. Windows supports up
4020 * to 1088 keys. Let's assume nobody opens more than 64 environments
4021 * in a single process, for now. They can override this if needed.
4023 #ifndef MAX_TLS_KEYS
4024 #define MAX_TLS_KEYS 64
4026 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4027 static int mdb_tls_nkeys;
4029 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4033 case DLL_PROCESS_ATTACH: break;
4034 case DLL_THREAD_ATTACH: break;
4035 case DLL_THREAD_DETACH:
4036 for (i=0; i<mdb_tls_nkeys; i++) {
4037 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4039 mdb_env_reader_dest(r);
4043 case DLL_PROCESS_DETACH: break;
4048 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4050 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4054 /* Force some symbol references.
4055 * _tls_used forces the linker to create the TLS directory if not already done
4056 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4058 #pragma comment(linker, "/INCLUDE:_tls_used")
4059 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4060 #pragma const_seg(".CRT$XLB")
4061 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4062 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4065 #pragma comment(linker, "/INCLUDE:__tls_used")
4066 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4067 #pragma data_seg(".CRT$XLB")
4068 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4070 #endif /* WIN 32/64 */
4071 #endif /* !__GNUC__ */
4074 /** Downgrade the exclusive lock on the region back to shared */
4076 mdb_env_share_locks(MDB_env *env, int *excl)
4078 int rc = 0, toggle = mdb_env_pick_meta(env);
4080 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4085 /* First acquire a shared lock. The Unlock will
4086 * then release the existing exclusive lock.
4088 memset(&ov, 0, sizeof(ov));
4089 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4092 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4098 struct flock lock_info;
4099 /* The shared lock replaces the existing lock */
4100 memset((void *)&lock_info, 0, sizeof(lock_info));
4101 lock_info.l_type = F_RDLCK;
4102 lock_info.l_whence = SEEK_SET;
4103 lock_info.l_start = 0;
4104 lock_info.l_len = 1;
4105 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4106 (rc = ErrCode()) == EINTR) ;
4107 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4114 /** Try to get exclusive lock, otherwise shared.
4115 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4118 mdb_env_excl_lock(MDB_env *env, int *excl)
4122 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4126 memset(&ov, 0, sizeof(ov));
4127 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4134 struct flock lock_info;
4135 memset((void *)&lock_info, 0, sizeof(lock_info));
4136 lock_info.l_type = F_WRLCK;
4137 lock_info.l_whence = SEEK_SET;
4138 lock_info.l_start = 0;
4139 lock_info.l_len = 1;
4140 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4141 (rc = ErrCode()) == EINTR) ;
4145 # ifdef MDB_USE_SYSV_SEM
4146 if (*excl < 0) /* always true when !MDB_USE_SYSV_SEM */
4149 lock_info.l_type = F_RDLCK;
4150 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4151 (rc = ErrCode()) == EINTR) ;
4161 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4163 * @(#) $Revision: 5.1 $
4164 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4165 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4167 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4171 * Please do not copyright this code. This code is in the public domain.
4173 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4174 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4175 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4176 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4177 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4178 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4179 * PERFORMANCE OF THIS SOFTWARE.
4182 * chongo <Landon Curt Noll> /\oo/\
4183 * http://www.isthe.com/chongo/
4185 * Share and Enjoy! :-)
4188 typedef unsigned long long mdb_hash_t;
4189 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4191 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4192 * @param[in] val value to hash
4193 * @param[in] hval initial value for hash
4194 * @return 64 bit hash
4196 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4197 * hval arg on the first call.
4200 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4202 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4203 unsigned char *end = s + val->mv_size;
4205 * FNV-1a hash each octet of the string
4208 /* xor the bottom with the current octet */
4209 hval ^= (mdb_hash_t)*s++;
4211 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4212 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4213 (hval << 7) + (hval << 8) + (hval << 40);
4215 /* return our new hash value */
4219 /** Hash the string and output the encoded hash.
4220 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4221 * very short name limits. We don't care about the encoding being reversible,
4222 * we just want to preserve as many bits of the input as possible in a
4223 * small printable string.
4224 * @param[in] str string to hash
4225 * @param[out] encbuf an array of 11 chars to hold the hash
4227 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4230 mdb_pack85(unsigned long l, char *out)
4234 for (i=0; i<5; i++) {
4235 *out++ = mdb_a85[l % 85];
4241 mdb_hash_enc(MDB_val *val, char *encbuf)
4243 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4245 mdb_pack85(h, encbuf);
4246 mdb_pack85(h>>32, encbuf+5);
4251 /** Open and/or initialize the lock region for the environment.
4252 * @param[in] env The LMDB environment.
4253 * @param[in] lpath The pathname of the file used for the lock region.
4254 * @param[in] mode The Unix permissions for the file, if we create it.
4255 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4256 * @return 0 on success, non-zero on failure.
4259 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4262 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4264 # define MDB_ERRCODE_ROFS EROFS
4265 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4266 # define MDB_CLOEXEC O_CLOEXEC
4269 # define MDB_CLOEXEC 0
4276 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4277 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4278 FILE_ATTRIBUTE_NORMAL, NULL);
4280 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4282 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4284 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4289 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4290 /* Lose record locks when exec*() */
4291 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4292 fcntl(env->me_lfd, F_SETFD, fdflags);
4295 if (!(env->me_flags & MDB_NOTLS)) {
4296 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4299 env->me_flags |= MDB_ENV_TXKEY;
4301 /* Windows TLS callbacks need help finding their TLS info. */
4302 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4306 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4310 /* Try to get exclusive lock. If we succeed, then
4311 * nobody is using the lock region and we should initialize it.
4313 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4316 size = GetFileSize(env->me_lfd, NULL);
4318 size = lseek(env->me_lfd, 0, SEEK_END);
4319 if (size == -1) goto fail_errno;
4321 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4322 if (size < rsize && *excl > 0) {
4324 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4325 || !SetEndOfFile(env->me_lfd))
4328 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4332 size = rsize - sizeof(MDB_txninfo);
4333 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4338 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4340 if (!mh) goto fail_errno;
4341 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4343 if (!env->me_txns) goto fail_errno;
4345 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4347 if (m == MAP_FAILED) goto fail_errno;
4353 BY_HANDLE_FILE_INFORMATION stbuf;
4362 if (!mdb_sec_inited) {
4363 InitializeSecurityDescriptor(&mdb_null_sd,
4364 SECURITY_DESCRIPTOR_REVISION);
4365 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4366 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4367 mdb_all_sa.bInheritHandle = FALSE;
4368 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4371 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4372 idbuf.volume = stbuf.dwVolumeSerialNumber;
4373 idbuf.nhigh = stbuf.nFileIndexHigh;
4374 idbuf.nlow = stbuf.nFileIndexLow;
4375 val.mv_data = &idbuf;
4376 val.mv_size = sizeof(idbuf);
4377 mdb_hash_enc(&val, encbuf);
4378 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4379 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4380 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4381 if (!env->me_rmutex) goto fail_errno;
4382 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4383 if (!env->me_wmutex) goto fail_errno;
4384 #elif defined(MDB_USE_SYSV_SEM)
4386 unsigned short vals[2] = {1, 1};
4387 int semid = semget(IPC_PRIVATE, 2, mode);
4391 env->me_rmutex.semid = semid;
4392 env->me_wmutex.semid = semid;
4393 env->me_rmutex.semnum = 0;
4394 env->me_wmutex.semnum = 1;
4397 if (semctl(semid, 0, SETALL, semu) < 0)
4399 env->me_txns->mti_semid = semid;
4400 #else /* MDB_USE_SYSV_SEM */
4401 pthread_mutexattr_t mattr;
4403 if ((rc = pthread_mutexattr_init(&mattr))
4404 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4405 #ifdef MDB_ROBUST_SUPPORTED
4406 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4408 || (rc = pthread_mutex_init(&env->me_txns->mti_rmutex, &mattr))
4409 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4411 pthread_mutexattr_destroy(&mattr);
4412 #endif /* _WIN32 || MDB_USE_SYSV_SEM */
4414 env->me_txns->mti_magic = MDB_MAGIC;
4415 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4416 env->me_txns->mti_txnid = 0;
4417 env->me_txns->mti_numreaders = 0;
4420 if (env->me_txns->mti_magic != MDB_MAGIC) {
4421 DPUTS("lock region has invalid magic");
4425 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4426 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4427 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4428 rc = MDB_VERSION_MISMATCH;
4432 if (rc && rc != EACCES && rc != EAGAIN) {
4436 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4437 if (!env->me_rmutex) goto fail_errno;
4438 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4439 if (!env->me_wmutex) goto fail_errno;
4440 #elif defined(MDB_USE_SYSV_SEM)
4441 struct semid_ds buf;
4443 int semid = env->me_txns->mti_semid;
4446 /* check for read access */
4447 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4449 /* check for write access */
4450 if (semctl(semid, 0, IPC_SET, semu) < 0)
4453 env->me_rmutex.semid = semid;
4454 env->me_wmutex.semid = semid;
4455 env->me_rmutex.semnum = 0;
4456 env->me_wmutex.semnum = 1;
4467 /** The name of the lock file in the DB environment */
4468 #define LOCKNAME "/lock.mdb"
4469 /** The name of the data file in the DB environment */
4470 #define DATANAME "/data.mdb"
4471 /** The suffix of the lock file when no subdir is used */
4472 #define LOCKSUFF "-lock"
4473 /** Only a subset of the @ref mdb_env flags can be changed
4474 * at runtime. Changing other flags requires closing the
4475 * environment and re-opening it with the new flags.
4477 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4478 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4479 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4481 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4482 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4486 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4488 int oflags, rc, len, excl = -1;
4489 char *lpath, *dpath;
4491 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4495 if (flags & MDB_NOSUBDIR) {
4496 rc = len + sizeof(LOCKSUFF) + len + 1;
4498 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4503 if (flags & MDB_NOSUBDIR) {
4504 dpath = lpath + len + sizeof(LOCKSUFF);
4505 sprintf(lpath, "%s" LOCKSUFF, path);
4506 strcpy(dpath, path);
4508 dpath = lpath + len + sizeof(LOCKNAME);
4509 sprintf(lpath, "%s" LOCKNAME, path);
4510 sprintf(dpath, "%s" DATANAME, path);
4514 flags |= env->me_flags;
4515 if (flags & MDB_RDONLY) {
4516 /* silently ignore WRITEMAP when we're only getting read access */
4517 flags &= ~MDB_WRITEMAP;
4519 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4520 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4523 env->me_flags = flags |= MDB_ENV_ACTIVE;
4527 env->me_path = strdup(path);
4528 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4529 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4530 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4531 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4536 /* For RDONLY, get lockfile after we know datafile exists */
4537 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4538 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4544 if (F_ISSET(flags, MDB_RDONLY)) {
4545 oflags = GENERIC_READ;
4546 len = OPEN_EXISTING;
4548 oflags = GENERIC_READ|GENERIC_WRITE;
4551 mode = FILE_ATTRIBUTE_NORMAL;
4552 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4553 NULL, len, mode, NULL);
4555 if (F_ISSET(flags, MDB_RDONLY))
4558 oflags = O_RDWR | O_CREAT;
4560 env->me_fd = open(dpath, oflags, mode);
4562 if (env->me_fd == INVALID_HANDLE_VALUE) {
4567 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4568 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4573 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4574 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4575 env->me_mfd = env->me_fd;
4577 /* Synchronous fd for meta writes. Needed even with
4578 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4581 len = OPEN_EXISTING;
4582 env->me_mfd = CreateFile(dpath, oflags,
4583 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4584 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4587 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4589 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4594 DPRINTF(("opened dbenv %p", (void *) env));
4596 rc = mdb_env_share_locks(env, &excl);
4600 if (!((flags & MDB_RDONLY) ||
4601 (env->me_pbuf = calloc(1, env->me_psize))))
4603 if (!(flags & MDB_RDONLY)) {
4605 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4606 (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1);
4607 txn = calloc(1, size);
4609 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4610 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4611 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4612 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4614 txn->mt_dbxs = env->me_dbxs;
4624 mdb_env_close0(env, excl);
4630 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4632 mdb_env_close0(MDB_env *env, int excl)
4636 if (!(env->me_flags & MDB_ENV_ACTIVE))
4639 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4640 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4641 free(env->me_dbxs[i].md_name.mv_data);
4644 free(env->me_dbiseqs);
4645 free(env->me_dbflags);
4648 free(env->me_dirty_list);
4650 mdb_midl_free(env->me_free_pgs);
4652 if (env->me_flags & MDB_ENV_TXKEY) {
4653 pthread_key_delete(env->me_txkey);
4655 /* Delete our key from the global list */
4656 for (i=0; i<mdb_tls_nkeys; i++)
4657 if (mdb_tls_keys[i] == env->me_txkey) {
4658 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4666 munmap(env->me_map, env->me_mapsize);
4668 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4669 (void) close(env->me_mfd);
4670 if (env->me_fd != INVALID_HANDLE_VALUE)
4671 (void) close(env->me_fd);
4673 MDB_PID_T pid = env->me_pid;
4674 /* Clearing readers is done in this function because
4675 * me_txkey with its destructor must be disabled first.
4677 for (i = env->me_numreaders; --i >= 0; )
4678 if (env->me_txns->mti_readers[i].mr_pid == pid)
4679 env->me_txns->mti_readers[i].mr_pid = 0;
4681 if (env->me_rmutex) {
4682 CloseHandle(env->me_rmutex);
4683 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4685 /* Windows automatically destroys the mutexes when
4686 * the last handle closes.
4688 #elif defined(MDB_USE_SYSV_SEM)
4689 if (env->me_rmutex.semid != -1) {
4690 /* If we have the filelock: If we are the
4691 * only remaining user, clean up semaphores.
4694 mdb_env_excl_lock(env, &excl);
4696 semctl(env->me_rmutex.semid, 0, IPC_RMID);
4699 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4701 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4704 /* Unlock the lockfile. Windows would have unlocked it
4705 * after closing anyway, but not necessarily at once.
4707 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4710 (void) close(env->me_lfd);
4713 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4717 mdb_env_close(MDB_env *env)
4724 VGMEMP_DESTROY(env);
4725 while ((dp = env->me_dpages) != NULL) {
4726 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4727 env->me_dpages = dp->mp_next;
4731 mdb_env_close0(env, 0);
4735 /** Compare two items pointing at aligned size_t's */
4737 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4739 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4740 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4743 /** Compare two items pointing at aligned unsigned int's */
4745 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4747 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4748 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4751 /** Compare two items pointing at unsigned ints of unknown alignment.
4752 * Nodes and keys are guaranteed to be 2-byte aligned.
4755 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4757 #if BYTE_ORDER == LITTLE_ENDIAN
4758 unsigned short *u, *c;
4761 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4762 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4765 } while(!x && u > (unsigned short *)a->mv_data);
4768 unsigned short *u, *c, *end;
4771 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4772 u = (unsigned short *)a->mv_data;
4773 c = (unsigned short *)b->mv_data;
4776 } while(!x && u < end);
4781 /** Compare two items pointing at size_t's of unknown alignment. */
4782 #ifdef MISALIGNED_OK
4783 # define mdb_cmp_clong mdb_cmp_long
4785 # define mdb_cmp_clong mdb_cmp_cint
4788 /** Compare two items lexically */
4790 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4797 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4803 diff = memcmp(a->mv_data, b->mv_data, len);
4804 return diff ? diff : len_diff<0 ? -1 : len_diff;
4807 /** Compare two items in reverse byte order */
4809 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4811 const unsigned char *p1, *p2, *p1_lim;
4815 p1_lim = (const unsigned char *)a->mv_data;
4816 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4817 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4819 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4825 while (p1 > p1_lim) {
4826 diff = *--p1 - *--p2;
4830 return len_diff<0 ? -1 : len_diff;
4833 /** Search for key within a page, using binary search.
4834 * Returns the smallest entry larger or equal to the key.
4835 * If exactp is non-null, stores whether the found entry was an exact match
4836 * in *exactp (1 or 0).
4837 * Updates the cursor index with the index of the found entry.
4838 * If no entry larger or equal to the key is found, returns NULL.
4841 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4843 unsigned int i = 0, nkeys;
4846 MDB_page *mp = mc->mc_pg[mc->mc_top];
4847 MDB_node *node = NULL;
4852 nkeys = NUMKEYS(mp);
4854 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4855 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4858 low = IS_LEAF(mp) ? 0 : 1;
4860 cmp = mc->mc_dbx->md_cmp;
4862 /* Branch pages have no data, so if using integer keys,
4863 * alignment is guaranteed. Use faster mdb_cmp_int.
4865 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4866 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4873 nodekey.mv_size = mc->mc_db->md_pad;
4874 node = NODEPTR(mp, 0); /* fake */
4875 while (low <= high) {
4876 i = (low + high) >> 1;
4877 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4878 rc = cmp(key, &nodekey);
4879 DPRINTF(("found leaf index %u [%s], rc = %i",
4880 i, DKEY(&nodekey), rc));
4889 while (low <= high) {
4890 i = (low + high) >> 1;
4892 node = NODEPTR(mp, i);
4893 nodekey.mv_size = NODEKSZ(node);
4894 nodekey.mv_data = NODEKEY(node);
4896 rc = cmp(key, &nodekey);
4899 DPRINTF(("found leaf index %u [%s], rc = %i",
4900 i, DKEY(&nodekey), rc));
4902 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4903 i, DKEY(&nodekey), NODEPGNO(node), rc));
4914 if (rc > 0) { /* Found entry is less than the key. */
4915 i++; /* Skip to get the smallest entry larger than key. */
4917 node = NODEPTR(mp, i);
4920 *exactp = (rc == 0 && nkeys > 0);
4921 /* store the key index */
4922 mc->mc_ki[mc->mc_top] = i;
4924 /* There is no entry larger or equal to the key. */
4927 /* nodeptr is fake for LEAF2 */
4933 mdb_cursor_adjust(MDB_cursor *mc, func)
4937 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4938 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4945 /** Pop a page off the top of the cursor's stack. */
4947 mdb_cursor_pop(MDB_cursor *mc)
4951 MDB_page *top = mc->mc_pg[mc->mc_top];
4957 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4958 DDBI(mc), (void *) mc));
4962 /** Push a page onto the top of the cursor's stack. */
4964 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4966 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4967 DDBI(mc), (void *) mc));
4969 if (mc->mc_snum >= CURSOR_STACK) {
4970 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4971 return MDB_CURSOR_FULL;
4974 mc->mc_top = mc->mc_snum++;
4975 mc->mc_pg[mc->mc_top] = mp;
4976 mc->mc_ki[mc->mc_top] = 0;
4981 /** Find the address of the page corresponding to a given page number.
4982 * @param[in] txn the transaction for this access.
4983 * @param[in] pgno the page number for the page to retrieve.
4984 * @param[out] ret address of a pointer where the page's address will be stored.
4985 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4986 * @return 0 on success, non-zero on failure.
4989 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4991 MDB_env *env = txn->mt_env;
4995 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4999 MDB_ID2L dl = tx2->mt_u.dirty_list;
5001 /* Spilled pages were dirtied in this txn and flushed
5002 * because the dirty list got full. Bring this page
5003 * back in from the map (but don't unspill it here,
5004 * leave that unless page_touch happens again).
5006 if (tx2->mt_spill_pgs) {
5007 MDB_ID pn = pgno << 1;
5008 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5009 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5010 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5015 unsigned x = mdb_mid2l_search(dl, pgno);
5016 if (x <= dl[0].mid && dl[x].mid == pgno) {
5022 } while ((tx2 = tx2->mt_parent) != NULL);
5025 if (pgno < txn->mt_next_pgno) {
5027 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5029 DPRINTF(("page %"Z"u not found", pgno));
5030 txn->mt_flags |= MDB_TXN_ERROR;
5031 return MDB_PAGE_NOTFOUND;
5041 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5042 * The cursor is at the root page, set up the rest of it.
5045 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5047 MDB_page *mp = mc->mc_pg[mc->mc_top];
5051 while (IS_BRANCH(mp)) {
5055 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5056 mdb_cassert(mc, NUMKEYS(mp) > 1);
5057 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5059 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5061 if (flags & MDB_PS_LAST)
5062 i = NUMKEYS(mp) - 1;
5065 node = mdb_node_search(mc, key, &exact);
5067 i = NUMKEYS(mp) - 1;
5069 i = mc->mc_ki[mc->mc_top];
5071 mdb_cassert(mc, i > 0);
5075 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5078 mdb_cassert(mc, i < NUMKEYS(mp));
5079 node = NODEPTR(mp, i);
5081 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5084 mc->mc_ki[mc->mc_top] = i;
5085 if ((rc = mdb_cursor_push(mc, mp)))
5088 if (flags & MDB_PS_MODIFY) {
5089 if ((rc = mdb_page_touch(mc)) != 0)
5091 mp = mc->mc_pg[mc->mc_top];
5096 DPRINTF(("internal error, index points to a %02X page!?",
5098 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5099 return MDB_CORRUPTED;
5102 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5103 key ? DKEY(key) : "null"));
5104 mc->mc_flags |= C_INITIALIZED;
5105 mc->mc_flags &= ~C_EOF;
5110 /** Search for the lowest key under the current branch page.
5111 * This just bypasses a NUMKEYS check in the current page
5112 * before calling mdb_page_search_root(), because the callers
5113 * are all in situations where the current page is known to
5117 mdb_page_search_lowest(MDB_cursor *mc)
5119 MDB_page *mp = mc->mc_pg[mc->mc_top];
5120 MDB_node *node = NODEPTR(mp, 0);
5123 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5126 mc->mc_ki[mc->mc_top] = 0;
5127 if ((rc = mdb_cursor_push(mc, mp)))
5129 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5132 /** Search for the page a given key should be in.
5133 * Push it and its parent pages on the cursor stack.
5134 * @param[in,out] mc the cursor for this operation.
5135 * @param[in] key the key to search for, or NULL for first/last page.
5136 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5137 * are touched (updated with new page numbers).
5138 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5139 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5140 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5141 * @return 0 on success, non-zero on failure.
5144 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5149 /* Make sure the txn is still viable, then find the root from
5150 * the txn's db table and set it as the root of the cursor's stack.
5152 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5153 DPUTS("transaction has failed, must abort");
5156 /* Make sure we're using an up-to-date root */
5157 if (*mc->mc_dbflag & DB_STALE) {
5159 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5161 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5162 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5169 MDB_node *leaf = mdb_node_search(&mc2,
5170 &mc->mc_dbx->md_name, &exact);
5172 return MDB_NOTFOUND;
5173 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5176 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5178 /* The txn may not know this DBI, or another process may
5179 * have dropped and recreated the DB with other flags.
5181 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5182 return MDB_INCOMPATIBLE;
5183 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5185 *mc->mc_dbflag &= ~DB_STALE;
5187 root = mc->mc_db->md_root;
5189 if (root == P_INVALID) { /* Tree is empty. */
5190 DPUTS("tree is empty");
5191 return MDB_NOTFOUND;
5195 mdb_cassert(mc, root > 1);
5196 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5197 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5203 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5204 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5206 if (flags & MDB_PS_MODIFY) {
5207 if ((rc = mdb_page_touch(mc)))
5211 if (flags & MDB_PS_ROOTONLY)
5214 return mdb_page_search_root(mc, key, flags);
5218 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5220 MDB_txn *txn = mc->mc_txn;
5221 pgno_t pg = mp->mp_pgno;
5222 unsigned x = 0, ovpages = mp->mp_pages;
5223 MDB_env *env = txn->mt_env;
5224 MDB_IDL sl = txn->mt_spill_pgs;
5225 MDB_ID pn = pg << 1;
5228 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5229 /* If the page is dirty or on the spill list we just acquired it,
5230 * so we should give it back to our current free list, if any.
5231 * Otherwise put it onto the list of pages we freed in this txn.
5233 * Won't create me_pghead: me_pglast must be inited along with it.
5234 * Unsupported in nested txns: They would need to hide the page
5235 * range in ancestor txns' dirty and spilled lists.
5237 if (env->me_pghead &&
5239 ((mp->mp_flags & P_DIRTY) ||
5240 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5244 MDB_ID2 *dl, ix, iy;
5245 rc = mdb_midl_need(&env->me_pghead, ovpages);
5248 if (!(mp->mp_flags & P_DIRTY)) {
5249 /* This page is no longer spilled */
5256 /* Remove from dirty list */
5257 dl = txn->mt_u.dirty_list;
5259 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5265 mdb_cassert(mc, x > 1);
5267 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5268 txn->mt_flags |= MDB_TXN_ERROR;
5269 return MDB_CORRUPTED;
5272 if (!(env->me_flags & MDB_WRITEMAP))
5273 mdb_dpage_free(env, mp);
5275 /* Insert in me_pghead */
5276 mop = env->me_pghead;
5277 j = mop[0] + ovpages;
5278 for (i = mop[0]; i && mop[i] < pg; i--)
5284 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5288 mc->mc_db->md_overflow_pages -= ovpages;
5292 /** Return the data associated with a given node.
5293 * @param[in] txn The transaction for this operation.
5294 * @param[in] leaf The node being read.
5295 * @param[out] data Updated to point to the node's data.
5296 * @return 0 on success, non-zero on failure.
5299 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5301 MDB_page *omp; /* overflow page */
5305 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5306 data->mv_size = NODEDSZ(leaf);
5307 data->mv_data = NODEDATA(leaf);
5311 /* Read overflow data.
5313 data->mv_size = NODEDSZ(leaf);
5314 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5315 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5316 DPRINTF(("read overflow page %"Z"u failed", pgno));
5319 data->mv_data = METADATA(omp);
5325 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5326 MDB_val *key, MDB_val *data)
5333 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5335 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5338 if (txn->mt_flags & MDB_TXN_ERROR)
5341 mdb_cursor_init(&mc, txn, dbi, &mx);
5342 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5345 /** Find a sibling for a page.
5346 * Replaces the page at the top of the cursor's stack with the
5347 * specified sibling, if one exists.
5348 * @param[in] mc The cursor for this operation.
5349 * @param[in] move_right Non-zero if the right sibling is requested,
5350 * otherwise the left sibling.
5351 * @return 0 on success, non-zero on failure.
5354 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5360 if (mc->mc_snum < 2) {
5361 return MDB_NOTFOUND; /* root has no siblings */
5365 DPRINTF(("parent page is page %"Z"u, index %u",
5366 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5368 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5369 : (mc->mc_ki[mc->mc_top] == 0)) {
5370 DPRINTF(("no more keys left, moving to %s sibling",
5371 move_right ? "right" : "left"));
5372 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5373 /* undo cursor_pop before returning */
5380 mc->mc_ki[mc->mc_top]++;
5382 mc->mc_ki[mc->mc_top]--;
5383 DPRINTF(("just moving to %s index key %u",
5384 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5386 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5388 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5389 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5390 /* mc will be inconsistent if caller does mc_snum++ as above */
5391 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5395 mdb_cursor_push(mc, mp);
5397 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5402 /** Move the cursor to the next data item. */
5404 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5410 if (mc->mc_flags & C_EOF) {
5411 return MDB_NOTFOUND;
5414 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5416 mp = mc->mc_pg[mc->mc_top];
5418 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5419 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5420 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5421 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5422 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5423 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5424 if (rc == MDB_SUCCESS)
5425 MDB_GET_KEY(leaf, key);
5430 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5431 if (op == MDB_NEXT_DUP)
5432 return MDB_NOTFOUND;
5436 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5437 mdb_dbg_pgno(mp), (void *) mc));
5438 if (mc->mc_flags & C_DEL)
5441 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5442 DPUTS("=====> move to next sibling page");
5443 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5444 mc->mc_flags |= C_EOF;
5447 mp = mc->mc_pg[mc->mc_top];
5448 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5450 mc->mc_ki[mc->mc_top]++;
5453 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5454 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5457 key->mv_size = mc->mc_db->md_pad;
5458 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5462 mdb_cassert(mc, IS_LEAF(mp));
5463 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5465 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5466 mdb_xcursor_init1(mc, leaf);
5469 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5472 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5473 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5474 if (rc != MDB_SUCCESS)
5479 MDB_GET_KEY(leaf, key);
5483 /** Move the cursor to the previous data item. */
5485 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5491 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5493 mp = mc->mc_pg[mc->mc_top];
5495 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5496 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5497 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5498 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5499 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5500 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5501 if (rc == MDB_SUCCESS) {
5502 MDB_GET_KEY(leaf, key);
5503 mc->mc_flags &= ~C_EOF;
5509 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5510 if (op == MDB_PREV_DUP)
5511 return MDB_NOTFOUND;
5515 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5516 mdb_dbg_pgno(mp), (void *) mc));
5518 if (mc->mc_ki[mc->mc_top] == 0) {
5519 DPUTS("=====> move to prev sibling page");
5520 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5523 mp = mc->mc_pg[mc->mc_top];
5524 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5525 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5527 mc->mc_ki[mc->mc_top]--;
5529 mc->mc_flags &= ~C_EOF;
5531 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5532 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5535 key->mv_size = mc->mc_db->md_pad;
5536 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5540 mdb_cassert(mc, IS_LEAF(mp));
5541 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5543 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5544 mdb_xcursor_init1(mc, leaf);
5547 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5550 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5551 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5552 if (rc != MDB_SUCCESS)
5557 MDB_GET_KEY(leaf, key);
5561 /** Set the cursor on a specific data item. */
5563 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5564 MDB_cursor_op op, int *exactp)
5568 MDB_node *leaf = NULL;
5571 if (key->mv_size == 0)
5572 return MDB_BAD_VALSIZE;
5575 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5577 /* See if we're already on the right page */
5578 if (mc->mc_flags & C_INITIALIZED) {
5581 mp = mc->mc_pg[mc->mc_top];
5583 mc->mc_ki[mc->mc_top] = 0;
5584 return MDB_NOTFOUND;
5586 if (mp->mp_flags & P_LEAF2) {
5587 nodekey.mv_size = mc->mc_db->md_pad;
5588 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5590 leaf = NODEPTR(mp, 0);
5591 MDB_GET_KEY2(leaf, nodekey);
5593 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5595 /* Probably happens rarely, but first node on the page
5596 * was the one we wanted.
5598 mc->mc_ki[mc->mc_top] = 0;
5605 unsigned int nkeys = NUMKEYS(mp);
5607 if (mp->mp_flags & P_LEAF2) {
5608 nodekey.mv_data = LEAF2KEY(mp,
5609 nkeys-1, nodekey.mv_size);
5611 leaf = NODEPTR(mp, nkeys-1);
5612 MDB_GET_KEY2(leaf, nodekey);
5614 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5616 /* last node was the one we wanted */
5617 mc->mc_ki[mc->mc_top] = nkeys-1;
5623 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5624 /* This is definitely the right page, skip search_page */
5625 if (mp->mp_flags & P_LEAF2) {
5626 nodekey.mv_data = LEAF2KEY(mp,
5627 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5629 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5630 MDB_GET_KEY2(leaf, nodekey);
5632 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5634 /* current node was the one we wanted */
5644 /* If any parents have right-sibs, search.
5645 * Otherwise, there's nothing further.
5647 for (i=0; i<mc->mc_top; i++)
5649 NUMKEYS(mc->mc_pg[i])-1)
5651 if (i == mc->mc_top) {
5652 /* There are no other pages */
5653 mc->mc_ki[mc->mc_top] = nkeys;
5654 return MDB_NOTFOUND;
5658 /* There are no other pages */
5659 mc->mc_ki[mc->mc_top] = 0;
5660 if (op == MDB_SET_RANGE && !exactp) {
5664 return MDB_NOTFOUND;
5668 rc = mdb_page_search(mc, key, 0);
5669 if (rc != MDB_SUCCESS)
5672 mp = mc->mc_pg[mc->mc_top];
5673 mdb_cassert(mc, IS_LEAF(mp));
5676 leaf = mdb_node_search(mc, key, exactp);
5677 if (exactp != NULL && !*exactp) {
5678 /* MDB_SET specified and not an exact match. */
5679 return MDB_NOTFOUND;
5683 DPUTS("===> inexact leaf not found, goto sibling");
5684 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5685 return rc; /* no entries matched */
5686 mp = mc->mc_pg[mc->mc_top];
5687 mdb_cassert(mc, IS_LEAF(mp));
5688 leaf = NODEPTR(mp, 0);
5692 mc->mc_flags |= C_INITIALIZED;
5693 mc->mc_flags &= ~C_EOF;
5696 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5697 key->mv_size = mc->mc_db->md_pad;
5698 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5703 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5704 mdb_xcursor_init1(mc, leaf);
5707 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5708 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5709 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5712 if (op == MDB_GET_BOTH) {
5718 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5719 if (rc != MDB_SUCCESS)
5722 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5724 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5726 rc = mc->mc_dbx->md_dcmp(data, &d2);
5728 if (op == MDB_GET_BOTH || rc > 0)
5729 return MDB_NOTFOUND;
5736 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5737 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5742 /* The key already matches in all other cases */
5743 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5744 MDB_GET_KEY(leaf, key);
5745 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5750 /** Move the cursor to the first item in the database. */
5752 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5758 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5760 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5761 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5762 if (rc != MDB_SUCCESS)
5765 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5767 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5768 mc->mc_flags |= C_INITIALIZED;
5769 mc->mc_flags &= ~C_EOF;
5771 mc->mc_ki[mc->mc_top] = 0;
5773 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5774 key->mv_size = mc->mc_db->md_pad;
5775 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5780 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5781 mdb_xcursor_init1(mc, leaf);
5782 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5786 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5790 MDB_GET_KEY(leaf, key);
5794 /** Move the cursor to the last item in the database. */
5796 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5802 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5804 if (!(mc->mc_flags & C_EOF)) {
5806 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5807 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5808 if (rc != MDB_SUCCESS)
5811 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5814 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5815 mc->mc_flags |= C_INITIALIZED|C_EOF;
5816 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5818 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5819 key->mv_size = mc->mc_db->md_pad;
5820 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5825 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5826 mdb_xcursor_init1(mc, leaf);
5827 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5831 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5836 MDB_GET_KEY(leaf, key);
5841 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5846 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5851 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5855 case MDB_GET_CURRENT:
5856 if (!(mc->mc_flags & C_INITIALIZED)) {
5859 MDB_page *mp = mc->mc_pg[mc->mc_top];
5860 int nkeys = NUMKEYS(mp);
5861 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5862 mc->mc_ki[mc->mc_top] = nkeys;
5868 key->mv_size = mc->mc_db->md_pad;
5869 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5871 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5872 MDB_GET_KEY(leaf, key);
5874 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5875 if (mc->mc_flags & C_DEL)
5876 mdb_xcursor_init1(mc, leaf);
5877 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5879 rc = mdb_node_read(mc->mc_txn, leaf, data);
5886 case MDB_GET_BOTH_RANGE:
5891 if (mc->mc_xcursor == NULL) {
5892 rc = MDB_INCOMPATIBLE;
5902 rc = mdb_cursor_set(mc, key, data, op,
5903 op == MDB_SET_RANGE ? NULL : &exact);
5906 case MDB_GET_MULTIPLE:
5907 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5911 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5912 rc = MDB_INCOMPATIBLE;
5916 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5917 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5920 case MDB_NEXT_MULTIPLE:
5925 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5926 rc = MDB_INCOMPATIBLE;
5929 if (!(mc->mc_flags & C_INITIALIZED))
5930 rc = mdb_cursor_first(mc, key, data);
5932 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5933 if (rc == MDB_SUCCESS) {
5934 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5937 mx = &mc->mc_xcursor->mx_cursor;
5938 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5940 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5941 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5949 case MDB_NEXT_NODUP:
5950 if (!(mc->mc_flags & C_INITIALIZED))
5951 rc = mdb_cursor_first(mc, key, data);
5953 rc = mdb_cursor_next(mc, key, data, op);
5957 case MDB_PREV_NODUP:
5958 if (!(mc->mc_flags & C_INITIALIZED)) {
5959 rc = mdb_cursor_last(mc, key, data);
5962 mc->mc_flags |= C_INITIALIZED;
5963 mc->mc_ki[mc->mc_top]++;
5965 rc = mdb_cursor_prev(mc, key, data, op);
5968 rc = mdb_cursor_first(mc, key, data);
5971 mfunc = mdb_cursor_first;
5973 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5977 if (mc->mc_xcursor == NULL) {
5978 rc = MDB_INCOMPATIBLE;
5982 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5983 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5984 MDB_GET_KEY(leaf, key);
5985 rc = mdb_node_read(mc->mc_txn, leaf, data);
5989 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5993 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5996 rc = mdb_cursor_last(mc, key, data);
5999 mfunc = mdb_cursor_last;
6002 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6007 if (mc->mc_flags & C_DEL)
6008 mc->mc_flags ^= C_DEL;
6013 /** Touch all the pages in the cursor stack. Set mc_top.
6014 * Makes sure all the pages are writable, before attempting a write operation.
6015 * @param[in] mc The cursor to operate on.
6018 mdb_cursor_touch(MDB_cursor *mc)
6020 int rc = MDB_SUCCESS;
6022 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6025 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6027 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6028 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6031 *mc->mc_dbflag |= DB_DIRTY;
6036 rc = mdb_page_touch(mc);
6037 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6038 mc->mc_top = mc->mc_snum-1;
6043 /** Do not spill pages to disk if txn is getting full, may fail instead */
6044 #define MDB_NOSPILL 0x8000
6047 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6050 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6052 MDB_node *leaf = NULL;
6055 MDB_val xdata, *rdata, dkey, olddata;
6057 int do_sub = 0, insert_key, insert_data;
6058 unsigned int mcount = 0, dcount = 0, nospill;
6061 unsigned int nflags;
6064 if (mc == NULL || key == NULL)
6067 env = mc->mc_txn->mt_env;
6069 /* Check this first so counter will always be zero on any
6072 if (flags & MDB_MULTIPLE) {
6073 dcount = data[1].mv_size;
6074 data[1].mv_size = 0;
6075 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6076 return MDB_INCOMPATIBLE;
6079 nospill = flags & MDB_NOSPILL;
6080 flags &= ~MDB_NOSPILL;
6082 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6083 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6085 if (key->mv_size-1 >= ENV_MAXKEY(env))
6086 return MDB_BAD_VALSIZE;
6088 #if SIZE_MAX > MAXDATASIZE
6089 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6090 return MDB_BAD_VALSIZE;
6092 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6093 return MDB_BAD_VALSIZE;
6096 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6097 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6101 if (flags == MDB_CURRENT) {
6102 if (!(mc->mc_flags & C_INITIALIZED))
6105 } else if (mc->mc_db->md_root == P_INVALID) {
6106 /* new database, cursor has nothing to point to */
6109 mc->mc_flags &= ~C_INITIALIZED;
6114 if (flags & MDB_APPEND) {
6116 rc = mdb_cursor_last(mc, &k2, &d2);
6118 rc = mc->mc_dbx->md_cmp(key, &k2);
6121 mc->mc_ki[mc->mc_top]++;
6123 /* new key is <= last key */
6128 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6130 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6131 DPRINTF(("duplicate key [%s]", DKEY(key)));
6133 return MDB_KEYEXIST;
6135 if (rc && rc != MDB_NOTFOUND)
6139 if (mc->mc_flags & C_DEL)
6140 mc->mc_flags ^= C_DEL;
6142 /* Cursor is positioned, check for room in the dirty list */
6144 if (flags & MDB_MULTIPLE) {
6146 xdata.mv_size = data->mv_size * dcount;
6150 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6154 if (rc == MDB_NO_ROOT) {
6156 /* new database, write a root leaf page */
6157 DPUTS("allocating new root leaf page");
6158 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6161 mdb_cursor_push(mc, np);
6162 mc->mc_db->md_root = np->mp_pgno;
6163 mc->mc_db->md_depth++;
6164 *mc->mc_dbflag |= DB_DIRTY;
6165 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6167 np->mp_flags |= P_LEAF2;
6168 mc->mc_flags |= C_INITIALIZED;
6170 /* make sure all cursor pages are writable */
6171 rc2 = mdb_cursor_touch(mc);
6176 insert_key = insert_data = rc;
6178 /* The key does not exist */
6179 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6180 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6181 LEAFSIZE(key, data) > env->me_nodemax)
6183 /* Too big for a node, insert in sub-DB. Set up an empty
6184 * "old sub-page" for prep_subDB to expand to a full page.
6186 fp_flags = P_LEAF|P_DIRTY;
6188 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6189 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6190 olddata.mv_size = PAGEHDRSZ;
6194 /* there's only a key anyway, so this is a no-op */
6195 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6197 unsigned int ksize = mc->mc_db->md_pad;
6198 if (key->mv_size != ksize)
6199 return MDB_BAD_VALSIZE;
6200 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6201 memcpy(ptr, key->mv_data, ksize);
6203 /* if overwriting slot 0 of leaf, need to
6204 * update branch key if there is a parent page
6206 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6207 unsigned short top = mc->mc_top;
6209 /* slot 0 is always an empty key, find real slot */
6210 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6212 if (mc->mc_ki[mc->mc_top])
6213 rc2 = mdb_update_key(mc, key);
6224 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6225 olddata.mv_size = NODEDSZ(leaf);
6226 olddata.mv_data = NODEDATA(leaf);
6229 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6230 /* Prepare (sub-)page/sub-DB to accept the new item,
6231 * if needed. fp: old sub-page or a header faking
6232 * it. mp: new (sub-)page. offset: growth in page
6233 * size. xdata: node data with new page or DB.
6235 unsigned i, offset = 0;
6236 mp = fp = xdata.mv_data = env->me_pbuf;
6237 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6239 /* Was a single item before, must convert now */
6240 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6241 /* Just overwrite the current item */
6242 if (flags == MDB_CURRENT)
6245 #if UINT_MAX < SIZE_MAX
6246 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6247 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6249 /* does data match? */
6250 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6251 if (flags & MDB_NODUPDATA)
6252 return MDB_KEYEXIST;
6257 /* Back up original data item */
6258 dkey.mv_size = olddata.mv_size;
6259 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6261 /* Make sub-page header for the dup items, with dummy body */
6262 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6263 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6264 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6265 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6266 fp->mp_flags |= P_LEAF2;
6267 fp->mp_pad = data->mv_size;
6268 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6270 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6271 (dkey.mv_size & 1) + (data->mv_size & 1);
6273 fp->mp_upper = xdata.mv_size - PAGEBASE;
6274 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6275 } else if (leaf->mn_flags & F_SUBDATA) {
6276 /* Data is on sub-DB, just store it */
6277 flags |= F_DUPDATA|F_SUBDATA;
6280 /* Data is on sub-page */
6281 fp = olddata.mv_data;
6284 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6285 offset = EVEN(NODESIZE + sizeof(indx_t) +
6289 offset = fp->mp_pad;
6290 if (SIZELEFT(fp) < offset) {
6291 offset *= 4; /* space for 4 more */
6294 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6296 fp->mp_flags |= P_DIRTY;
6297 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6298 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6302 xdata.mv_size = olddata.mv_size + offset;
6305 fp_flags = fp->mp_flags;
6306 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6307 /* Too big for a sub-page, convert to sub-DB */
6308 fp_flags &= ~P_SUBP;
6310 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6311 fp_flags |= P_LEAF2;
6312 dummy.md_pad = fp->mp_pad;
6313 dummy.md_flags = MDB_DUPFIXED;
6314 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6315 dummy.md_flags |= MDB_INTEGERKEY;
6321 dummy.md_branch_pages = 0;
6322 dummy.md_leaf_pages = 1;
6323 dummy.md_overflow_pages = 0;
6324 dummy.md_entries = NUMKEYS(fp);
6325 xdata.mv_size = sizeof(MDB_db);
6326 xdata.mv_data = &dummy;
6327 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6329 offset = env->me_psize - olddata.mv_size;
6330 flags |= F_DUPDATA|F_SUBDATA;
6331 dummy.md_root = mp->mp_pgno;
6334 mp->mp_flags = fp_flags | P_DIRTY;
6335 mp->mp_pad = fp->mp_pad;
6336 mp->mp_lower = fp->mp_lower;
6337 mp->mp_upper = fp->mp_upper + offset;
6338 if (fp_flags & P_LEAF2) {
6339 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6341 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6342 olddata.mv_size - fp->mp_upper - PAGEBASE);
6343 for (i=0; i<NUMKEYS(fp); i++)
6344 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6352 mdb_node_del(mc, 0);
6356 /* overflow page overwrites need special handling */
6357 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6360 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6362 memcpy(&pg, olddata.mv_data, sizeof(pg));
6363 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6365 ovpages = omp->mp_pages;
6367 /* Is the ov page large enough? */
6368 if (ovpages >= dpages) {
6369 if (!(omp->mp_flags & P_DIRTY) &&
6370 (level || (env->me_flags & MDB_WRITEMAP)))
6372 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6375 level = 0; /* dirty in this txn or clean */
6378 if (omp->mp_flags & P_DIRTY) {
6379 /* yes, overwrite it. Note in this case we don't
6380 * bother to try shrinking the page if the new data
6381 * is smaller than the overflow threshold.
6384 /* It is writable only in a parent txn */
6385 size_t sz = (size_t) env->me_psize * ovpages, off;
6386 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6392 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6393 mdb_cassert(mc, rc2 == 0);
6394 if (!(flags & MDB_RESERVE)) {
6395 /* Copy end of page, adjusting alignment so
6396 * compiler may copy words instead of bytes.
6398 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6399 memcpy((size_t *)((char *)np + off),
6400 (size_t *)((char *)omp + off), sz - off);
6403 memcpy(np, omp, sz); /* Copy beginning of page */
6406 SETDSZ(leaf, data->mv_size);
6407 if (F_ISSET(flags, MDB_RESERVE))
6408 data->mv_data = METADATA(omp);
6410 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6414 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6416 } else if (data->mv_size == olddata.mv_size) {
6417 /* same size, just replace it. Note that we could
6418 * also reuse this node if the new data is smaller,
6419 * but instead we opt to shrink the node in that case.
6421 if (F_ISSET(flags, MDB_RESERVE))
6422 data->mv_data = olddata.mv_data;
6423 else if (!(mc->mc_flags & C_SUB))
6424 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6426 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6431 mdb_node_del(mc, 0);
6437 nflags = flags & NODE_ADD_FLAGS;
6438 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6439 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6440 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6441 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6443 nflags |= MDB_SPLIT_REPLACE;
6444 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6446 /* There is room already in this leaf page. */
6447 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6448 if (rc == 0 && insert_key) {
6449 /* Adjust other cursors pointing to mp */
6450 MDB_cursor *m2, *m3;
6451 MDB_dbi dbi = mc->mc_dbi;
6452 unsigned i = mc->mc_top;
6453 MDB_page *mp = mc->mc_pg[i];
6455 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6456 if (mc->mc_flags & C_SUB)
6457 m3 = &m2->mc_xcursor->mx_cursor;
6460 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6461 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6468 if (rc == MDB_SUCCESS) {
6469 /* Now store the actual data in the child DB. Note that we're
6470 * storing the user data in the keys field, so there are strict
6471 * size limits on dupdata. The actual data fields of the child
6472 * DB are all zero size.
6480 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6481 if (flags & MDB_CURRENT) {
6482 xflags = MDB_CURRENT|MDB_NOSPILL;
6484 mdb_xcursor_init1(mc, leaf);
6485 xflags = (flags & MDB_NODUPDATA) ?
6486 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6488 /* converted, write the original data first */
6490 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6494 /* Adjust other cursors pointing to mp */
6496 unsigned i = mc->mc_top;
6497 MDB_page *mp = mc->mc_pg[i];
6499 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6500 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6501 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6502 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6503 mdb_xcursor_init1(m2, leaf);
6507 /* we've done our job */
6510 ecount = mc->mc_xcursor->mx_db.md_entries;
6511 if (flags & MDB_APPENDDUP)
6512 xflags |= MDB_APPEND;
6513 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6514 if (flags & F_SUBDATA) {
6515 void *db = NODEDATA(leaf);
6516 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6518 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6520 /* Increment count unless we just replaced an existing item. */
6522 mc->mc_db->md_entries++;
6524 /* Invalidate txn if we created an empty sub-DB */
6527 /* If we succeeded and the key didn't exist before,
6528 * make sure the cursor is marked valid.
6530 mc->mc_flags |= C_INITIALIZED;
6532 if (flags & MDB_MULTIPLE) {
6535 /* let caller know how many succeeded, if any */
6536 data[1].mv_size = mcount;
6537 if (mcount < dcount) {
6538 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6539 insert_key = insert_data = 0;
6546 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6549 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6554 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6560 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6561 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6563 if (!(mc->mc_flags & C_INITIALIZED))
6566 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6567 return MDB_NOTFOUND;
6569 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6572 rc = mdb_cursor_touch(mc);
6576 mp = mc->mc_pg[mc->mc_top];
6579 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6581 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6582 if (flags & MDB_NODUPDATA) {
6583 /* mdb_cursor_del0() will subtract the final entry */
6584 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6586 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6587 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6589 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6592 /* If sub-DB still has entries, we're done */
6593 if (mc->mc_xcursor->mx_db.md_entries) {
6594 if (leaf->mn_flags & F_SUBDATA) {
6595 /* update subDB info */
6596 void *db = NODEDATA(leaf);
6597 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6600 /* shrink fake page */
6601 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6602 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6603 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6604 /* fix other sub-DB cursors pointed at this fake page */
6605 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6606 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6607 if (m2->mc_pg[mc->mc_top] == mp &&
6608 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6609 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6612 mc->mc_db->md_entries--;
6613 mc->mc_flags |= C_DEL;
6616 /* otherwise fall thru and delete the sub-DB */
6619 if (leaf->mn_flags & F_SUBDATA) {
6620 /* add all the child DB's pages to the free list */
6621 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6627 /* add overflow pages to free list */
6628 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6632 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6633 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6634 (rc = mdb_ovpage_free(mc, omp)))
6639 return mdb_cursor_del0(mc);
6642 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6646 /** Allocate and initialize new pages for a database.
6647 * @param[in] mc a cursor on the database being added to.
6648 * @param[in] flags flags defining what type of page is being allocated.
6649 * @param[in] num the number of pages to allocate. This is usually 1,
6650 * unless allocating overflow pages for a large record.
6651 * @param[out] mp Address of a page, or NULL on failure.
6652 * @return 0 on success, non-zero on failure.
6655 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6660 if ((rc = mdb_page_alloc(mc, num, &np)))
6662 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6663 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6664 np->mp_flags = flags | P_DIRTY;
6665 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6666 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6669 mc->mc_db->md_branch_pages++;
6670 else if (IS_LEAF(np))
6671 mc->mc_db->md_leaf_pages++;
6672 else if (IS_OVERFLOW(np)) {
6673 mc->mc_db->md_overflow_pages += num;
6681 /** Calculate the size of a leaf node.
6682 * The size depends on the environment's page size; if a data item
6683 * is too large it will be put onto an overflow page and the node
6684 * size will only include the key and not the data. Sizes are always
6685 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6686 * of the #MDB_node headers.
6687 * @param[in] env The environment handle.
6688 * @param[in] key The key for the node.
6689 * @param[in] data The data for the node.
6690 * @return The number of bytes needed to store the node.
6693 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6697 sz = LEAFSIZE(key, data);
6698 if (sz > env->me_nodemax) {
6699 /* put on overflow page */
6700 sz -= data->mv_size - sizeof(pgno_t);
6703 return EVEN(sz + sizeof(indx_t));
6706 /** Calculate the size of a branch node.
6707 * The size should depend on the environment's page size but since
6708 * we currently don't support spilling large keys onto overflow
6709 * pages, it's simply the size of the #MDB_node header plus the
6710 * size of the key. Sizes are always rounded up to an even number
6711 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6712 * @param[in] env The environment handle.
6713 * @param[in] key The key for the node.
6714 * @return The number of bytes needed to store the node.
6717 mdb_branch_size(MDB_env *env, MDB_val *key)
6722 if (sz > env->me_nodemax) {
6723 /* put on overflow page */
6724 /* not implemented */
6725 /* sz -= key->size - sizeof(pgno_t); */
6728 return sz + sizeof(indx_t);
6731 /** Add a node to the page pointed to by the cursor.
6732 * @param[in] mc The cursor for this operation.
6733 * @param[in] indx The index on the page where the new node should be added.
6734 * @param[in] key The key for the new node.
6735 * @param[in] data The data for the new node, if any.
6736 * @param[in] pgno The page number, if adding a branch node.
6737 * @param[in] flags Flags for the node.
6738 * @return 0 on success, non-zero on failure. Possible errors are:
6740 * <li>ENOMEM - failed to allocate overflow pages for the node.
6741 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6742 * should never happen since all callers already calculate the
6743 * page's free space before calling this function.
6747 mdb_node_add(MDB_cursor *mc, indx_t indx,
6748 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6751 size_t node_size = NODESIZE;
6755 MDB_page *mp = mc->mc_pg[mc->mc_top];
6756 MDB_page *ofp = NULL; /* overflow page */
6759 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6761 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6762 IS_LEAF(mp) ? "leaf" : "branch",
6763 IS_SUBP(mp) ? "sub-" : "",
6764 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6765 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6768 /* Move higher keys up one slot. */
6769 int ksize = mc->mc_db->md_pad, dif;
6770 char *ptr = LEAF2KEY(mp, indx, ksize);
6771 dif = NUMKEYS(mp) - indx;
6773 memmove(ptr+ksize, ptr, dif*ksize);
6774 /* insert new key */
6775 memcpy(ptr, key->mv_data, ksize);
6777 /* Just using these for counting */
6778 mp->mp_lower += sizeof(indx_t);
6779 mp->mp_upper -= ksize - sizeof(indx_t);
6783 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6785 node_size += key->mv_size;
6787 mdb_cassert(mc, data);
6788 if (F_ISSET(flags, F_BIGDATA)) {
6789 /* Data already on overflow page. */
6790 node_size += sizeof(pgno_t);
6791 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6792 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6794 /* Put data on overflow page. */
6795 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6796 data->mv_size, node_size+data->mv_size));
6797 node_size = EVEN(node_size + sizeof(pgno_t));
6798 if ((ssize_t)node_size > room)
6800 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6802 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6806 node_size += data->mv_size;
6809 node_size = EVEN(node_size);
6810 if ((ssize_t)node_size > room)
6814 /* Move higher pointers up one slot. */
6815 for (i = NUMKEYS(mp); i > indx; i--)
6816 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6818 /* Adjust free space offsets. */
6819 ofs = mp->mp_upper - node_size;
6820 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6821 mp->mp_ptrs[indx] = ofs;
6823 mp->mp_lower += sizeof(indx_t);
6825 /* Write the node data. */
6826 node = NODEPTR(mp, indx);
6827 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6828 node->mn_flags = flags;
6830 SETDSZ(node,data->mv_size);
6835 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6838 mdb_cassert(mc, key);
6840 if (F_ISSET(flags, F_BIGDATA))
6841 memcpy(node->mn_data + key->mv_size, data->mv_data,
6843 else if (F_ISSET(flags, MDB_RESERVE))
6844 data->mv_data = node->mn_data + key->mv_size;
6846 memcpy(node->mn_data + key->mv_size, data->mv_data,
6849 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6851 if (F_ISSET(flags, MDB_RESERVE))
6852 data->mv_data = METADATA(ofp);
6854 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6861 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6862 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6863 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6864 DPRINTF(("node size = %"Z"u", node_size));
6865 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6866 return MDB_PAGE_FULL;
6869 /** Delete the specified node from a page.
6870 * @param[in] mc Cursor pointing to the node to delete.
6871 * @param[in] ksize The size of a node. Only used if the page is
6872 * part of a #MDB_DUPFIXED database.
6875 mdb_node_del(MDB_cursor *mc, int ksize)
6877 MDB_page *mp = mc->mc_pg[mc->mc_top];
6878 indx_t indx = mc->mc_ki[mc->mc_top];
6880 indx_t i, j, numkeys, ptr;
6884 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6885 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6886 numkeys = NUMKEYS(mp);
6887 mdb_cassert(mc, indx < numkeys);
6890 int x = numkeys - 1 - indx;
6891 base = LEAF2KEY(mp, indx, ksize);
6893 memmove(base, base + ksize, x * ksize);
6894 mp->mp_lower -= sizeof(indx_t);
6895 mp->mp_upper += ksize - sizeof(indx_t);
6899 node = NODEPTR(mp, indx);
6900 sz = NODESIZE + node->mn_ksize;
6902 if (F_ISSET(node->mn_flags, F_BIGDATA))
6903 sz += sizeof(pgno_t);
6905 sz += NODEDSZ(node);
6909 ptr = mp->mp_ptrs[indx];
6910 for (i = j = 0; i < numkeys; i++) {
6912 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6913 if (mp->mp_ptrs[i] < ptr)
6914 mp->mp_ptrs[j] += sz;
6919 base = (char *)mp + mp->mp_upper + PAGEBASE;
6920 memmove(base + sz, base, ptr - mp->mp_upper);
6922 mp->mp_lower -= sizeof(indx_t);
6926 /** Compact the main page after deleting a node on a subpage.
6927 * @param[in] mp The main page to operate on.
6928 * @param[in] indx The index of the subpage on the main page.
6931 mdb_node_shrink(MDB_page *mp, indx_t indx)
6937 indx_t i, numkeys, ptr;
6939 node = NODEPTR(mp, indx);
6940 sp = (MDB_page *)NODEDATA(node);
6941 delta = SIZELEFT(sp);
6942 xp = (MDB_page *)((char *)sp + delta);
6944 /* shift subpage upward */
6946 nsize = NUMKEYS(sp) * sp->mp_pad;
6948 return; /* do not make the node uneven-sized */
6949 memmove(METADATA(xp), METADATA(sp), nsize);
6952 numkeys = NUMKEYS(sp);
6953 for (i=numkeys-1; i>=0; i--)
6954 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6956 xp->mp_upper = sp->mp_lower;
6957 xp->mp_lower = sp->mp_lower;
6958 xp->mp_flags = sp->mp_flags;
6959 xp->mp_pad = sp->mp_pad;
6960 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6962 nsize = NODEDSZ(node) - delta;
6963 SETDSZ(node, nsize);
6965 /* shift lower nodes upward */
6966 ptr = mp->mp_ptrs[indx];
6967 numkeys = NUMKEYS(mp);
6968 for (i = 0; i < numkeys; i++) {
6969 if (mp->mp_ptrs[i] <= ptr)
6970 mp->mp_ptrs[i] += delta;
6973 base = (char *)mp + mp->mp_upper + PAGEBASE;
6974 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6975 mp->mp_upper += delta;
6978 /** Initial setup of a sorted-dups cursor.
6979 * Sorted duplicates are implemented as a sub-database for the given key.
6980 * The duplicate data items are actually keys of the sub-database.
6981 * Operations on the duplicate data items are performed using a sub-cursor
6982 * initialized when the sub-database is first accessed. This function does
6983 * the preliminary setup of the sub-cursor, filling in the fields that
6984 * depend only on the parent DB.
6985 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6988 mdb_xcursor_init0(MDB_cursor *mc)
6990 MDB_xcursor *mx = mc->mc_xcursor;
6992 mx->mx_cursor.mc_xcursor = NULL;
6993 mx->mx_cursor.mc_txn = mc->mc_txn;
6994 mx->mx_cursor.mc_db = &mx->mx_db;
6995 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6996 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6997 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6998 mx->mx_cursor.mc_snum = 0;
6999 mx->mx_cursor.mc_top = 0;
7000 mx->mx_cursor.mc_flags = C_SUB;
7001 mx->mx_dbx.md_name.mv_size = 0;
7002 mx->mx_dbx.md_name.mv_data = NULL;
7003 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7004 mx->mx_dbx.md_dcmp = NULL;
7005 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7008 /** Final setup of a sorted-dups cursor.
7009 * Sets up the fields that depend on the data from the main cursor.
7010 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7011 * @param[in] node The data containing the #MDB_db record for the
7012 * sorted-dup database.
7015 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7017 MDB_xcursor *mx = mc->mc_xcursor;
7019 if (node->mn_flags & F_SUBDATA) {
7020 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7021 mx->mx_cursor.mc_pg[0] = 0;
7022 mx->mx_cursor.mc_snum = 0;
7023 mx->mx_cursor.mc_top = 0;
7024 mx->mx_cursor.mc_flags = C_SUB;
7026 MDB_page *fp = NODEDATA(node);
7027 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7028 mx->mx_db.md_flags = 0;
7029 mx->mx_db.md_depth = 1;
7030 mx->mx_db.md_branch_pages = 0;
7031 mx->mx_db.md_leaf_pages = 1;
7032 mx->mx_db.md_overflow_pages = 0;
7033 mx->mx_db.md_entries = NUMKEYS(fp);
7034 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7035 mx->mx_cursor.mc_snum = 1;
7036 mx->mx_cursor.mc_top = 0;
7037 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7038 mx->mx_cursor.mc_pg[0] = fp;
7039 mx->mx_cursor.mc_ki[0] = 0;
7040 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7041 mx->mx_db.md_flags = MDB_DUPFIXED;
7042 mx->mx_db.md_pad = fp->mp_pad;
7043 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7044 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7047 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7048 mx->mx_db.md_root));
7049 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7050 #if UINT_MAX < SIZE_MAX
7051 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7052 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7056 /** Initialize a cursor for a given transaction and database. */
7058 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7061 mc->mc_backup = NULL;
7064 mc->mc_db = &txn->mt_dbs[dbi];
7065 mc->mc_dbx = &txn->mt_dbxs[dbi];
7066 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7071 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7072 mdb_tassert(txn, mx != NULL);
7073 mc->mc_xcursor = mx;
7074 mdb_xcursor_init0(mc);
7076 mc->mc_xcursor = NULL;
7078 if (*mc->mc_dbflag & DB_STALE) {
7079 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7084 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7087 size_t size = sizeof(MDB_cursor);
7089 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7092 if (txn->mt_flags & MDB_TXN_ERROR)
7095 /* Allow read access to the freelist */
7096 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7099 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7100 size += sizeof(MDB_xcursor);
7102 if ((mc = malloc(size)) != NULL) {
7103 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7104 if (txn->mt_cursors) {
7105 mc->mc_next = txn->mt_cursors[dbi];
7106 txn->mt_cursors[dbi] = mc;
7107 mc->mc_flags |= C_UNTRACK;
7119 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7121 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7124 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7127 if (txn->mt_flags & MDB_TXN_ERROR)
7130 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7134 /* Return the count of duplicate data items for the current key */
7136 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7140 if (mc == NULL || countp == NULL)
7143 if (mc->mc_xcursor == NULL)
7144 return MDB_INCOMPATIBLE;
7146 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7149 if (!(mc->mc_flags & C_INITIALIZED))
7152 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7153 return MDB_NOTFOUND;
7155 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7156 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7159 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7162 *countp = mc->mc_xcursor->mx_db.md_entries;
7168 mdb_cursor_close(MDB_cursor *mc)
7170 if (mc && !mc->mc_backup) {
7171 /* remove from txn, if tracked */
7172 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7173 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7174 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7176 *prev = mc->mc_next;
7183 mdb_cursor_txn(MDB_cursor *mc)
7185 if (!mc) return NULL;
7190 mdb_cursor_dbi(MDB_cursor *mc)
7195 /** Replace the key for a branch node with a new key.
7196 * @param[in] mc Cursor pointing to the node to operate on.
7197 * @param[in] key The new key to use.
7198 * @return 0 on success, non-zero on failure.
7201 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7207 int delta, ksize, oksize;
7208 indx_t ptr, i, numkeys, indx;
7211 indx = mc->mc_ki[mc->mc_top];
7212 mp = mc->mc_pg[mc->mc_top];
7213 node = NODEPTR(mp, indx);
7214 ptr = mp->mp_ptrs[indx];
7218 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7219 k2.mv_data = NODEKEY(node);
7220 k2.mv_size = node->mn_ksize;
7221 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7223 mdb_dkey(&k2, kbuf2),
7229 /* Sizes must be 2-byte aligned. */
7230 ksize = EVEN(key->mv_size);
7231 oksize = EVEN(node->mn_ksize);
7232 delta = ksize - oksize;
7234 /* Shift node contents if EVEN(key length) changed. */
7236 if (delta > 0 && SIZELEFT(mp) < delta) {
7238 /* not enough space left, do a delete and split */
7239 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7240 pgno = NODEPGNO(node);
7241 mdb_node_del(mc, 0);
7242 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7245 numkeys = NUMKEYS(mp);
7246 for (i = 0; i < numkeys; i++) {
7247 if (mp->mp_ptrs[i] <= ptr)
7248 mp->mp_ptrs[i] -= delta;
7251 base = (char *)mp + mp->mp_upper + PAGEBASE;
7252 len = ptr - mp->mp_upper + NODESIZE;
7253 memmove(base - delta, base, len);
7254 mp->mp_upper -= delta;
7256 node = NODEPTR(mp, indx);
7259 /* But even if no shift was needed, update ksize */
7260 if (node->mn_ksize != key->mv_size)
7261 node->mn_ksize = key->mv_size;
7264 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7270 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7272 /** Move a node from csrc to cdst.
7275 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7282 unsigned short flags;
7286 /* Mark src and dst as dirty. */
7287 if ((rc = mdb_page_touch(csrc)) ||
7288 (rc = mdb_page_touch(cdst)))
7291 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7292 key.mv_size = csrc->mc_db->md_pad;
7293 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7295 data.mv_data = NULL;
7299 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7300 mdb_cassert(csrc, !((size_t)srcnode & 1));
7301 srcpg = NODEPGNO(srcnode);
7302 flags = srcnode->mn_flags;
7303 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7304 unsigned int snum = csrc->mc_snum;
7306 /* must find the lowest key below src */
7307 rc = mdb_page_search_lowest(csrc);
7310 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7311 key.mv_size = csrc->mc_db->md_pad;
7312 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7314 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7315 key.mv_size = NODEKSZ(s2);
7316 key.mv_data = NODEKEY(s2);
7318 csrc->mc_snum = snum--;
7319 csrc->mc_top = snum;
7321 key.mv_size = NODEKSZ(srcnode);
7322 key.mv_data = NODEKEY(srcnode);
7324 data.mv_size = NODEDSZ(srcnode);
7325 data.mv_data = NODEDATA(srcnode);
7327 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7328 unsigned int snum = cdst->mc_snum;
7331 /* must find the lowest key below dst */
7332 mdb_cursor_copy(cdst, &mn);
7333 rc = mdb_page_search_lowest(&mn);
7336 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7337 bkey.mv_size = mn.mc_db->md_pad;
7338 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7340 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7341 bkey.mv_size = NODEKSZ(s2);
7342 bkey.mv_data = NODEKEY(s2);
7344 mn.mc_snum = snum--;
7347 rc = mdb_update_key(&mn, &bkey);
7352 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7353 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7354 csrc->mc_ki[csrc->mc_top],
7356 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7357 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7359 /* Add the node to the destination page.
7361 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7362 if (rc != MDB_SUCCESS)
7365 /* Delete the node from the source page.
7367 mdb_node_del(csrc, key.mv_size);
7370 /* Adjust other cursors pointing to mp */
7371 MDB_cursor *m2, *m3;
7372 MDB_dbi dbi = csrc->mc_dbi;
7373 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7375 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7376 if (csrc->mc_flags & C_SUB)
7377 m3 = &m2->mc_xcursor->mx_cursor;
7380 if (m3 == csrc) continue;
7381 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7382 csrc->mc_ki[csrc->mc_top]) {
7383 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7384 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7389 /* Update the parent separators.
7391 if (csrc->mc_ki[csrc->mc_top] == 0) {
7392 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7393 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7394 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7396 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7397 key.mv_size = NODEKSZ(srcnode);
7398 key.mv_data = NODEKEY(srcnode);
7400 DPRINTF(("update separator for source page %"Z"u to [%s]",
7401 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7402 mdb_cursor_copy(csrc, &mn);
7405 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7408 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7410 indx_t ix = csrc->mc_ki[csrc->mc_top];
7411 nullkey.mv_size = 0;
7412 csrc->mc_ki[csrc->mc_top] = 0;
7413 rc = mdb_update_key(csrc, &nullkey);
7414 csrc->mc_ki[csrc->mc_top] = ix;
7415 mdb_cassert(csrc, rc == MDB_SUCCESS);
7419 if (cdst->mc_ki[cdst->mc_top] == 0) {
7420 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7421 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7422 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7424 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7425 key.mv_size = NODEKSZ(srcnode);
7426 key.mv_data = NODEKEY(srcnode);
7428 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7429 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7430 mdb_cursor_copy(cdst, &mn);
7433 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7436 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7438 indx_t ix = cdst->mc_ki[cdst->mc_top];
7439 nullkey.mv_size = 0;
7440 cdst->mc_ki[cdst->mc_top] = 0;
7441 rc = mdb_update_key(cdst, &nullkey);
7442 cdst->mc_ki[cdst->mc_top] = ix;
7443 mdb_cassert(csrc, rc == MDB_SUCCESS);
7450 /** Merge one page into another.
7451 * The nodes from the page pointed to by \b csrc will
7452 * be copied to the page pointed to by \b cdst and then
7453 * the \b csrc page will be freed.
7454 * @param[in] csrc Cursor pointing to the source page.
7455 * @param[in] cdst Cursor pointing to the destination page.
7456 * @return 0 on success, non-zero on failure.
7459 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7461 MDB_page *psrc, *pdst;
7468 psrc = csrc->mc_pg[csrc->mc_top];
7469 pdst = cdst->mc_pg[cdst->mc_top];
7471 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7473 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7474 mdb_cassert(csrc, cdst->mc_snum > 1);
7476 /* Mark dst as dirty. */
7477 if ((rc = mdb_page_touch(cdst)))
7480 /* Move all nodes from src to dst.
7482 j = nkeys = NUMKEYS(pdst);
7483 if (IS_LEAF2(psrc)) {
7484 key.mv_size = csrc->mc_db->md_pad;
7485 key.mv_data = METADATA(psrc);
7486 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7487 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7488 if (rc != MDB_SUCCESS)
7490 key.mv_data = (char *)key.mv_data + key.mv_size;
7493 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7494 srcnode = NODEPTR(psrc, i);
7495 if (i == 0 && IS_BRANCH(psrc)) {
7498 mdb_cursor_copy(csrc, &mn);
7499 /* must find the lowest key below src */
7500 rc = mdb_page_search_lowest(&mn);
7503 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7504 key.mv_size = mn.mc_db->md_pad;
7505 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7507 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7508 key.mv_size = NODEKSZ(s2);
7509 key.mv_data = NODEKEY(s2);
7512 key.mv_size = srcnode->mn_ksize;
7513 key.mv_data = NODEKEY(srcnode);
7516 data.mv_size = NODEDSZ(srcnode);
7517 data.mv_data = NODEDATA(srcnode);
7518 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7519 if (rc != MDB_SUCCESS)
7524 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7525 pdst->mp_pgno, NUMKEYS(pdst),
7526 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7528 /* Unlink the src page from parent and add to free list.
7531 mdb_node_del(csrc, 0);
7532 if (csrc->mc_ki[csrc->mc_top] == 0) {
7534 rc = mdb_update_key(csrc, &key);
7542 psrc = csrc->mc_pg[csrc->mc_top];
7543 /* If not operating on FreeDB, allow this page to be reused
7544 * in this txn. Otherwise just add to free list.
7546 rc = mdb_page_loose(csrc, psrc);
7550 csrc->mc_db->md_leaf_pages--;
7552 csrc->mc_db->md_branch_pages--;
7554 /* Adjust other cursors pointing to mp */
7555 MDB_cursor *m2, *m3;
7556 MDB_dbi dbi = csrc->mc_dbi;
7558 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7559 if (csrc->mc_flags & C_SUB)
7560 m3 = &m2->mc_xcursor->mx_cursor;
7563 if (m3 == csrc) continue;
7564 if (m3->mc_snum < csrc->mc_snum) continue;
7565 if (m3->mc_pg[csrc->mc_top] == psrc) {
7566 m3->mc_pg[csrc->mc_top] = pdst;
7567 m3->mc_ki[csrc->mc_top] += nkeys;
7572 unsigned int snum = cdst->mc_snum;
7573 uint16_t depth = cdst->mc_db->md_depth;
7574 mdb_cursor_pop(cdst);
7575 rc = mdb_rebalance(cdst);
7576 /* Did the tree shrink? */
7577 if (depth > cdst->mc_db->md_depth)
7579 cdst->mc_snum = snum;
7580 cdst->mc_top = snum-1;
7585 /** Copy the contents of a cursor.
7586 * @param[in] csrc The cursor to copy from.
7587 * @param[out] cdst The cursor to copy to.
7590 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7594 cdst->mc_txn = csrc->mc_txn;
7595 cdst->mc_dbi = csrc->mc_dbi;
7596 cdst->mc_db = csrc->mc_db;
7597 cdst->mc_dbx = csrc->mc_dbx;
7598 cdst->mc_snum = csrc->mc_snum;
7599 cdst->mc_top = csrc->mc_top;
7600 cdst->mc_flags = csrc->mc_flags;
7602 for (i=0; i<csrc->mc_snum; i++) {
7603 cdst->mc_pg[i] = csrc->mc_pg[i];
7604 cdst->mc_ki[i] = csrc->mc_ki[i];
7608 /** Rebalance the tree after a delete operation.
7609 * @param[in] mc Cursor pointing to the page where rebalancing
7611 * @return 0 on success, non-zero on failure.
7614 mdb_rebalance(MDB_cursor *mc)
7618 unsigned int ptop, minkeys;
7622 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7623 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7624 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7625 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7626 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7628 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7629 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7630 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7631 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7635 if (mc->mc_snum < 2) {
7636 MDB_page *mp = mc->mc_pg[0];
7638 DPUTS("Can't rebalance a subpage, ignoring");
7641 if (NUMKEYS(mp) == 0) {
7642 DPUTS("tree is completely empty");
7643 mc->mc_db->md_root = P_INVALID;
7644 mc->mc_db->md_depth = 0;
7645 mc->mc_db->md_leaf_pages = 0;
7646 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7649 /* Adjust cursors pointing to mp */
7652 mc->mc_flags &= ~C_INITIALIZED;
7654 MDB_cursor *m2, *m3;
7655 MDB_dbi dbi = mc->mc_dbi;
7657 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7658 if (mc->mc_flags & C_SUB)
7659 m3 = &m2->mc_xcursor->mx_cursor;
7662 if (m3->mc_snum < mc->mc_snum) continue;
7663 if (m3->mc_pg[0] == mp) {
7666 m3->mc_flags &= ~C_INITIALIZED;
7670 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7672 DPUTS("collapsing root page!");
7673 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7676 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7677 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7680 mc->mc_db->md_depth--;
7681 mc->mc_db->md_branch_pages--;
7682 mc->mc_ki[0] = mc->mc_ki[1];
7683 for (i = 1; i<mc->mc_db->md_depth; i++) {
7684 mc->mc_pg[i] = mc->mc_pg[i+1];
7685 mc->mc_ki[i] = mc->mc_ki[i+1];
7688 /* Adjust other cursors pointing to mp */
7689 MDB_cursor *m2, *m3;
7690 MDB_dbi dbi = mc->mc_dbi;
7692 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7693 if (mc->mc_flags & C_SUB)
7694 m3 = &m2->mc_xcursor->mx_cursor;
7697 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7698 if (m3->mc_pg[0] == mp) {
7701 for (i=0; i<m3->mc_snum; i++) {
7702 m3->mc_pg[i] = m3->mc_pg[i+1];
7703 m3->mc_ki[i] = m3->mc_ki[i+1];
7709 DPUTS("root page doesn't need rebalancing");
7713 /* The parent (branch page) must have at least 2 pointers,
7714 * otherwise the tree is invalid.
7716 ptop = mc->mc_top-1;
7717 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7719 /* Leaf page fill factor is below the threshold.
7720 * Try to move keys from left or right neighbor, or
7721 * merge with a neighbor page.
7726 mdb_cursor_copy(mc, &mn);
7727 mn.mc_xcursor = NULL;
7729 oldki = mc->mc_ki[mc->mc_top];
7730 if (mc->mc_ki[ptop] == 0) {
7731 /* We're the leftmost leaf in our parent.
7733 DPUTS("reading right neighbor");
7735 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7736 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7739 mn.mc_ki[mn.mc_top] = 0;
7740 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7742 /* There is at least one neighbor to the left.
7744 DPUTS("reading left neighbor");
7746 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7747 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7750 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7751 mc->mc_ki[mc->mc_top] = 0;
7754 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7755 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7756 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7758 /* If the neighbor page is above threshold and has enough keys,
7759 * move one key from it. Otherwise we should try to merge them.
7760 * (A branch page must never have less than 2 keys.)
7762 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7763 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7764 rc = mdb_node_move(&mn, mc);
7765 if (mc->mc_ki[ptop]) {
7769 if (mc->mc_ki[ptop] == 0) {
7770 rc = mdb_page_merge(&mn, mc);
7772 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7773 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7774 rc = mdb_page_merge(mc, &mn);
7775 mdb_cursor_copy(&mn, mc);
7777 mc->mc_flags &= ~C_EOF;
7779 mc->mc_ki[mc->mc_top] = oldki;
7783 /** Complete a delete operation started by #mdb_cursor_del(). */
7785 mdb_cursor_del0(MDB_cursor *mc)
7792 ki = mc->mc_ki[mc->mc_top];
7793 mdb_node_del(mc, mc->mc_db->md_pad);
7794 mc->mc_db->md_entries--;
7795 rc = mdb_rebalance(mc);
7797 if (rc == MDB_SUCCESS) {
7798 MDB_cursor *m2, *m3;
7799 MDB_dbi dbi = mc->mc_dbi;
7801 mp = mc->mc_pg[mc->mc_top];
7802 nkeys = NUMKEYS(mp);
7804 /* if mc points past last node in page, find next sibling */
7805 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7806 rc = mdb_cursor_sibling(mc, 1);
7807 if (rc == MDB_NOTFOUND) {
7808 mc->mc_flags |= C_EOF;
7813 /* Adjust other cursors pointing to mp */
7814 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7815 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7816 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7818 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7820 if (m3->mc_pg[mc->mc_top] == mp) {
7821 if (m3->mc_ki[mc->mc_top] >= ki) {
7822 m3->mc_flags |= C_DEL;
7823 if (m3->mc_ki[mc->mc_top] > ki)
7824 m3->mc_ki[mc->mc_top]--;
7825 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7826 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7828 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7829 rc = mdb_cursor_sibling(m3, 1);
7830 if (rc == MDB_NOTFOUND) {
7831 m3->mc_flags |= C_EOF;
7837 mc->mc_flags |= C_DEL;
7841 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7846 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7847 MDB_val *key, MDB_val *data)
7849 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7852 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7853 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7855 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7856 /* must ignore any data */
7860 return mdb_del0(txn, dbi, key, data, 0);
7864 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7865 MDB_val *key, MDB_val *data, unsigned flags)
7870 MDB_val rdata, *xdata;
7874 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7876 mdb_cursor_init(&mc, txn, dbi, &mx);
7885 flags |= MDB_NODUPDATA;
7887 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7889 /* let mdb_page_split know about this cursor if needed:
7890 * delete will trigger a rebalance; if it needs to move
7891 * a node from one page to another, it will have to
7892 * update the parent's separator key(s). If the new sepkey
7893 * is larger than the current one, the parent page may
7894 * run out of space, triggering a split. We need this
7895 * cursor to be consistent until the end of the rebalance.
7897 mc.mc_flags |= C_UNTRACK;
7898 mc.mc_next = txn->mt_cursors[dbi];
7899 txn->mt_cursors[dbi] = &mc;
7900 rc = mdb_cursor_del(&mc, flags);
7901 txn->mt_cursors[dbi] = mc.mc_next;
7906 /** Split a page and insert a new node.
7907 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7908 * The cursor will be updated to point to the actual page and index where
7909 * the node got inserted after the split.
7910 * @param[in] newkey The key for the newly inserted node.
7911 * @param[in] newdata The data for the newly inserted node.
7912 * @param[in] newpgno The page number, if the new node is a branch node.
7913 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7914 * @return 0 on success, non-zero on failure.
7917 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7918 unsigned int nflags)
7921 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7924 int i, j, split_indx, nkeys, pmax;
7925 MDB_env *env = mc->mc_txn->mt_env;
7927 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7928 MDB_page *copy = NULL;
7929 MDB_page *mp, *rp, *pp;
7934 mp = mc->mc_pg[mc->mc_top];
7935 newindx = mc->mc_ki[mc->mc_top];
7936 nkeys = NUMKEYS(mp);
7938 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7939 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7940 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7942 /* Create a right sibling. */
7943 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7945 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7947 if (mc->mc_snum < 2) {
7948 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7950 /* shift current top to make room for new parent */
7951 mc->mc_pg[1] = mc->mc_pg[0];
7952 mc->mc_ki[1] = mc->mc_ki[0];
7955 mc->mc_db->md_root = pp->mp_pgno;
7956 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7957 mc->mc_db->md_depth++;
7960 /* Add left (implicit) pointer. */
7961 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7962 /* undo the pre-push */
7963 mc->mc_pg[0] = mc->mc_pg[1];
7964 mc->mc_ki[0] = mc->mc_ki[1];
7965 mc->mc_db->md_root = mp->mp_pgno;
7966 mc->mc_db->md_depth--;
7973 ptop = mc->mc_top-1;
7974 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7977 mc->mc_flags |= C_SPLITTING;
7978 mdb_cursor_copy(mc, &mn);
7979 mn.mc_pg[mn.mc_top] = rp;
7980 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7982 if (nflags & MDB_APPEND) {
7983 mn.mc_ki[mn.mc_top] = 0;
7985 split_indx = newindx;
7989 split_indx = (nkeys+1) / 2;
7994 unsigned int lsize, rsize, ksize;
7995 /* Move half of the keys to the right sibling */
7996 x = mc->mc_ki[mc->mc_top] - split_indx;
7997 ksize = mc->mc_db->md_pad;
7998 split = LEAF2KEY(mp, split_indx, ksize);
7999 rsize = (nkeys - split_indx) * ksize;
8000 lsize = (nkeys - split_indx) * sizeof(indx_t);
8001 mp->mp_lower -= lsize;
8002 rp->mp_lower += lsize;
8003 mp->mp_upper += rsize - lsize;
8004 rp->mp_upper -= rsize - lsize;
8005 sepkey.mv_size = ksize;
8006 if (newindx == split_indx) {
8007 sepkey.mv_data = newkey->mv_data;
8009 sepkey.mv_data = split;
8012 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8013 memcpy(rp->mp_ptrs, split, rsize);
8014 sepkey.mv_data = rp->mp_ptrs;
8015 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8016 memcpy(ins, newkey->mv_data, ksize);
8017 mp->mp_lower += sizeof(indx_t);
8018 mp->mp_upper -= ksize - sizeof(indx_t);
8021 memcpy(rp->mp_ptrs, split, x * ksize);
8022 ins = LEAF2KEY(rp, x, ksize);
8023 memcpy(ins, newkey->mv_data, ksize);
8024 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8025 rp->mp_lower += sizeof(indx_t);
8026 rp->mp_upper -= ksize - sizeof(indx_t);
8027 mc->mc_ki[mc->mc_top] = x;
8028 mc->mc_pg[mc->mc_top] = rp;
8031 int psize, nsize, k;
8032 /* Maximum free space in an empty page */
8033 pmax = env->me_psize - PAGEHDRSZ;
8035 nsize = mdb_leaf_size(env, newkey, newdata);
8037 nsize = mdb_branch_size(env, newkey);
8038 nsize = EVEN(nsize);
8040 /* grab a page to hold a temporary copy */
8041 copy = mdb_page_malloc(mc->mc_txn, 1);
8046 copy->mp_pgno = mp->mp_pgno;
8047 copy->mp_flags = mp->mp_flags;
8048 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8049 copy->mp_upper = env->me_psize - PAGEBASE;
8051 /* prepare to insert */
8052 for (i=0, j=0; i<nkeys; i++) {
8054 copy->mp_ptrs[j++] = 0;
8056 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8059 /* When items are relatively large the split point needs
8060 * to be checked, because being off-by-one will make the
8061 * difference between success or failure in mdb_node_add.
8063 * It's also relevant if a page happens to be laid out
8064 * such that one half of its nodes are all "small" and
8065 * the other half of its nodes are "large." If the new
8066 * item is also "large" and falls on the half with
8067 * "large" nodes, it also may not fit.
8069 * As a final tweak, if the new item goes on the last
8070 * spot on the page (and thus, onto the new page), bias
8071 * the split so the new page is emptier than the old page.
8072 * This yields better packing during sequential inserts.
8074 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8075 /* Find split point */
8077 if (newindx <= split_indx || newindx >= nkeys) {
8079 k = newindx >= nkeys ? nkeys : split_indx+2;
8084 for (; i!=k; i+=j) {
8089 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8090 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8092 if (F_ISSET(node->mn_flags, F_BIGDATA))
8093 psize += sizeof(pgno_t);
8095 psize += NODEDSZ(node);
8097 psize = EVEN(psize);
8099 if (psize > pmax || i == k-j) {
8100 split_indx = i + (j<0);
8105 if (split_indx == newindx) {
8106 sepkey.mv_size = newkey->mv_size;
8107 sepkey.mv_data = newkey->mv_data;
8109 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8110 sepkey.mv_size = node->mn_ksize;
8111 sepkey.mv_data = NODEKEY(node);
8116 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8118 /* Copy separator key to the parent.
8120 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8124 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8129 if (mn.mc_snum == mc->mc_snum) {
8130 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8131 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8132 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8133 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8138 /* Right page might now have changed parent.
8139 * Check if left page also changed parent.
8141 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8142 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8143 for (i=0; i<ptop; i++) {
8144 mc->mc_pg[i] = mn.mc_pg[i];
8145 mc->mc_ki[i] = mn.mc_ki[i];
8147 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8148 if (mn.mc_ki[ptop]) {
8149 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8151 /* find right page's left sibling */
8152 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8153 mdb_cursor_sibling(mc, 0);
8158 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8161 mc->mc_flags ^= C_SPLITTING;
8162 if (rc != MDB_SUCCESS) {
8165 if (nflags & MDB_APPEND) {
8166 mc->mc_pg[mc->mc_top] = rp;
8167 mc->mc_ki[mc->mc_top] = 0;
8168 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8171 for (i=0; i<mc->mc_top; i++)
8172 mc->mc_ki[i] = mn.mc_ki[i];
8173 } else if (!IS_LEAF2(mp)) {
8175 mc->mc_pg[mc->mc_top] = rp;
8180 rkey.mv_data = newkey->mv_data;
8181 rkey.mv_size = newkey->mv_size;
8187 /* Update index for the new key. */
8188 mc->mc_ki[mc->mc_top] = j;
8190 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8191 rkey.mv_data = NODEKEY(node);
8192 rkey.mv_size = node->mn_ksize;
8194 xdata.mv_data = NODEDATA(node);
8195 xdata.mv_size = NODEDSZ(node);
8198 pgno = NODEPGNO(node);
8199 flags = node->mn_flags;
8202 if (!IS_LEAF(mp) && j == 0) {
8203 /* First branch index doesn't need key data. */
8207 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8213 mc->mc_pg[mc->mc_top] = copy;
8218 } while (i != split_indx);
8220 nkeys = NUMKEYS(copy);
8221 for (i=0; i<nkeys; i++)
8222 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8223 mp->mp_lower = copy->mp_lower;
8224 mp->mp_upper = copy->mp_upper;
8225 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8226 env->me_psize - copy->mp_upper - PAGEBASE);
8228 /* reset back to original page */
8229 if (newindx < split_indx) {
8230 mc->mc_pg[mc->mc_top] = mp;
8231 if (nflags & MDB_RESERVE) {
8232 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8233 if (!(node->mn_flags & F_BIGDATA))
8234 newdata->mv_data = NODEDATA(node);
8237 mc->mc_pg[mc->mc_top] = rp;
8239 /* Make sure mc_ki is still valid.
8241 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8242 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8243 for (i=0; i<=ptop; i++) {
8244 mc->mc_pg[i] = mn.mc_pg[i];
8245 mc->mc_ki[i] = mn.mc_ki[i];
8252 /* Adjust other cursors pointing to mp */
8253 MDB_cursor *m2, *m3;
8254 MDB_dbi dbi = mc->mc_dbi;
8255 int fixup = NUMKEYS(mp);
8257 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8258 if (mc->mc_flags & C_SUB)
8259 m3 = &m2->mc_xcursor->mx_cursor;
8264 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8266 if (m3->mc_flags & C_SPLITTING)
8271 for (k=m3->mc_top; k>=0; k--) {
8272 m3->mc_ki[k+1] = m3->mc_ki[k];
8273 m3->mc_pg[k+1] = m3->mc_pg[k];
8275 if (m3->mc_ki[0] >= split_indx) {
8280 m3->mc_pg[0] = mc->mc_pg[0];
8284 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8285 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8286 m3->mc_ki[mc->mc_top]++;
8287 if (m3->mc_ki[mc->mc_top] >= fixup) {
8288 m3->mc_pg[mc->mc_top] = rp;
8289 m3->mc_ki[mc->mc_top] -= fixup;
8290 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8292 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8293 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8298 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8301 if (copy) /* tmp page */
8302 mdb_page_free(env, copy);
8304 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8309 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8310 MDB_val *key, MDB_val *data, unsigned int flags)
8315 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8318 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8321 mdb_cursor_init(&mc, txn, dbi, &mx);
8322 return mdb_cursor_put(&mc, key, data, flags);
8326 #define MDB_WBUF (1024*1024)
8329 /** State needed for a compacting copy. */
8330 typedef struct mdb_copy {
8331 pthread_mutex_t mc_mutex;
8332 pthread_cond_t mc_cond;
8339 pgno_t mc_next_pgno;
8342 volatile int mc_new;
8347 /** Dedicated writer thread for compacting copy. */
8348 static THREAD_RET ESECT
8349 mdb_env_copythr(void *arg)
8353 int toggle = 0, wsize, rc;
8356 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8359 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8362 pthread_mutex_lock(&my->mc_mutex);
8364 pthread_cond_signal(&my->mc_cond);
8367 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8368 if (my->mc_new < 0) {
8373 wsize = my->mc_wlen[toggle];
8374 ptr = my->mc_wbuf[toggle];
8377 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8381 } else if (len > 0) {
8395 /* If there's an overflow page tail, write it too */
8396 if (my->mc_olen[toggle]) {
8397 wsize = my->mc_olen[toggle];
8398 ptr = my->mc_over[toggle];
8399 my->mc_olen[toggle] = 0;
8402 my->mc_wlen[toggle] = 0;
8404 pthread_cond_signal(&my->mc_cond);
8406 pthread_cond_signal(&my->mc_cond);
8407 pthread_mutex_unlock(&my->mc_mutex);
8408 return (THREAD_RET)0;
8412 /** Tell the writer thread there's a buffer ready to write */
8414 mdb_env_cthr_toggle(mdb_copy *my, int st)
8416 int toggle = my->mc_toggle ^ 1;
8417 pthread_mutex_lock(&my->mc_mutex);
8418 if (my->mc_status) {
8419 pthread_mutex_unlock(&my->mc_mutex);
8420 return my->mc_status;
8422 while (my->mc_new == 1)
8423 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8425 my->mc_toggle = toggle;
8426 pthread_cond_signal(&my->mc_cond);
8427 pthread_mutex_unlock(&my->mc_mutex);
8431 /** Depth-first tree traversal for compacting copy. */
8433 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8436 MDB_txn *txn = my->mc_txn;
8438 MDB_page *mo, *mp, *leaf;
8443 /* Empty DB, nothing to do */
8444 if (*pg == P_INVALID)
8451 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8454 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8458 /* Make cursor pages writable */
8459 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8463 for (i=0; i<mc.mc_top; i++) {
8464 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8465 mc.mc_pg[i] = (MDB_page *)ptr;
8466 ptr += my->mc_env->me_psize;
8469 /* This is writable space for a leaf page. Usually not needed. */
8470 leaf = (MDB_page *)ptr;
8472 toggle = my->mc_toggle;
8473 while (mc.mc_snum > 0) {
8475 mp = mc.mc_pg[mc.mc_top];
8479 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8480 for (i=0; i<n; i++) {
8481 ni = NODEPTR(mp, i);
8482 if (ni->mn_flags & F_BIGDATA) {
8486 /* Need writable leaf */
8488 mc.mc_pg[mc.mc_top] = leaf;
8489 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8491 ni = NODEPTR(mp, i);
8494 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8495 rc = mdb_page_get(txn, pg, &omp, NULL);
8498 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8499 rc = mdb_env_cthr_toggle(my, 1);
8502 toggle = my->mc_toggle;
8504 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8505 memcpy(mo, omp, my->mc_env->me_psize);
8506 mo->mp_pgno = my->mc_next_pgno;
8507 my->mc_next_pgno += omp->mp_pages;
8508 my->mc_wlen[toggle] += my->mc_env->me_psize;
8509 if (omp->mp_pages > 1) {
8510 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8511 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8512 rc = mdb_env_cthr_toggle(my, 1);
8515 toggle = my->mc_toggle;
8517 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8518 } else if (ni->mn_flags & F_SUBDATA) {
8521 /* Need writable leaf */
8523 mc.mc_pg[mc.mc_top] = leaf;
8524 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8526 ni = NODEPTR(mp, i);
8529 memcpy(&db, NODEDATA(ni), sizeof(db));
8530 my->mc_toggle = toggle;
8531 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8534 toggle = my->mc_toggle;
8535 memcpy(NODEDATA(ni), &db, sizeof(db));
8540 mc.mc_ki[mc.mc_top]++;
8541 if (mc.mc_ki[mc.mc_top] < n) {
8544 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8546 rc = mdb_page_get(txn, pg, &mp, NULL);
8551 mc.mc_ki[mc.mc_top] = 0;
8552 if (IS_BRANCH(mp)) {
8553 /* Whenever we advance to a sibling branch page,
8554 * we must proceed all the way down to its first leaf.
8556 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8559 mc.mc_pg[mc.mc_top] = mp;
8563 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8564 rc = mdb_env_cthr_toggle(my, 1);
8567 toggle = my->mc_toggle;
8569 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8570 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8571 mo->mp_pgno = my->mc_next_pgno++;
8572 my->mc_wlen[toggle] += my->mc_env->me_psize;
8574 /* Update parent if there is one */
8575 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8576 SETPGNO(ni, mo->mp_pgno);
8577 mdb_cursor_pop(&mc);
8579 /* Otherwise we're done */
8589 /** Copy environment with compaction. */
8591 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8596 MDB_txn *txn = NULL;
8601 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8602 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8603 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8604 if (my.mc_wbuf[0] == NULL)
8607 pthread_mutex_init(&my.mc_mutex, NULL);
8608 pthread_cond_init(&my.mc_cond, NULL);
8609 #ifdef HAVE_MEMALIGN
8610 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8611 if (my.mc_wbuf[0] == NULL)
8614 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8619 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8620 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8625 my.mc_next_pgno = 2;
8631 THREAD_CREATE(thr, mdb_env_copythr, &my);
8633 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8637 mp = (MDB_page *)my.mc_wbuf[0];
8638 memset(mp, 0, 2*env->me_psize);
8640 mp->mp_flags = P_META;
8641 mm = (MDB_meta *)METADATA(mp);
8642 mdb_env_init_meta0(env, mm);
8643 mm->mm_address = env->me_metas[0]->mm_address;
8645 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8647 mp->mp_flags = P_META;
8648 *(MDB_meta *)METADATA(mp) = *mm;
8649 mm = (MDB_meta *)METADATA(mp);
8651 /* Count the number of free pages, subtract from lastpg to find
8652 * number of active pages
8655 MDB_ID freecount = 0;
8658 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8659 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8660 freecount += *(MDB_ID *)data.mv_data;
8661 freecount += txn->mt_dbs[0].md_branch_pages +
8662 txn->mt_dbs[0].md_leaf_pages +
8663 txn->mt_dbs[0].md_overflow_pages;
8665 /* Set metapage 1 */
8666 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8667 mm->mm_dbs[1] = txn->mt_dbs[1];
8668 if (mm->mm_last_pg > 1) {
8669 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8672 mm->mm_dbs[1].md_root = P_INVALID;
8675 my.mc_wlen[0] = env->me_psize * 2;
8677 pthread_mutex_lock(&my.mc_mutex);
8679 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8680 pthread_mutex_unlock(&my.mc_mutex);
8681 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8682 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8683 rc = mdb_env_cthr_toggle(&my, 1);
8684 mdb_env_cthr_toggle(&my, -1);
8685 pthread_mutex_lock(&my.mc_mutex);
8687 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8688 pthread_mutex_unlock(&my.mc_mutex);
8693 CloseHandle(my.mc_cond);
8694 CloseHandle(my.mc_mutex);
8695 _aligned_free(my.mc_wbuf[0]);
8697 pthread_cond_destroy(&my.mc_cond);
8698 pthread_mutex_destroy(&my.mc_mutex);
8699 free(my.mc_wbuf[0]);
8704 /** Copy environment as-is. */
8706 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8708 MDB_txn *txn = NULL;
8709 mdb_mutex_t *wmutex = NULL;
8715 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8719 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8722 /* Do the lock/unlock of the reader mutex before starting the
8723 * write txn. Otherwise other read txns could block writers.
8725 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8730 /* We must start the actual read txn after blocking writers */
8731 mdb_txn_reset0(txn, "reset-stage1");
8733 /* Temporarily block writers until we snapshot the meta pages */
8734 wmutex = MDB_MUTEX(env, w);
8735 if (LOCK_MUTEX(rc, env, wmutex))
8738 rc = mdb_txn_renew0(txn);
8740 UNLOCK_MUTEX(wmutex);
8745 wsize = env->me_psize * 2;
8749 DO_WRITE(rc, fd, ptr, w2, len);
8753 } else if (len > 0) {
8759 /* Non-blocking or async handles are not supported */
8765 UNLOCK_MUTEX(wmutex);
8770 w2 = txn->mt_next_pgno * env->me_psize;
8773 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8780 if (wsize > MAX_WRITE)
8784 DO_WRITE(rc, fd, ptr, w2, len);
8788 } else if (len > 0) {
8805 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8807 if (flags & MDB_CP_COMPACT)
8808 return mdb_env_copyfd1(env, fd);
8810 return mdb_env_copyfd0(env, fd);
8814 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8816 return mdb_env_copyfd2(env, fd, 0);
8820 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8824 HANDLE newfd = INVALID_HANDLE_VALUE;
8826 if (env->me_flags & MDB_NOSUBDIR) {
8827 lpath = (char *)path;
8830 len += sizeof(DATANAME);
8831 lpath = malloc(len);
8834 sprintf(lpath, "%s" DATANAME, path);
8837 /* The destination path must exist, but the destination file must not.
8838 * We don't want the OS to cache the writes, since the source data is
8839 * already in the OS cache.
8842 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8843 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8845 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8847 if (newfd == INVALID_HANDLE_VALUE) {
8852 if (env->me_psize >= env->me_os_psize) {
8854 /* Set O_DIRECT if the file system supports it */
8855 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8856 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8858 #ifdef F_NOCACHE /* __APPLE__ */
8859 rc = fcntl(newfd, F_NOCACHE, 1);
8867 rc = mdb_env_copyfd2(env, newfd, flags);
8870 if (!(env->me_flags & MDB_NOSUBDIR))
8872 if (newfd != INVALID_HANDLE_VALUE)
8873 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8880 mdb_env_copy(MDB_env *env, const char *path)
8882 return mdb_env_copy2(env, path, 0);
8886 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8888 if (flag & (env->me_map ? ~CHANGEABLE : ~(CHANGEABLE|CHANGELESS)))
8891 env->me_flags |= flag;
8893 env->me_flags &= ~flag;
8898 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8903 *arg = env->me_flags;
8908 mdb_env_set_userctx(MDB_env *env, void *ctx)
8912 env->me_userctx = ctx;
8917 mdb_env_get_userctx(MDB_env *env)
8919 return env ? env->me_userctx : NULL;
8923 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8928 env->me_assert_func = func;
8934 mdb_env_get_path(MDB_env *env, const char **arg)
8939 *arg = env->me_path;
8944 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8953 /** Common code for #mdb_stat() and #mdb_env_stat().
8954 * @param[in] env the environment to operate in.
8955 * @param[in] db the #MDB_db record containing the stats to return.
8956 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8957 * @return 0, this function always succeeds.
8960 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8962 arg->ms_psize = env->me_psize;
8963 arg->ms_depth = db->md_depth;
8964 arg->ms_branch_pages = db->md_branch_pages;
8965 arg->ms_leaf_pages = db->md_leaf_pages;
8966 arg->ms_overflow_pages = db->md_overflow_pages;
8967 arg->ms_entries = db->md_entries;
8973 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8977 if (env == NULL || arg == NULL)
8980 toggle = mdb_env_pick_meta(env);
8982 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8986 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8990 if (env == NULL || arg == NULL)
8993 toggle = mdb_env_pick_meta(env);
8994 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8995 arg->me_mapsize = env->me_mapsize;
8996 arg->me_maxreaders = env->me_maxreaders;
8998 /* me_numreaders may be zero if this process never used any readers. Use
8999 * the shared numreader count if it exists.
9001 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
9003 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9004 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9008 /** Set the default comparison functions for a database.
9009 * Called immediately after a database is opened to set the defaults.
9010 * The user can then override them with #mdb_set_compare() or
9011 * #mdb_set_dupsort().
9012 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9013 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9016 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9018 uint16_t f = txn->mt_dbs[dbi].md_flags;
9020 txn->mt_dbxs[dbi].md_cmp =
9021 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9022 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9024 txn->mt_dbxs[dbi].md_dcmp =
9025 !(f & MDB_DUPSORT) ? 0 :
9026 ((f & MDB_INTEGERDUP)
9027 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9028 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9031 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9037 int rc, dbflag, exact;
9038 unsigned int unused = 0, seq;
9041 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9042 mdb_default_cmp(txn, FREE_DBI);
9045 if ((flags & VALID_FLAGS) != flags)
9047 if (txn->mt_flags & MDB_TXN_ERROR)
9053 if (flags & PERSISTENT_FLAGS) {
9054 uint16_t f2 = flags & PERSISTENT_FLAGS;
9055 /* make sure flag changes get committed */
9056 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9057 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9058 txn->mt_flags |= MDB_TXN_DIRTY;
9061 mdb_default_cmp(txn, MAIN_DBI);
9065 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9066 mdb_default_cmp(txn, MAIN_DBI);
9069 /* Is the DB already open? */
9071 for (i=2; i<txn->mt_numdbs; i++) {
9072 if (!txn->mt_dbxs[i].md_name.mv_size) {
9073 /* Remember this free slot */
9074 if (!unused) unused = i;
9077 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9078 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9084 /* If no free slot and max hit, fail */
9085 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9086 return MDB_DBS_FULL;
9088 /* Cannot mix named databases with some mainDB flags */
9089 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9090 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9092 /* Find the DB info */
9093 dbflag = DB_NEW|DB_VALID;
9096 key.mv_data = (void *)name;
9097 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9098 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9099 if (rc == MDB_SUCCESS) {
9100 /* make sure this is actually a DB */
9101 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9102 if (!(node->mn_flags & F_SUBDATA))
9103 return MDB_INCOMPATIBLE;
9104 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9105 /* Create if requested */
9106 data.mv_size = sizeof(MDB_db);
9107 data.mv_data = &dummy;
9108 memset(&dummy, 0, sizeof(dummy));
9109 dummy.md_root = P_INVALID;
9110 dummy.md_flags = flags & PERSISTENT_FLAGS;
9111 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9115 /* OK, got info, add to table */
9116 if (rc == MDB_SUCCESS) {
9117 unsigned int slot = unused ? unused : txn->mt_numdbs;
9118 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9119 txn->mt_dbxs[slot].md_name.mv_size = len;
9120 txn->mt_dbxs[slot].md_rel = NULL;
9121 txn->mt_dbflags[slot] = dbflag;
9122 /* txn-> and env-> are the same in read txns, use
9123 * tmp variable to avoid undefined assignment
9125 seq = ++txn->mt_env->me_dbiseqs[slot];
9126 txn->mt_dbiseqs[slot] = seq;
9128 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9130 mdb_default_cmp(txn, slot);
9139 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9141 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9144 if (txn->mt_flags & MDB_TXN_ERROR)
9147 if (txn->mt_dbflags[dbi] & DB_STALE) {
9150 /* Stale, must read the DB's root. cursor_init does it for us. */
9151 mdb_cursor_init(&mc, txn, dbi, &mx);
9153 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9156 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9159 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9161 ptr = env->me_dbxs[dbi].md_name.mv_data;
9162 /* If there was no name, this was already closed */
9164 env->me_dbxs[dbi].md_name.mv_data = NULL;
9165 env->me_dbxs[dbi].md_name.mv_size = 0;
9166 env->me_dbflags[dbi] = 0;
9167 env->me_dbiseqs[dbi]++;
9172 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9174 /* We could return the flags for the FREE_DBI too but what's the point? */
9175 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9177 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9181 /** Add all the DB's pages to the free list.
9182 * @param[in] mc Cursor on the DB to free.
9183 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9184 * @return 0 on success, non-zero on failure.
9187 mdb_drop0(MDB_cursor *mc, int subs)
9191 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9192 if (rc == MDB_SUCCESS) {
9193 MDB_txn *txn = mc->mc_txn;
9198 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9199 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9202 mdb_cursor_copy(mc, &mx);
9203 while (mc->mc_snum > 0) {
9204 MDB_page *mp = mc->mc_pg[mc->mc_top];
9205 unsigned n = NUMKEYS(mp);
9207 for (i=0; i<n; i++) {
9208 ni = NODEPTR(mp, i);
9209 if (ni->mn_flags & F_BIGDATA) {
9212 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9213 rc = mdb_page_get(txn, pg, &omp, NULL);
9216 mdb_cassert(mc, IS_OVERFLOW(omp));
9217 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9221 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9222 mdb_xcursor_init1(mc, ni);
9223 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9229 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9231 for (i=0; i<n; i++) {
9233 ni = NODEPTR(mp, i);
9236 mdb_midl_xappend(txn->mt_free_pgs, pg);
9241 mc->mc_ki[mc->mc_top] = i;
9242 rc = mdb_cursor_sibling(mc, 1);
9244 if (rc != MDB_NOTFOUND)
9246 /* no more siblings, go back to beginning
9247 * of previous level.
9251 for (i=1; i<mc->mc_snum; i++) {
9253 mc->mc_pg[i] = mx.mc_pg[i];
9258 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9261 txn->mt_flags |= MDB_TXN_ERROR;
9262 } else if (rc == MDB_NOTFOUND) {
9268 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9270 MDB_cursor *mc, *m2;
9273 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9276 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9279 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9282 rc = mdb_cursor_open(txn, dbi, &mc);
9286 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9287 /* Invalidate the dropped DB's cursors */
9288 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9289 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9293 /* Can't delete the main DB */
9294 if (del && dbi > MAIN_DBI) {
9295 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9297 txn->mt_dbflags[dbi] = DB_STALE;
9298 mdb_dbi_close(txn->mt_env, dbi);
9300 txn->mt_flags |= MDB_TXN_ERROR;
9303 /* reset the DB record, mark it dirty */
9304 txn->mt_dbflags[dbi] |= DB_DIRTY;
9305 txn->mt_dbs[dbi].md_depth = 0;
9306 txn->mt_dbs[dbi].md_branch_pages = 0;
9307 txn->mt_dbs[dbi].md_leaf_pages = 0;
9308 txn->mt_dbs[dbi].md_overflow_pages = 0;
9309 txn->mt_dbs[dbi].md_entries = 0;
9310 txn->mt_dbs[dbi].md_root = P_INVALID;
9312 txn->mt_flags |= MDB_TXN_DIRTY;
9315 mdb_cursor_close(mc);
9319 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9321 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9324 txn->mt_dbxs[dbi].md_cmp = cmp;
9328 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9330 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9333 txn->mt_dbxs[dbi].md_dcmp = cmp;
9337 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9339 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9342 txn->mt_dbxs[dbi].md_rel = rel;
9346 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9348 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9351 txn->mt_dbxs[dbi].md_relctx = ctx;
9356 mdb_env_get_maxkeysize(MDB_env *env)
9358 return ENV_MAXKEY(env);
9362 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9364 unsigned int i, rdrs;
9367 int rc = 0, first = 1;
9371 if (!env->me_txns) {
9372 return func("(no reader locks)\n", ctx);
9374 rdrs = env->me_txns->mti_numreaders;
9375 mr = env->me_txns->mti_readers;
9376 for (i=0; i<rdrs; i++) {
9378 txnid_t txnid = mr[i].mr_txnid;
9379 sprintf(buf, txnid == (txnid_t)-1 ?
9380 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9381 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9384 rc = func(" pid thread txnid\n", ctx);
9388 rc = func(buf, ctx);
9394 rc = func("(no active readers)\n", ctx);
9399 /** Insert pid into list if not already present.
9400 * return -1 if already present.
9403 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9405 /* binary search of pid in list */
9407 unsigned cursor = 1;
9409 unsigned n = ids[0];
9412 unsigned pivot = n >> 1;
9413 cursor = base + pivot + 1;
9414 val = pid - ids[cursor];
9419 } else if ( val > 0 ) {
9424 /* found, so it's a duplicate */
9433 for (n = ids[0]; n > cursor; n--)
9440 mdb_reader_check(MDB_env *env, int *dead)
9446 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9449 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9450 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9452 mdb_mutex_t *rmutex = rlocked ? NULL : MDB_MUTEX(env, r);
9453 unsigned int i, j, rdrs;
9455 MDB_PID_T *pids, pid;
9456 int rc = MDB_SUCCESS, count = 0;
9458 rdrs = env->me_txns->mti_numreaders;
9459 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9463 mr = env->me_txns->mti_readers;
9464 for (i=0; i<rdrs; i++) {
9466 if (pid && pid != env->me_pid) {
9467 if (mdb_pid_insert(pids, pid) == 0) {
9468 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9469 /* Stale reader found */
9472 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9473 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9475 rdrs = 0; /* the above checked all readers */
9477 /* Recheck, a new process may have reused pid */
9478 if (mdb_reader_pid(env, Pidcheck, pid))
9483 if (mr[j].mr_pid == pid) {
9484 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9485 (unsigned) pid, mr[j].mr_txnid));
9490 UNLOCK_MUTEX(rmutex);
9501 #ifdef MDB_ROBUST_SUPPORTED
9502 /** Handle #LOCK_MUTEX0() failure.
9503 * With #MDB_ROBUST, try to repair the lock file if the mutex owner died.
9504 * @param[in] env the environment handle
9505 * @param[in] mutex LOCK_MUTEX0() mutex
9506 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9507 * @return 0 on success with the mutex locked, or an error code on failure.
9509 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc)
9511 int toggle, rlocked, rc2;
9513 enum { WAIT_ABANDONED = EOWNERDEAD };
9516 if (rc == (int) WAIT_ABANDONED) {
9517 /* We own the mutex. Clean up after dead previous owner. */
9519 rlocked = (mutex == MDB_MUTEX(env, r));
9521 /* Keep mti_txnid updated, otherwise next writer can
9522 * overwrite data which latest meta page refers to.
9524 toggle = mdb_env_pick_meta(env);
9525 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9526 /* env is hosed if the dead thread was ours */
9528 env->me_flags |= MDB_FATAL_ERROR;
9533 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9534 (rc ? "this process' env is hosed" : "recovering")));
9535 rc2 = mdb_reader_check0(env, rlocked, NULL);
9537 rc2 = pthread_mutex_consistent(mutex);
9538 if (rc || (rc = rc2)) {
9539 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9540 UNLOCK_MUTEX(mutex);
9546 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9551 #endif /* MDB_ROBUST_SUPPORTED */
projects / platform / upstream / lmdb.git / blob