2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2015 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments. You can safely
85 * define MDB_FDATASYNC_WORKS if this code will only be run
86 * on kernels 3.6 and newer.
88 #define BROKEN_FDATASYNC
101 #if defined(__sun) || defined(ANDROID)
102 /* Most platforms have posix_memalign, older may only have memalign */
103 #define HAVE_MEMALIGN 1
107 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
108 #include <netinet/in.h>
109 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
112 #if defined(__APPLE__) || defined (BSD)
113 # define MDB_USE_SYSV_SEM 1
114 # define MDB_FDATASYNC fsync
115 #elif defined(ANDROID)
116 # define MDB_FDATASYNC fsync
121 #ifdef MDB_USE_SYSV_SEM
124 #ifdef _SEM_SEMUN_UNDEFINED
127 struct semid_ds *buf;
128 unsigned short *array;
130 #endif /* _SEM_SEMUN_UNDEFINED */
132 #define MDB_USE_POSIX_MUTEX 1
133 #endif /* MDB_USE_SYSV_SEM */
136 #if defined(_WIN32) + defined(MDB_USE_SYSV_SEM) \
137 + defined(MDB_USE_POSIX_MUTEX) != 1
138 # error "Ambiguous shared-lock implementation"
142 #include <valgrind/memcheck.h>
143 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
144 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
145 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
146 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
147 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
149 #define VGMEMP_CREATE(h,r,z)
150 #define VGMEMP_ALLOC(h,a,s)
151 #define VGMEMP_FREE(h,a)
152 #define VGMEMP_DESTROY(h)
153 #define VGMEMP_DEFINED(a,s)
157 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
158 /* Solaris just defines one or the other */
159 # define LITTLE_ENDIAN 1234
160 # define BIG_ENDIAN 4321
161 # ifdef _LITTLE_ENDIAN
162 # define BYTE_ORDER LITTLE_ENDIAN
164 # define BYTE_ORDER BIG_ENDIAN
167 # define BYTE_ORDER __BYTE_ORDER
171 #ifndef LITTLE_ENDIAN
172 #define LITTLE_ENDIAN __LITTLE_ENDIAN
175 #define BIG_ENDIAN __BIG_ENDIAN
178 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
179 #define MISALIGNED_OK 1
185 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
186 # error "Unknown or unsupported endianness (BYTE_ORDER)"
187 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
188 # error "Two's complement, reasonably sized integer types, please"
192 /** Put infrequently used env functions in separate section */
194 # define ESECT __attribute__ ((section("__TEXT,text_env")))
196 # define ESECT __attribute__ ((section("text_env")))
202 /** @defgroup internal LMDB Internals
205 /** @defgroup compat Compatibility Macros
206 * A bunch of macros to minimize the amount of platform-specific ifdefs
207 * needed throughout the rest of the code. When the features this library
208 * needs are similar enough to POSIX to be hidden in a one-or-two line
209 * replacement, this macro approach is used.
213 /** Features under development */
218 /** Wrapper around __func__, which is a C99 feature */
219 #if __STDC_VERSION__ >= 199901L
220 # define mdb_func_ __func__
221 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
222 # define mdb_func_ __FUNCTION__
224 /* If a debug message says <mdb_unknown>(), update the #if statements above */
225 # define mdb_func_ "<mdb_unknown>"
228 /* Internal error codes, not exposed outside liblmdb */
229 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
231 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
232 #elif defined MDB_USE_SYSV_SEM
233 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
234 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
235 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
239 #define MDB_ROBUST_SUPPORTED 1
243 #define MDB_USE_HASH 1
244 #define MDB_PIDLOCK 0
245 #define THREAD_RET DWORD
246 #define pthread_t HANDLE
247 #define pthread_mutex_t HANDLE
248 #define pthread_cond_t HANDLE
249 typedef HANDLE mdb_mutex_t;
250 #define pthread_key_t DWORD
251 #define pthread_self() GetCurrentThreadId()
252 #define pthread_key_create(x,y) \
253 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
254 #define pthread_key_delete(x) TlsFree(x)
255 #define pthread_getspecific(x) TlsGetValue(x)
256 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
257 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
258 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
259 #define pthread_cond_signal(x) SetEvent(*x)
260 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
261 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
262 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
263 #define MDB_MUTEX(env, rw) ((env)->me_##rw##mutex)
264 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
265 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
266 #define mdb_mutex_consistent(mutex) 0
267 #define getpid() GetCurrentProcessId()
268 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
269 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
270 #define ErrCode() GetLastError()
271 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
272 #define close(fd) (CloseHandle(fd) ? 0 : -1)
273 #define munmap(ptr,len) UnmapViewOfFile(ptr)
274 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
275 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
277 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
281 #define THREAD_RET void *
282 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
283 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
284 #define Z "z" /**< printf format modifier for size_t */
286 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
287 #define MDB_PIDLOCK 1
289 #ifdef MDB_USE_SYSV_SEM
291 typedef struct mdb_mutex {
297 #define MDB_MUTEX(env, rw) (&(env)->me_##rw##mutex)
298 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
299 #define UNLOCK_MUTEX(mutex) do { \
300 struct sembuf sb = { 0, 1, SEM_UNDO }; \
301 sb.sem_num = (mutex)->semnum; \
302 *(mutex)->locked = 0; \
303 semop((mutex)->semid, &sb, 1); \
307 mdb_sem_wait(mdb_mutex_t *sem)
309 int rc, *locked = sem->locked;
310 struct sembuf sb = { 0, -1, SEM_UNDO };
311 sb.sem_num = sem->semnum;
313 if (!semop(sem->semid, &sb, 1)) {
314 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
318 } while ((rc = errno) == EINTR);
322 #define mdb_mutex_consistent(mutex) 0
324 #else /* MDB_USE_POSIX_MUTEX: */
325 /** Pointer/HANDLE type of shared mutex/semaphore.
327 typedef pthread_mutex_t mdb_mutex_t;
328 /** Mutex for the reader table (rw = r) or write transaction (rw = w).
330 #define MDB_MUTEX(env, rw) (&(env)->me_txns->mti_##rw##mutex)
331 /** Lock the reader or writer mutex.
332 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
334 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
335 /** Unlock the reader or writer mutex.
337 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
338 /** Mark mutex-protected data as repaired, after death of previous owner.
340 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
341 #endif /* MDB_USE_SYSV_SEM */
343 /** Get the error code for the last failed system function.
345 #define ErrCode() errno
347 /** An abstraction for a file handle.
348 * On POSIX systems file handles are small integers. On Windows
349 * they're opaque pointers.
353 /** A value for an invalid file handle.
354 * Mainly used to initialize file variables and signify that they are
357 #define INVALID_HANDLE_VALUE (-1)
359 /** Get the size of a memory page for the system.
360 * This is the basic size that the platform's memory manager uses, and is
361 * fundamental to the use of memory-mapped files.
363 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
368 #elif defined(MDB_USE_SYSV_SEM)
369 #define MNAME_LEN (sizeof(int))
371 #define MNAME_LEN (sizeof(pthread_mutex_t))
374 #ifdef MDB_USE_SYSV_SEM
375 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
377 #define SYSV_SEM_FLAG 0
382 #ifdef MDB_ROBUST_SUPPORTED
383 /** Lock mutex, handle any error, set rc = result.
384 * Return 0 on success, nonzero (not rc) on error.
386 #define LOCK_MUTEX(rc, env, mutex) \
387 (((rc) = LOCK_MUTEX0(mutex)) && \
388 ((rc) = mdb_mutex_failed(env, mutex, rc)))
389 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc);
391 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
392 #define mdb_mutex_failed(env, mutex, rc) (rc)
396 /** A flag for opening a file and requesting synchronous data writes.
397 * This is only used when writing a meta page. It's not strictly needed;
398 * we could just do a normal write and then immediately perform a flush.
399 * But if this flag is available it saves us an extra system call.
401 * @note If O_DSYNC is undefined but exists in /usr/include,
402 * preferably set some compiler flag to get the definition.
403 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
406 # define MDB_DSYNC O_DSYNC
410 /** Function for flushing the data of a file. Define this to fsync
411 * if fdatasync() is not supported.
413 #ifndef MDB_FDATASYNC
414 # define MDB_FDATASYNC fdatasync
418 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
429 /** A page number in the database.
430 * Note that 64 bit page numbers are overkill, since pages themselves
431 * already represent 12-13 bits of addressable memory, and the OS will
432 * always limit applications to a maximum of 63 bits of address space.
434 * @note In the #MDB_node structure, we only store 48 bits of this value,
435 * which thus limits us to only 60 bits of addressable data.
437 typedef MDB_ID pgno_t;
439 /** A transaction ID.
440 * See struct MDB_txn.mt_txnid for details.
442 typedef MDB_ID txnid_t;
444 /** @defgroup debug Debug Macros
448 /** Enable debug output. Needs variable argument macros (a C99 feature).
449 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
450 * read from and written to the database (used for free space management).
456 static int mdb_debug;
457 static txnid_t mdb_debug_start;
459 /** Print a debug message with printf formatting.
460 * Requires double parenthesis around 2 or more args.
462 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
463 # define DPRINTF0(fmt, ...) \
464 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
466 # define DPRINTF(args) ((void) 0)
468 /** Print a debug string.
469 * The string is printed literally, with no format processing.
471 #define DPUTS(arg) DPRINTF(("%s", arg))
472 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
474 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
477 /** @brief The maximum size of a database page.
479 * It is 32k or 64k, since value-PAGEBASE must fit in
480 * #MDB_page.%mp_upper.
482 * LMDB will use database pages < OS pages if needed.
483 * That causes more I/O in write transactions: The OS must
484 * know (read) the whole page before writing a partial page.
486 * Note that we don't currently support Huge pages. On Linux,
487 * regular data files cannot use Huge pages, and in general
488 * Huge pages aren't actually pageable. We rely on the OS
489 * demand-pager to read our data and page it out when memory
490 * pressure from other processes is high. So until OSs have
491 * actual paging support for Huge pages, they're not viable.
493 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
495 /** The minimum number of keys required in a database page.
496 * Setting this to a larger value will place a smaller bound on the
497 * maximum size of a data item. Data items larger than this size will
498 * be pushed into overflow pages instead of being stored directly in
499 * the B-tree node. This value used to default to 4. With a page size
500 * of 4096 bytes that meant that any item larger than 1024 bytes would
501 * go into an overflow page. That also meant that on average 2-3KB of
502 * each overflow page was wasted space. The value cannot be lower than
503 * 2 because then there would no longer be a tree structure. With this
504 * value, items larger than 2KB will go into overflow pages, and on
505 * average only 1KB will be wasted.
507 #define MDB_MINKEYS 2
509 /** A stamp that identifies a file as an LMDB file.
510 * There's nothing special about this value other than that it is easily
511 * recognizable, and it will reflect any byte order mismatches.
513 #define MDB_MAGIC 0xBEEFC0DE
515 /** The version number for a database's datafile format. */
516 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
517 /** The version number for a database's lockfile format. */
518 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
520 /** @brief The max size of a key we can write, or 0 for computed max.
522 * This macro should normally be left alone or set to 0.
523 * Note that a database with big keys or dupsort data cannot be
524 * reliably modified by a liblmdb which uses a smaller max.
525 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
527 * Other values are allowed, for backwards compat. However:
528 * A value bigger than the computed max can break if you do not
529 * know what you are doing, and liblmdb <= 0.9.10 can break when
530 * modifying a DB with keys/dupsort data bigger than its max.
532 * Data items in an #MDB_DUPSORT database are also limited to
533 * this size, since they're actually keys of a sub-DB. Keys and
534 * #MDB_DUPSORT data items must fit on a node in a regular page.
536 #ifndef MDB_MAXKEYSIZE
537 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
540 /** The maximum size of a key we can write to the environment. */
542 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
544 #define ENV_MAXKEY(env) ((env)->me_maxkey)
547 /** @brief The maximum size of a data item.
549 * We only store a 32 bit value for node sizes.
551 #define MAXDATASIZE 0xffffffffUL
554 /** Key size which fits in a #DKBUF.
557 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
560 * This is used for printing a hex dump of a key's contents.
562 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
563 /** Display a key in hex.
565 * Invoke a function to display a key in hex.
567 #define DKEY(x) mdb_dkey(x, kbuf)
573 /** An invalid page number.
574 * Mainly used to denote an empty tree.
576 #define P_INVALID (~(pgno_t)0)
578 /** Test if the flags \b f are set in a flag word \b w. */
579 #define F_ISSET(w, f) (((w) & (f)) == (f))
581 /** Round \b n up to an even number. */
582 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
584 /** Used for offsets within a single page.
585 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
588 typedef uint16_t indx_t;
590 /** Default size of memory map.
591 * This is certainly too small for any actual applications. Apps should always set
592 * the size explicitly using #mdb_env_set_mapsize().
594 #define DEFAULT_MAPSIZE 1048576
596 /** @defgroup readers Reader Lock Table
597 * Readers don't acquire any locks for their data access. Instead, they
598 * simply record their transaction ID in the reader table. The reader
599 * mutex is needed just to find an empty slot in the reader table. The
600 * slot's address is saved in thread-specific data so that subsequent read
601 * transactions started by the same thread need no further locking to proceed.
603 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
605 * No reader table is used if the database is on a read-only filesystem, or
606 * if #MDB_NOLOCK is set.
608 * Since the database uses multi-version concurrency control, readers don't
609 * actually need any locking. This table is used to keep track of which
610 * readers are using data from which old transactions, so that we'll know
611 * when a particular old transaction is no longer in use. Old transactions
612 * that have discarded any data pages can then have those pages reclaimed
613 * for use by a later write transaction.
615 * The lock table is constructed such that reader slots are aligned with the
616 * processor's cache line size. Any slot is only ever used by one thread.
617 * This alignment guarantees that there will be no contention or cache
618 * thrashing as threads update their own slot info, and also eliminates
619 * any need for locking when accessing a slot.
621 * A writer thread will scan every slot in the table to determine the oldest
622 * outstanding reader transaction. Any freed pages older than this will be
623 * reclaimed by the writer. The writer doesn't use any locks when scanning
624 * this table. This means that there's no guarantee that the writer will
625 * see the most up-to-date reader info, but that's not required for correct
626 * operation - all we need is to know the upper bound on the oldest reader,
627 * we don't care at all about the newest reader. So the only consequence of
628 * reading stale information here is that old pages might hang around a
629 * while longer before being reclaimed. That's actually good anyway, because
630 * the longer we delay reclaiming old pages, the more likely it is that a
631 * string of contiguous pages can be found after coalescing old pages from
632 * many old transactions together.
635 /** Number of slots in the reader table.
636 * This value was chosen somewhat arbitrarily. 126 readers plus a
637 * couple mutexes fit exactly into 8KB on my development machine.
638 * Applications should set the table size using #mdb_env_set_maxreaders().
640 #define DEFAULT_READERS 126
642 /** The size of a CPU cache line in bytes. We want our lock structures
643 * aligned to this size to avoid false cache line sharing in the
645 * This value works for most CPUs. For Itanium this should be 128.
651 /** The information we store in a single slot of the reader table.
652 * In addition to a transaction ID, we also record the process and
653 * thread ID that owns a slot, so that we can detect stale information,
654 * e.g. threads or processes that went away without cleaning up.
655 * @note We currently don't check for stale records. We simply re-init
656 * the table when we know that we're the only process opening the
659 typedef struct MDB_rxbody {
660 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
661 * Multiple readers that start at the same time will probably have the
662 * same ID here. Again, it's not important to exclude them from
663 * anything; all we need to know is which version of the DB they
664 * started from so we can avoid overwriting any data used in that
665 * particular version.
667 volatile txnid_t mrb_txnid;
668 /** The process ID of the process owning this reader txn. */
669 volatile MDB_PID_T mrb_pid;
670 /** The thread ID of the thread owning this txn. */
671 volatile MDB_THR_T mrb_tid;
674 /** The actual reader record, with cacheline padding. */
675 typedef struct MDB_reader {
678 /** shorthand for mrb_txnid */
679 #define mr_txnid mru.mrx.mrb_txnid
680 #define mr_pid mru.mrx.mrb_pid
681 #define mr_tid mru.mrx.mrb_tid
682 /** cache line alignment */
683 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
687 /** The header for the reader table.
688 * The table resides in a memory-mapped file. (This is a different file
689 * than is used for the main database.)
691 * For POSIX the actual mutexes reside in the shared memory of this
692 * mapped file. On Windows, mutexes are named objects allocated by the
693 * kernel; we store the mutex names in this mapped file so that other
694 * processes can grab them. This same approach is also used on
695 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
696 * process-shared POSIX mutexes. For these cases where a named object
697 * is used, the object name is derived from a 64 bit FNV hash of the
698 * environment pathname. As such, naming collisions are extremely
699 * unlikely. If a collision occurs, the results are unpredictable.
701 typedef struct MDB_txbody {
702 /** Stamp identifying this as an LMDB file. It must be set
705 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
708 char mtb_rmname[MNAME_LEN];
709 #elif defined(MDB_USE_SYSV_SEM)
713 /** Mutex protecting access to this table.
714 * This is the #MDB_MUTEX(env,r) reader table lock.
716 pthread_mutex_t mtb_rmutex;
718 /** The ID of the last transaction committed to the database.
719 * This is recorded here only for convenience; the value can always
720 * be determined by reading the main database meta pages.
722 volatile txnid_t mtb_txnid;
723 /** The number of slots that have been used in the reader table.
724 * This always records the maximum count, it is not decremented
725 * when readers release their slots.
727 volatile unsigned mtb_numreaders;
730 /** The actual reader table definition. */
731 typedef struct MDB_txninfo {
734 #define mti_magic mt1.mtb.mtb_magic
735 #define mti_format mt1.mtb.mtb_format
736 #define mti_rmutex mt1.mtb.mtb_rmutex
737 #define mti_rmname mt1.mtb.mtb_rmname
738 #define mti_txnid mt1.mtb.mtb_txnid
739 #define mti_numreaders mt1.mtb.mtb_numreaders
740 #ifdef MDB_USE_SYSV_SEM
741 #define mti_semid mt1.mtb.mtb_semid
742 #define mti_rlocked mt1.mtb.mtb_rlocked
744 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
748 char mt2_wmname[MNAME_LEN];
749 #define mti_wmname mt2.mt2_wmname
750 #elif defined MDB_USE_SYSV_SEM
752 #define mti_wlocked mt2.mt2_wlocked
754 pthread_mutex_t mt2_wmutex;
755 #define mti_wmutex mt2.mt2_wmutex
757 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
759 MDB_reader mti_readers[1];
762 /** Lockfile format signature: version, features and field layout */
763 #define MDB_LOCK_FORMAT \
765 ((MDB_LOCK_VERSION) \
766 /* Flags which describe functionality */ \
767 + (SYSV_SEM_FLAG << 18) \
768 + (((MDB_PIDLOCK) != 0) << 16)))
771 /** Common header for all page types.
772 * Overflow records occupy a number of contiguous pages with no
773 * headers on any page after the first.
775 typedef struct MDB_page {
776 #define mp_pgno mp_p.p_pgno
777 #define mp_next mp_p.p_next
779 pgno_t p_pgno; /**< page number */
780 struct MDB_page *p_next; /**< for in-memory list of freed pages */
783 /** @defgroup mdb_page Page Flags
785 * Flags for the page headers.
788 #define P_BRANCH 0x01 /**< branch page */
789 #define P_LEAF 0x02 /**< leaf page */
790 #define P_OVERFLOW 0x04 /**< overflow page */
791 #define P_META 0x08 /**< meta page */
792 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
793 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
794 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
795 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
796 #define P_KEEP 0x8000 /**< leave this page alone during spill */
798 uint16_t mp_flags; /**< @ref mdb_page */
799 #define mp_lower mp_pb.pb.pb_lower
800 #define mp_upper mp_pb.pb.pb_upper
801 #define mp_pages mp_pb.pb_pages
804 indx_t pb_lower; /**< lower bound of free space */
805 indx_t pb_upper; /**< upper bound of free space */
807 uint32_t pb_pages; /**< number of overflow pages */
809 indx_t mp_ptrs[1]; /**< dynamic size */
812 /** Size of the page header, excluding dynamic data at the end */
813 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
815 /** Address of first usable data byte in a page, after the header */
816 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
818 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
819 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
821 /** Number of nodes on a page */
822 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
824 /** The amount of space remaining in the page */
825 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
827 /** The percentage of space used in the page, in tenths of a percent. */
828 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
829 ((env)->me_psize - PAGEHDRSZ))
830 /** The minimum page fill factor, in tenths of a percent.
831 * Pages emptier than this are candidates for merging.
833 #define FILL_THRESHOLD 250
835 /** Test if a page is a leaf page */
836 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
837 /** Test if a page is a LEAF2 page */
838 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
839 /** Test if a page is a branch page */
840 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
841 /** Test if a page is an overflow page */
842 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
843 /** Test if a page is a sub page */
844 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
846 /** The number of overflow pages needed to store the given size. */
847 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
849 /** Link in #MDB_txn.%mt_loose_pgs list */
850 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
852 /** Header for a single key/data pair within a page.
853 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
854 * We guarantee 2-byte alignment for 'MDB_node's.
856 typedef struct MDB_node {
857 /** lo and hi are used for data size on leaf nodes and for
858 * child pgno on branch nodes. On 64 bit platforms, flags
859 * is also used for pgno. (Branch nodes have no flags).
860 * They are in host byte order in case that lets some
861 * accesses be optimized into a 32-bit word access.
863 #if BYTE_ORDER == LITTLE_ENDIAN
864 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
866 unsigned short mn_hi, mn_lo;
868 /** @defgroup mdb_node Node Flags
870 * Flags for node headers.
873 #define F_BIGDATA 0x01 /**< data put on overflow page */
874 #define F_SUBDATA 0x02 /**< data is a sub-database */
875 #define F_DUPDATA 0x04 /**< data has duplicates */
877 /** valid flags for #mdb_node_add() */
878 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
881 unsigned short mn_flags; /**< @ref mdb_node */
882 unsigned short mn_ksize; /**< key size */
883 char mn_data[1]; /**< key and data are appended here */
886 /** Size of the node header, excluding dynamic data at the end */
887 #define NODESIZE offsetof(MDB_node, mn_data)
889 /** Bit position of top word in page number, for shifting mn_flags */
890 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
892 /** Size of a node in a branch page with a given key.
893 * This is just the node header plus the key, there is no data.
895 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
897 /** Size of a node in a leaf page with a given key and data.
898 * This is node header plus key plus data size.
900 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
902 /** Address of node \b i in page \b p */
903 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
905 /** Address of the key for the node */
906 #define NODEKEY(node) (void *)((node)->mn_data)
908 /** Address of the data for a node */
909 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
911 /** Get the page number pointed to by a branch node */
912 #define NODEPGNO(node) \
913 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
914 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
915 /** Set the page number in a branch node */
916 #define SETPGNO(node,pgno) do { \
917 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
918 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
920 /** Get the size of the data in a leaf node */
921 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
922 /** Set the size of the data for a leaf node */
923 #define SETDSZ(node,size) do { \
924 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
925 /** The size of a key in a node */
926 #define NODEKSZ(node) ((node)->mn_ksize)
928 /** Copy a page number from src to dst */
930 #define COPY_PGNO(dst,src) dst = src
932 #if SIZE_MAX > 4294967295UL
933 #define COPY_PGNO(dst,src) do { \
934 unsigned short *s, *d; \
935 s = (unsigned short *)&(src); \
936 d = (unsigned short *)&(dst); \
943 #define COPY_PGNO(dst,src) do { \
944 unsigned short *s, *d; \
945 s = (unsigned short *)&(src); \
946 d = (unsigned short *)&(dst); \
952 /** The address of a key in a LEAF2 page.
953 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
954 * There are no node headers, keys are stored contiguously.
956 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
958 /** Set the \b node's key into \b keyptr, if requested. */
959 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
960 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
962 /** Set the \b node's key into \b key. */
963 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
965 /** Information about a single database in the environment. */
966 typedef struct MDB_db {
967 uint32_t md_pad; /**< also ksize for LEAF2 pages */
968 uint16_t md_flags; /**< @ref mdb_dbi_open */
969 uint16_t md_depth; /**< depth of this tree */
970 pgno_t md_branch_pages; /**< number of internal pages */
971 pgno_t md_leaf_pages; /**< number of leaf pages */
972 pgno_t md_overflow_pages; /**< number of overflow pages */
973 size_t md_entries; /**< number of data items */
974 pgno_t md_root; /**< the root page of this tree */
977 /** mdb_dbi_open flags */
978 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
979 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
980 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
981 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
983 /** Handle for the DB used to track free pages. */
985 /** Handle for the default DB. */
988 /** Meta page content.
989 * A meta page is the start point for accessing a database snapshot.
990 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
992 typedef struct MDB_meta {
993 /** Stamp identifying this as an LMDB file. It must be set
996 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
998 void *mm_address; /**< address for fixed mapping */
999 size_t mm_mapsize; /**< size of mmap region */
1000 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
1001 /** The size of pages used in this DB */
1002 #define mm_psize mm_dbs[0].md_pad
1003 /** Any persistent environment flags. @ref mdb_env */
1004 #define mm_flags mm_dbs[0].md_flags
1005 pgno_t mm_last_pg; /**< last used page in file */
1006 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1009 /** Buffer for a stack-allocated meta page.
1010 * The members define size and alignment, and silence type
1011 * aliasing warnings. They are not used directly; that could
1012 * mean incorrectly using several union members in parallel.
1014 typedef union MDB_metabuf {
1017 char mm_pad[PAGEHDRSZ];
1022 /** Auxiliary DB info.
1023 * The information here is mostly static/read-only. There is
1024 * only a single copy of this record in the environment.
1026 typedef struct MDB_dbx {
1027 MDB_val md_name; /**< name of the database */
1028 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1029 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1030 MDB_rel_func *md_rel; /**< user relocate function */
1031 void *md_relctx; /**< user-provided context for md_rel */
1034 /** A database transaction.
1035 * Every operation requires a transaction handle.
1038 MDB_txn *mt_parent; /**< parent of a nested txn */
1039 MDB_txn *mt_child; /**< nested txn under this txn */
1040 pgno_t mt_next_pgno; /**< next unallocated page */
1041 /** The ID of this transaction. IDs are integers incrementing from 1.
1042 * Only committed write transactions increment the ID. If a transaction
1043 * aborts, the ID may be re-used by the next writer.
1046 MDB_env *mt_env; /**< the DB environment */
1047 /** The list of pages that became unused during this transaction.
1049 MDB_IDL mt_free_pgs;
1050 /** The list of loose pages that became unused and may be reused
1051 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1053 MDB_page *mt_loose_pgs;
1054 /* #Number of loose pages (#mt_loose_pgs) */
1056 /** The sorted list of dirty pages we temporarily wrote to disk
1057 * because the dirty list was full. page numbers in here are
1058 * shifted left by 1, deleted slots have the LSB set.
1060 MDB_IDL mt_spill_pgs;
1062 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1063 MDB_ID2L dirty_list;
1064 /** For read txns: This thread/txn's reader table slot, or NULL. */
1067 /** Array of records for each DB known in the environment. */
1069 /** Array of MDB_db records for each known DB */
1071 /** Array of sequence numbers for each DB handle */
1072 unsigned int *mt_dbiseqs;
1073 /** @defgroup mt_dbflag Transaction DB Flags
1077 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1078 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1079 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1080 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1082 /** In write txns, array of cursors for each DB */
1083 MDB_cursor **mt_cursors;
1084 /** Array of flags for each DB */
1085 unsigned char *mt_dbflags;
1086 /** Number of DB records in use. This number only ever increments;
1087 * we don't decrement it when individual DB handles are closed.
1091 /** @defgroup mdb_txn Transaction Flags
1095 /** #mdb_txn_begin() flags */
1096 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1097 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1098 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1099 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1100 /* internal txn flags */
1101 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1102 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1103 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1104 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1106 unsigned int mt_flags; /**< @ref mdb_txn */
1107 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1108 * Includes ancestor txns' dirty pages not hidden by other txns'
1109 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1110 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1112 unsigned int mt_dirty_room;
1115 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1116 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1117 * raise this on a 64 bit machine.
1119 #define CURSOR_STACK 32
1123 /** Cursors are used for all DB operations.
1124 * A cursor holds a path of (page pointer, key index) from the DB
1125 * root to a position in the DB, plus other state. #MDB_DUPSORT
1126 * cursors include an xcursor to the current data item. Write txns
1127 * track their cursors and keep them up to date when data moves.
1128 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1129 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1132 /** Next cursor on this DB in this txn */
1133 MDB_cursor *mc_next;
1134 /** Backup of the original cursor if this cursor is a shadow */
1135 MDB_cursor *mc_backup;
1136 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1137 struct MDB_xcursor *mc_xcursor;
1138 /** The transaction that owns this cursor */
1140 /** The database handle this cursor operates on */
1142 /** The database record for this cursor */
1144 /** The database auxiliary record for this cursor */
1146 /** The @ref mt_dbflag for this database */
1147 unsigned char *mc_dbflag;
1148 unsigned short mc_snum; /**< number of pushed pages */
1149 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1150 /** @defgroup mdb_cursor Cursor Flags
1152 * Cursor state flags.
1155 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1156 #define C_EOF 0x02 /**< No more data */
1157 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1158 #define C_DEL 0x08 /**< last op was a cursor_del */
1159 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1160 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1162 unsigned int mc_flags; /**< @ref mdb_cursor */
1163 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1164 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1167 /** Context for sorted-dup records.
1168 * We could have gone to a fully recursive design, with arbitrarily
1169 * deep nesting of sub-databases. But for now we only handle these
1170 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1172 typedef struct MDB_xcursor {
1173 /** A sub-cursor for traversing the Dup DB */
1174 MDB_cursor mx_cursor;
1175 /** The database record for this Dup DB */
1177 /** The auxiliary DB record for this Dup DB */
1179 /** The @ref mt_dbflag for this Dup DB */
1180 unsigned char mx_dbflag;
1183 /** State of FreeDB old pages, stored in the MDB_env */
1184 typedef struct MDB_pgstate {
1185 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1186 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1189 /** The database environment. */
1191 HANDLE me_fd; /**< The main data file */
1192 HANDLE me_lfd; /**< The lock file */
1193 HANDLE me_mfd; /**< just for writing the meta pages */
1194 /** Failed to update the meta page. Probably an I/O error. */
1195 #define MDB_FATAL_ERROR 0x80000000U
1196 /** Some fields are initialized. */
1197 #define MDB_ENV_ACTIVE 0x20000000U
1198 /** me_txkey is set */
1199 #define MDB_ENV_TXKEY 0x10000000U
1200 /** fdatasync is unreliable */
1201 #define MDB_FSYNCONLY 0x08000000U
1202 uint32_t me_flags; /**< @ref mdb_env */
1203 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1204 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1205 unsigned int me_maxreaders; /**< size of the reader table */
1206 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1207 volatile int me_close_readers;
1208 MDB_dbi me_numdbs; /**< number of DBs opened */
1209 MDB_dbi me_maxdbs; /**< size of the DB table */
1210 MDB_PID_T me_pid; /**< process ID of this env */
1211 char *me_path; /**< path to the DB files */
1212 char *me_map; /**< the memory map of the data file */
1213 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1214 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1215 void *me_pbuf; /**< scratch area for DUPSORT put() */
1216 MDB_txn *me_txn; /**< current write transaction */
1217 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1218 size_t me_mapsize; /**< size of the data memory map */
1219 off_t me_size; /**< current file size */
1220 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1221 MDB_dbx *me_dbxs; /**< array of static DB info */
1222 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1223 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1224 pthread_key_t me_txkey; /**< thread-key for readers */
1225 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1226 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1227 # define me_pglast me_pgstate.mf_pglast
1228 # define me_pghead me_pgstate.mf_pghead
1229 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1230 /** IDL of pages that became unused in a write txn */
1231 MDB_IDL me_free_pgs;
1232 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1233 MDB_ID2L me_dirty_list;
1234 /** Max number of freelist items that can fit in a single overflow page */
1236 /** Max size of a node on a page */
1237 unsigned int me_nodemax;
1238 #if !(MDB_MAXKEYSIZE)
1239 unsigned int me_maxkey; /**< max size of a key */
1241 int me_live_reader; /**< have liveness lock in reader table */
1243 int me_pidquery; /**< Used in OpenProcess */
1245 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM)
1246 /* Windows mutexes/SysV semaphores do not reside in shared mem */
1247 mdb_mutex_t me_rmutex;
1248 mdb_mutex_t me_wmutex;
1250 void *me_userctx; /**< User-settable context */
1251 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1254 /** Nested transaction */
1255 typedef struct MDB_ntxn {
1256 MDB_txn mnt_txn; /**< the transaction */
1257 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1260 /** max number of pages to commit in one writev() call */
1261 #define MDB_COMMIT_PAGES 64
1262 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1263 #undef MDB_COMMIT_PAGES
1264 #define MDB_COMMIT_PAGES IOV_MAX
1267 /** max bytes to write in one call */
1268 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1270 /** Check \b txn and \b dbi arguments to a function */
1271 #define TXN_DBI_EXIST(txn, dbi) \
1272 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1274 /** Check for misused \b dbi handles */
1275 #define TXN_DBI_CHANGED(txn, dbi) \
1276 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1278 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1279 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1280 static int mdb_page_touch(MDB_cursor *mc);
1282 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1283 static int mdb_page_search_root(MDB_cursor *mc,
1284 MDB_val *key, int modify);
1285 #define MDB_PS_MODIFY 1
1286 #define MDB_PS_ROOTONLY 2
1287 #define MDB_PS_FIRST 4
1288 #define MDB_PS_LAST 8
1289 static int mdb_page_search(MDB_cursor *mc,
1290 MDB_val *key, int flags);
1291 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1293 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1294 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1295 pgno_t newpgno, unsigned int nflags);
1297 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1298 static int mdb_env_pick_meta(const MDB_env *env);
1299 static int mdb_env_write_meta(MDB_txn *txn);
1300 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1301 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1303 static void mdb_env_close0(MDB_env *env, int excl);
1305 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1306 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1307 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1308 static void mdb_node_del(MDB_cursor *mc, int ksize);
1309 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1310 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1311 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1312 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1313 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1315 static int mdb_rebalance(MDB_cursor *mc);
1316 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1318 static void mdb_cursor_pop(MDB_cursor *mc);
1319 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1321 static int mdb_cursor_del0(MDB_cursor *mc);
1322 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1323 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1324 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1325 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1326 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1328 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1329 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1331 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1332 static void mdb_xcursor_init0(MDB_cursor *mc);
1333 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1335 static int mdb_drop0(MDB_cursor *mc, int subs);
1336 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1337 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1340 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1343 /** Compare two items pointing at size_t's of unknown alignment. */
1344 #ifdef MISALIGNED_OK
1345 # define mdb_cmp_clong mdb_cmp_long
1347 # define mdb_cmp_clong mdb_cmp_cint
1351 static SECURITY_DESCRIPTOR mdb_null_sd;
1352 static SECURITY_ATTRIBUTES mdb_all_sa;
1353 static int mdb_sec_inited;
1356 /** Return the library version info. */
1358 mdb_version(int *major, int *minor, int *patch)
1360 if (major) *major = MDB_VERSION_MAJOR;
1361 if (minor) *minor = MDB_VERSION_MINOR;
1362 if (patch) *patch = MDB_VERSION_PATCH;
1363 return MDB_VERSION_STRING;
1366 /** Table of descriptions for LMDB @ref errors */
1367 static char *const mdb_errstr[] = {
1368 "MDB_KEYEXIST: Key/data pair already exists",
1369 "MDB_NOTFOUND: No matching key/data pair found",
1370 "MDB_PAGE_NOTFOUND: Requested page not found",
1371 "MDB_CORRUPTED: Located page was wrong type",
1372 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1373 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1374 "MDB_INVALID: File is not an LMDB file",
1375 "MDB_MAP_FULL: Environment mapsize limit reached",
1376 "MDB_DBS_FULL: Environment maxdbs limit reached",
1377 "MDB_READERS_FULL: Environment maxreaders limit reached",
1378 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1379 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1380 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1381 "MDB_PAGE_FULL: Internal error - page has no more space",
1382 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1383 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1384 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1385 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1386 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1387 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1391 mdb_strerror(int err)
1394 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1395 * This works as long as no function between the call to mdb_strerror
1396 * and the actual use of the message uses more than 4K of stack.
1399 char buf[1024], *ptr = buf;
1403 return ("Successful return: 0");
1405 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1406 i = err - MDB_KEYEXIST;
1407 return mdb_errstr[i];
1411 /* These are the C-runtime error codes we use. The comment indicates
1412 * their numeric value, and the Win32 error they would correspond to
1413 * if the error actually came from a Win32 API. A major mess, we should
1414 * have used LMDB-specific error codes for everything.
1417 case ENOENT: /* 2, FILE_NOT_FOUND */
1418 case EIO: /* 5, ACCESS_DENIED */
1419 case ENOMEM: /* 12, INVALID_ACCESS */
1420 case EACCES: /* 13, INVALID_DATA */
1421 case EBUSY: /* 16, CURRENT_DIRECTORY */
1422 case EINVAL: /* 22, BAD_COMMAND */
1423 case ENOSPC: /* 28, OUT_OF_PAPER */
1424 return strerror(err);
1429 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1430 FORMAT_MESSAGE_IGNORE_INSERTS,
1431 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1434 return strerror(err);
1438 /** assert(3) variant in cursor context */
1439 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1440 /** assert(3) variant in transaction context */
1441 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1442 /** assert(3) variant in environment context */
1443 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1446 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1447 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1450 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1451 const char *func, const char *file, int line)
1454 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1455 file, line, expr_txt, func);
1456 if (env->me_assert_func)
1457 env->me_assert_func(env, buf);
1458 fprintf(stderr, "%s\n", buf);
1462 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1466 /** Return the page number of \b mp which may be sub-page, for debug output */
1468 mdb_dbg_pgno(MDB_page *mp)
1471 COPY_PGNO(ret, mp->mp_pgno);
1475 /** Display a key in hexadecimal and return the address of the result.
1476 * @param[in] key the key to display
1477 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1478 * @return The key in hexadecimal form.
1481 mdb_dkey(MDB_val *key, char *buf)
1484 unsigned char *c = key->mv_data;
1490 if (key->mv_size > DKBUF_MAXKEYSIZE)
1491 return "MDB_MAXKEYSIZE";
1492 /* may want to make this a dynamic check: if the key is mostly
1493 * printable characters, print it as-is instead of converting to hex.
1497 for (i=0; i<key->mv_size; i++)
1498 ptr += sprintf(ptr, "%02x", *c++);
1500 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1506 mdb_leafnode_type(MDB_node *n)
1508 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1509 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1510 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1513 /** Display all the keys in the page. */
1515 mdb_page_list(MDB_page *mp)
1517 pgno_t pgno = mdb_dbg_pgno(mp);
1518 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1520 unsigned int i, nkeys, nsize, total = 0;
1524 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1525 case P_BRANCH: type = "Branch page"; break;
1526 case P_LEAF: type = "Leaf page"; break;
1527 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1528 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1529 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1531 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1532 pgno, mp->mp_pages, state);
1535 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1536 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1539 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1543 nkeys = NUMKEYS(mp);
1544 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1546 for (i=0; i<nkeys; i++) {
1547 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1548 key.mv_size = nsize = mp->mp_pad;
1549 key.mv_data = LEAF2KEY(mp, i, nsize);
1551 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1554 node = NODEPTR(mp, i);
1555 key.mv_size = node->mn_ksize;
1556 key.mv_data = node->mn_data;
1557 nsize = NODESIZE + key.mv_size;
1558 if (IS_BRANCH(mp)) {
1559 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1563 if (F_ISSET(node->mn_flags, F_BIGDATA))
1564 nsize += sizeof(pgno_t);
1566 nsize += NODEDSZ(node);
1568 nsize += sizeof(indx_t);
1569 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1570 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1572 total = EVEN(total);
1574 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1575 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1579 mdb_cursor_chk(MDB_cursor *mc)
1585 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1586 for (i=0; i<mc->mc_top; i++) {
1588 node = NODEPTR(mp, mc->mc_ki[i]);
1589 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1592 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1598 /** Count all the pages in each DB and in the freelist
1599 * and make sure it matches the actual number of pages
1601 * All named DBs must be open for a correct count.
1603 static void mdb_audit(MDB_txn *txn)
1607 MDB_ID freecount, count;
1612 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1613 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1614 freecount += *(MDB_ID *)data.mv_data;
1615 mdb_tassert(txn, rc == MDB_NOTFOUND);
1618 for (i = 0; i<txn->mt_numdbs; i++) {
1620 if (!(txn->mt_dbflags[i] & DB_VALID))
1622 mdb_cursor_init(&mc, txn, i, &mx);
1623 if (txn->mt_dbs[i].md_root == P_INVALID)
1625 count += txn->mt_dbs[i].md_branch_pages +
1626 txn->mt_dbs[i].md_leaf_pages +
1627 txn->mt_dbs[i].md_overflow_pages;
1628 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1629 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1630 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1633 mp = mc.mc_pg[mc.mc_top];
1634 for (j=0; j<NUMKEYS(mp); j++) {
1635 MDB_node *leaf = NODEPTR(mp, j);
1636 if (leaf->mn_flags & F_SUBDATA) {
1638 memcpy(&db, NODEDATA(leaf), sizeof(db));
1639 count += db.md_branch_pages + db.md_leaf_pages +
1640 db.md_overflow_pages;
1644 mdb_tassert(txn, rc == MDB_NOTFOUND);
1647 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1648 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1649 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1655 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1657 return txn->mt_dbxs[dbi].md_cmp(a, b);
1661 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1663 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1664 #if UINT_MAX < SIZE_MAX
1665 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1666 dcmp = mdb_cmp_clong;
1671 /** Allocate memory for a page.
1672 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1675 mdb_page_malloc(MDB_txn *txn, unsigned num)
1677 MDB_env *env = txn->mt_env;
1678 MDB_page *ret = env->me_dpages;
1679 size_t psize = env->me_psize, sz = psize, off;
1680 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1681 * For a single page alloc, we init everything after the page header.
1682 * For multi-page, we init the final page; if the caller needed that
1683 * many pages they will be filling in at least up to the last page.
1687 VGMEMP_ALLOC(env, ret, sz);
1688 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1689 env->me_dpages = ret->mp_next;
1692 psize -= off = PAGEHDRSZ;
1697 if ((ret = malloc(sz)) != NULL) {
1698 VGMEMP_ALLOC(env, ret, sz);
1699 if (!(env->me_flags & MDB_NOMEMINIT)) {
1700 memset((char *)ret + off, 0, psize);
1704 txn->mt_flags |= MDB_TXN_ERROR;
1708 /** Free a single page.
1709 * Saves single pages to a list, for future reuse.
1710 * (This is not used for multi-page overflow pages.)
1713 mdb_page_free(MDB_env *env, MDB_page *mp)
1715 mp->mp_next = env->me_dpages;
1716 VGMEMP_FREE(env, mp);
1717 env->me_dpages = mp;
1720 /** Free a dirty page */
1722 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1724 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1725 mdb_page_free(env, dp);
1727 /* large pages just get freed directly */
1728 VGMEMP_FREE(env, dp);
1733 /** Return all dirty pages to dpage list */
1735 mdb_dlist_free(MDB_txn *txn)
1737 MDB_env *env = txn->mt_env;
1738 MDB_ID2L dl = txn->mt_u.dirty_list;
1739 unsigned i, n = dl[0].mid;
1741 for (i = 1; i <= n; i++) {
1742 mdb_dpage_free(env, dl[i].mptr);
1747 /** Loosen or free a single page.
1748 * Saves single pages to a list for future reuse
1749 * in this same txn. It has been pulled from the freeDB
1750 * and already resides on the dirty list, but has been
1751 * deleted. Use these pages first before pulling again
1754 * If the page wasn't dirtied in this txn, just add it
1755 * to this txn's free list.
1758 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1761 pgno_t pgno = mp->mp_pgno;
1762 MDB_txn *txn = mc->mc_txn;
1764 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1765 if (txn->mt_parent) {
1766 MDB_ID2 *dl = txn->mt_u.dirty_list;
1767 /* If txn has a parent, make sure the page is in our
1771 unsigned x = mdb_mid2l_search(dl, pgno);
1772 if (x <= dl[0].mid && dl[x].mid == pgno) {
1773 if (mp != dl[x].mptr) { /* bad cursor? */
1774 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1775 txn->mt_flags |= MDB_TXN_ERROR;
1776 return MDB_CORRUPTED;
1783 /* no parent txn, so it's just ours */
1788 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1790 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1791 txn->mt_loose_pgs = mp;
1792 txn->mt_loose_count++;
1793 mp->mp_flags |= P_LOOSE;
1795 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1803 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1804 * @param[in] mc A cursor handle for the current operation.
1805 * @param[in] pflags Flags of the pages to update:
1806 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1807 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1808 * @return 0 on success, non-zero on failure.
1811 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1813 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1814 MDB_txn *txn = mc->mc_txn;
1820 int rc = MDB_SUCCESS, level;
1822 /* Mark pages seen by cursors */
1823 if (mc->mc_flags & C_UNTRACK)
1824 mc = NULL; /* will find mc in mt_cursors */
1825 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1826 for (; mc; mc=mc->mc_next) {
1827 if (!(mc->mc_flags & C_INITIALIZED))
1829 for (m3 = mc;; m3 = &mx->mx_cursor) {
1831 for (j=0; j<m3->mc_snum; j++) {
1833 if ((mp->mp_flags & Mask) == pflags)
1834 mp->mp_flags ^= P_KEEP;
1836 mx = m3->mc_xcursor;
1837 /* Proceed to mx if it is at a sub-database */
1838 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1840 if (! (mp && (mp->mp_flags & P_LEAF)))
1842 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1843 if (!(leaf->mn_flags & F_SUBDATA))
1852 /* Mark dirty root pages */
1853 for (i=0; i<txn->mt_numdbs; i++) {
1854 if (txn->mt_dbflags[i] & DB_DIRTY) {
1855 pgno_t pgno = txn->mt_dbs[i].md_root;
1856 if (pgno == P_INVALID)
1858 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1860 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1861 dp->mp_flags ^= P_KEEP;
1869 static int mdb_page_flush(MDB_txn *txn, int keep);
1871 /** Spill pages from the dirty list back to disk.
1872 * This is intended to prevent running into #MDB_TXN_FULL situations,
1873 * but note that they may still occur in a few cases:
1874 * 1) our estimate of the txn size could be too small. Currently this
1875 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1876 * 2) child txns may run out of space if their parents dirtied a
1877 * lot of pages and never spilled them. TODO: we probably should do
1878 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1879 * the parent's dirty_room is below a given threshold.
1881 * Otherwise, if not using nested txns, it is expected that apps will
1882 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1883 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1884 * If the txn never references them again, they can be left alone.
1885 * If the txn only reads them, they can be used without any fuss.
1886 * If the txn writes them again, they can be dirtied immediately without
1887 * going thru all of the work of #mdb_page_touch(). Such references are
1888 * handled by #mdb_page_unspill().
1890 * Also note, we never spill DB root pages, nor pages of active cursors,
1891 * because we'll need these back again soon anyway. And in nested txns,
1892 * we can't spill a page in a child txn if it was already spilled in a
1893 * parent txn. That would alter the parent txns' data even though
1894 * the child hasn't committed yet, and we'd have no way to undo it if
1895 * the child aborted.
1897 * @param[in] m0 cursor A cursor handle identifying the transaction and
1898 * database for which we are checking space.
1899 * @param[in] key For a put operation, the key being stored.
1900 * @param[in] data For a put operation, the data being stored.
1901 * @return 0 on success, non-zero on failure.
1904 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1906 MDB_txn *txn = m0->mc_txn;
1908 MDB_ID2L dl = txn->mt_u.dirty_list;
1909 unsigned int i, j, need;
1912 if (m0->mc_flags & C_SUB)
1915 /* Estimate how much space this op will take */
1916 i = m0->mc_db->md_depth;
1917 /* Named DBs also dirty the main DB */
1918 if (m0->mc_dbi > MAIN_DBI)
1919 i += txn->mt_dbs[MAIN_DBI].md_depth;
1920 /* For puts, roughly factor in the key+data size */
1922 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1923 i += i; /* double it for good measure */
1926 if (txn->mt_dirty_room > i)
1929 if (!txn->mt_spill_pgs) {
1930 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1931 if (!txn->mt_spill_pgs)
1934 /* purge deleted slots */
1935 MDB_IDL sl = txn->mt_spill_pgs;
1936 unsigned int num = sl[0];
1938 for (i=1; i<=num; i++) {
1945 /* Preserve pages which may soon be dirtied again */
1946 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1949 /* Less aggressive spill - we originally spilled the entire dirty list,
1950 * with a few exceptions for cursor pages and DB root pages. But this
1951 * turns out to be a lot of wasted effort because in a large txn many
1952 * of those pages will need to be used again. So now we spill only 1/8th
1953 * of the dirty pages. Testing revealed this to be a good tradeoff,
1954 * better than 1/2, 1/4, or 1/10.
1956 if (need < MDB_IDL_UM_MAX / 8)
1957 need = MDB_IDL_UM_MAX / 8;
1959 /* Save the page IDs of all the pages we're flushing */
1960 /* flush from the tail forward, this saves a lot of shifting later on. */
1961 for (i=dl[0].mid; i && need; i--) {
1962 MDB_ID pn = dl[i].mid << 1;
1964 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1966 /* Can't spill twice, make sure it's not already in a parent's
1969 if (txn->mt_parent) {
1971 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1972 if (tx2->mt_spill_pgs) {
1973 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1974 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1975 dp->mp_flags |= P_KEEP;
1983 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1987 mdb_midl_sort(txn->mt_spill_pgs);
1989 /* Flush the spilled part of dirty list */
1990 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1993 /* Reset any dirty pages we kept that page_flush didn't see */
1994 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1997 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2001 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2003 mdb_find_oldest(MDB_txn *txn)
2006 txnid_t mr, oldest = txn->mt_txnid - 1;
2007 if (txn->mt_env->me_txns) {
2008 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2009 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2020 /** Add a page to the txn's dirty list */
2022 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2025 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2027 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2028 insert = mdb_mid2l_append;
2030 insert = mdb_mid2l_insert;
2032 mid.mid = mp->mp_pgno;
2034 rc = insert(txn->mt_u.dirty_list, &mid);
2035 mdb_tassert(txn, rc == 0);
2036 txn->mt_dirty_room--;
2039 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2040 * me_pghead and mt_next_pgno.
2042 * If there are free pages available from older transactions, they
2043 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2044 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2045 * and move me_pglast to say which records were consumed. Only this
2046 * function can create me_pghead and move me_pglast/mt_next_pgno.
2047 * @param[in] mc cursor A cursor handle identifying the transaction and
2048 * database for which we are allocating.
2049 * @param[in] num the number of pages to allocate.
2050 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2051 * will always be satisfied by a single contiguous chunk of memory.
2052 * @return 0 on success, non-zero on failure.
2055 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2057 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2058 /* Get at most <Max_retries> more freeDB records once me_pghead
2059 * has enough pages. If not enough, use new pages from the map.
2060 * If <Paranoid> and mc is updating the freeDB, only get new
2061 * records if me_pghead is empty. Then the freelist cannot play
2062 * catch-up with itself by growing while trying to save it.
2064 enum { Paranoid = 1, Max_retries = 500 };
2066 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2068 int rc, retry = num * 60;
2069 MDB_txn *txn = mc->mc_txn;
2070 MDB_env *env = txn->mt_env;
2071 pgno_t pgno, *mop = env->me_pghead;
2072 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2074 txnid_t oldest = 0, last;
2079 /* If there are any loose pages, just use them */
2080 if (num == 1 && txn->mt_loose_pgs) {
2081 np = txn->mt_loose_pgs;
2082 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2083 txn->mt_loose_count--;
2084 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2092 /* If our dirty list is already full, we can't do anything */
2093 if (txn->mt_dirty_room == 0) {
2098 for (op = MDB_FIRST;; op = MDB_NEXT) {
2103 /* Seek a big enough contiguous page range. Prefer
2104 * pages at the tail, just truncating the list.
2110 if (mop[i-n2] == pgno+n2)
2117 if (op == MDB_FIRST) { /* 1st iteration */
2118 /* Prepare to fetch more and coalesce */
2119 last = env->me_pglast;
2120 oldest = env->me_pgoldest;
2121 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2124 key.mv_data = &last; /* will look up last+1 */
2125 key.mv_size = sizeof(last);
2127 if (Paranoid && mc->mc_dbi == FREE_DBI)
2130 if (Paranoid && retry < 0 && mop_len)
2134 /* Do not fetch more if the record will be too recent */
2135 if (oldest <= last) {
2137 oldest = mdb_find_oldest(txn);
2138 env->me_pgoldest = oldest;
2144 rc = mdb_cursor_get(&m2, &key, NULL, op);
2146 if (rc == MDB_NOTFOUND)
2150 last = *(txnid_t*)key.mv_data;
2151 if (oldest <= last) {
2153 oldest = mdb_find_oldest(txn);
2154 env->me_pgoldest = oldest;
2160 np = m2.mc_pg[m2.mc_top];
2161 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2162 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2165 idl = (MDB_ID *) data.mv_data;
2168 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2173 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2175 mop = env->me_pghead;
2177 env->me_pglast = last;
2179 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2180 last, txn->mt_dbs[FREE_DBI].md_root, i));
2182 DPRINTF(("IDL %"Z"u", idl[j]));
2184 /* Merge in descending sorted order */
2185 mdb_midl_xmerge(mop, idl);
2189 /* Use new pages from the map when nothing suitable in the freeDB */
2191 pgno = txn->mt_next_pgno;
2192 if (pgno + num >= env->me_maxpg) {
2193 DPUTS("DB size maxed out");
2199 if (env->me_flags & MDB_WRITEMAP) {
2200 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2202 if (!(np = mdb_page_malloc(txn, num))) {
2208 mop[0] = mop_len -= num;
2209 /* Move any stragglers down */
2210 for (j = i-num; j < mop_len; )
2211 mop[++j] = mop[++i];
2213 txn->mt_next_pgno = pgno + num;
2216 mdb_page_dirty(txn, np);
2222 txn->mt_flags |= MDB_TXN_ERROR;
2226 /** Copy the used portions of a non-overflow page.
2227 * @param[in] dst page to copy into
2228 * @param[in] src page to copy from
2229 * @param[in] psize size of a page
2232 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2234 enum { Align = sizeof(pgno_t) };
2235 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2237 /* If page isn't full, just copy the used portion. Adjust
2238 * alignment so memcpy may copy words instead of bytes.
2240 if ((unused &= -Align) && !IS_LEAF2(src)) {
2241 upper = (upper + PAGEBASE) & -Align;
2242 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2243 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2246 memcpy(dst, src, psize - unused);
2250 /** Pull a page off the txn's spill list, if present.
2251 * If a page being referenced was spilled to disk in this txn, bring
2252 * it back and make it dirty/writable again.
2253 * @param[in] txn the transaction handle.
2254 * @param[in] mp the page being referenced. It must not be dirty.
2255 * @param[out] ret the writable page, if any. ret is unchanged if
2256 * mp wasn't spilled.
2259 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2261 MDB_env *env = txn->mt_env;
2264 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2266 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2267 if (!tx2->mt_spill_pgs)
2269 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2270 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2273 if (txn->mt_dirty_room == 0)
2274 return MDB_TXN_FULL;
2275 if (IS_OVERFLOW(mp))
2279 if (env->me_flags & MDB_WRITEMAP) {
2282 np = mdb_page_malloc(txn, num);
2286 memcpy(np, mp, num * env->me_psize);
2288 mdb_page_copy(np, mp, env->me_psize);
2291 /* If in current txn, this page is no longer spilled.
2292 * If it happens to be the last page, truncate the spill list.
2293 * Otherwise mark it as deleted by setting the LSB.
2295 if (x == txn->mt_spill_pgs[0])
2296 txn->mt_spill_pgs[0]--;
2298 txn->mt_spill_pgs[x] |= 1;
2299 } /* otherwise, if belonging to a parent txn, the
2300 * page remains spilled until child commits
2303 mdb_page_dirty(txn, np);
2304 np->mp_flags |= P_DIRTY;
2312 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2313 * @param[in] mc cursor pointing to the page to be touched
2314 * @return 0 on success, non-zero on failure.
2317 mdb_page_touch(MDB_cursor *mc)
2319 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2320 MDB_txn *txn = mc->mc_txn;
2321 MDB_cursor *m2, *m3;
2325 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2326 if (txn->mt_flags & MDB_TXN_SPILLS) {
2328 rc = mdb_page_unspill(txn, mp, &np);
2334 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2335 (rc = mdb_page_alloc(mc, 1, &np)))
2338 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2339 mp->mp_pgno, pgno));
2340 mdb_cassert(mc, mp->mp_pgno != pgno);
2341 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2342 /* Update the parent page, if any, to point to the new page */
2344 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2345 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2346 SETPGNO(node, pgno);
2348 mc->mc_db->md_root = pgno;
2350 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2351 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2353 /* If txn has a parent, make sure the page is in our
2357 unsigned x = mdb_mid2l_search(dl, pgno);
2358 if (x <= dl[0].mid && dl[x].mid == pgno) {
2359 if (mp != dl[x].mptr) { /* bad cursor? */
2360 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2361 txn->mt_flags |= MDB_TXN_ERROR;
2362 return MDB_CORRUPTED;
2367 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2369 np = mdb_page_malloc(txn, 1);
2374 rc = mdb_mid2l_insert(dl, &mid);
2375 mdb_cassert(mc, rc == 0);
2380 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2382 np->mp_flags |= P_DIRTY;
2385 /* Adjust cursors pointing to mp */
2386 mc->mc_pg[mc->mc_top] = np;
2387 m2 = txn->mt_cursors[mc->mc_dbi];
2388 if (mc->mc_flags & C_SUB) {
2389 for (; m2; m2=m2->mc_next) {
2390 m3 = &m2->mc_xcursor->mx_cursor;
2391 if (m3->mc_snum < mc->mc_snum) continue;
2392 if (m3->mc_pg[mc->mc_top] == mp)
2393 m3->mc_pg[mc->mc_top] = np;
2396 for (; m2; m2=m2->mc_next) {
2397 if (m2->mc_snum < mc->mc_snum) continue;
2398 if (m2->mc_pg[mc->mc_top] == mp) {
2399 m2->mc_pg[mc->mc_top] = np;
2400 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2402 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2404 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2405 if (!(leaf->mn_flags & F_SUBDATA))
2406 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2414 txn->mt_flags |= MDB_TXN_ERROR;
2419 mdb_env_sync(MDB_env *env, int force)
2422 if (env->me_flags & MDB_RDONLY)
2424 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2425 if (env->me_flags & MDB_WRITEMAP) {
2426 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2427 ? MS_ASYNC : MS_SYNC;
2428 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2431 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2435 #ifdef BROKEN_FDATASYNC
2436 if (env->me_flags & MDB_FSYNCONLY) {
2437 if (fsync(env->me_fd))
2441 if (MDB_FDATASYNC(env->me_fd))
2448 /** Back up parent txn's cursors, then grab the originals for tracking */
2450 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2452 MDB_cursor *mc, *bk;
2457 for (i = src->mt_numdbs; --i >= 0; ) {
2458 if ((mc = src->mt_cursors[i]) != NULL) {
2459 size = sizeof(MDB_cursor);
2461 size += sizeof(MDB_xcursor);
2462 for (; mc; mc = bk->mc_next) {
2468 mc->mc_db = &dst->mt_dbs[i];
2469 /* Kill pointers into src - and dst to reduce abuse: The
2470 * user may not use mc until dst ends. Otherwise we'd...
2472 mc->mc_txn = NULL; /* ...set this to dst */
2473 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2474 if ((mx = mc->mc_xcursor) != NULL) {
2475 *(MDB_xcursor *)(bk+1) = *mx;
2476 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2478 mc->mc_next = dst->mt_cursors[i];
2479 dst->mt_cursors[i] = mc;
2486 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2487 * @param[in] txn the transaction handle.
2488 * @param[in] merge true to keep changes to parent cursors, false to revert.
2489 * @return 0 on success, non-zero on failure.
2492 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2494 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2498 for (i = txn->mt_numdbs; --i >= 0; ) {
2499 for (mc = cursors[i]; mc; mc = next) {
2501 if ((bk = mc->mc_backup) != NULL) {
2503 /* Commit changes to parent txn */
2504 mc->mc_next = bk->mc_next;
2505 mc->mc_backup = bk->mc_backup;
2506 mc->mc_txn = bk->mc_txn;
2507 mc->mc_db = bk->mc_db;
2508 mc->mc_dbflag = bk->mc_dbflag;
2509 if ((mx = mc->mc_xcursor) != NULL)
2510 mx->mx_cursor.mc_txn = bk->mc_txn;
2512 /* Abort nested txn */
2514 if ((mx = mc->mc_xcursor) != NULL)
2515 *mx = *(MDB_xcursor *)(bk+1);
2519 /* Only malloced cursors are permanently tracked. */
2527 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2530 mdb_txn_reset0(MDB_txn *txn, const char *act);
2532 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2538 Pidset = F_SETLK, Pidcheck = F_GETLK
2542 /** Set or check a pid lock. Set returns 0 on success.
2543 * Check returns 0 if the process is certainly dead, nonzero if it may
2544 * be alive (the lock exists or an error happened so we do not know).
2546 * On Windows Pidset is a no-op, we merely check for the existence
2547 * of the process with the given pid. On POSIX we use a single byte
2548 * lock on the lockfile, set at an offset equal to the pid.
2551 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2553 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2556 if (op == Pidcheck) {
2557 h = OpenProcess(env->me_pidquery, FALSE, pid);
2558 /* No documented "no such process" code, but other program use this: */
2560 return ErrCode() != ERROR_INVALID_PARAMETER;
2561 /* A process exists until all handles to it close. Has it exited? */
2562 ret = WaitForSingleObject(h, 0) != 0;
2569 struct flock lock_info;
2570 memset(&lock_info, 0, sizeof(lock_info));
2571 lock_info.l_type = F_WRLCK;
2572 lock_info.l_whence = SEEK_SET;
2573 lock_info.l_start = pid;
2574 lock_info.l_len = 1;
2575 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2576 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2578 } else if ((rc = ErrCode()) == EINTR) {
2586 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2587 * @param[in] txn the transaction handle to initialize
2588 * @return 0 on success, non-zero on failure.
2591 mdb_txn_renew0(MDB_txn *txn)
2593 MDB_env *env = txn->mt_env;
2594 MDB_txninfo *ti = env->me_txns;
2596 unsigned int i, nr, flags = txn->mt_flags;
2598 int rc, new_notls = 0;
2600 if ((flags &= MDB_TXN_RDONLY) != 0) {
2602 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2603 txn->mt_txnid = meta->mm_txnid;
2604 txn->mt_u.reader = NULL;
2606 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2607 pthread_getspecific(env->me_txkey);
2609 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2610 return MDB_BAD_RSLOT;
2612 MDB_PID_T pid = env->me_pid;
2613 MDB_THR_T tid = pthread_self();
2614 mdb_mutex_t *rmutex = MDB_MUTEX(env, r);
2616 if (!env->me_live_reader) {
2617 rc = mdb_reader_pid(env, Pidset, pid);
2620 env->me_live_reader = 1;
2623 if (LOCK_MUTEX(rc, env, rmutex))
2625 nr = ti->mti_numreaders;
2626 for (i=0; i<nr; i++)
2627 if (ti->mti_readers[i].mr_pid == 0)
2629 if (i == env->me_maxreaders) {
2630 UNLOCK_MUTEX(rmutex);
2631 return MDB_READERS_FULL;
2633 r = &ti->mti_readers[i];
2634 /* Claim the reader slot, carefully since other code
2635 * uses the reader table un-mutexed: First reset the
2636 * slot, next publish it in mti_numreaders. After
2637 * that, it is safe for mdb_env_close() to touch it.
2638 * When it will be closed, we can finally claim it.
2641 r->mr_txnid = (txnid_t)-1;
2644 ti->mti_numreaders = ++nr;
2645 env->me_close_readers = nr;
2647 UNLOCK_MUTEX(rmutex);
2649 new_notls = (env->me_flags & MDB_NOTLS);
2650 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2655 do /* LY: Retry on a race, ITS#7970. */
2656 r->mr_txnid = ti->mti_txnid;
2657 while(r->mr_txnid != ti->mti_txnid);
2658 txn->mt_txnid = r->mr_txnid;
2659 txn->mt_u.reader = r;
2660 meta = env->me_metas[txn->mt_txnid & 1];
2662 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2664 /* Not yet touching txn == env->me_txn0, it may be active */
2666 if (LOCK_MUTEX(rc, env, MDB_MUTEX(env, w)))
2668 txn->mt_txnid = ti->mti_txnid;
2669 meta = env->me_metas[txn->mt_txnid & 1];
2671 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2672 txn->mt_txnid = meta->mm_txnid;
2676 if (txn->mt_txnid == mdb_debug_start)
2679 txn->mt_child = NULL;
2680 txn->mt_loose_pgs = NULL;
2681 txn->mt_loose_count = 0;
2682 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2683 txn->mt_u.dirty_list = env->me_dirty_list;
2684 txn->mt_u.dirty_list[0].mid = 0;
2685 txn->mt_free_pgs = env->me_free_pgs;
2686 txn->mt_free_pgs[0] = 0;
2687 txn->mt_spill_pgs = NULL;
2689 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2692 /* Copy the DB info and flags */
2693 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2695 /* Moved to here to avoid a data race in read TXNs */
2696 txn->mt_next_pgno = meta->mm_last_pg+1;
2698 txn->mt_flags = flags;
2701 txn->mt_numdbs = env->me_numdbs;
2702 for (i=2; i<txn->mt_numdbs; i++) {
2703 x = env->me_dbflags[i];
2704 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2705 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2707 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2709 if (env->me_maxpg < txn->mt_next_pgno) {
2710 mdb_txn_reset0(txn, "renew0-mapfail");
2712 txn->mt_u.reader->mr_pid = 0;
2713 txn->mt_u.reader = NULL;
2715 return MDB_MAP_RESIZED;
2722 mdb_txn_renew(MDB_txn *txn)
2726 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2729 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2730 DPUTS("environment had fatal error, must shutdown!");
2734 rc = mdb_txn_renew0(txn);
2735 if (rc == MDB_SUCCESS) {
2736 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2737 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2738 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2744 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2748 int rc, size, tsize;
2750 flags &= MDB_TXN_BEGIN_FLAGS;
2751 flags |= env->me_flags & MDB_WRITEMAP;
2753 if (env->me_flags & MDB_FATAL_ERROR) {
2754 DPUTS("environment had fatal error, must shutdown!");
2757 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2760 size = tsize = sizeof(MDB_txn);
2762 /* Nested transactions: Max 1 child, write txns only, no writemap */
2763 flags |= parent->mt_flags;
2764 if (parent->mt_child ||
2765 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2767 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2769 /* Child txns save MDB_pgstate and use own copy of cursors */
2770 size = tsize = sizeof(MDB_ntxn);
2771 size += env->me_maxdbs * sizeof(MDB_cursor *);
2772 } else if (!(flags & MDB_RDONLY)) {
2773 /* Reuse preallocated write txn. However, do not touch it until
2774 * mdb_txn_renew0() succeeds, since it currently may be active.
2779 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2781 if ((txn = calloc(1, size)) == NULL) {
2782 DPRINTF(("calloc: %s", strerror(errno)));
2785 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2786 if (flags & MDB_RDONLY) {
2787 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2788 txn->mt_dbiseqs = env->me_dbiseqs;
2790 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2792 txn->mt_dbiseqs = parent->mt_dbiseqs;
2793 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2795 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2796 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2799 txn->mt_flags = flags;
2804 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2805 if (!txn->mt_u.dirty_list ||
2806 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2808 free(txn->mt_u.dirty_list);
2812 txn->mt_txnid = parent->mt_txnid;
2813 txn->mt_dirty_room = parent->mt_dirty_room;
2814 txn->mt_u.dirty_list[0].mid = 0;
2815 txn->mt_spill_pgs = NULL;
2816 txn->mt_next_pgno = parent->mt_next_pgno;
2817 parent->mt_child = txn;
2818 txn->mt_parent = parent;
2819 txn->mt_numdbs = parent->mt_numdbs;
2820 txn->mt_dbxs = parent->mt_dbxs;
2821 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2822 /* Copy parent's mt_dbflags, but clear DB_NEW */
2823 for (i=0; i<txn->mt_numdbs; i++)
2824 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2826 ntxn = (MDB_ntxn *)txn;
2827 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2828 if (env->me_pghead) {
2829 size = MDB_IDL_SIZEOF(env->me_pghead);
2830 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2832 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2837 rc = mdb_cursor_shadow(parent, txn);
2839 mdb_txn_reset0(txn, "beginchild-fail");
2842 rc = mdb_txn_renew0(txn);
2845 if (txn != env->me_txn0)
2848 txn->mt_flags |= flags; /* for txn==me_txn0, no effect otherwise */
2850 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2851 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2852 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2859 mdb_txn_env(MDB_txn *txn)
2861 if(!txn) return NULL;
2866 mdb_txn_id(MDB_txn *txn)
2869 return txn->mt_txnid;
2872 /** Export or close DBI handles opened in this txn. */
2874 mdb_dbis_update(MDB_txn *txn, int keep)
2877 MDB_dbi n = txn->mt_numdbs;
2878 MDB_env *env = txn->mt_env;
2879 unsigned char *tdbflags = txn->mt_dbflags;
2881 for (i = n; --i >= 2;) {
2882 if (tdbflags[i] & DB_NEW) {
2884 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2886 char *ptr = env->me_dbxs[i].md_name.mv_data;
2888 env->me_dbxs[i].md_name.mv_data = NULL;
2889 env->me_dbxs[i].md_name.mv_size = 0;
2890 env->me_dbflags[i] = 0;
2891 env->me_dbiseqs[i]++;
2897 if (keep && env->me_numdbs < n)
2901 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2902 * May be called twice for readonly txns: First reset it, then abort.
2903 * @param[in] txn the transaction handle to reset
2904 * @param[in] act why the transaction is being reset
2907 mdb_txn_reset0(MDB_txn *txn, const char *act)
2909 MDB_env *env = txn->mt_env;
2911 /* Close any DBI handles opened in this txn */
2912 mdb_dbis_update(txn, 0);
2914 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2915 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2916 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2918 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2919 if (txn->mt_u.reader) {
2920 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2921 if (!(env->me_flags & MDB_NOTLS))
2922 txn->mt_u.reader = NULL; /* txn does not own reader */
2924 txn->mt_numdbs = 0; /* close nothing if called again */
2925 txn->mt_dbxs = NULL; /* mark txn as reset */
2927 pgno_t *pghead = env->me_pghead;
2929 mdb_cursors_close(txn, 0);
2930 if (!(env->me_flags & MDB_WRITEMAP)) {
2931 mdb_dlist_free(txn);
2934 if (!txn->mt_parent) {
2935 if (mdb_midl_shrink(&txn->mt_free_pgs))
2936 env->me_free_pgs = txn->mt_free_pgs;
2938 env->me_pghead = NULL;
2942 /* The writer mutex was locked in mdb_txn_begin. */
2944 UNLOCK_MUTEX(MDB_MUTEX(env, w));
2946 txn->mt_parent->mt_child = NULL;
2947 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2948 mdb_midl_free(txn->mt_free_pgs);
2949 mdb_midl_free(txn->mt_spill_pgs);
2950 free(txn->mt_u.dirty_list);
2953 mdb_midl_free(pghead);
2958 mdb_txn_reset(MDB_txn *txn)
2963 /* This call is only valid for read-only txns */
2964 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2967 mdb_txn_reset0(txn, "reset");
2971 mdb_txn_abort(MDB_txn *txn)
2977 mdb_txn_abort(txn->mt_child);
2979 mdb_txn_reset0(txn, "abort");
2980 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2981 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2982 txn->mt_u.reader->mr_pid = 0;
2984 if (txn != txn->mt_env->me_txn0)
2988 /** Save the freelist as of this transaction to the freeDB.
2989 * This changes the freelist. Keep trying until it stabilizes.
2992 mdb_freelist_save(MDB_txn *txn)
2994 /* env->me_pghead[] can grow and shrink during this call.
2995 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2996 * Page numbers cannot disappear from txn->mt_free_pgs[].
2999 MDB_env *env = txn->mt_env;
3000 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3001 txnid_t pglast = 0, head_id = 0;
3002 pgno_t freecnt = 0, *free_pgs, *mop;
3003 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3005 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3007 if (env->me_pghead) {
3008 /* Make sure first page of freeDB is touched and on freelist */
3009 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3010 if (rc && rc != MDB_NOTFOUND)
3014 if (!env->me_pghead && txn->mt_loose_pgs) {
3015 /* Put loose page numbers in mt_free_pgs, since
3016 * we may be unable to return them to me_pghead.
3018 MDB_page *mp = txn->mt_loose_pgs;
3019 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3021 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3022 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3023 txn->mt_loose_pgs = NULL;
3024 txn->mt_loose_count = 0;
3027 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3028 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3029 ? SSIZE_MAX : maxfree_1pg;
3032 /* Come back here after each Put() in case freelist changed */
3037 /* If using records from freeDB which we have not yet
3038 * deleted, delete them and any we reserved for me_pghead.
3040 while (pglast < env->me_pglast) {
3041 rc = mdb_cursor_first(&mc, &key, NULL);
3044 pglast = head_id = *(txnid_t *)key.mv_data;
3045 total_room = head_room = 0;
3046 mdb_tassert(txn, pglast <= env->me_pglast);
3047 rc = mdb_cursor_del(&mc, 0);
3052 /* Save the IDL of pages freed by this txn, to a single record */
3053 if (freecnt < txn->mt_free_pgs[0]) {
3055 /* Make sure last page of freeDB is touched and on freelist */
3056 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3057 if (rc && rc != MDB_NOTFOUND)
3060 free_pgs = txn->mt_free_pgs;
3061 /* Write to last page of freeDB */
3062 key.mv_size = sizeof(txn->mt_txnid);
3063 key.mv_data = &txn->mt_txnid;
3065 freecnt = free_pgs[0];
3066 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3067 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3070 /* Retry if mt_free_pgs[] grew during the Put() */
3071 free_pgs = txn->mt_free_pgs;
3072 } while (freecnt < free_pgs[0]);
3073 mdb_midl_sort(free_pgs);
3074 memcpy(data.mv_data, free_pgs, data.mv_size);
3077 unsigned int i = free_pgs[0];
3078 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3079 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3081 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3087 mop = env->me_pghead;
3088 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3090 /* Reserve records for me_pghead[]. Split it if multi-page,
3091 * to avoid searching freeDB for a page range. Use keys in
3092 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3094 if (total_room >= mop_len) {
3095 if (total_room == mop_len || --more < 0)
3097 } else if (head_room >= maxfree_1pg && head_id > 1) {
3098 /* Keep current record (overflow page), add a new one */
3102 /* (Re)write {key = head_id, IDL length = head_room} */
3103 total_room -= head_room;
3104 head_room = mop_len - total_room;
3105 if (head_room > maxfree_1pg && head_id > 1) {
3106 /* Overflow multi-page for part of me_pghead */
3107 head_room /= head_id; /* amortize page sizes */
3108 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3109 } else if (head_room < 0) {
3110 /* Rare case, not bothering to delete this record */
3113 key.mv_size = sizeof(head_id);
3114 key.mv_data = &head_id;
3115 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3116 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3119 /* IDL is initially empty, zero out at least the length */
3120 pgs = (pgno_t *)data.mv_data;
3121 j = head_room > clean_limit ? head_room : 0;
3125 total_room += head_room;
3128 /* Return loose page numbers to me_pghead, though usually none are
3129 * left at this point. The pages themselves remain in dirty_list.
3131 if (txn->mt_loose_pgs) {
3132 MDB_page *mp = txn->mt_loose_pgs;
3133 unsigned count = txn->mt_loose_count;
3135 /* Room for loose pages + temp IDL with same */
3136 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3138 mop = env->me_pghead;
3139 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3140 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3141 loose[ ++count ] = mp->mp_pgno;
3143 mdb_midl_sort(loose);
3144 mdb_midl_xmerge(mop, loose);
3145 txn->mt_loose_pgs = NULL;
3146 txn->mt_loose_count = 0;
3150 /* Fill in the reserved me_pghead records */
3156 rc = mdb_cursor_first(&mc, &key, &data);
3157 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3158 txnid_t id = *(txnid_t *)key.mv_data;
3159 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3162 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3164 if (len > mop_len) {
3166 data.mv_size = (len + 1) * sizeof(MDB_ID);
3168 data.mv_data = mop -= len;
3171 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3173 if (rc || !(mop_len -= len))
3180 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3181 * @param[in] txn the transaction that's being committed
3182 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3183 * @return 0 on success, non-zero on failure.
3186 mdb_page_flush(MDB_txn *txn, int keep)
3188 MDB_env *env = txn->mt_env;
3189 MDB_ID2L dl = txn->mt_u.dirty_list;
3190 unsigned psize = env->me_psize, j;
3191 int i, pagecount = dl[0].mid, rc;
3192 size_t size = 0, pos = 0;
3194 MDB_page *dp = NULL;
3198 struct iovec iov[MDB_COMMIT_PAGES];
3199 ssize_t wpos = 0, wsize = 0, wres;
3200 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3206 if (env->me_flags & MDB_WRITEMAP) {
3207 /* Clear dirty flags */
3208 while (++i <= pagecount) {
3210 /* Don't flush this page yet */
3211 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3212 dp->mp_flags &= ~P_KEEP;
3216 dp->mp_flags &= ~P_DIRTY;
3221 /* Write the pages */
3223 if (++i <= pagecount) {
3225 /* Don't flush this page yet */
3226 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3227 dp->mp_flags &= ~P_KEEP;
3232 /* clear dirty flag */
3233 dp->mp_flags &= ~P_DIRTY;
3236 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3241 /* Windows actually supports scatter/gather I/O, but only on
3242 * unbuffered file handles. Since we're relying on the OS page
3243 * cache for all our data, that's self-defeating. So we just
3244 * write pages one at a time. We use the ov structure to set
3245 * the write offset, to at least save the overhead of a Seek
3248 DPRINTF(("committing page %"Z"u", pgno));
3249 memset(&ov, 0, sizeof(ov));
3250 ov.Offset = pos & 0xffffffff;
3251 ov.OffsetHigh = pos >> 16 >> 16;
3252 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3254 DPRINTF(("WriteFile: %d", rc));
3258 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3259 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3262 /* Write previous page(s) */
3263 #ifdef MDB_USE_PWRITEV
3264 wres = pwritev(env->me_fd, iov, n, wpos);
3267 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3270 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3274 DPRINTF(("lseek: %s", strerror(rc)));
3277 wres = writev(env->me_fd, iov, n);
3280 if (wres != wsize) {
3285 DPRINTF(("Write error: %s", strerror(rc)));
3287 rc = EIO; /* TODO: Use which error code? */
3288 DPUTS("short write, filesystem full?");
3299 DPRINTF(("committing page %"Z"u", pgno));
3300 next_pos = pos + size;
3301 iov[n].iov_len = size;
3302 iov[n].iov_base = (char *)dp;
3308 /* MIPS has cache coherency issues, this is a no-op everywhere else
3309 * Note: for any size >= on-chip cache size, entire on-chip cache is
3312 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3314 for (i = keep; ++i <= pagecount; ) {
3316 /* This is a page we skipped above */
3319 dl[j].mid = dp->mp_pgno;
3322 mdb_dpage_free(env, dp);
3327 txn->mt_dirty_room += i - j;
3333 mdb_txn_commit(MDB_txn *txn)
3339 if (txn == NULL || txn->mt_env == NULL)
3342 if (txn->mt_child) {
3343 rc = mdb_txn_commit(txn->mt_child);
3344 txn->mt_child = NULL;
3351 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3352 mdb_dbis_update(txn, 1);
3353 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3358 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3359 DPUTS("error flag is set, can't commit");
3361 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3366 if (txn->mt_parent) {
3367 MDB_txn *parent = txn->mt_parent;
3371 unsigned x, y, len, ps_len;
3373 /* Append our free list to parent's */
3374 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3377 mdb_midl_free(txn->mt_free_pgs);
3378 /* Failures after this must either undo the changes
3379 * to the parent or set MDB_TXN_ERROR in the parent.
3382 parent->mt_next_pgno = txn->mt_next_pgno;
3383 parent->mt_flags = txn->mt_flags;
3385 /* Merge our cursors into parent's and close them */
3386 mdb_cursors_close(txn, 1);
3388 /* Update parent's DB table. */
3389 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3390 parent->mt_numdbs = txn->mt_numdbs;
3391 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3392 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3393 for (i=2; i<txn->mt_numdbs; i++) {
3394 /* preserve parent's DB_NEW status */
3395 x = parent->mt_dbflags[i] & DB_NEW;
3396 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3399 dst = parent->mt_u.dirty_list;
3400 src = txn->mt_u.dirty_list;
3401 /* Remove anything in our dirty list from parent's spill list */
3402 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3404 pspill[0] = (pgno_t)-1;
3405 /* Mark our dirty pages as deleted in parent spill list */
3406 for (i=0, len=src[0].mid; ++i <= len; ) {
3407 MDB_ID pn = src[i].mid << 1;
3408 while (pn > pspill[x])
3410 if (pn == pspill[x]) {
3415 /* Squash deleted pagenums if we deleted any */
3416 for (x=y; ++x <= ps_len; )
3417 if (!(pspill[x] & 1))
3418 pspill[++y] = pspill[x];
3422 /* Find len = length of merging our dirty list with parent's */
3424 dst[0].mid = 0; /* simplify loops */
3425 if (parent->mt_parent) {
3426 len = x + src[0].mid;
3427 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3428 for (i = x; y && i; y--) {
3429 pgno_t yp = src[y].mid;
3430 while (yp < dst[i].mid)
3432 if (yp == dst[i].mid) {
3437 } else { /* Simplify the above for single-ancestor case */
3438 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3440 /* Merge our dirty list with parent's */
3442 for (i = len; y; dst[i--] = src[y--]) {
3443 pgno_t yp = src[y].mid;
3444 while (yp < dst[x].mid)
3445 dst[i--] = dst[x--];
3446 if (yp == dst[x].mid)
3447 free(dst[x--].mptr);
3449 mdb_tassert(txn, i == x);
3451 free(txn->mt_u.dirty_list);
3452 parent->mt_dirty_room = txn->mt_dirty_room;
3453 if (txn->mt_spill_pgs) {
3454 if (parent->mt_spill_pgs) {
3455 /* TODO: Prevent failure here, so parent does not fail */
3456 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3458 parent->mt_flags |= MDB_TXN_ERROR;
3459 mdb_midl_free(txn->mt_spill_pgs);
3460 mdb_midl_sort(parent->mt_spill_pgs);
3462 parent->mt_spill_pgs = txn->mt_spill_pgs;
3466 /* Append our loose page list to parent's */
3467 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3469 *lp = txn->mt_loose_pgs;
3470 parent->mt_loose_count += txn->mt_loose_count;
3472 parent->mt_child = NULL;
3473 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3478 if (txn != env->me_txn) {
3479 DPUTS("attempt to commit unknown transaction");
3484 mdb_cursors_close(txn, 0);
3486 if (!txn->mt_u.dirty_list[0].mid &&
3487 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3490 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3491 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3493 /* Update DB root pointers */
3494 if (txn->mt_numdbs > 2) {
3498 data.mv_size = sizeof(MDB_db);
3500 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3501 for (i = 2; i < txn->mt_numdbs; i++) {
3502 if (txn->mt_dbflags[i] & DB_DIRTY) {
3503 if (TXN_DBI_CHANGED(txn, i)) {
3507 data.mv_data = &txn->mt_dbs[i];
3508 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3515 rc = mdb_freelist_save(txn);
3519 mdb_midl_free(env->me_pghead);
3520 env->me_pghead = NULL;
3521 if (mdb_midl_shrink(&txn->mt_free_pgs))
3522 env->me_free_pgs = txn->mt_free_pgs;
3528 if ((rc = mdb_page_flush(txn, 0)))
3530 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3531 (rc = mdb_env_sync(env, 0)))
3533 if ((rc = mdb_env_write_meta(txn)))
3536 /* Free P_LOOSE pages left behind in dirty_list */
3537 if (!(env->me_flags & MDB_WRITEMAP))
3538 mdb_dlist_free(txn);
3543 mdb_dbis_update(txn, 1);
3546 UNLOCK_MUTEX(MDB_MUTEX(env, w));
3547 if (txn != env->me_txn0)
3557 /** Read the environment parameters of a DB environment before
3558 * mapping it into memory.
3559 * @param[in] env the environment handle
3560 * @param[out] meta address of where to store the meta information
3561 * @return 0 on success, non-zero on failure.
3564 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3570 enum { Size = sizeof(pbuf) };
3572 /* We don't know the page size yet, so use a minimum value.
3573 * Read both meta pages so we can use the latest one.
3576 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3580 memset(&ov, 0, sizeof(ov));
3582 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3583 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3586 rc = pread(env->me_fd, &pbuf, Size, off);
3589 if (rc == 0 && off == 0)
3591 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3592 DPRINTF(("read: %s", mdb_strerror(rc)));
3596 p = (MDB_page *)&pbuf;
3598 if (!F_ISSET(p->mp_flags, P_META)) {
3599 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3604 if (m->mm_magic != MDB_MAGIC) {
3605 DPUTS("meta has invalid magic");
3609 if (m->mm_version != MDB_DATA_VERSION) {
3610 DPRINTF(("database is version %u, expected version %u",
3611 m->mm_version, MDB_DATA_VERSION));
3612 return MDB_VERSION_MISMATCH;
3615 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3621 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3623 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3625 meta->mm_magic = MDB_MAGIC;
3626 meta->mm_version = MDB_DATA_VERSION;
3627 meta->mm_mapsize = env->me_mapsize;
3628 meta->mm_psize = env->me_psize;
3629 meta->mm_last_pg = 1;
3630 meta->mm_flags = env->me_flags & 0xffff;
3631 meta->mm_flags |= MDB_INTEGERKEY;
3632 meta->mm_dbs[0].md_root = P_INVALID;
3633 meta->mm_dbs[1].md_root = P_INVALID;
3636 /** Write the environment parameters of a freshly created DB environment.
3637 * @param[in] env the environment handle
3638 * @param[in] meta the #MDB_meta to write
3639 * @return 0 on success, non-zero on failure.
3642 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3650 memset(&ov, 0, sizeof(ov));
3651 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3653 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3656 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3657 len = pwrite(fd, ptr, size, pos); \
3658 if (len == -1 && ErrCode() == EINTR) continue; \
3659 rc = (len >= 0); break; } while(1)
3662 DPUTS("writing new meta page");
3664 psize = env->me_psize;
3666 p = calloc(2, psize);
3668 p->mp_flags = P_META;
3669 *(MDB_meta *)METADATA(p) = *meta;
3671 q = (MDB_page *)((char *)p + psize);
3673 q->mp_flags = P_META;
3674 *(MDB_meta *)METADATA(q) = *meta;
3676 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3679 else if ((unsigned) len == psize * 2)
3687 /** Update the environment info to commit a transaction.
3688 * @param[in] txn the transaction that's being committed
3689 * @return 0 on success, non-zero on failure.
3692 mdb_env_write_meta(MDB_txn *txn)
3695 MDB_meta meta, metab, *mp;
3699 int rc, len, toggle;
3708 toggle = txn->mt_txnid & 1;
3709 DPRINTF(("writing meta page %d for root page %"Z"u",
3710 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3713 flags = txn->mt_flags & env->me_flags;
3714 mp = env->me_metas[toggle];
3715 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3716 /* Persist any increases of mapsize config */
3717 if (mapsize < env->me_mapsize)
3718 mapsize = env->me_mapsize;
3720 if (flags & MDB_WRITEMAP) {
3721 mp->mm_mapsize = mapsize;
3722 mp->mm_dbs[0] = txn->mt_dbs[0];
3723 mp->mm_dbs[1] = txn->mt_dbs[1];
3724 mp->mm_last_pg = txn->mt_next_pgno - 1;
3725 #if !(defined(_MSC_VER) || defined(__i386__) || defined(__x86_64__))
3726 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3727 __sync_synchronize();
3729 mp->mm_txnid = txn->mt_txnid;
3730 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3731 unsigned meta_size = env->me_psize;
3732 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3735 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3736 if (meta_size < env->me_os_psize)
3737 meta_size += meta_size;
3742 if (MDB_MSYNC(ptr, meta_size, rc)) {
3749 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3750 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3752 meta.mm_mapsize = mapsize;
3753 meta.mm_dbs[0] = txn->mt_dbs[0];
3754 meta.mm_dbs[1] = txn->mt_dbs[1];
3755 meta.mm_last_pg = txn->mt_next_pgno - 1;
3756 meta.mm_txnid = txn->mt_txnid;
3758 off = offsetof(MDB_meta, mm_mapsize);
3759 ptr = (char *)&meta + off;
3760 len = sizeof(MDB_meta) - off;
3762 off += env->me_psize;
3765 /* Write to the SYNC fd */
3766 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3770 memset(&ov, 0, sizeof(ov));
3772 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3776 rc = pwrite(mfd, ptr, len, off);
3779 rc = rc < 0 ? ErrCode() : EIO;
3782 DPUTS("write failed, disk error?");
3783 /* On a failure, the pagecache still contains the new data.
3784 * Write some old data back, to prevent it from being used.
3785 * Use the non-SYNC fd; we know it will fail anyway.
3787 meta.mm_last_pg = metab.mm_last_pg;
3788 meta.mm_txnid = metab.mm_txnid;
3790 memset(&ov, 0, sizeof(ov));
3792 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3794 r2 = pwrite(env->me_fd, ptr, len, off);
3795 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3798 env->me_flags |= MDB_FATAL_ERROR;
3801 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3802 CACHEFLUSH(env->me_map + off, len, DCACHE);
3804 /* Memory ordering issues are irrelevant; since the entire writer
3805 * is wrapped by wmutex, all of these changes will become visible
3806 * after the wmutex is unlocked. Since the DB is multi-version,
3807 * readers will get consistent data regardless of how fresh or
3808 * how stale their view of these values is.
3811 env->me_txns->mti_txnid = txn->mt_txnid;
3816 /** Check both meta pages to see which one is newer.
3817 * @param[in] env the environment handle
3818 * @return meta toggle (0 or 1).
3821 mdb_env_pick_meta(const MDB_env *env)
3823 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3827 mdb_env_create(MDB_env **env)
3831 e = calloc(1, sizeof(MDB_env));
3835 e->me_maxreaders = DEFAULT_READERS;
3836 e->me_maxdbs = e->me_numdbs = 2;
3837 e->me_fd = INVALID_HANDLE_VALUE;
3838 e->me_lfd = INVALID_HANDLE_VALUE;
3839 e->me_mfd = INVALID_HANDLE_VALUE;
3840 #ifdef MDB_USE_SYSV_SEM
3841 e->me_rmutex.semid = -1;
3842 e->me_wmutex.semid = -1;
3844 e->me_pid = getpid();
3845 GET_PAGESIZE(e->me_os_psize);
3846 VGMEMP_CREATE(e,0,0);
3852 mdb_env_map(MDB_env *env, void *addr)
3855 unsigned int flags = env->me_flags;
3859 LONG sizelo, sizehi;
3862 if (flags & MDB_RDONLY) {
3863 /* Don't set explicit map size, use whatever exists */
3868 msize = env->me_mapsize;
3869 sizelo = msize & 0xffffffff;
3870 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3872 /* Windows won't create mappings for zero length files.
3873 * and won't map more than the file size.
3874 * Just set the maxsize right now.
3876 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3877 || !SetEndOfFile(env->me_fd)
3878 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3882 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3883 PAGE_READWRITE : PAGE_READONLY,
3884 sizehi, sizelo, NULL);
3887 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3888 FILE_MAP_WRITE : FILE_MAP_READ,
3890 rc = env->me_map ? 0 : ErrCode();
3895 int prot = PROT_READ;
3896 if (flags & MDB_WRITEMAP) {
3898 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3901 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3903 if (env->me_map == MAP_FAILED) {
3908 if (flags & MDB_NORDAHEAD) {
3909 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3911 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3913 #ifdef POSIX_MADV_RANDOM
3914 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3915 #endif /* POSIX_MADV_RANDOM */
3916 #endif /* MADV_RANDOM */
3920 /* Can happen because the address argument to mmap() is just a
3921 * hint. mmap() can pick another, e.g. if the range is in use.
3922 * The MAP_FIXED flag would prevent that, but then mmap could
3923 * instead unmap existing pages to make room for the new map.
3925 if (addr && env->me_map != addr)
3926 return EBUSY; /* TODO: Make a new MDB_* error code? */
3928 p = (MDB_page *)env->me_map;
3929 env->me_metas[0] = METADATA(p);
3930 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3936 mdb_env_set_mapsize(MDB_env *env, size_t size)
3938 /* If env is already open, caller is responsible for making
3939 * sure there are no active txns.
3947 meta = env->me_metas[mdb_env_pick_meta(env)];
3949 size = meta->mm_mapsize;
3951 /* Silently round up to minimum if the size is too small */
3952 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3956 munmap(env->me_map, env->me_mapsize);
3957 env->me_mapsize = size;
3958 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3959 rc = mdb_env_map(env, old);
3963 env->me_mapsize = size;
3965 env->me_maxpg = env->me_mapsize / env->me_psize;
3970 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3974 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3979 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3981 if (env->me_map || readers < 1)
3983 env->me_maxreaders = readers;
3988 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3990 if (!env || !readers)
3992 *readers = env->me_maxreaders;
3997 mdb_fsize(HANDLE fd, size_t *size)
4000 LARGE_INTEGER fsize;
4002 if (!GetFileSizeEx(fd, &fsize))
4005 *size = fsize.QuadPart;
4017 #ifdef BROKEN_FDATASYNC
4018 #include <sys/utsname.h>
4019 #include <sys/vfs.h>
4022 /** Further setup required for opening an LMDB environment
4025 mdb_env_open2(MDB_env *env)
4027 unsigned int flags = env->me_flags;
4028 int i, newenv = 0, rc;
4032 /* See if we should use QueryLimited */
4034 if ((rc & 0xff) > 5)
4035 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4037 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4040 #ifdef BROKEN_FDATASYNC
4041 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4042 * https://lkml.org/lkml/2012/9/3/83
4043 * Kernels after 3.6-rc6 are known good.
4044 * https://lkml.org/lkml/2012/9/10/556
4045 * See if the DB is on ext3/ext4, then check for new enough kernel
4046 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4051 fstatfs(env->me_fd, &st);
4052 while (st.f_type == 0xEF53) {
4056 if (uts.release[0] < '3') {
4057 if (!strncmp(uts.release, "2.6.32.", 7)) {
4058 i = atoi(uts.release+7);
4060 break; /* 2.6.32.60 and newer is OK */
4061 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4062 i = atoi(uts.release+7);
4064 break; /* 2.6.34.15 and newer is OK */
4066 } else if (uts.release[0] == '3') {
4067 i = atoi(uts.release+2);
4069 break; /* 3.6 and newer is OK */
4071 i = atoi(uts.release+4);
4073 break; /* 3.5.4 and newer is OK */
4074 } else if (i == 2) {
4075 i = atoi(uts.release+4);
4077 break; /* 3.2.30 and newer is OK */
4079 } else { /* 4.x and newer is OK */
4082 env->me_flags |= MDB_FSYNCONLY;
4088 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4091 DPUTS("new mdbenv");
4093 env->me_psize = env->me_os_psize;
4094 if (env->me_psize > MAX_PAGESIZE)
4095 env->me_psize = MAX_PAGESIZE;
4096 memset(&meta, 0, sizeof(meta));
4097 mdb_env_init_meta0(env, &meta);
4098 meta.mm_mapsize = DEFAULT_MAPSIZE;
4100 env->me_psize = meta.mm_psize;
4103 /* Was a mapsize configured? */
4104 if (!env->me_mapsize) {
4105 env->me_mapsize = meta.mm_mapsize;
4108 /* Make sure mapsize >= committed data size. Even when using
4109 * mm_mapsize, which could be broken in old files (ITS#7789).
4111 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4112 if (env->me_mapsize < minsize)
4113 env->me_mapsize = minsize;
4115 meta.mm_mapsize = env->me_mapsize;
4117 if (newenv && !(flags & MDB_FIXEDMAP)) {
4118 /* mdb_env_map() may grow the datafile. Write the metapages
4119 * first, so the file will be valid if initialization fails.
4120 * Except with FIXEDMAP, since we do not yet know mm_address.
4121 * We could fill in mm_address later, but then a different
4122 * program might end up doing that - one with a memory layout
4123 * and map address which does not suit the main program.
4125 rc = mdb_env_init_meta(env, &meta);
4131 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4136 if (flags & MDB_FIXEDMAP)
4137 meta.mm_address = env->me_map;
4138 i = mdb_env_init_meta(env, &meta);
4139 if (i != MDB_SUCCESS) {
4144 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4145 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4147 #if !(MDB_MAXKEYSIZE)
4148 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4150 env->me_maxpg = env->me_mapsize / env->me_psize;
4154 int toggle = mdb_env_pick_meta(env);
4155 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4157 DPRINTF(("opened database version %u, pagesize %u",
4158 env->me_metas[0]->mm_version, env->me_psize));
4159 DPRINTF(("using meta page %d", toggle));
4160 DPRINTF(("depth: %u", db->md_depth));
4161 DPRINTF(("entries: %"Z"u", db->md_entries));
4162 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4163 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4164 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4165 DPRINTF(("root: %"Z"u", db->md_root));
4173 /** Release a reader thread's slot in the reader lock table.
4174 * This function is called automatically when a thread exits.
4175 * @param[in] ptr This points to the slot in the reader lock table.
4178 mdb_env_reader_dest(void *ptr)
4180 MDB_reader *reader = ptr;
4186 /** Junk for arranging thread-specific callbacks on Windows. This is
4187 * necessarily platform and compiler-specific. Windows supports up
4188 * to 1088 keys. Let's assume nobody opens more than 64 environments
4189 * in a single process, for now. They can override this if needed.
4191 #ifndef MAX_TLS_KEYS
4192 #define MAX_TLS_KEYS 64
4194 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4195 static int mdb_tls_nkeys;
4197 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4201 case DLL_PROCESS_ATTACH: break;
4202 case DLL_THREAD_ATTACH: break;
4203 case DLL_THREAD_DETACH:
4204 for (i=0; i<mdb_tls_nkeys; i++) {
4205 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4207 mdb_env_reader_dest(r);
4211 case DLL_PROCESS_DETACH: break;
4216 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4218 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4222 /* Force some symbol references.
4223 * _tls_used forces the linker to create the TLS directory if not already done
4224 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4226 #pragma comment(linker, "/INCLUDE:_tls_used")
4227 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4228 #pragma const_seg(".CRT$XLB")
4229 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4230 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4233 #pragma comment(linker, "/INCLUDE:__tls_used")
4234 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4235 #pragma data_seg(".CRT$XLB")
4236 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4238 #endif /* WIN 32/64 */
4239 #endif /* !__GNUC__ */
4242 /** Downgrade the exclusive lock on the region back to shared */
4244 mdb_env_share_locks(MDB_env *env, int *excl)
4246 int rc = 0, toggle = mdb_env_pick_meta(env);
4248 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4253 /* First acquire a shared lock. The Unlock will
4254 * then release the existing exclusive lock.
4256 memset(&ov, 0, sizeof(ov));
4257 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4260 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4266 struct flock lock_info;
4267 /* The shared lock replaces the existing lock */
4268 memset((void *)&lock_info, 0, sizeof(lock_info));
4269 lock_info.l_type = F_RDLCK;
4270 lock_info.l_whence = SEEK_SET;
4271 lock_info.l_start = 0;
4272 lock_info.l_len = 1;
4273 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4274 (rc = ErrCode()) == EINTR) ;
4275 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4282 /** Try to get exclusive lock, otherwise shared.
4283 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4286 mdb_env_excl_lock(MDB_env *env, int *excl)
4290 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4294 memset(&ov, 0, sizeof(ov));
4295 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4302 struct flock lock_info;
4303 memset((void *)&lock_info, 0, sizeof(lock_info));
4304 lock_info.l_type = F_WRLCK;
4305 lock_info.l_whence = SEEK_SET;
4306 lock_info.l_start = 0;
4307 lock_info.l_len = 1;
4308 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4309 (rc = ErrCode()) == EINTR) ;
4313 # ifndef MDB_USE_POSIX_MUTEX
4314 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4317 lock_info.l_type = F_RDLCK;
4318 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4319 (rc = ErrCode()) == EINTR) ;
4329 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4331 * @(#) $Revision: 5.1 $
4332 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4333 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4335 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4339 * Please do not copyright this code. This code is in the public domain.
4341 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4342 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4343 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4344 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4345 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4346 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4347 * PERFORMANCE OF THIS SOFTWARE.
4350 * chongo <Landon Curt Noll> /\oo/\
4351 * http://www.isthe.com/chongo/
4353 * Share and Enjoy! :-)
4356 typedef unsigned long long mdb_hash_t;
4357 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4359 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4360 * @param[in] val value to hash
4361 * @param[in] hval initial value for hash
4362 * @return 64 bit hash
4364 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4365 * hval arg on the first call.
4368 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4370 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4371 unsigned char *end = s + val->mv_size;
4373 * FNV-1a hash each octet of the string
4376 /* xor the bottom with the current octet */
4377 hval ^= (mdb_hash_t)*s++;
4379 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4380 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4381 (hval << 7) + (hval << 8) + (hval << 40);
4383 /* return our new hash value */
4387 /** Hash the string and output the encoded hash.
4388 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4389 * very short name limits. We don't care about the encoding being reversible,
4390 * we just want to preserve as many bits of the input as possible in a
4391 * small printable string.
4392 * @param[in] str string to hash
4393 * @param[out] encbuf an array of 11 chars to hold the hash
4395 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4398 mdb_pack85(unsigned long l, char *out)
4402 for (i=0; i<5; i++) {
4403 *out++ = mdb_a85[l % 85];
4409 mdb_hash_enc(MDB_val *val, char *encbuf)
4411 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4413 mdb_pack85(h, encbuf);
4414 mdb_pack85(h>>32, encbuf+5);
4419 /** Open and/or initialize the lock region for the environment.
4420 * @param[in] env The LMDB environment.
4421 * @param[in] lpath The pathname of the file used for the lock region.
4422 * @param[in] mode The Unix permissions for the file, if we create it.
4423 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4424 * @return 0 on success, non-zero on failure.
4427 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4430 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4432 # define MDB_ERRCODE_ROFS EROFS
4433 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4434 # define MDB_CLOEXEC O_CLOEXEC
4437 # define MDB_CLOEXEC 0
4440 #ifdef MDB_USE_SYSV_SEM
4448 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4449 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4450 FILE_ATTRIBUTE_NORMAL, NULL);
4452 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4454 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4456 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4461 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4462 /* Lose record locks when exec*() */
4463 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4464 fcntl(env->me_lfd, F_SETFD, fdflags);
4467 if (!(env->me_flags & MDB_NOTLS)) {
4468 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4471 env->me_flags |= MDB_ENV_TXKEY;
4473 /* Windows TLS callbacks need help finding their TLS info. */
4474 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4478 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4482 /* Try to get exclusive lock. If we succeed, then
4483 * nobody is using the lock region and we should initialize it.
4485 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4488 size = GetFileSize(env->me_lfd, NULL);
4490 size = lseek(env->me_lfd, 0, SEEK_END);
4491 if (size == -1) goto fail_errno;
4493 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4494 if (size < rsize && *excl > 0) {
4496 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4497 || !SetEndOfFile(env->me_lfd))
4500 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4504 size = rsize - sizeof(MDB_txninfo);
4505 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4510 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4512 if (!mh) goto fail_errno;
4513 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4515 if (!env->me_txns) goto fail_errno;
4517 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4519 if (m == MAP_FAILED) goto fail_errno;
4525 BY_HANDLE_FILE_INFORMATION stbuf;
4534 if (!mdb_sec_inited) {
4535 InitializeSecurityDescriptor(&mdb_null_sd,
4536 SECURITY_DESCRIPTOR_REVISION);
4537 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4538 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4539 mdb_all_sa.bInheritHandle = FALSE;
4540 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4543 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4544 idbuf.volume = stbuf.dwVolumeSerialNumber;
4545 idbuf.nhigh = stbuf.nFileIndexHigh;
4546 idbuf.nlow = stbuf.nFileIndexLow;
4547 val.mv_data = &idbuf;
4548 val.mv_size = sizeof(idbuf);
4549 mdb_hash_enc(&val, encbuf);
4550 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4551 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4552 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4553 if (!env->me_rmutex) goto fail_errno;
4554 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4555 if (!env->me_wmutex) goto fail_errno;
4556 #elif defined(MDB_USE_SYSV_SEM)
4557 unsigned short vals[2] = {1, 1};
4558 key_t key = ftok(lpath, 'M');
4561 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4565 if (semctl(semid, 0, SETALL, semu) < 0)
4567 env->me_txns->mti_semid = semid;
4568 #else /* MDB_USE_POSIX_MUTEX: */
4569 pthread_mutexattr_t mattr;
4571 if ((rc = pthread_mutexattr_init(&mattr))
4572 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4573 #ifdef MDB_ROBUST_SUPPORTED
4574 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4576 || (rc = pthread_mutex_init(&env->me_txns->mti_rmutex, &mattr))
4577 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4579 pthread_mutexattr_destroy(&mattr);
4580 #endif /* _WIN32 || MDB_USE_SYSV_SEM */
4582 env->me_txns->mti_magic = MDB_MAGIC;
4583 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4584 env->me_txns->mti_txnid = 0;
4585 env->me_txns->mti_numreaders = 0;
4588 #ifdef MDB_USE_SYSV_SEM
4589 struct semid_ds buf;
4591 if (env->me_txns->mti_magic != MDB_MAGIC) {
4592 DPUTS("lock region has invalid magic");
4596 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4597 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4598 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4599 rc = MDB_VERSION_MISMATCH;
4603 if (rc && rc != EACCES && rc != EAGAIN) {
4607 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4608 if (!env->me_rmutex) goto fail_errno;
4609 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4610 if (!env->me_wmutex) goto fail_errno;
4611 #elif defined(MDB_USE_SYSV_SEM)
4612 semid = env->me_txns->mti_semid;
4614 /* check for read access */
4615 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4617 /* check for write access */
4618 if (semctl(semid, 0, IPC_SET, semu) < 0)
4622 #ifdef MDB_USE_SYSV_SEM
4623 env->me_rmutex.semid = semid;
4624 env->me_wmutex.semid = semid;
4625 env->me_rmutex.semnum = 0;
4626 env->me_wmutex.semnum = 1;
4627 env->me_rmutex.locked = &env->me_txns->mti_rlocked;
4628 env->me_wmutex.locked = &env->me_txns->mti_wlocked;
4639 /** The name of the lock file in the DB environment */
4640 #define LOCKNAME "/lock.mdb"
4641 /** The name of the data file in the DB environment */
4642 #define DATANAME "/data.mdb"
4643 /** The suffix of the lock file when no subdir is used */
4644 #define LOCKSUFF "-lock"
4645 /** Only a subset of the @ref mdb_env flags can be changed
4646 * at runtime. Changing other flags requires closing the
4647 * environment and re-opening it with the new flags.
4649 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4650 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4651 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4653 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4654 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4658 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4660 int oflags, rc, len, excl = -1;
4661 char *lpath, *dpath;
4663 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4667 if (flags & MDB_NOSUBDIR) {
4668 rc = len + sizeof(LOCKSUFF) + len + 1;
4670 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4675 if (flags & MDB_NOSUBDIR) {
4676 dpath = lpath + len + sizeof(LOCKSUFF);
4677 sprintf(lpath, "%s" LOCKSUFF, path);
4678 strcpy(dpath, path);
4680 dpath = lpath + len + sizeof(LOCKNAME);
4681 sprintf(lpath, "%s" LOCKNAME, path);
4682 sprintf(dpath, "%s" DATANAME, path);
4686 flags |= env->me_flags;
4687 if (flags & MDB_RDONLY) {
4688 /* silently ignore WRITEMAP when we're only getting read access */
4689 flags &= ~MDB_WRITEMAP;
4691 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4692 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4695 env->me_flags = flags |= MDB_ENV_ACTIVE;
4699 env->me_path = strdup(path);
4700 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4701 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4702 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4703 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4708 /* For RDONLY, get lockfile after we know datafile exists */
4709 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4710 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4716 if (F_ISSET(flags, MDB_RDONLY)) {
4717 oflags = GENERIC_READ;
4718 len = OPEN_EXISTING;
4720 oflags = GENERIC_READ|GENERIC_WRITE;
4723 mode = FILE_ATTRIBUTE_NORMAL;
4724 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4725 NULL, len, mode, NULL);
4727 if (F_ISSET(flags, MDB_RDONLY))
4730 oflags = O_RDWR | O_CREAT;
4732 env->me_fd = open(dpath, oflags, mode);
4734 if (env->me_fd == INVALID_HANDLE_VALUE) {
4739 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4740 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4745 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4746 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4747 env->me_mfd = env->me_fd;
4749 /* Synchronous fd for meta writes. Needed even with
4750 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4753 len = OPEN_EXISTING;
4754 env->me_mfd = CreateFile(dpath, oflags,
4755 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4756 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4759 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4761 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4766 DPRINTF(("opened dbenv %p", (void *) env));
4768 rc = mdb_env_share_locks(env, &excl);
4772 if (!(flags & MDB_RDONLY)) {
4774 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4775 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4776 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4777 (txn = calloc(1, size)))
4779 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4780 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4781 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4782 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4784 txn->mt_dbxs = env->me_dbxs;
4794 mdb_env_close0(env, excl);
4800 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4802 mdb_env_close0(MDB_env *env, int excl)
4806 if (!(env->me_flags & MDB_ENV_ACTIVE))
4809 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4811 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4812 free(env->me_dbxs[i].md_name.mv_data);
4817 free(env->me_dbiseqs);
4818 free(env->me_dbflags);
4820 free(env->me_dirty_list);
4822 mdb_midl_free(env->me_free_pgs);
4824 if (env->me_flags & MDB_ENV_TXKEY) {
4825 pthread_key_delete(env->me_txkey);
4827 /* Delete our key from the global list */
4828 for (i=0; i<mdb_tls_nkeys; i++)
4829 if (mdb_tls_keys[i] == env->me_txkey) {
4830 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4838 munmap(env->me_map, env->me_mapsize);
4840 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4841 (void) close(env->me_mfd);
4842 if (env->me_fd != INVALID_HANDLE_VALUE)
4843 (void) close(env->me_fd);
4845 MDB_PID_T pid = env->me_pid;
4846 /* Clearing readers is done in this function because
4847 * me_txkey with its destructor must be disabled first.
4849 * We skip the the reader mutex, so we touch only
4850 * data owned by this process (me_close_readers and
4851 * our readers), and clear each reader atomically.
4853 for (i = env->me_close_readers; --i >= 0; )
4854 if (env->me_txns->mti_readers[i].mr_pid == pid)
4855 env->me_txns->mti_readers[i].mr_pid = 0;
4857 if (env->me_rmutex) {
4858 CloseHandle(env->me_rmutex);
4859 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4861 /* Windows automatically destroys the mutexes when
4862 * the last handle closes.
4864 #elif defined(MDB_USE_SYSV_SEM)
4865 if (env->me_rmutex.semid != -1) {
4866 /* If we have the filelock: If we are the
4867 * only remaining user, clean up semaphores.
4870 mdb_env_excl_lock(env, &excl);
4872 semctl(env->me_rmutex.semid, 0, IPC_RMID);
4875 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4877 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4880 /* Unlock the lockfile. Windows would have unlocked it
4881 * after closing anyway, but not necessarily at once.
4883 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4886 (void) close(env->me_lfd);
4889 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4893 mdb_env_close(MDB_env *env)
4900 VGMEMP_DESTROY(env);
4901 while ((dp = env->me_dpages) != NULL) {
4902 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4903 env->me_dpages = dp->mp_next;
4907 mdb_env_close0(env, 0);
4911 /** Compare two items pointing at aligned size_t's */
4913 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4915 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4916 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4919 /** Compare two items pointing at aligned unsigned int's.
4921 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4922 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4925 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4927 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4928 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4931 /** Compare two items pointing at unsigned ints of unknown alignment.
4932 * Nodes and keys are guaranteed to be 2-byte aligned.
4935 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4937 #if BYTE_ORDER == LITTLE_ENDIAN
4938 unsigned short *u, *c;
4941 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4942 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4945 } while(!x && u > (unsigned short *)a->mv_data);
4948 unsigned short *u, *c, *end;
4951 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4952 u = (unsigned short *)a->mv_data;
4953 c = (unsigned short *)b->mv_data;
4956 } while(!x && u < end);
4961 /** Compare two items lexically */
4963 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4970 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4976 diff = memcmp(a->mv_data, b->mv_data, len);
4977 return diff ? diff : len_diff<0 ? -1 : len_diff;
4980 /** Compare two items in reverse byte order */
4982 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4984 const unsigned char *p1, *p2, *p1_lim;
4988 p1_lim = (const unsigned char *)a->mv_data;
4989 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4990 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4992 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4998 while (p1 > p1_lim) {
4999 diff = *--p1 - *--p2;
5003 return len_diff<0 ? -1 : len_diff;
5006 /** Search for key within a page, using binary search.
5007 * Returns the smallest entry larger or equal to the key.
5008 * If exactp is non-null, stores whether the found entry was an exact match
5009 * in *exactp (1 or 0).
5010 * Updates the cursor index with the index of the found entry.
5011 * If no entry larger or equal to the key is found, returns NULL.
5014 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5016 unsigned int i = 0, nkeys;
5019 MDB_page *mp = mc->mc_pg[mc->mc_top];
5020 MDB_node *node = NULL;
5025 nkeys = NUMKEYS(mp);
5027 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5028 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5031 low = IS_LEAF(mp) ? 0 : 1;
5033 cmp = mc->mc_dbx->md_cmp;
5035 /* Branch pages have no data, so if using integer keys,
5036 * alignment is guaranteed. Use faster mdb_cmp_int.
5038 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5039 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5046 nodekey.mv_size = mc->mc_db->md_pad;
5047 node = NODEPTR(mp, 0); /* fake */
5048 while (low <= high) {
5049 i = (low + high) >> 1;
5050 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5051 rc = cmp(key, &nodekey);
5052 DPRINTF(("found leaf index %u [%s], rc = %i",
5053 i, DKEY(&nodekey), rc));
5062 while (low <= high) {
5063 i = (low + high) >> 1;
5065 node = NODEPTR(mp, i);
5066 nodekey.mv_size = NODEKSZ(node);
5067 nodekey.mv_data = NODEKEY(node);
5069 rc = cmp(key, &nodekey);
5072 DPRINTF(("found leaf index %u [%s], rc = %i",
5073 i, DKEY(&nodekey), rc));
5075 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5076 i, DKEY(&nodekey), NODEPGNO(node), rc));
5087 if (rc > 0) { /* Found entry is less than the key. */
5088 i++; /* Skip to get the smallest entry larger than key. */
5090 node = NODEPTR(mp, i);
5093 *exactp = (rc == 0 && nkeys > 0);
5094 /* store the key index */
5095 mc->mc_ki[mc->mc_top] = i;
5097 /* There is no entry larger or equal to the key. */
5100 /* nodeptr is fake for LEAF2 */
5106 mdb_cursor_adjust(MDB_cursor *mc, func)
5110 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5111 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5118 /** Pop a page off the top of the cursor's stack. */
5120 mdb_cursor_pop(MDB_cursor *mc)
5124 MDB_page *top = mc->mc_pg[mc->mc_top];
5130 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5131 DDBI(mc), (void *) mc));
5135 /** Push a page onto the top of the cursor's stack. */
5137 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5139 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5140 DDBI(mc), (void *) mc));
5142 if (mc->mc_snum >= CURSOR_STACK) {
5143 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5144 return MDB_CURSOR_FULL;
5147 mc->mc_top = mc->mc_snum++;
5148 mc->mc_pg[mc->mc_top] = mp;
5149 mc->mc_ki[mc->mc_top] = 0;
5154 /** Find the address of the page corresponding to a given page number.
5155 * @param[in] txn the transaction for this access.
5156 * @param[in] pgno the page number for the page to retrieve.
5157 * @param[out] ret address of a pointer where the page's address will be stored.
5158 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5159 * @return 0 on success, non-zero on failure.
5162 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5164 MDB_env *env = txn->mt_env;
5168 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5172 MDB_ID2L dl = tx2->mt_u.dirty_list;
5174 /* Spilled pages were dirtied in this txn and flushed
5175 * because the dirty list got full. Bring this page
5176 * back in from the map (but don't unspill it here,
5177 * leave that unless page_touch happens again).
5179 if (tx2->mt_spill_pgs) {
5180 MDB_ID pn = pgno << 1;
5181 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5182 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5183 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5188 unsigned x = mdb_mid2l_search(dl, pgno);
5189 if (x <= dl[0].mid && dl[x].mid == pgno) {
5195 } while ((tx2 = tx2->mt_parent) != NULL);
5198 if (pgno < txn->mt_next_pgno) {
5200 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5202 DPRINTF(("page %"Z"u not found", pgno));
5203 txn->mt_flags |= MDB_TXN_ERROR;
5204 return MDB_PAGE_NOTFOUND;
5214 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5215 * The cursor is at the root page, set up the rest of it.
5218 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5220 MDB_page *mp = mc->mc_pg[mc->mc_top];
5224 while (IS_BRANCH(mp)) {
5228 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5229 mdb_cassert(mc, NUMKEYS(mp) > 1);
5230 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5232 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5234 if (flags & MDB_PS_LAST)
5235 i = NUMKEYS(mp) - 1;
5238 node = mdb_node_search(mc, key, &exact);
5240 i = NUMKEYS(mp) - 1;
5242 i = mc->mc_ki[mc->mc_top];
5244 mdb_cassert(mc, i > 0);
5248 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5251 mdb_cassert(mc, i < NUMKEYS(mp));
5252 node = NODEPTR(mp, i);
5254 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5257 mc->mc_ki[mc->mc_top] = i;
5258 if ((rc = mdb_cursor_push(mc, mp)))
5261 if (flags & MDB_PS_MODIFY) {
5262 if ((rc = mdb_page_touch(mc)) != 0)
5264 mp = mc->mc_pg[mc->mc_top];
5269 DPRINTF(("internal error, index points to a %02X page!?",
5271 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5272 return MDB_CORRUPTED;
5275 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5276 key ? DKEY(key) : "null"));
5277 mc->mc_flags |= C_INITIALIZED;
5278 mc->mc_flags &= ~C_EOF;
5283 /** Search for the lowest key under the current branch page.
5284 * This just bypasses a NUMKEYS check in the current page
5285 * before calling mdb_page_search_root(), because the callers
5286 * are all in situations where the current page is known to
5290 mdb_page_search_lowest(MDB_cursor *mc)
5292 MDB_page *mp = mc->mc_pg[mc->mc_top];
5293 MDB_node *node = NODEPTR(mp, 0);
5296 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5299 mc->mc_ki[mc->mc_top] = 0;
5300 if ((rc = mdb_cursor_push(mc, mp)))
5302 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5305 /** Search for the page a given key should be in.
5306 * Push it and its parent pages on the cursor stack.
5307 * @param[in,out] mc the cursor for this operation.
5308 * @param[in] key the key to search for, or NULL for first/last page.
5309 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5310 * are touched (updated with new page numbers).
5311 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5312 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5313 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5314 * @return 0 on success, non-zero on failure.
5317 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5322 /* Make sure the txn is still viable, then find the root from
5323 * the txn's db table and set it as the root of the cursor's stack.
5325 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5326 DPUTS("transaction has failed, must abort");
5329 /* Make sure we're using an up-to-date root */
5330 if (*mc->mc_dbflag & DB_STALE) {
5332 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5334 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5335 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5342 MDB_node *leaf = mdb_node_search(&mc2,
5343 &mc->mc_dbx->md_name, &exact);
5345 return MDB_NOTFOUND;
5346 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5349 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5351 /* The txn may not know this DBI, or another process may
5352 * have dropped and recreated the DB with other flags.
5354 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5355 return MDB_INCOMPATIBLE;
5356 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5358 *mc->mc_dbflag &= ~DB_STALE;
5360 root = mc->mc_db->md_root;
5362 if (root == P_INVALID) { /* Tree is empty. */
5363 DPUTS("tree is empty");
5364 return MDB_NOTFOUND;
5368 mdb_cassert(mc, root > 1);
5369 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5370 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5376 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5377 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5379 if (flags & MDB_PS_MODIFY) {
5380 if ((rc = mdb_page_touch(mc)))
5384 if (flags & MDB_PS_ROOTONLY)
5387 return mdb_page_search_root(mc, key, flags);
5391 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5393 MDB_txn *txn = mc->mc_txn;
5394 pgno_t pg = mp->mp_pgno;
5395 unsigned x = 0, ovpages = mp->mp_pages;
5396 MDB_env *env = txn->mt_env;
5397 MDB_IDL sl = txn->mt_spill_pgs;
5398 MDB_ID pn = pg << 1;
5401 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5402 /* If the page is dirty or on the spill list we just acquired it,
5403 * so we should give it back to our current free list, if any.
5404 * Otherwise put it onto the list of pages we freed in this txn.
5406 * Won't create me_pghead: me_pglast must be inited along with it.
5407 * Unsupported in nested txns: They would need to hide the page
5408 * range in ancestor txns' dirty and spilled lists.
5410 if (env->me_pghead &&
5412 ((mp->mp_flags & P_DIRTY) ||
5413 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5417 MDB_ID2 *dl, ix, iy;
5418 rc = mdb_midl_need(&env->me_pghead, ovpages);
5421 if (!(mp->mp_flags & P_DIRTY)) {
5422 /* This page is no longer spilled */
5429 /* Remove from dirty list */
5430 dl = txn->mt_u.dirty_list;
5432 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5438 mdb_cassert(mc, x > 1);
5440 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5441 txn->mt_flags |= MDB_TXN_ERROR;
5442 return MDB_CORRUPTED;
5445 if (!(env->me_flags & MDB_WRITEMAP))
5446 mdb_dpage_free(env, mp);
5448 /* Insert in me_pghead */
5449 mop = env->me_pghead;
5450 j = mop[0] + ovpages;
5451 for (i = mop[0]; i && mop[i] < pg; i--)
5457 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5461 mc->mc_db->md_overflow_pages -= ovpages;
5465 /** Return the data associated with a given node.
5466 * @param[in] txn The transaction for this operation.
5467 * @param[in] leaf The node being read.
5468 * @param[out] data Updated to point to the node's data.
5469 * @return 0 on success, non-zero on failure.
5472 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5474 MDB_page *omp; /* overflow page */
5478 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5479 data->mv_size = NODEDSZ(leaf);
5480 data->mv_data = NODEDATA(leaf);
5484 /* Read overflow data.
5486 data->mv_size = NODEDSZ(leaf);
5487 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5488 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5489 DPRINTF(("read overflow page %"Z"u failed", pgno));
5492 data->mv_data = METADATA(omp);
5498 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5499 MDB_val *key, MDB_val *data)
5506 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5508 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5511 if (txn->mt_flags & MDB_TXN_ERROR)
5514 mdb_cursor_init(&mc, txn, dbi, &mx);
5515 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5518 /** Find a sibling for a page.
5519 * Replaces the page at the top of the cursor's stack with the
5520 * specified sibling, if one exists.
5521 * @param[in] mc The cursor for this operation.
5522 * @param[in] move_right Non-zero if the right sibling is requested,
5523 * otherwise the left sibling.
5524 * @return 0 on success, non-zero on failure.
5527 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5533 if (mc->mc_snum < 2) {
5534 return MDB_NOTFOUND; /* root has no siblings */
5538 DPRINTF(("parent page is page %"Z"u, index %u",
5539 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5541 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5542 : (mc->mc_ki[mc->mc_top] == 0)) {
5543 DPRINTF(("no more keys left, moving to %s sibling",
5544 move_right ? "right" : "left"));
5545 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5546 /* undo cursor_pop before returning */
5553 mc->mc_ki[mc->mc_top]++;
5555 mc->mc_ki[mc->mc_top]--;
5556 DPRINTF(("just moving to %s index key %u",
5557 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5559 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5561 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5562 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5563 /* mc will be inconsistent if caller does mc_snum++ as above */
5564 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5568 mdb_cursor_push(mc, mp);
5570 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5575 /** Move the cursor to the next data item. */
5577 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5583 if (mc->mc_flags & C_EOF) {
5584 return MDB_NOTFOUND;
5587 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5589 mp = mc->mc_pg[mc->mc_top];
5591 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5592 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5593 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5594 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5595 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5596 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5597 if (rc == MDB_SUCCESS)
5598 MDB_GET_KEY(leaf, key);
5603 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5604 if (op == MDB_NEXT_DUP)
5605 return MDB_NOTFOUND;
5609 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5610 mdb_dbg_pgno(mp), (void *) mc));
5611 if (mc->mc_flags & C_DEL)
5614 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5615 DPUTS("=====> move to next sibling page");
5616 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5617 mc->mc_flags |= C_EOF;
5620 mp = mc->mc_pg[mc->mc_top];
5621 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5623 mc->mc_ki[mc->mc_top]++;
5626 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5627 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5630 key->mv_size = mc->mc_db->md_pad;
5631 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5635 mdb_cassert(mc, IS_LEAF(mp));
5636 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5638 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5639 mdb_xcursor_init1(mc, leaf);
5642 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5645 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5646 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5647 if (rc != MDB_SUCCESS)
5652 MDB_GET_KEY(leaf, key);
5656 /** Move the cursor to the previous data item. */
5658 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5664 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5666 mp = mc->mc_pg[mc->mc_top];
5668 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5669 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5670 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5671 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5672 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5673 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5674 if (rc == MDB_SUCCESS) {
5675 MDB_GET_KEY(leaf, key);
5676 mc->mc_flags &= ~C_EOF;
5682 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5683 if (op == MDB_PREV_DUP)
5684 return MDB_NOTFOUND;
5688 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5689 mdb_dbg_pgno(mp), (void *) mc));
5691 if (mc->mc_ki[mc->mc_top] == 0) {
5692 DPUTS("=====> move to prev sibling page");
5693 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5696 mp = mc->mc_pg[mc->mc_top];
5697 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5698 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5700 mc->mc_ki[mc->mc_top]--;
5702 mc->mc_flags &= ~C_EOF;
5704 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5705 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5708 key->mv_size = mc->mc_db->md_pad;
5709 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5713 mdb_cassert(mc, IS_LEAF(mp));
5714 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5716 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5717 mdb_xcursor_init1(mc, leaf);
5720 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5723 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5724 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5725 if (rc != MDB_SUCCESS)
5730 MDB_GET_KEY(leaf, key);
5734 /** Set the cursor on a specific data item. */
5736 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5737 MDB_cursor_op op, int *exactp)
5741 MDB_node *leaf = NULL;
5744 if (key->mv_size == 0)
5745 return MDB_BAD_VALSIZE;
5748 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5750 /* See if we're already on the right page */
5751 if (mc->mc_flags & C_INITIALIZED) {
5754 mp = mc->mc_pg[mc->mc_top];
5756 mc->mc_ki[mc->mc_top] = 0;
5757 return MDB_NOTFOUND;
5759 if (mp->mp_flags & P_LEAF2) {
5760 nodekey.mv_size = mc->mc_db->md_pad;
5761 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5763 leaf = NODEPTR(mp, 0);
5764 MDB_GET_KEY2(leaf, nodekey);
5766 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5768 /* Probably happens rarely, but first node on the page
5769 * was the one we wanted.
5771 mc->mc_ki[mc->mc_top] = 0;
5778 unsigned int nkeys = NUMKEYS(mp);
5780 if (mp->mp_flags & P_LEAF2) {
5781 nodekey.mv_data = LEAF2KEY(mp,
5782 nkeys-1, nodekey.mv_size);
5784 leaf = NODEPTR(mp, nkeys-1);
5785 MDB_GET_KEY2(leaf, nodekey);
5787 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5789 /* last node was the one we wanted */
5790 mc->mc_ki[mc->mc_top] = nkeys-1;
5796 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5797 /* This is definitely the right page, skip search_page */
5798 if (mp->mp_flags & P_LEAF2) {
5799 nodekey.mv_data = LEAF2KEY(mp,
5800 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5802 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5803 MDB_GET_KEY2(leaf, nodekey);
5805 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5807 /* current node was the one we wanted */
5817 /* If any parents have right-sibs, search.
5818 * Otherwise, there's nothing further.
5820 for (i=0; i<mc->mc_top; i++)
5822 NUMKEYS(mc->mc_pg[i])-1)
5824 if (i == mc->mc_top) {
5825 /* There are no other pages */
5826 mc->mc_ki[mc->mc_top] = nkeys;
5827 return MDB_NOTFOUND;
5831 /* There are no other pages */
5832 mc->mc_ki[mc->mc_top] = 0;
5833 if (op == MDB_SET_RANGE && !exactp) {
5837 return MDB_NOTFOUND;
5841 rc = mdb_page_search(mc, key, 0);
5842 if (rc != MDB_SUCCESS)
5845 mp = mc->mc_pg[mc->mc_top];
5846 mdb_cassert(mc, IS_LEAF(mp));
5849 leaf = mdb_node_search(mc, key, exactp);
5850 if (exactp != NULL && !*exactp) {
5851 /* MDB_SET specified and not an exact match. */
5852 return MDB_NOTFOUND;
5856 DPUTS("===> inexact leaf not found, goto sibling");
5857 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5858 return rc; /* no entries matched */
5859 mp = mc->mc_pg[mc->mc_top];
5860 mdb_cassert(mc, IS_LEAF(mp));
5861 leaf = NODEPTR(mp, 0);
5865 mc->mc_flags |= C_INITIALIZED;
5866 mc->mc_flags &= ~C_EOF;
5869 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5870 key->mv_size = mc->mc_db->md_pad;
5871 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5876 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5877 mdb_xcursor_init1(mc, leaf);
5880 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5881 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5882 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5885 if (op == MDB_GET_BOTH) {
5891 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5892 if (rc != MDB_SUCCESS)
5895 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5898 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5900 dcmp = mc->mc_dbx->md_dcmp;
5901 #if UINT_MAX < SIZE_MAX
5902 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5903 dcmp = mdb_cmp_clong;
5905 rc = dcmp(data, &olddata);
5907 if (op == MDB_GET_BOTH || rc > 0)
5908 return MDB_NOTFOUND;
5915 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5916 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5921 /* The key already matches in all other cases */
5922 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5923 MDB_GET_KEY(leaf, key);
5924 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5929 /** Move the cursor to the first item in the database. */
5931 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5937 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5939 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5940 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5941 if (rc != MDB_SUCCESS)
5944 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5946 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5947 mc->mc_flags |= C_INITIALIZED;
5948 mc->mc_flags &= ~C_EOF;
5950 mc->mc_ki[mc->mc_top] = 0;
5952 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5953 key->mv_size = mc->mc_db->md_pad;
5954 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5959 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5960 mdb_xcursor_init1(mc, leaf);
5961 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5965 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5969 MDB_GET_KEY(leaf, key);
5973 /** Move the cursor to the last item in the database. */
5975 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5981 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5983 if (!(mc->mc_flags & C_EOF)) {
5985 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5986 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5987 if (rc != MDB_SUCCESS)
5990 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5993 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5994 mc->mc_flags |= C_INITIALIZED|C_EOF;
5995 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5997 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5998 key->mv_size = mc->mc_db->md_pad;
5999 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6004 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6005 mdb_xcursor_init1(mc, leaf);
6006 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6010 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6015 MDB_GET_KEY(leaf, key);
6020 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6025 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6030 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6034 case MDB_GET_CURRENT:
6035 if (!(mc->mc_flags & C_INITIALIZED)) {
6038 MDB_page *mp = mc->mc_pg[mc->mc_top];
6039 int nkeys = NUMKEYS(mp);
6040 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6041 mc->mc_ki[mc->mc_top] = nkeys;
6047 key->mv_size = mc->mc_db->md_pad;
6048 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6050 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6051 MDB_GET_KEY(leaf, key);
6053 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6054 if (mc->mc_flags & C_DEL)
6055 mdb_xcursor_init1(mc, leaf);
6056 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6058 rc = mdb_node_read(mc->mc_txn, leaf, data);
6065 case MDB_GET_BOTH_RANGE:
6070 if (mc->mc_xcursor == NULL) {
6071 rc = MDB_INCOMPATIBLE;
6081 rc = mdb_cursor_set(mc, key, data, op,
6082 op == MDB_SET_RANGE ? NULL : &exact);
6085 case MDB_GET_MULTIPLE:
6086 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6090 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6091 rc = MDB_INCOMPATIBLE;
6095 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6096 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6099 case MDB_NEXT_MULTIPLE:
6104 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6105 rc = MDB_INCOMPATIBLE;
6108 if (!(mc->mc_flags & C_INITIALIZED))
6109 rc = mdb_cursor_first(mc, key, data);
6111 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6112 if (rc == MDB_SUCCESS) {
6113 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6116 mx = &mc->mc_xcursor->mx_cursor;
6117 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6119 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6120 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6128 case MDB_NEXT_NODUP:
6129 if (!(mc->mc_flags & C_INITIALIZED))
6130 rc = mdb_cursor_first(mc, key, data);
6132 rc = mdb_cursor_next(mc, key, data, op);
6136 case MDB_PREV_NODUP:
6137 if (!(mc->mc_flags & C_INITIALIZED)) {
6138 rc = mdb_cursor_last(mc, key, data);
6141 mc->mc_flags |= C_INITIALIZED;
6142 mc->mc_ki[mc->mc_top]++;
6144 rc = mdb_cursor_prev(mc, key, data, op);
6147 rc = mdb_cursor_first(mc, key, data);
6150 mfunc = mdb_cursor_first;
6152 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6156 if (mc->mc_xcursor == NULL) {
6157 rc = MDB_INCOMPATIBLE;
6161 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6162 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6163 MDB_GET_KEY(leaf, key);
6164 rc = mdb_node_read(mc->mc_txn, leaf, data);
6168 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6172 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6175 rc = mdb_cursor_last(mc, key, data);
6178 mfunc = mdb_cursor_last;
6181 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6186 if (mc->mc_flags & C_DEL)
6187 mc->mc_flags ^= C_DEL;
6192 /** Touch all the pages in the cursor stack. Set mc_top.
6193 * Makes sure all the pages are writable, before attempting a write operation.
6194 * @param[in] mc The cursor to operate on.
6197 mdb_cursor_touch(MDB_cursor *mc)
6199 int rc = MDB_SUCCESS;
6201 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6204 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6206 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6207 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6210 *mc->mc_dbflag |= DB_DIRTY;
6215 rc = mdb_page_touch(mc);
6216 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6217 mc->mc_top = mc->mc_snum-1;
6222 /** Do not spill pages to disk if txn is getting full, may fail instead */
6223 #define MDB_NOSPILL 0x8000
6226 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6230 MDB_node *leaf = NULL;
6233 MDB_val xdata, *rdata, dkey, olddata;
6235 int do_sub = 0, insert_key, insert_data;
6236 unsigned int mcount = 0, dcount = 0, nospill;
6239 unsigned int nflags;
6242 if (mc == NULL || key == NULL)
6245 env = mc->mc_txn->mt_env;
6247 /* Check this first so counter will always be zero on any
6250 if (flags & MDB_MULTIPLE) {
6251 dcount = data[1].mv_size;
6252 data[1].mv_size = 0;
6253 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6254 return MDB_INCOMPATIBLE;
6257 nospill = flags & MDB_NOSPILL;
6258 flags &= ~MDB_NOSPILL;
6260 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6261 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6263 if (key->mv_size-1 >= ENV_MAXKEY(env))
6264 return MDB_BAD_VALSIZE;
6266 #if SIZE_MAX > MAXDATASIZE
6267 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6268 return MDB_BAD_VALSIZE;
6270 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6271 return MDB_BAD_VALSIZE;
6274 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6275 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6279 if (flags == MDB_CURRENT) {
6280 if (!(mc->mc_flags & C_INITIALIZED))
6283 } else if (mc->mc_db->md_root == P_INVALID) {
6284 /* new database, cursor has nothing to point to */
6287 mc->mc_flags &= ~C_INITIALIZED;
6292 if (flags & MDB_APPEND) {
6294 rc = mdb_cursor_last(mc, &k2, &d2);
6296 rc = mc->mc_dbx->md_cmp(key, &k2);
6299 mc->mc_ki[mc->mc_top]++;
6301 /* new key is <= last key */
6306 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6308 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6309 DPRINTF(("duplicate key [%s]", DKEY(key)));
6311 return MDB_KEYEXIST;
6313 if (rc && rc != MDB_NOTFOUND)
6317 if (mc->mc_flags & C_DEL)
6318 mc->mc_flags ^= C_DEL;
6320 /* Cursor is positioned, check for room in the dirty list */
6322 if (flags & MDB_MULTIPLE) {
6324 xdata.mv_size = data->mv_size * dcount;
6328 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6332 if (rc == MDB_NO_ROOT) {
6334 /* new database, write a root leaf page */
6335 DPUTS("allocating new root leaf page");
6336 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6339 mdb_cursor_push(mc, np);
6340 mc->mc_db->md_root = np->mp_pgno;
6341 mc->mc_db->md_depth++;
6342 *mc->mc_dbflag |= DB_DIRTY;
6343 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6345 np->mp_flags |= P_LEAF2;
6346 mc->mc_flags |= C_INITIALIZED;
6348 /* make sure all cursor pages are writable */
6349 rc2 = mdb_cursor_touch(mc);
6354 insert_key = insert_data = rc;
6356 /* The key does not exist */
6357 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6358 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6359 LEAFSIZE(key, data) > env->me_nodemax)
6361 /* Too big for a node, insert in sub-DB. Set up an empty
6362 * "old sub-page" for prep_subDB to expand to a full page.
6364 fp_flags = P_LEAF|P_DIRTY;
6366 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6367 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6368 olddata.mv_size = PAGEHDRSZ;
6372 /* there's only a key anyway, so this is a no-op */
6373 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6375 unsigned int ksize = mc->mc_db->md_pad;
6376 if (key->mv_size != ksize)
6377 return MDB_BAD_VALSIZE;
6378 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6379 memcpy(ptr, key->mv_data, ksize);
6381 /* if overwriting slot 0 of leaf, need to
6382 * update branch key if there is a parent page
6384 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6385 unsigned short top = mc->mc_top;
6387 /* slot 0 is always an empty key, find real slot */
6388 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6390 if (mc->mc_ki[mc->mc_top])
6391 rc2 = mdb_update_key(mc, key);
6402 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6403 olddata.mv_size = NODEDSZ(leaf);
6404 olddata.mv_data = NODEDATA(leaf);
6407 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6408 /* Prepare (sub-)page/sub-DB to accept the new item,
6409 * if needed. fp: old sub-page or a header faking
6410 * it. mp: new (sub-)page. offset: growth in page
6411 * size. xdata: node data with new page or DB.
6413 unsigned i, offset = 0;
6414 mp = fp = xdata.mv_data = env->me_pbuf;
6415 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6417 /* Was a single item before, must convert now */
6418 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6420 /* Just overwrite the current item */
6421 if (flags == MDB_CURRENT)
6423 dcmp = mc->mc_dbx->md_dcmp;
6424 #if UINT_MAX < SIZE_MAX
6425 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6426 dcmp = mdb_cmp_clong;
6428 /* does data match? */
6429 if (!dcmp(data, &olddata)) {
6430 if (flags & MDB_NODUPDATA)
6431 return MDB_KEYEXIST;
6436 /* Back up original data item */
6437 dkey.mv_size = olddata.mv_size;
6438 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6440 /* Make sub-page header for the dup items, with dummy body */
6441 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6442 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6443 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6444 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6445 fp->mp_flags |= P_LEAF2;
6446 fp->mp_pad = data->mv_size;
6447 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6449 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6450 (dkey.mv_size & 1) + (data->mv_size & 1);
6452 fp->mp_upper = xdata.mv_size - PAGEBASE;
6453 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6454 } else if (leaf->mn_flags & F_SUBDATA) {
6455 /* Data is on sub-DB, just store it */
6456 flags |= F_DUPDATA|F_SUBDATA;
6459 /* Data is on sub-page */
6460 fp = olddata.mv_data;
6463 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6464 offset = EVEN(NODESIZE + sizeof(indx_t) +
6468 offset = fp->mp_pad;
6469 if (SIZELEFT(fp) < offset) {
6470 offset *= 4; /* space for 4 more */
6473 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6475 fp->mp_flags |= P_DIRTY;
6476 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6477 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6481 xdata.mv_size = olddata.mv_size + offset;
6484 fp_flags = fp->mp_flags;
6485 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6486 /* Too big for a sub-page, convert to sub-DB */
6487 fp_flags &= ~P_SUBP;
6489 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6490 fp_flags |= P_LEAF2;
6491 dummy.md_pad = fp->mp_pad;
6492 dummy.md_flags = MDB_DUPFIXED;
6493 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6494 dummy.md_flags |= MDB_INTEGERKEY;
6500 dummy.md_branch_pages = 0;
6501 dummy.md_leaf_pages = 1;
6502 dummy.md_overflow_pages = 0;
6503 dummy.md_entries = NUMKEYS(fp);
6504 xdata.mv_size = sizeof(MDB_db);
6505 xdata.mv_data = &dummy;
6506 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6508 offset = env->me_psize - olddata.mv_size;
6509 flags |= F_DUPDATA|F_SUBDATA;
6510 dummy.md_root = mp->mp_pgno;
6513 mp->mp_flags = fp_flags | P_DIRTY;
6514 mp->mp_pad = fp->mp_pad;
6515 mp->mp_lower = fp->mp_lower;
6516 mp->mp_upper = fp->mp_upper + offset;
6517 if (fp_flags & P_LEAF2) {
6518 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6520 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6521 olddata.mv_size - fp->mp_upper - PAGEBASE);
6522 for (i=0; i<NUMKEYS(fp); i++)
6523 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6531 mdb_node_del(mc, 0);
6535 /* overflow page overwrites need special handling */
6536 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6539 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6541 memcpy(&pg, olddata.mv_data, sizeof(pg));
6542 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6544 ovpages = omp->mp_pages;
6546 /* Is the ov page large enough? */
6547 if (ovpages >= dpages) {
6548 if (!(omp->mp_flags & P_DIRTY) &&
6549 (level || (env->me_flags & MDB_WRITEMAP)))
6551 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6554 level = 0; /* dirty in this txn or clean */
6557 if (omp->mp_flags & P_DIRTY) {
6558 /* yes, overwrite it. Note in this case we don't
6559 * bother to try shrinking the page if the new data
6560 * is smaller than the overflow threshold.
6563 /* It is writable only in a parent txn */
6564 size_t sz = (size_t) env->me_psize * ovpages, off;
6565 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6571 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6572 mdb_cassert(mc, rc2 == 0);
6573 if (!(flags & MDB_RESERVE)) {
6574 /* Copy end of page, adjusting alignment so
6575 * compiler may copy words instead of bytes.
6577 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6578 memcpy((size_t *)((char *)np + off),
6579 (size_t *)((char *)omp + off), sz - off);
6582 memcpy(np, omp, sz); /* Copy beginning of page */
6585 SETDSZ(leaf, data->mv_size);
6586 if (F_ISSET(flags, MDB_RESERVE))
6587 data->mv_data = METADATA(omp);
6589 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6593 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6595 } else if (data->mv_size == olddata.mv_size) {
6596 /* same size, just replace it. Note that we could
6597 * also reuse this node if the new data is smaller,
6598 * but instead we opt to shrink the node in that case.
6600 if (F_ISSET(flags, MDB_RESERVE))
6601 data->mv_data = olddata.mv_data;
6602 else if (!(mc->mc_flags & C_SUB))
6603 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6605 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6610 mdb_node_del(mc, 0);
6616 nflags = flags & NODE_ADD_FLAGS;
6617 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6618 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6619 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6620 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6622 nflags |= MDB_SPLIT_REPLACE;
6623 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6625 /* There is room already in this leaf page. */
6626 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6627 if (rc == 0 && insert_key) {
6628 /* Adjust other cursors pointing to mp */
6629 MDB_cursor *m2, *m3;
6630 MDB_dbi dbi = mc->mc_dbi;
6631 unsigned i = mc->mc_top;
6632 MDB_page *mp = mc->mc_pg[i];
6634 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6635 if (mc->mc_flags & C_SUB)
6636 m3 = &m2->mc_xcursor->mx_cursor;
6639 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6640 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6647 if (rc == MDB_SUCCESS) {
6648 /* Now store the actual data in the child DB. Note that we're
6649 * storing the user data in the keys field, so there are strict
6650 * size limits on dupdata. The actual data fields of the child
6651 * DB are all zero size.
6659 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6660 if (flags & MDB_CURRENT) {
6661 xflags = MDB_CURRENT|MDB_NOSPILL;
6663 mdb_xcursor_init1(mc, leaf);
6664 xflags = (flags & MDB_NODUPDATA) ?
6665 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6667 /* converted, write the original data first */
6669 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6673 /* Adjust other cursors pointing to mp */
6675 unsigned i = mc->mc_top;
6676 MDB_page *mp = mc->mc_pg[i];
6678 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6679 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6680 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6681 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6682 mdb_xcursor_init1(m2, leaf);
6686 /* we've done our job */
6689 ecount = mc->mc_xcursor->mx_db.md_entries;
6690 if (flags & MDB_APPENDDUP)
6691 xflags |= MDB_APPEND;
6692 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6693 if (flags & F_SUBDATA) {
6694 void *db = NODEDATA(leaf);
6695 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6697 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6699 /* Increment count unless we just replaced an existing item. */
6701 mc->mc_db->md_entries++;
6703 /* Invalidate txn if we created an empty sub-DB */
6706 /* If we succeeded and the key didn't exist before,
6707 * make sure the cursor is marked valid.
6709 mc->mc_flags |= C_INITIALIZED;
6711 if (flags & MDB_MULTIPLE) {
6714 /* let caller know how many succeeded, if any */
6715 data[1].mv_size = mcount;
6716 if (mcount < dcount) {
6717 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6718 insert_key = insert_data = 0;
6725 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6728 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6733 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6739 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6740 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6742 if (!(mc->mc_flags & C_INITIALIZED))
6745 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6746 return MDB_NOTFOUND;
6748 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6751 rc = mdb_cursor_touch(mc);
6755 mp = mc->mc_pg[mc->mc_top];
6758 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6760 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6761 if (flags & MDB_NODUPDATA) {
6762 /* mdb_cursor_del0() will subtract the final entry */
6763 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6765 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6766 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6768 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6771 /* If sub-DB still has entries, we're done */
6772 if (mc->mc_xcursor->mx_db.md_entries) {
6773 if (leaf->mn_flags & F_SUBDATA) {
6774 /* update subDB info */
6775 void *db = NODEDATA(leaf);
6776 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6779 /* shrink fake page */
6780 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6781 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6782 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6783 /* fix other sub-DB cursors pointed at this fake page */
6784 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6785 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6786 if (m2->mc_pg[mc->mc_top] == mp &&
6787 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6788 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6791 mc->mc_db->md_entries--;
6792 mc->mc_flags |= C_DEL;
6795 /* otherwise fall thru and delete the sub-DB */
6798 if (leaf->mn_flags & F_SUBDATA) {
6799 /* add all the child DB's pages to the free list */
6800 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6806 /* add overflow pages to free list */
6807 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6811 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6812 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6813 (rc = mdb_ovpage_free(mc, omp)))
6818 return mdb_cursor_del0(mc);
6821 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6825 /** Allocate and initialize new pages for a database.
6826 * @param[in] mc a cursor on the database being added to.
6827 * @param[in] flags flags defining what type of page is being allocated.
6828 * @param[in] num the number of pages to allocate. This is usually 1,
6829 * unless allocating overflow pages for a large record.
6830 * @param[out] mp Address of a page, or NULL on failure.
6831 * @return 0 on success, non-zero on failure.
6834 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6839 if ((rc = mdb_page_alloc(mc, num, &np)))
6841 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6842 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6843 np->mp_flags = flags | P_DIRTY;
6844 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6845 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6848 mc->mc_db->md_branch_pages++;
6849 else if (IS_LEAF(np))
6850 mc->mc_db->md_leaf_pages++;
6851 else if (IS_OVERFLOW(np)) {
6852 mc->mc_db->md_overflow_pages += num;
6860 /** Calculate the size of a leaf node.
6861 * The size depends on the environment's page size; if a data item
6862 * is too large it will be put onto an overflow page and the node
6863 * size will only include the key and not the data. Sizes are always
6864 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6865 * of the #MDB_node headers.
6866 * @param[in] env The environment handle.
6867 * @param[in] key The key for the node.
6868 * @param[in] data The data for the node.
6869 * @return The number of bytes needed to store the node.
6872 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6876 sz = LEAFSIZE(key, data);
6877 if (sz > env->me_nodemax) {
6878 /* put on overflow page */
6879 sz -= data->mv_size - sizeof(pgno_t);
6882 return EVEN(sz + sizeof(indx_t));
6885 /** Calculate the size of a branch node.
6886 * The size should depend on the environment's page size but since
6887 * we currently don't support spilling large keys onto overflow
6888 * pages, it's simply the size of the #MDB_node header plus the
6889 * size of the key. Sizes are always rounded up to an even number
6890 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6891 * @param[in] env The environment handle.
6892 * @param[in] key The key for the node.
6893 * @return The number of bytes needed to store the node.
6896 mdb_branch_size(MDB_env *env, MDB_val *key)
6901 if (sz > env->me_nodemax) {
6902 /* put on overflow page */
6903 /* not implemented */
6904 /* sz -= key->size - sizeof(pgno_t); */
6907 return sz + sizeof(indx_t);
6910 /** Add a node to the page pointed to by the cursor.
6911 * @param[in] mc The cursor for this operation.
6912 * @param[in] indx The index on the page where the new node should be added.
6913 * @param[in] key The key for the new node.
6914 * @param[in] data The data for the new node, if any.
6915 * @param[in] pgno The page number, if adding a branch node.
6916 * @param[in] flags Flags for the node.
6917 * @return 0 on success, non-zero on failure. Possible errors are:
6919 * <li>ENOMEM - failed to allocate overflow pages for the node.
6920 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6921 * should never happen since all callers already calculate the
6922 * page's free space before calling this function.
6926 mdb_node_add(MDB_cursor *mc, indx_t indx,
6927 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6930 size_t node_size = NODESIZE;
6934 MDB_page *mp = mc->mc_pg[mc->mc_top];
6935 MDB_page *ofp = NULL; /* overflow page */
6938 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6940 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6941 IS_LEAF(mp) ? "leaf" : "branch",
6942 IS_SUBP(mp) ? "sub-" : "",
6943 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6944 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6947 /* Move higher keys up one slot. */
6948 int ksize = mc->mc_db->md_pad, dif;
6949 char *ptr = LEAF2KEY(mp, indx, ksize);
6950 dif = NUMKEYS(mp) - indx;
6952 memmove(ptr+ksize, ptr, dif*ksize);
6953 /* insert new key */
6954 memcpy(ptr, key->mv_data, ksize);
6956 /* Just using these for counting */
6957 mp->mp_lower += sizeof(indx_t);
6958 mp->mp_upper -= ksize - sizeof(indx_t);
6962 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6964 node_size += key->mv_size;
6966 mdb_cassert(mc, data);
6967 if (F_ISSET(flags, F_BIGDATA)) {
6968 /* Data already on overflow page. */
6969 node_size += sizeof(pgno_t);
6970 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6971 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6973 /* Put data on overflow page. */
6974 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6975 data->mv_size, node_size+data->mv_size));
6976 node_size = EVEN(node_size + sizeof(pgno_t));
6977 if ((ssize_t)node_size > room)
6979 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6981 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6985 node_size += data->mv_size;
6988 node_size = EVEN(node_size);
6989 if ((ssize_t)node_size > room)
6993 /* Move higher pointers up one slot. */
6994 for (i = NUMKEYS(mp); i > indx; i--)
6995 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6997 /* Adjust free space offsets. */
6998 ofs = mp->mp_upper - node_size;
6999 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7000 mp->mp_ptrs[indx] = ofs;
7002 mp->mp_lower += sizeof(indx_t);
7004 /* Write the node data. */
7005 node = NODEPTR(mp, indx);
7006 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7007 node->mn_flags = flags;
7009 SETDSZ(node,data->mv_size);
7014 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7017 mdb_cassert(mc, key);
7019 if (F_ISSET(flags, F_BIGDATA))
7020 memcpy(node->mn_data + key->mv_size, data->mv_data,
7022 else if (F_ISSET(flags, MDB_RESERVE))
7023 data->mv_data = node->mn_data + key->mv_size;
7025 memcpy(node->mn_data + key->mv_size, data->mv_data,
7028 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
7030 if (F_ISSET(flags, MDB_RESERVE))
7031 data->mv_data = METADATA(ofp);
7033 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
7040 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7041 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7042 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7043 DPRINTF(("node size = %"Z"u", node_size));
7044 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7045 return MDB_PAGE_FULL;
7048 /** Delete the specified node from a page.
7049 * @param[in] mc Cursor pointing to the node to delete.
7050 * @param[in] ksize The size of a node. Only used if the page is
7051 * part of a #MDB_DUPFIXED database.
7054 mdb_node_del(MDB_cursor *mc, int ksize)
7056 MDB_page *mp = mc->mc_pg[mc->mc_top];
7057 indx_t indx = mc->mc_ki[mc->mc_top];
7059 indx_t i, j, numkeys, ptr;
7063 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7064 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7065 numkeys = NUMKEYS(mp);
7066 mdb_cassert(mc, indx < numkeys);
7069 int x = numkeys - 1 - indx;
7070 base = LEAF2KEY(mp, indx, ksize);
7072 memmove(base, base + ksize, x * ksize);
7073 mp->mp_lower -= sizeof(indx_t);
7074 mp->mp_upper += ksize - sizeof(indx_t);
7078 node = NODEPTR(mp, indx);
7079 sz = NODESIZE + node->mn_ksize;
7081 if (F_ISSET(node->mn_flags, F_BIGDATA))
7082 sz += sizeof(pgno_t);
7084 sz += NODEDSZ(node);
7088 ptr = mp->mp_ptrs[indx];
7089 for (i = j = 0; i < numkeys; i++) {
7091 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7092 if (mp->mp_ptrs[i] < ptr)
7093 mp->mp_ptrs[j] += sz;
7098 base = (char *)mp + mp->mp_upper + PAGEBASE;
7099 memmove(base + sz, base, ptr - mp->mp_upper);
7101 mp->mp_lower -= sizeof(indx_t);
7105 /** Compact the main page after deleting a node on a subpage.
7106 * @param[in] mp The main page to operate on.
7107 * @param[in] indx The index of the subpage on the main page.
7110 mdb_node_shrink(MDB_page *mp, indx_t indx)
7116 indx_t i, numkeys, ptr;
7118 node = NODEPTR(mp, indx);
7119 sp = (MDB_page *)NODEDATA(node);
7120 delta = SIZELEFT(sp);
7121 xp = (MDB_page *)((char *)sp + delta);
7123 /* shift subpage upward */
7125 nsize = NUMKEYS(sp) * sp->mp_pad;
7127 return; /* do not make the node uneven-sized */
7128 memmove(METADATA(xp), METADATA(sp), nsize);
7131 numkeys = NUMKEYS(sp);
7132 for (i=numkeys-1; i>=0; i--)
7133 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7135 xp->mp_upper = sp->mp_lower;
7136 xp->mp_lower = sp->mp_lower;
7137 xp->mp_flags = sp->mp_flags;
7138 xp->mp_pad = sp->mp_pad;
7139 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7141 nsize = NODEDSZ(node) - delta;
7142 SETDSZ(node, nsize);
7144 /* shift lower nodes upward */
7145 ptr = mp->mp_ptrs[indx];
7146 numkeys = NUMKEYS(mp);
7147 for (i = 0; i < numkeys; i++) {
7148 if (mp->mp_ptrs[i] <= ptr)
7149 mp->mp_ptrs[i] += delta;
7152 base = (char *)mp + mp->mp_upper + PAGEBASE;
7153 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7154 mp->mp_upper += delta;
7157 /** Initial setup of a sorted-dups cursor.
7158 * Sorted duplicates are implemented as a sub-database for the given key.
7159 * The duplicate data items are actually keys of the sub-database.
7160 * Operations on the duplicate data items are performed using a sub-cursor
7161 * initialized when the sub-database is first accessed. This function does
7162 * the preliminary setup of the sub-cursor, filling in the fields that
7163 * depend only on the parent DB.
7164 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7167 mdb_xcursor_init0(MDB_cursor *mc)
7169 MDB_xcursor *mx = mc->mc_xcursor;
7171 mx->mx_cursor.mc_xcursor = NULL;
7172 mx->mx_cursor.mc_txn = mc->mc_txn;
7173 mx->mx_cursor.mc_db = &mx->mx_db;
7174 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7175 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7176 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7177 mx->mx_cursor.mc_snum = 0;
7178 mx->mx_cursor.mc_top = 0;
7179 mx->mx_cursor.mc_flags = C_SUB;
7180 mx->mx_dbx.md_name.mv_size = 0;
7181 mx->mx_dbx.md_name.mv_data = NULL;
7182 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7183 mx->mx_dbx.md_dcmp = NULL;
7184 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7187 /** Final setup of a sorted-dups cursor.
7188 * Sets up the fields that depend on the data from the main cursor.
7189 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7190 * @param[in] node The data containing the #MDB_db record for the
7191 * sorted-dup database.
7194 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7196 MDB_xcursor *mx = mc->mc_xcursor;
7198 if (node->mn_flags & F_SUBDATA) {
7199 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7200 mx->mx_cursor.mc_pg[0] = 0;
7201 mx->mx_cursor.mc_snum = 0;
7202 mx->mx_cursor.mc_top = 0;
7203 mx->mx_cursor.mc_flags = C_SUB;
7205 MDB_page *fp = NODEDATA(node);
7206 mx->mx_db.md_pad = 0;
7207 mx->mx_db.md_flags = 0;
7208 mx->mx_db.md_depth = 1;
7209 mx->mx_db.md_branch_pages = 0;
7210 mx->mx_db.md_leaf_pages = 1;
7211 mx->mx_db.md_overflow_pages = 0;
7212 mx->mx_db.md_entries = NUMKEYS(fp);
7213 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7214 mx->mx_cursor.mc_snum = 1;
7215 mx->mx_cursor.mc_top = 0;
7216 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7217 mx->mx_cursor.mc_pg[0] = fp;
7218 mx->mx_cursor.mc_ki[0] = 0;
7219 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7220 mx->mx_db.md_flags = MDB_DUPFIXED;
7221 mx->mx_db.md_pad = fp->mp_pad;
7222 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7223 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7226 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7227 mx->mx_db.md_root));
7228 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7229 #if UINT_MAX < SIZE_MAX
7230 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7231 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7235 /** Initialize a cursor for a given transaction and database. */
7237 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7240 mc->mc_backup = NULL;
7243 mc->mc_db = &txn->mt_dbs[dbi];
7244 mc->mc_dbx = &txn->mt_dbxs[dbi];
7245 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7251 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7252 mdb_tassert(txn, mx != NULL);
7253 mc->mc_xcursor = mx;
7254 mdb_xcursor_init0(mc);
7256 mc->mc_xcursor = NULL;
7258 if (*mc->mc_dbflag & DB_STALE) {
7259 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7264 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7267 size_t size = sizeof(MDB_cursor);
7269 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7272 if (txn->mt_flags & MDB_TXN_ERROR)
7275 /* Allow read access to the freelist */
7276 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7279 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7280 size += sizeof(MDB_xcursor);
7282 if ((mc = malloc(size)) != NULL) {
7283 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7284 if (txn->mt_cursors) {
7285 mc->mc_next = txn->mt_cursors[dbi];
7286 txn->mt_cursors[dbi] = mc;
7287 mc->mc_flags |= C_UNTRACK;
7299 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7301 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7304 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7307 if (txn->mt_flags & MDB_TXN_ERROR)
7310 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7314 /* Return the count of duplicate data items for the current key */
7316 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7320 if (mc == NULL || countp == NULL)
7323 if (mc->mc_xcursor == NULL)
7324 return MDB_INCOMPATIBLE;
7326 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7329 if (!(mc->mc_flags & C_INITIALIZED))
7332 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7333 return MDB_NOTFOUND;
7335 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7336 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7339 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7342 *countp = mc->mc_xcursor->mx_db.md_entries;
7348 mdb_cursor_close(MDB_cursor *mc)
7350 if (mc && !mc->mc_backup) {
7351 /* remove from txn, if tracked */
7352 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7353 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7354 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7356 *prev = mc->mc_next;
7363 mdb_cursor_txn(MDB_cursor *mc)
7365 if (!mc) return NULL;
7370 mdb_cursor_dbi(MDB_cursor *mc)
7375 /** Replace the key for a branch node with a new key.
7376 * @param[in] mc Cursor pointing to the node to operate on.
7377 * @param[in] key The new key to use.
7378 * @return 0 on success, non-zero on failure.
7381 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7387 int delta, ksize, oksize;
7388 indx_t ptr, i, numkeys, indx;
7391 indx = mc->mc_ki[mc->mc_top];
7392 mp = mc->mc_pg[mc->mc_top];
7393 node = NODEPTR(mp, indx);
7394 ptr = mp->mp_ptrs[indx];
7398 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7399 k2.mv_data = NODEKEY(node);
7400 k2.mv_size = node->mn_ksize;
7401 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7403 mdb_dkey(&k2, kbuf2),
7409 /* Sizes must be 2-byte aligned. */
7410 ksize = EVEN(key->mv_size);
7411 oksize = EVEN(node->mn_ksize);
7412 delta = ksize - oksize;
7414 /* Shift node contents if EVEN(key length) changed. */
7416 if (delta > 0 && SIZELEFT(mp) < delta) {
7418 /* not enough space left, do a delete and split */
7419 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7420 pgno = NODEPGNO(node);
7421 mdb_node_del(mc, 0);
7422 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7425 numkeys = NUMKEYS(mp);
7426 for (i = 0; i < numkeys; i++) {
7427 if (mp->mp_ptrs[i] <= ptr)
7428 mp->mp_ptrs[i] -= delta;
7431 base = (char *)mp + mp->mp_upper + PAGEBASE;
7432 len = ptr - mp->mp_upper + NODESIZE;
7433 memmove(base - delta, base, len);
7434 mp->mp_upper -= delta;
7436 node = NODEPTR(mp, indx);
7439 /* But even if no shift was needed, update ksize */
7440 if (node->mn_ksize != key->mv_size)
7441 node->mn_ksize = key->mv_size;
7444 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7450 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7452 /** Move a node from csrc to cdst.
7455 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7462 unsigned short flags;
7466 /* Mark src and dst as dirty. */
7467 if ((rc = mdb_page_touch(csrc)) ||
7468 (rc = mdb_page_touch(cdst)))
7471 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7472 key.mv_size = csrc->mc_db->md_pad;
7473 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7475 data.mv_data = NULL;
7479 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7480 mdb_cassert(csrc, !((size_t)srcnode & 1));
7481 srcpg = NODEPGNO(srcnode);
7482 flags = srcnode->mn_flags;
7483 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7484 unsigned int snum = csrc->mc_snum;
7486 /* must find the lowest key below src */
7487 rc = mdb_page_search_lowest(csrc);
7490 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7491 key.mv_size = csrc->mc_db->md_pad;
7492 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7494 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7495 key.mv_size = NODEKSZ(s2);
7496 key.mv_data = NODEKEY(s2);
7498 csrc->mc_snum = snum--;
7499 csrc->mc_top = snum;
7501 key.mv_size = NODEKSZ(srcnode);
7502 key.mv_data = NODEKEY(srcnode);
7504 data.mv_size = NODEDSZ(srcnode);
7505 data.mv_data = NODEDATA(srcnode);
7507 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7508 unsigned int snum = cdst->mc_snum;
7511 /* must find the lowest key below dst */
7512 mdb_cursor_copy(cdst, &mn);
7513 rc = mdb_page_search_lowest(&mn);
7516 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7517 bkey.mv_size = mn.mc_db->md_pad;
7518 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7520 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7521 bkey.mv_size = NODEKSZ(s2);
7522 bkey.mv_data = NODEKEY(s2);
7524 mn.mc_snum = snum--;
7527 rc = mdb_update_key(&mn, &bkey);
7532 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7533 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7534 csrc->mc_ki[csrc->mc_top],
7536 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7537 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7539 /* Add the node to the destination page.
7541 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7542 if (rc != MDB_SUCCESS)
7545 /* Delete the node from the source page.
7547 mdb_node_del(csrc, key.mv_size);
7550 /* Adjust other cursors pointing to mp */
7551 MDB_cursor *m2, *m3;
7552 MDB_dbi dbi = csrc->mc_dbi;
7553 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7555 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7556 if (csrc->mc_flags & C_SUB)
7557 m3 = &m2->mc_xcursor->mx_cursor;
7560 if (m3 == csrc) continue;
7561 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7562 csrc->mc_ki[csrc->mc_top]) {
7563 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7564 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7569 /* Update the parent separators.
7571 if (csrc->mc_ki[csrc->mc_top] == 0) {
7572 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7573 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7574 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7576 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7577 key.mv_size = NODEKSZ(srcnode);
7578 key.mv_data = NODEKEY(srcnode);
7580 DPRINTF(("update separator for source page %"Z"u to [%s]",
7581 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7582 mdb_cursor_copy(csrc, &mn);
7585 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7588 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7590 indx_t ix = csrc->mc_ki[csrc->mc_top];
7591 nullkey.mv_size = 0;
7592 csrc->mc_ki[csrc->mc_top] = 0;
7593 rc = mdb_update_key(csrc, &nullkey);
7594 csrc->mc_ki[csrc->mc_top] = ix;
7595 mdb_cassert(csrc, rc == MDB_SUCCESS);
7599 if (cdst->mc_ki[cdst->mc_top] == 0) {
7600 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7601 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7602 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7604 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7605 key.mv_size = NODEKSZ(srcnode);
7606 key.mv_data = NODEKEY(srcnode);
7608 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7609 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7610 mdb_cursor_copy(cdst, &mn);
7613 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7616 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7618 indx_t ix = cdst->mc_ki[cdst->mc_top];
7619 nullkey.mv_size = 0;
7620 cdst->mc_ki[cdst->mc_top] = 0;
7621 rc = mdb_update_key(cdst, &nullkey);
7622 cdst->mc_ki[cdst->mc_top] = ix;
7623 mdb_cassert(cdst, rc == MDB_SUCCESS);
7630 /** Merge one page into another.
7631 * The nodes from the page pointed to by \b csrc will
7632 * be copied to the page pointed to by \b cdst and then
7633 * the \b csrc page will be freed.
7634 * @param[in] csrc Cursor pointing to the source page.
7635 * @param[in] cdst Cursor pointing to the destination page.
7636 * @return 0 on success, non-zero on failure.
7639 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7641 MDB_page *psrc, *pdst;
7648 psrc = csrc->mc_pg[csrc->mc_top];
7649 pdst = cdst->mc_pg[cdst->mc_top];
7651 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7653 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7654 mdb_cassert(csrc, cdst->mc_snum > 1);
7656 /* Mark dst as dirty. */
7657 if ((rc = mdb_page_touch(cdst)))
7660 /* Move all nodes from src to dst.
7662 j = nkeys = NUMKEYS(pdst);
7663 if (IS_LEAF2(psrc)) {
7664 key.mv_size = csrc->mc_db->md_pad;
7665 key.mv_data = METADATA(psrc);
7666 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7667 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7668 if (rc != MDB_SUCCESS)
7670 key.mv_data = (char *)key.mv_data + key.mv_size;
7673 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7674 srcnode = NODEPTR(psrc, i);
7675 if (i == 0 && IS_BRANCH(psrc)) {
7678 mdb_cursor_copy(csrc, &mn);
7679 /* must find the lowest key below src */
7680 rc = mdb_page_search_lowest(&mn);
7683 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7684 key.mv_size = mn.mc_db->md_pad;
7685 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7687 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7688 key.mv_size = NODEKSZ(s2);
7689 key.mv_data = NODEKEY(s2);
7692 key.mv_size = srcnode->mn_ksize;
7693 key.mv_data = NODEKEY(srcnode);
7696 data.mv_size = NODEDSZ(srcnode);
7697 data.mv_data = NODEDATA(srcnode);
7698 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7699 if (rc != MDB_SUCCESS)
7704 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7705 pdst->mp_pgno, NUMKEYS(pdst),
7706 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7708 /* Unlink the src page from parent and add to free list.
7711 mdb_node_del(csrc, 0);
7712 if (csrc->mc_ki[csrc->mc_top] == 0) {
7714 rc = mdb_update_key(csrc, &key);
7722 psrc = csrc->mc_pg[csrc->mc_top];
7723 /* If not operating on FreeDB, allow this page to be reused
7724 * in this txn. Otherwise just add to free list.
7726 rc = mdb_page_loose(csrc, psrc);
7730 csrc->mc_db->md_leaf_pages--;
7732 csrc->mc_db->md_branch_pages--;
7734 /* Adjust other cursors pointing to mp */
7735 MDB_cursor *m2, *m3;
7736 MDB_dbi dbi = csrc->mc_dbi;
7738 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7739 if (csrc->mc_flags & C_SUB)
7740 m3 = &m2->mc_xcursor->mx_cursor;
7743 if (m3 == csrc) continue;
7744 if (m3->mc_snum < csrc->mc_snum) continue;
7745 if (m3->mc_pg[csrc->mc_top] == psrc) {
7746 m3->mc_pg[csrc->mc_top] = pdst;
7747 m3->mc_ki[csrc->mc_top] += nkeys;
7752 unsigned int snum = cdst->mc_snum;
7753 uint16_t depth = cdst->mc_db->md_depth;
7754 mdb_cursor_pop(cdst);
7755 rc = mdb_rebalance(cdst);
7756 /* Did the tree shrink? */
7757 if (depth > cdst->mc_db->md_depth)
7759 cdst->mc_snum = snum;
7760 cdst->mc_top = snum-1;
7765 /** Copy the contents of a cursor.
7766 * @param[in] csrc The cursor to copy from.
7767 * @param[out] cdst The cursor to copy to.
7770 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7774 cdst->mc_txn = csrc->mc_txn;
7775 cdst->mc_dbi = csrc->mc_dbi;
7776 cdst->mc_db = csrc->mc_db;
7777 cdst->mc_dbx = csrc->mc_dbx;
7778 cdst->mc_snum = csrc->mc_snum;
7779 cdst->mc_top = csrc->mc_top;
7780 cdst->mc_flags = csrc->mc_flags;
7782 for (i=0; i<csrc->mc_snum; i++) {
7783 cdst->mc_pg[i] = csrc->mc_pg[i];
7784 cdst->mc_ki[i] = csrc->mc_ki[i];
7788 /** Rebalance the tree after a delete operation.
7789 * @param[in] mc Cursor pointing to the page where rebalancing
7791 * @return 0 on success, non-zero on failure.
7794 mdb_rebalance(MDB_cursor *mc)
7798 unsigned int ptop, minkeys;
7802 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7803 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7804 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7805 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7806 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7808 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7809 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7810 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7811 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7815 if (mc->mc_snum < 2) {
7816 MDB_page *mp = mc->mc_pg[0];
7818 DPUTS("Can't rebalance a subpage, ignoring");
7821 if (NUMKEYS(mp) == 0) {
7822 DPUTS("tree is completely empty");
7823 mc->mc_db->md_root = P_INVALID;
7824 mc->mc_db->md_depth = 0;
7825 mc->mc_db->md_leaf_pages = 0;
7826 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7829 /* Adjust cursors pointing to mp */
7832 mc->mc_flags &= ~C_INITIALIZED;
7834 MDB_cursor *m2, *m3;
7835 MDB_dbi dbi = mc->mc_dbi;
7837 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7838 if (mc->mc_flags & C_SUB)
7839 m3 = &m2->mc_xcursor->mx_cursor;
7842 if (m3->mc_snum < mc->mc_snum) continue;
7843 if (m3->mc_pg[0] == mp) {
7846 m3->mc_flags &= ~C_INITIALIZED;
7850 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7852 DPUTS("collapsing root page!");
7853 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7856 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7857 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7860 mc->mc_db->md_depth--;
7861 mc->mc_db->md_branch_pages--;
7862 mc->mc_ki[0] = mc->mc_ki[1];
7863 for (i = 1; i<mc->mc_db->md_depth; i++) {
7864 mc->mc_pg[i] = mc->mc_pg[i+1];
7865 mc->mc_ki[i] = mc->mc_ki[i+1];
7868 /* Adjust other cursors pointing to mp */
7869 MDB_cursor *m2, *m3;
7870 MDB_dbi dbi = mc->mc_dbi;
7872 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7873 if (mc->mc_flags & C_SUB)
7874 m3 = &m2->mc_xcursor->mx_cursor;
7877 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7878 if (m3->mc_pg[0] == mp) {
7879 for (i=0; i<m3->mc_snum; i++) {
7880 m3->mc_pg[i] = m3->mc_pg[i+1];
7881 m3->mc_ki[i] = m3->mc_ki[i+1];
7889 DPUTS("root page doesn't need rebalancing");
7893 /* The parent (branch page) must have at least 2 pointers,
7894 * otherwise the tree is invalid.
7896 ptop = mc->mc_top-1;
7897 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7899 /* Leaf page fill factor is below the threshold.
7900 * Try to move keys from left or right neighbor, or
7901 * merge with a neighbor page.
7906 mdb_cursor_copy(mc, &mn);
7907 mn.mc_xcursor = NULL;
7909 oldki = mc->mc_ki[mc->mc_top];
7910 if (mc->mc_ki[ptop] == 0) {
7911 /* We're the leftmost leaf in our parent.
7913 DPUTS("reading right neighbor");
7915 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7916 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7919 mn.mc_ki[mn.mc_top] = 0;
7920 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7922 /* There is at least one neighbor to the left.
7924 DPUTS("reading left neighbor");
7926 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7927 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7930 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7931 mc->mc_ki[mc->mc_top] = 0;
7934 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7935 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7936 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7938 /* If the neighbor page is above threshold and has enough keys,
7939 * move one key from it. Otherwise we should try to merge them.
7940 * (A branch page must never have less than 2 keys.)
7942 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7943 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7944 rc = mdb_node_move(&mn, mc);
7945 if (mc->mc_ki[ptop]) {
7949 if (mc->mc_ki[ptop] == 0) {
7950 rc = mdb_page_merge(&mn, mc);
7953 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7954 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7955 /* We want mdb_rebalance to find mn when doing fixups */
7956 if (mc->mc_flags & C_SUB) {
7957 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7958 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
7959 dummy.mc_xcursor = (MDB_xcursor *)&mn;
7961 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7962 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
7964 rc = mdb_page_merge(mc, &mn);
7965 if (mc->mc_flags & C_SUB)
7966 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
7968 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
7969 mdb_cursor_copy(&mn, mc);
7971 mc->mc_flags &= ~C_EOF;
7973 mc->mc_ki[mc->mc_top] = oldki;
7977 /** Complete a delete operation started by #mdb_cursor_del(). */
7979 mdb_cursor_del0(MDB_cursor *mc)
7986 ki = mc->mc_ki[mc->mc_top];
7987 mdb_node_del(mc, mc->mc_db->md_pad);
7988 mc->mc_db->md_entries--;
7989 rc = mdb_rebalance(mc);
7991 if (rc == MDB_SUCCESS) {
7992 MDB_cursor *m2, *m3;
7993 MDB_dbi dbi = mc->mc_dbi;
7995 /* DB is totally empty now, just bail out.
7996 * Other cursors adjustments were already done
7997 * by mdb_rebalance and aren't needed here.
8002 mp = mc->mc_pg[mc->mc_top];
8003 nkeys = NUMKEYS(mp);
8005 /* if mc points past last node in page, find next sibling */
8006 if (mc->mc_ki[mc->mc_top] >= nkeys) {
8007 rc = mdb_cursor_sibling(mc, 1);
8008 if (rc == MDB_NOTFOUND) {
8009 mc->mc_flags |= C_EOF;
8014 /* Adjust other cursors pointing to mp */
8015 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8016 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8017 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8019 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8021 if (m3->mc_pg[mc->mc_top] == mp) {
8022 if (m3->mc_ki[mc->mc_top] >= ki) {
8023 m3->mc_flags |= C_DEL;
8024 if (m3->mc_ki[mc->mc_top] > ki)
8025 m3->mc_ki[mc->mc_top]--;
8026 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8027 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8029 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8030 rc = mdb_cursor_sibling(m3, 1);
8031 if (rc == MDB_NOTFOUND) {
8032 m3->mc_flags |= C_EOF;
8038 mc->mc_flags |= C_DEL;
8042 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8047 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8048 MDB_val *key, MDB_val *data)
8050 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8053 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8054 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8056 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8057 /* must ignore any data */
8061 return mdb_del0(txn, dbi, key, data, 0);
8065 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8066 MDB_val *key, MDB_val *data, unsigned flags)
8071 MDB_val rdata, *xdata;
8075 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8077 mdb_cursor_init(&mc, txn, dbi, &mx);
8086 flags |= MDB_NODUPDATA;
8088 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8090 /* let mdb_page_split know about this cursor if needed:
8091 * delete will trigger a rebalance; if it needs to move
8092 * a node from one page to another, it will have to
8093 * update the parent's separator key(s). If the new sepkey
8094 * is larger than the current one, the parent page may
8095 * run out of space, triggering a split. We need this
8096 * cursor to be consistent until the end of the rebalance.
8098 mc.mc_flags |= C_UNTRACK;
8099 mc.mc_next = txn->mt_cursors[dbi];
8100 txn->mt_cursors[dbi] = &mc;
8101 rc = mdb_cursor_del(&mc, flags);
8102 txn->mt_cursors[dbi] = mc.mc_next;
8107 /** Split a page and insert a new node.
8108 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8109 * The cursor will be updated to point to the actual page and index where
8110 * the node got inserted after the split.
8111 * @param[in] newkey The key for the newly inserted node.
8112 * @param[in] newdata The data for the newly inserted node.
8113 * @param[in] newpgno The page number, if the new node is a branch node.
8114 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8115 * @return 0 on success, non-zero on failure.
8118 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8119 unsigned int nflags)
8122 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8125 int i, j, split_indx, nkeys, pmax;
8126 MDB_env *env = mc->mc_txn->mt_env;
8128 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8129 MDB_page *copy = NULL;
8130 MDB_page *mp, *rp, *pp;
8135 mp = mc->mc_pg[mc->mc_top];
8136 newindx = mc->mc_ki[mc->mc_top];
8137 nkeys = NUMKEYS(mp);
8139 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8140 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8141 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8143 /* Create a right sibling. */
8144 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8146 rp->mp_pad = mp->mp_pad;
8147 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8149 if (mc->mc_snum < 2) {
8150 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8152 /* shift current top to make room for new parent */
8153 mc->mc_pg[1] = mc->mc_pg[0];
8154 mc->mc_ki[1] = mc->mc_ki[0];
8157 mc->mc_db->md_root = pp->mp_pgno;
8158 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8159 mc->mc_db->md_depth++;
8162 /* Add left (implicit) pointer. */
8163 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8164 /* undo the pre-push */
8165 mc->mc_pg[0] = mc->mc_pg[1];
8166 mc->mc_ki[0] = mc->mc_ki[1];
8167 mc->mc_db->md_root = mp->mp_pgno;
8168 mc->mc_db->md_depth--;
8175 ptop = mc->mc_top-1;
8176 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8179 mc->mc_flags |= C_SPLITTING;
8180 mdb_cursor_copy(mc, &mn);
8181 mn.mc_pg[mn.mc_top] = rp;
8182 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8184 if (nflags & MDB_APPEND) {
8185 mn.mc_ki[mn.mc_top] = 0;
8187 split_indx = newindx;
8191 split_indx = (nkeys+1) / 2;
8196 unsigned int lsize, rsize, ksize;
8197 /* Move half of the keys to the right sibling */
8198 x = mc->mc_ki[mc->mc_top] - split_indx;
8199 ksize = mc->mc_db->md_pad;
8200 split = LEAF2KEY(mp, split_indx, ksize);
8201 rsize = (nkeys - split_indx) * ksize;
8202 lsize = (nkeys - split_indx) * sizeof(indx_t);
8203 mp->mp_lower -= lsize;
8204 rp->mp_lower += lsize;
8205 mp->mp_upper += rsize - lsize;
8206 rp->mp_upper -= rsize - lsize;
8207 sepkey.mv_size = ksize;
8208 if (newindx == split_indx) {
8209 sepkey.mv_data = newkey->mv_data;
8211 sepkey.mv_data = split;
8214 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8215 memcpy(rp->mp_ptrs, split, rsize);
8216 sepkey.mv_data = rp->mp_ptrs;
8217 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8218 memcpy(ins, newkey->mv_data, ksize);
8219 mp->mp_lower += sizeof(indx_t);
8220 mp->mp_upper -= ksize - sizeof(indx_t);
8223 memcpy(rp->mp_ptrs, split, x * ksize);
8224 ins = LEAF2KEY(rp, x, ksize);
8225 memcpy(ins, newkey->mv_data, ksize);
8226 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8227 rp->mp_lower += sizeof(indx_t);
8228 rp->mp_upper -= ksize - sizeof(indx_t);
8229 mc->mc_ki[mc->mc_top] = x;
8230 mc->mc_pg[mc->mc_top] = rp;
8233 int psize, nsize, k;
8234 /* Maximum free space in an empty page */
8235 pmax = env->me_psize - PAGEHDRSZ;
8237 nsize = mdb_leaf_size(env, newkey, newdata);
8239 nsize = mdb_branch_size(env, newkey);
8240 nsize = EVEN(nsize);
8242 /* grab a page to hold a temporary copy */
8243 copy = mdb_page_malloc(mc->mc_txn, 1);
8248 copy->mp_pgno = mp->mp_pgno;
8249 copy->mp_flags = mp->mp_flags;
8250 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8251 copy->mp_upper = env->me_psize - PAGEBASE;
8253 /* prepare to insert */
8254 for (i=0, j=0; i<nkeys; i++) {
8256 copy->mp_ptrs[j++] = 0;
8258 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8261 /* When items are relatively large the split point needs
8262 * to be checked, because being off-by-one will make the
8263 * difference between success or failure in mdb_node_add.
8265 * It's also relevant if a page happens to be laid out
8266 * such that one half of its nodes are all "small" and
8267 * the other half of its nodes are "large." If the new
8268 * item is also "large" and falls on the half with
8269 * "large" nodes, it also may not fit.
8271 * As a final tweak, if the new item goes on the last
8272 * spot on the page (and thus, onto the new page), bias
8273 * the split so the new page is emptier than the old page.
8274 * This yields better packing during sequential inserts.
8276 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8277 /* Find split point */
8279 if (newindx <= split_indx || newindx >= nkeys) {
8281 k = newindx >= nkeys ? nkeys : split_indx+2;
8286 for (; i!=k; i+=j) {
8291 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8292 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8294 if (F_ISSET(node->mn_flags, F_BIGDATA))
8295 psize += sizeof(pgno_t);
8297 psize += NODEDSZ(node);
8299 psize = EVEN(psize);
8301 if (psize > pmax || i == k-j) {
8302 split_indx = i + (j<0);
8307 if (split_indx == newindx) {
8308 sepkey.mv_size = newkey->mv_size;
8309 sepkey.mv_data = newkey->mv_data;
8311 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8312 sepkey.mv_size = node->mn_ksize;
8313 sepkey.mv_data = NODEKEY(node);
8318 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8320 /* Copy separator key to the parent.
8322 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8326 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8331 if (mn.mc_snum == mc->mc_snum) {
8332 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8333 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8334 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8335 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8340 /* Right page might now have changed parent.
8341 * Check if left page also changed parent.
8343 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8344 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8345 for (i=0; i<ptop; i++) {
8346 mc->mc_pg[i] = mn.mc_pg[i];
8347 mc->mc_ki[i] = mn.mc_ki[i];
8349 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8350 if (mn.mc_ki[ptop]) {
8351 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8353 /* find right page's left sibling */
8354 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8355 mdb_cursor_sibling(mc, 0);
8360 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8363 mc->mc_flags ^= C_SPLITTING;
8364 if (rc != MDB_SUCCESS) {
8367 if (nflags & MDB_APPEND) {
8368 mc->mc_pg[mc->mc_top] = rp;
8369 mc->mc_ki[mc->mc_top] = 0;
8370 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8373 for (i=0; i<mc->mc_top; i++)
8374 mc->mc_ki[i] = mn.mc_ki[i];
8375 } else if (!IS_LEAF2(mp)) {
8377 mc->mc_pg[mc->mc_top] = rp;
8382 rkey.mv_data = newkey->mv_data;
8383 rkey.mv_size = newkey->mv_size;
8389 /* Update index for the new key. */
8390 mc->mc_ki[mc->mc_top] = j;
8392 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8393 rkey.mv_data = NODEKEY(node);
8394 rkey.mv_size = node->mn_ksize;
8396 xdata.mv_data = NODEDATA(node);
8397 xdata.mv_size = NODEDSZ(node);
8400 pgno = NODEPGNO(node);
8401 flags = node->mn_flags;
8404 if (!IS_LEAF(mp) && j == 0) {
8405 /* First branch index doesn't need key data. */
8409 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8415 mc->mc_pg[mc->mc_top] = copy;
8420 } while (i != split_indx);
8422 nkeys = NUMKEYS(copy);
8423 for (i=0; i<nkeys; i++)
8424 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8425 mp->mp_lower = copy->mp_lower;
8426 mp->mp_upper = copy->mp_upper;
8427 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8428 env->me_psize - copy->mp_upper - PAGEBASE);
8430 /* reset back to original page */
8431 if (newindx < split_indx) {
8432 mc->mc_pg[mc->mc_top] = mp;
8433 if (nflags & MDB_RESERVE) {
8434 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8435 if (!(node->mn_flags & F_BIGDATA))
8436 newdata->mv_data = NODEDATA(node);
8439 mc->mc_pg[mc->mc_top] = rp;
8441 /* Make sure mc_ki is still valid.
8443 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8444 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8445 for (i=0; i<=ptop; i++) {
8446 mc->mc_pg[i] = mn.mc_pg[i];
8447 mc->mc_ki[i] = mn.mc_ki[i];
8454 /* Adjust other cursors pointing to mp */
8455 MDB_cursor *m2, *m3;
8456 MDB_dbi dbi = mc->mc_dbi;
8457 int fixup = NUMKEYS(mp);
8459 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8460 if (mc->mc_flags & C_SUB)
8461 m3 = &m2->mc_xcursor->mx_cursor;
8466 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8468 if (m3->mc_flags & C_SPLITTING)
8473 for (k=m3->mc_top; k>=0; k--) {
8474 m3->mc_ki[k+1] = m3->mc_ki[k];
8475 m3->mc_pg[k+1] = m3->mc_pg[k];
8477 if (m3->mc_ki[0] >= split_indx) {
8482 m3->mc_pg[0] = mc->mc_pg[0];
8486 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8487 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8488 m3->mc_ki[mc->mc_top]++;
8489 if (m3->mc_ki[mc->mc_top] >= fixup) {
8490 m3->mc_pg[mc->mc_top] = rp;
8491 m3->mc_ki[mc->mc_top] -= fixup;
8492 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8494 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8495 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8500 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8503 if (copy) /* tmp page */
8504 mdb_page_free(env, copy);
8506 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8511 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8512 MDB_val *key, MDB_val *data, unsigned int flags)
8517 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8520 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8523 mdb_cursor_init(&mc, txn, dbi, &mx);
8524 return mdb_cursor_put(&mc, key, data, flags);
8528 #define MDB_WBUF (1024*1024)
8531 /** State needed for a compacting copy. */
8532 typedef struct mdb_copy {
8533 pthread_mutex_t mc_mutex;
8534 pthread_cond_t mc_cond;
8541 pgno_t mc_next_pgno;
8544 volatile int mc_new;
8549 /** Dedicated writer thread for compacting copy. */
8550 static THREAD_RET ESECT
8551 mdb_env_copythr(void *arg)
8555 int toggle = 0, wsize, rc;
8558 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8561 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8564 pthread_mutex_lock(&my->mc_mutex);
8566 pthread_cond_signal(&my->mc_cond);
8569 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8570 if (my->mc_new < 0) {
8575 wsize = my->mc_wlen[toggle];
8576 ptr = my->mc_wbuf[toggle];
8579 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8583 } else if (len > 0) {
8597 /* If there's an overflow page tail, write it too */
8598 if (my->mc_olen[toggle]) {
8599 wsize = my->mc_olen[toggle];
8600 ptr = my->mc_over[toggle];
8601 my->mc_olen[toggle] = 0;
8604 my->mc_wlen[toggle] = 0;
8606 pthread_cond_signal(&my->mc_cond);
8608 pthread_cond_signal(&my->mc_cond);
8609 pthread_mutex_unlock(&my->mc_mutex);
8610 return (THREAD_RET)0;
8614 /** Tell the writer thread there's a buffer ready to write */
8616 mdb_env_cthr_toggle(mdb_copy *my, int st)
8618 int toggle = my->mc_toggle ^ 1;
8619 pthread_mutex_lock(&my->mc_mutex);
8620 if (my->mc_status) {
8621 pthread_mutex_unlock(&my->mc_mutex);
8622 return my->mc_status;
8624 while (my->mc_new == 1)
8625 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8627 my->mc_toggle = toggle;
8628 pthread_cond_signal(&my->mc_cond);
8629 pthread_mutex_unlock(&my->mc_mutex);
8633 /** Depth-first tree traversal for compacting copy. */
8635 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8638 MDB_txn *txn = my->mc_txn;
8640 MDB_page *mo, *mp, *leaf;
8645 /* Empty DB, nothing to do */
8646 if (*pg == P_INVALID)
8653 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8656 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8660 /* Make cursor pages writable */
8661 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8665 for (i=0; i<mc.mc_top; i++) {
8666 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8667 mc.mc_pg[i] = (MDB_page *)ptr;
8668 ptr += my->mc_env->me_psize;
8671 /* This is writable space for a leaf page. Usually not needed. */
8672 leaf = (MDB_page *)ptr;
8674 toggle = my->mc_toggle;
8675 while (mc.mc_snum > 0) {
8677 mp = mc.mc_pg[mc.mc_top];
8681 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8682 for (i=0; i<n; i++) {
8683 ni = NODEPTR(mp, i);
8684 if (ni->mn_flags & F_BIGDATA) {
8688 /* Need writable leaf */
8690 mc.mc_pg[mc.mc_top] = leaf;
8691 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8693 ni = NODEPTR(mp, i);
8696 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8697 rc = mdb_page_get(txn, pg, &omp, NULL);
8700 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8701 rc = mdb_env_cthr_toggle(my, 1);
8704 toggle = my->mc_toggle;
8706 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8707 memcpy(mo, omp, my->mc_env->me_psize);
8708 mo->mp_pgno = my->mc_next_pgno;
8709 my->mc_next_pgno += omp->mp_pages;
8710 my->mc_wlen[toggle] += my->mc_env->me_psize;
8711 if (omp->mp_pages > 1) {
8712 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8713 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8714 rc = mdb_env_cthr_toggle(my, 1);
8717 toggle = my->mc_toggle;
8719 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8720 } else if (ni->mn_flags & F_SUBDATA) {
8723 /* Need writable leaf */
8725 mc.mc_pg[mc.mc_top] = leaf;
8726 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8728 ni = NODEPTR(mp, i);
8731 memcpy(&db, NODEDATA(ni), sizeof(db));
8732 my->mc_toggle = toggle;
8733 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8736 toggle = my->mc_toggle;
8737 memcpy(NODEDATA(ni), &db, sizeof(db));
8742 mc.mc_ki[mc.mc_top]++;
8743 if (mc.mc_ki[mc.mc_top] < n) {
8746 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8748 rc = mdb_page_get(txn, pg, &mp, NULL);
8753 mc.mc_ki[mc.mc_top] = 0;
8754 if (IS_BRANCH(mp)) {
8755 /* Whenever we advance to a sibling branch page,
8756 * we must proceed all the way down to its first leaf.
8758 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8761 mc.mc_pg[mc.mc_top] = mp;
8765 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8766 rc = mdb_env_cthr_toggle(my, 1);
8769 toggle = my->mc_toggle;
8771 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8772 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8773 mo->mp_pgno = my->mc_next_pgno++;
8774 my->mc_wlen[toggle] += my->mc_env->me_psize;
8776 /* Update parent if there is one */
8777 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8778 SETPGNO(ni, mo->mp_pgno);
8779 mdb_cursor_pop(&mc);
8781 /* Otherwise we're done */
8791 /** Copy environment with compaction. */
8793 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8798 MDB_txn *txn = NULL;
8803 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8804 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8805 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8806 if (my.mc_wbuf[0] == NULL)
8809 pthread_mutex_init(&my.mc_mutex, NULL);
8810 pthread_cond_init(&my.mc_cond, NULL);
8811 #ifdef HAVE_MEMALIGN
8812 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8813 if (my.mc_wbuf[0] == NULL)
8816 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8821 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8822 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8827 my.mc_next_pgno = 2;
8833 THREAD_CREATE(thr, mdb_env_copythr, &my);
8835 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8839 mp = (MDB_page *)my.mc_wbuf[0];
8840 memset(mp, 0, 2*env->me_psize);
8842 mp->mp_flags = P_META;
8843 mm = (MDB_meta *)METADATA(mp);
8844 mdb_env_init_meta0(env, mm);
8845 mm->mm_address = env->me_metas[0]->mm_address;
8847 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8849 mp->mp_flags = P_META;
8850 *(MDB_meta *)METADATA(mp) = *mm;
8851 mm = (MDB_meta *)METADATA(mp);
8853 /* Count the number of free pages, subtract from lastpg to find
8854 * number of active pages
8857 MDB_ID freecount = 0;
8860 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8861 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8862 freecount += *(MDB_ID *)data.mv_data;
8863 freecount += txn->mt_dbs[0].md_branch_pages +
8864 txn->mt_dbs[0].md_leaf_pages +
8865 txn->mt_dbs[0].md_overflow_pages;
8867 /* Set metapage 1 */
8868 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8869 mm->mm_dbs[1] = txn->mt_dbs[1];
8870 if (mm->mm_last_pg > 1) {
8871 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8874 mm->mm_dbs[1].md_root = P_INVALID;
8877 my.mc_wlen[0] = env->me_psize * 2;
8879 pthread_mutex_lock(&my.mc_mutex);
8881 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8882 pthread_mutex_unlock(&my.mc_mutex);
8883 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8884 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8885 rc = mdb_env_cthr_toggle(&my, 1);
8886 mdb_env_cthr_toggle(&my, -1);
8887 pthread_mutex_lock(&my.mc_mutex);
8889 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8890 pthread_mutex_unlock(&my.mc_mutex);
8895 CloseHandle(my.mc_cond);
8896 CloseHandle(my.mc_mutex);
8897 _aligned_free(my.mc_wbuf[0]);
8899 pthread_cond_destroy(&my.mc_cond);
8900 pthread_mutex_destroy(&my.mc_mutex);
8901 free(my.mc_wbuf[0]);
8906 /** Copy environment as-is. */
8908 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8910 MDB_txn *txn = NULL;
8911 mdb_mutex_t *wmutex = NULL;
8917 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8921 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8924 /* Do the lock/unlock of the reader mutex before starting the
8925 * write txn. Otherwise other read txns could block writers.
8927 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8932 /* We must start the actual read txn after blocking writers */
8933 mdb_txn_reset0(txn, "reset-stage1");
8935 /* Temporarily block writers until we snapshot the meta pages */
8936 wmutex = MDB_MUTEX(env, w);
8937 if (LOCK_MUTEX(rc, env, wmutex))
8940 rc = mdb_txn_renew0(txn);
8942 UNLOCK_MUTEX(wmutex);
8947 wsize = env->me_psize * 2;
8951 DO_WRITE(rc, fd, ptr, w2, len);
8955 } else if (len > 0) {
8961 /* Non-blocking or async handles are not supported */
8967 UNLOCK_MUTEX(wmutex);
8972 w2 = txn->mt_next_pgno * env->me_psize;
8975 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8982 if (wsize > MAX_WRITE)
8986 DO_WRITE(rc, fd, ptr, w2, len);
8990 } else if (len > 0) {
9007 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9009 if (flags & MDB_CP_COMPACT)
9010 return mdb_env_copyfd1(env, fd);
9012 return mdb_env_copyfd0(env, fd);
9016 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9018 return mdb_env_copyfd2(env, fd, 0);
9022 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9026 HANDLE newfd = INVALID_HANDLE_VALUE;
9028 if (env->me_flags & MDB_NOSUBDIR) {
9029 lpath = (char *)path;
9032 len += sizeof(DATANAME);
9033 lpath = malloc(len);
9036 sprintf(lpath, "%s" DATANAME, path);
9039 /* The destination path must exist, but the destination file must not.
9040 * We don't want the OS to cache the writes, since the source data is
9041 * already in the OS cache.
9044 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9045 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9047 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9049 if (newfd == INVALID_HANDLE_VALUE) {
9054 if (env->me_psize >= env->me_os_psize) {
9056 /* Set O_DIRECT if the file system supports it */
9057 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9058 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9060 #ifdef F_NOCACHE /* __APPLE__ */
9061 rc = fcntl(newfd, F_NOCACHE, 1);
9069 rc = mdb_env_copyfd2(env, newfd, flags);
9072 if (!(env->me_flags & MDB_NOSUBDIR))
9074 if (newfd != INVALID_HANDLE_VALUE)
9075 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9082 mdb_env_copy(MDB_env *env, const char *path)
9084 return mdb_env_copy2(env, path, 0);
9088 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9090 if ((flag & CHANGEABLE) != flag)
9093 env->me_flags |= flag;
9095 env->me_flags &= ~flag;
9100 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9105 *arg = env->me_flags;
9110 mdb_env_set_userctx(MDB_env *env, void *ctx)
9114 env->me_userctx = ctx;
9119 mdb_env_get_userctx(MDB_env *env)
9121 return env ? env->me_userctx : NULL;
9125 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9130 env->me_assert_func = func;
9136 mdb_env_get_path(MDB_env *env, const char **arg)
9141 *arg = env->me_path;
9146 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9155 /** Common code for #mdb_stat() and #mdb_env_stat().
9156 * @param[in] env the environment to operate in.
9157 * @param[in] db the #MDB_db record containing the stats to return.
9158 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9159 * @return 0, this function always succeeds.
9162 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9164 arg->ms_psize = env->me_psize;
9165 arg->ms_depth = db->md_depth;
9166 arg->ms_branch_pages = db->md_branch_pages;
9167 arg->ms_leaf_pages = db->md_leaf_pages;
9168 arg->ms_overflow_pages = db->md_overflow_pages;
9169 arg->ms_entries = db->md_entries;
9175 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9179 if (env == NULL || arg == NULL)
9182 toggle = mdb_env_pick_meta(env);
9184 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9188 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9192 if (env == NULL || arg == NULL)
9195 toggle = mdb_env_pick_meta(env);
9196 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9197 arg->me_mapsize = env->me_mapsize;
9198 arg->me_maxreaders = env->me_maxreaders;
9199 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9201 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9202 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9206 /** Set the default comparison functions for a database.
9207 * Called immediately after a database is opened to set the defaults.
9208 * The user can then override them with #mdb_set_compare() or
9209 * #mdb_set_dupsort().
9210 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9211 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9214 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9216 uint16_t f = txn->mt_dbs[dbi].md_flags;
9218 txn->mt_dbxs[dbi].md_cmp =
9219 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9220 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9222 txn->mt_dbxs[dbi].md_dcmp =
9223 !(f & MDB_DUPSORT) ? 0 :
9224 ((f & MDB_INTEGERDUP)
9225 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9226 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9229 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9235 int rc, dbflag, exact;
9236 unsigned int unused = 0, seq;
9239 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9240 mdb_default_cmp(txn, FREE_DBI);
9243 if ((flags & VALID_FLAGS) != flags)
9245 if (txn->mt_flags & MDB_TXN_ERROR)
9251 if (flags & PERSISTENT_FLAGS) {
9252 uint16_t f2 = flags & PERSISTENT_FLAGS;
9253 /* make sure flag changes get committed */
9254 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9255 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9256 txn->mt_flags |= MDB_TXN_DIRTY;
9259 mdb_default_cmp(txn, MAIN_DBI);
9263 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9264 mdb_default_cmp(txn, MAIN_DBI);
9267 /* Is the DB already open? */
9269 for (i=2; i<txn->mt_numdbs; i++) {
9270 if (!txn->mt_dbxs[i].md_name.mv_size) {
9271 /* Remember this free slot */
9272 if (!unused) unused = i;
9275 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9276 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9282 /* If no free slot and max hit, fail */
9283 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9284 return MDB_DBS_FULL;
9286 /* Cannot mix named databases with some mainDB flags */
9287 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9288 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9290 /* Find the DB info */
9291 dbflag = DB_NEW|DB_VALID;
9294 key.mv_data = (void *)name;
9295 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9296 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9297 if (rc == MDB_SUCCESS) {
9298 /* make sure this is actually a DB */
9299 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9300 if (!(node->mn_flags & F_SUBDATA))
9301 return MDB_INCOMPATIBLE;
9302 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9303 /* Create if requested */
9304 data.mv_size = sizeof(MDB_db);
9305 data.mv_data = &dummy;
9306 memset(&dummy, 0, sizeof(dummy));
9307 dummy.md_root = P_INVALID;
9308 dummy.md_flags = flags & PERSISTENT_FLAGS;
9309 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9313 /* OK, got info, add to table */
9314 if (rc == MDB_SUCCESS) {
9315 unsigned int slot = unused ? unused : txn->mt_numdbs;
9316 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9317 txn->mt_dbxs[slot].md_name.mv_size = len;
9318 txn->mt_dbxs[slot].md_rel = NULL;
9319 txn->mt_dbflags[slot] = dbflag;
9320 /* txn-> and env-> are the same in read txns, use
9321 * tmp variable to avoid undefined assignment
9323 seq = ++txn->mt_env->me_dbiseqs[slot];
9324 txn->mt_dbiseqs[slot] = seq;
9326 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9328 mdb_default_cmp(txn, slot);
9337 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9339 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9342 if (txn->mt_flags & MDB_TXN_ERROR)
9345 if (txn->mt_dbflags[dbi] & DB_STALE) {
9348 /* Stale, must read the DB's root. cursor_init does it for us. */
9349 mdb_cursor_init(&mc, txn, dbi, &mx);
9351 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9354 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9357 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9359 ptr = env->me_dbxs[dbi].md_name.mv_data;
9360 /* If there was no name, this was already closed */
9362 env->me_dbxs[dbi].md_name.mv_data = NULL;
9363 env->me_dbxs[dbi].md_name.mv_size = 0;
9364 env->me_dbflags[dbi] = 0;
9365 env->me_dbiseqs[dbi]++;
9370 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9372 /* We could return the flags for the FREE_DBI too but what's the point? */
9373 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9375 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9379 /** Add all the DB's pages to the free list.
9380 * @param[in] mc Cursor on the DB to free.
9381 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9382 * @return 0 on success, non-zero on failure.
9385 mdb_drop0(MDB_cursor *mc, int subs)
9389 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9390 if (rc == MDB_SUCCESS) {
9391 MDB_txn *txn = mc->mc_txn;
9396 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9397 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9400 mdb_cursor_copy(mc, &mx);
9401 while (mc->mc_snum > 0) {
9402 MDB_page *mp = mc->mc_pg[mc->mc_top];
9403 unsigned n = NUMKEYS(mp);
9405 for (i=0; i<n; i++) {
9406 ni = NODEPTR(mp, i);
9407 if (ni->mn_flags & F_BIGDATA) {
9410 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9411 rc = mdb_page_get(txn, pg, &omp, NULL);
9414 mdb_cassert(mc, IS_OVERFLOW(omp));
9415 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9419 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9420 mdb_xcursor_init1(mc, ni);
9421 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9427 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9429 for (i=0; i<n; i++) {
9431 ni = NODEPTR(mp, i);
9434 mdb_midl_xappend(txn->mt_free_pgs, pg);
9439 mc->mc_ki[mc->mc_top] = i;
9440 rc = mdb_cursor_sibling(mc, 1);
9442 if (rc != MDB_NOTFOUND)
9444 /* no more siblings, go back to beginning
9445 * of previous level.
9449 for (i=1; i<mc->mc_snum; i++) {
9451 mc->mc_pg[i] = mx.mc_pg[i];
9456 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9459 txn->mt_flags |= MDB_TXN_ERROR;
9460 } else if (rc == MDB_NOTFOUND) {
9466 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9468 MDB_cursor *mc, *m2;
9471 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9474 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9477 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9480 rc = mdb_cursor_open(txn, dbi, &mc);
9484 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9485 /* Invalidate the dropped DB's cursors */
9486 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9487 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9491 /* Can't delete the main DB */
9492 if (del && dbi > MAIN_DBI) {
9493 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9495 txn->mt_dbflags[dbi] = DB_STALE;
9496 mdb_dbi_close(txn->mt_env, dbi);
9498 txn->mt_flags |= MDB_TXN_ERROR;
9501 /* reset the DB record, mark it dirty */
9502 txn->mt_dbflags[dbi] |= DB_DIRTY;
9503 txn->mt_dbs[dbi].md_depth = 0;
9504 txn->mt_dbs[dbi].md_branch_pages = 0;
9505 txn->mt_dbs[dbi].md_leaf_pages = 0;
9506 txn->mt_dbs[dbi].md_overflow_pages = 0;
9507 txn->mt_dbs[dbi].md_entries = 0;
9508 txn->mt_dbs[dbi].md_root = P_INVALID;
9510 txn->mt_flags |= MDB_TXN_DIRTY;
9513 mdb_cursor_close(mc);
9517 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9519 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9522 txn->mt_dbxs[dbi].md_cmp = cmp;
9526 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9528 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9531 txn->mt_dbxs[dbi].md_dcmp = cmp;
9535 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9537 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9540 txn->mt_dbxs[dbi].md_rel = rel;
9544 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9546 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9549 txn->mt_dbxs[dbi].md_relctx = ctx;
9554 mdb_env_get_maxkeysize(MDB_env *env)
9556 return ENV_MAXKEY(env);
9560 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9562 unsigned int i, rdrs;
9565 int rc = 0, first = 1;
9569 if (!env->me_txns) {
9570 return func("(no reader locks)\n", ctx);
9572 rdrs = env->me_txns->mti_numreaders;
9573 mr = env->me_txns->mti_readers;
9574 for (i=0; i<rdrs; i++) {
9576 txnid_t txnid = mr[i].mr_txnid;
9577 sprintf(buf, txnid == (txnid_t)-1 ?
9578 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9579 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9582 rc = func(" pid thread txnid\n", ctx);
9586 rc = func(buf, ctx);
9592 rc = func("(no active readers)\n", ctx);
9597 /** Insert pid into list if not already present.
9598 * return -1 if already present.
9601 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9603 /* binary search of pid in list */
9605 unsigned cursor = 1;
9607 unsigned n = ids[0];
9610 unsigned pivot = n >> 1;
9611 cursor = base + pivot + 1;
9612 val = pid - ids[cursor];
9617 } else if ( val > 0 ) {
9622 /* found, so it's a duplicate */
9631 for (n = ids[0]; n > cursor; n--)
9638 mdb_reader_check(MDB_env *env, int *dead)
9644 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9647 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9648 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9650 mdb_mutex_t *rmutex = rlocked ? NULL : MDB_MUTEX(env, r);
9651 unsigned int i, j, rdrs;
9653 MDB_PID_T *pids, pid;
9654 int rc = MDB_SUCCESS, count = 0;
9656 rdrs = env->me_txns->mti_numreaders;
9657 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9661 mr = env->me_txns->mti_readers;
9662 for (i=0; i<rdrs; i++) {
9664 if (pid && pid != env->me_pid) {
9665 if (mdb_pid_insert(pids, pid) == 0) {
9666 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9667 /* Stale reader found */
9670 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9671 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9673 rdrs = 0; /* the above checked all readers */
9675 /* Recheck, a new process may have reused pid */
9676 if (mdb_reader_pid(env, Pidcheck, pid))
9681 if (mr[j].mr_pid == pid) {
9682 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9683 (unsigned) pid, mr[j].mr_txnid));
9688 UNLOCK_MUTEX(rmutex);
9699 #ifdef MDB_ROBUST_SUPPORTED
9700 /** Handle #LOCK_MUTEX0() failure.
9701 * Try to repair the lock file if the mutex owner died.
9702 * @param[in] env the environment handle
9703 * @param[in] mutex LOCK_MUTEX0() mutex
9704 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9705 * @return 0 on success with the mutex locked, or an error code on failure.
9707 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc)
9709 int toggle, rlocked, rc2;
9711 if (rc == MDB_OWNERDEAD) {
9712 /* We own the mutex. Clean up after dead previous owner. */
9714 rlocked = (mutex == MDB_MUTEX(env, r));
9716 /* Keep mti_txnid updated, otherwise next writer can
9717 * overwrite data which latest meta page refers to.
9719 toggle = mdb_env_pick_meta(env);
9720 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9721 /* env is hosed if the dead thread was ours */
9723 env->me_flags |= MDB_FATAL_ERROR;
9728 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9729 (rc ? "this process' env is hosed" : "recovering")));
9730 rc2 = mdb_reader_check0(env, rlocked, NULL);
9732 rc2 = mdb_mutex_consistent(mutex);
9733 if (rc || (rc = rc2)) {
9734 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9735 UNLOCK_MUTEX(mutex);
9741 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9746 #endif /* MDB_ROBUST_SUPPORTED */