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
102 typedef SSIZE_T ssize_t;
107 #if defined(__sun) || defined(ANDROID)
108 /* Most platforms have posix_memalign, older may only have memalign */
109 #define HAVE_MEMALIGN 1
113 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
114 #include <netinet/in.h>
115 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
118 #if defined(__APPLE__) || defined (BSD)
119 # define MDB_USE_POSIX_SEM 1
120 # define MDB_FDATASYNC fsync
121 #elif defined(ANDROID)
122 # define MDB_FDATASYNC fsync
127 #ifdef MDB_USE_POSIX_SEM
128 # define MDB_USE_HASH 1
129 #include <semaphore.h>
131 #define MDB_USE_POSIX_MUTEX 1
135 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \
136 + defined(MDB_USE_POSIX_MUTEX) != 1
137 # error "Ambiguous shared-lock implementation"
141 #include <valgrind/memcheck.h>
142 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
143 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
144 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
145 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
146 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
148 #define VGMEMP_CREATE(h,r,z)
149 #define VGMEMP_ALLOC(h,a,s)
150 #define VGMEMP_FREE(h,a)
151 #define VGMEMP_DESTROY(h)
152 #define VGMEMP_DEFINED(a,s)
156 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
157 /* Solaris just defines one or the other */
158 # define LITTLE_ENDIAN 1234
159 # define BIG_ENDIAN 4321
160 # ifdef _LITTLE_ENDIAN
161 # define BYTE_ORDER LITTLE_ENDIAN
163 # define BYTE_ORDER BIG_ENDIAN
166 # define BYTE_ORDER __BYTE_ORDER
170 #ifndef LITTLE_ENDIAN
171 #define LITTLE_ENDIAN __LITTLE_ENDIAN
174 #define BIG_ENDIAN __BIG_ENDIAN
177 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
178 #define MISALIGNED_OK 1
184 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
185 # error "Unknown or unsupported endianness (BYTE_ORDER)"
186 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
187 # error "Two's complement, reasonably sized integer types, please"
191 /** Put infrequently used env functions in separate section */
193 # define ESECT __attribute__ ((section("__TEXT,text_env")))
195 # define ESECT __attribute__ ((section("text_env")))
202 #define CALL_CONV WINAPI
207 /** @defgroup internal LMDB Internals
210 /** @defgroup compat Compatibility Macros
211 * A bunch of macros to minimize the amount of platform-specific ifdefs
212 * needed throughout the rest of the code. When the features this library
213 * needs are similar enough to POSIX to be hidden in a one-or-two line
214 * replacement, this macro approach is used.
218 /** Features under development */
223 /** Wrapper around __func__, which is a C99 feature */
224 #if __STDC_VERSION__ >= 199901L
225 # define mdb_func_ __func__
226 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
227 # define mdb_func_ __FUNCTION__
229 /* If a debug message says <mdb_unknown>(), update the #if statements above */
230 # define mdb_func_ "<mdb_unknown>"
233 /* Internal error codes, not exposed outside liblmdb */
234 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
236 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
237 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
238 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
242 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
244 /** Some platforms define the EOWNERDEAD error code
245 * even though they don't support Robust Mutexes.
246 * Compile with -DMDB_USE_ROBUST=0, or use some other
247 * mechanism like -DMDB_USE_SYSV_SEM instead of
248 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
249 * also Robust, but some systems don't support them
252 #ifndef MDB_USE_ROBUST
253 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
254 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
255 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
256 # define MDB_USE_ROBUST 0
258 # define MDB_USE_ROBUST 1
259 /* glibc < 2.10 only provided _np API */
260 # if defined(__GLIBC__) && GLIBC_VER < 0x02000a
261 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
262 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
263 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
266 #endif /* MDB_USE_ROBUST */
268 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
269 #define MDB_ROBUST_SUPPORTED 1
273 #define MDB_USE_HASH 1
274 #define MDB_PIDLOCK 0
275 #define THREAD_RET DWORD
276 #define pthread_t HANDLE
277 #define pthread_mutex_t HANDLE
278 #define pthread_cond_t HANDLE
279 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
280 #define pthread_key_t DWORD
281 #define pthread_self() GetCurrentThreadId()
282 #define pthread_key_create(x,y) \
283 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
284 #define pthread_key_delete(x) TlsFree(x)
285 #define pthread_getspecific(x) TlsGetValue(x)
286 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
287 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
288 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
289 #define pthread_cond_signal(x) SetEvent(*x)
290 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
291 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
292 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
293 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
294 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
295 #define mdb_mutex_consistent(mutex) 0
296 #define getpid() GetCurrentProcessId()
297 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
298 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
299 #define ErrCode() GetLastError()
300 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
301 #define close(fd) (CloseHandle(fd) ? 0 : -1)
302 #define munmap(ptr,len) UnmapViewOfFile(ptr)
303 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
304 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
306 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
310 #define THREAD_RET void *
311 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
312 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
313 #define Z "z" /**< printf format modifier for size_t */
315 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
316 #define MDB_PIDLOCK 1
318 #ifdef MDB_USE_POSIX_SEM
320 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
321 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
322 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
325 mdb_sem_wait(sem_t *sem)
328 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
332 #else /* MDB_USE_POSIX_MUTEX: */
333 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
334 * local variables keep it (mdb_mutexref_t).
336 * When #mdb_mutexref_t is a pointer declaration and #mdb_mutex_t is
337 * not, then it is array[size 1] so it can be assigned to a pointer.
340 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
342 /** Lock the reader or writer mutex.
343 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
345 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
346 /** Unlock the reader or writer mutex.
348 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
349 /** Mark mutex-protected data as repaired, after death of previous owner.
351 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
352 #endif /* MDB_USE_POSIX_SEM */
354 /** Get the error code for the last failed system function.
356 #define ErrCode() errno
358 /** An abstraction for a file handle.
359 * On POSIX systems file handles are small integers. On Windows
360 * they're opaque pointers.
364 /** A value for an invalid file handle.
365 * Mainly used to initialize file variables and signify that they are
368 #define INVALID_HANDLE_VALUE (-1)
370 /** Get the size of a memory page for the system.
371 * This is the basic size that the platform's memory manager uses, and is
372 * fundamental to the use of memory-mapped files.
374 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
377 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
380 #define MNAME_LEN (sizeof(pthread_mutex_t))
385 #ifdef MDB_ROBUST_SUPPORTED
386 /** Lock mutex, handle any error, set rc = result.
387 * Return 0 on success, nonzero (not rc) on error.
389 #define LOCK_MUTEX(rc, env, mutex) \
390 (((rc) = LOCK_MUTEX0(mutex)) && \
391 ((rc) = mdb_mutex_failed(env, mutex, rc)))
392 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
394 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
395 #define mdb_mutex_failed(env, mutex, rc) (rc)
399 /** A flag for opening a file and requesting synchronous data writes.
400 * This is only used when writing a meta page. It's not strictly needed;
401 * we could just do a normal write and then immediately perform a flush.
402 * But if this flag is available it saves us an extra system call.
404 * @note If O_DSYNC is undefined but exists in /usr/include,
405 * preferably set some compiler flag to get the definition.
409 # define MDB_DSYNC O_DSYNC
411 # define MDB_DSYNC O_SYNC
416 /** Function for flushing the data of a file. Define this to fsync
417 * if fdatasync() is not supported.
419 #ifndef MDB_FDATASYNC
420 # define MDB_FDATASYNC fdatasync
424 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
435 /** A page number in the database.
436 * Note that 64 bit page numbers are overkill, since pages themselves
437 * already represent 12-13 bits of addressable memory, and the OS will
438 * always limit applications to a maximum of 63 bits of address space.
440 * @note In the #MDB_node structure, we only store 48 bits of this value,
441 * which thus limits us to only 60 bits of addressable data.
443 typedef MDB_ID pgno_t;
445 /** A transaction ID.
446 * See struct MDB_txn.mt_txnid for details.
448 typedef MDB_ID txnid_t;
450 /** @defgroup debug Debug Macros
454 /** Enable debug output. Needs variable argument macros (a C99 feature).
455 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
456 * read from and written to the database (used for free space management).
462 static int mdb_debug;
463 static txnid_t mdb_debug_start;
465 /** Print a debug message with printf formatting.
466 * Requires double parenthesis around 2 or more args.
468 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
469 # define DPRINTF0(fmt, ...) \
470 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
472 # define DPRINTF(args) ((void) 0)
474 /** Print a debug string.
475 * The string is printed literally, with no format processing.
477 #define DPUTS(arg) DPRINTF(("%s", arg))
478 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
480 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
483 /** @brief The maximum size of a database page.
485 * It is 32k or 64k, since value-PAGEBASE must fit in
486 * #MDB_page.%mp_upper.
488 * LMDB will use database pages < OS pages if needed.
489 * That causes more I/O in write transactions: The OS must
490 * know (read) the whole page before writing a partial page.
492 * Note that we don't currently support Huge pages. On Linux,
493 * regular data files cannot use Huge pages, and in general
494 * Huge pages aren't actually pageable. We rely on the OS
495 * demand-pager to read our data and page it out when memory
496 * pressure from other processes is high. So until OSs have
497 * actual paging support for Huge pages, they're not viable.
499 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
501 /** The minimum number of keys required in a database page.
502 * Setting this to a larger value will place a smaller bound on the
503 * maximum size of a data item. Data items larger than this size will
504 * be pushed into overflow pages instead of being stored directly in
505 * the B-tree node. This value used to default to 4. With a page size
506 * of 4096 bytes that meant that any item larger than 1024 bytes would
507 * go into an overflow page. That also meant that on average 2-3KB of
508 * each overflow page was wasted space. The value cannot be lower than
509 * 2 because then there would no longer be a tree structure. With this
510 * value, items larger than 2KB will go into overflow pages, and on
511 * average only 1KB will be wasted.
513 #define MDB_MINKEYS 2
515 /** A stamp that identifies a file as an LMDB file.
516 * There's nothing special about this value other than that it is easily
517 * recognizable, and it will reflect any byte order mismatches.
519 #define MDB_MAGIC 0xBEEFC0DE
521 /** The version number for a database's datafile format. */
522 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
523 /** The version number for a database's lockfile format. */
524 #define MDB_LOCK_VERSION 1
526 /** @brief The max size of a key we can write, or 0 for computed max.
528 * This macro should normally be left alone or set to 0.
529 * Note that a database with big keys or dupsort data cannot be
530 * reliably modified by a liblmdb which uses a smaller max.
531 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
533 * Other values are allowed, for backwards compat. However:
534 * A value bigger than the computed max can break if you do not
535 * know what you are doing, and liblmdb <= 0.9.10 can break when
536 * modifying a DB with keys/dupsort data bigger than its max.
538 * Data items in an #MDB_DUPSORT database are also limited to
539 * this size, since they're actually keys of a sub-DB. Keys and
540 * #MDB_DUPSORT data items must fit on a node in a regular page.
542 #ifndef MDB_MAXKEYSIZE
543 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
546 /** The maximum size of a key we can write to the environment. */
548 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
550 #define ENV_MAXKEY(env) ((env)->me_maxkey)
553 /** @brief The maximum size of a data item.
555 * We only store a 32 bit value for node sizes.
557 #define MAXDATASIZE 0xffffffffUL
560 /** Key size which fits in a #DKBUF.
563 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
566 * This is used for printing a hex dump of a key's contents.
568 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
569 /** Display a key in hex.
571 * Invoke a function to display a key in hex.
573 #define DKEY(x) mdb_dkey(x, kbuf)
579 /** An invalid page number.
580 * Mainly used to denote an empty tree.
582 #define P_INVALID (~(pgno_t)0)
584 /** Test if the flags \b f are set in a flag word \b w. */
585 #define F_ISSET(w, f) (((w) & (f)) == (f))
587 /** Round \b n up to an even number. */
588 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
590 /** Used for offsets within a single page.
591 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
594 typedef uint16_t indx_t;
596 /** Default size of memory map.
597 * This is certainly too small for any actual applications. Apps should always set
598 * the size explicitly using #mdb_env_set_mapsize().
600 #define DEFAULT_MAPSIZE 1048576
602 /** @defgroup readers Reader Lock Table
603 * Readers don't acquire any locks for their data access. Instead, they
604 * simply record their transaction ID in the reader table. The reader
605 * mutex is needed just to find an empty slot in the reader table. The
606 * slot's address is saved in thread-specific data so that subsequent read
607 * transactions started by the same thread need no further locking to proceed.
609 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
611 * No reader table is used if the database is on a read-only filesystem, or
612 * if #MDB_NOLOCK is set.
614 * Since the database uses multi-version concurrency control, readers don't
615 * actually need any locking. This table is used to keep track of which
616 * readers are using data from which old transactions, so that we'll know
617 * when a particular old transaction is no longer in use. Old transactions
618 * that have discarded any data pages can then have those pages reclaimed
619 * for use by a later write transaction.
621 * The lock table is constructed such that reader slots are aligned with the
622 * processor's cache line size. Any slot is only ever used by one thread.
623 * This alignment guarantees that there will be no contention or cache
624 * thrashing as threads update their own slot info, and also eliminates
625 * any need for locking when accessing a slot.
627 * A writer thread will scan every slot in the table to determine the oldest
628 * outstanding reader transaction. Any freed pages older than this will be
629 * reclaimed by the writer. The writer doesn't use any locks when scanning
630 * this table. This means that there's no guarantee that the writer will
631 * see the most up-to-date reader info, but that's not required for correct
632 * operation - all we need is to know the upper bound on the oldest reader,
633 * we don't care at all about the newest reader. So the only consequence of
634 * reading stale information here is that old pages might hang around a
635 * while longer before being reclaimed. That's actually good anyway, because
636 * the longer we delay reclaiming old pages, the more likely it is that a
637 * string of contiguous pages can be found after coalescing old pages from
638 * many old transactions together.
641 /** Number of slots in the reader table.
642 * This value was chosen somewhat arbitrarily. 126 readers plus a
643 * couple mutexes fit exactly into 8KB on my development machine.
644 * Applications should set the table size using #mdb_env_set_maxreaders().
646 #define DEFAULT_READERS 126
648 /** The size of a CPU cache line in bytes. We want our lock structures
649 * aligned to this size to avoid false cache line sharing in the
651 * This value works for most CPUs. For Itanium this should be 128.
657 /** The information we store in a single slot of the reader table.
658 * In addition to a transaction ID, we also record the process and
659 * thread ID that owns a slot, so that we can detect stale information,
660 * e.g. threads or processes that went away without cleaning up.
661 * @note We currently don't check for stale records. We simply re-init
662 * the table when we know that we're the only process opening the
665 typedef struct MDB_rxbody {
666 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
667 * Multiple readers that start at the same time will probably have the
668 * same ID here. Again, it's not important to exclude them from
669 * anything; all we need to know is which version of the DB they
670 * started from so we can avoid overwriting any data used in that
671 * particular version.
673 volatile txnid_t mrb_txnid;
674 /** The process ID of the process owning this reader txn. */
675 volatile MDB_PID_T mrb_pid;
676 /** The thread ID of the thread owning this txn. */
677 volatile MDB_THR_T mrb_tid;
680 /** The actual reader record, with cacheline padding. */
681 typedef struct MDB_reader {
684 /** shorthand for mrb_txnid */
685 #define mr_txnid mru.mrx.mrb_txnid
686 #define mr_pid mru.mrx.mrb_pid
687 #define mr_tid mru.mrx.mrb_tid
688 /** cache line alignment */
689 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
693 /** The header for the reader table.
694 * The table resides in a memory-mapped file. (This is a different file
695 * than is used for the main database.)
697 * For POSIX the actual mutexes reside in the shared memory of this
698 * mapped file. On Windows, mutexes are named objects allocated by the
699 * kernel; we store the mutex names in this mapped file so that other
700 * processes can grab them. This same approach is also used on
701 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
702 * process-shared POSIX mutexes. For these cases where a named object
703 * is used, the object name is derived from a 64 bit FNV hash of the
704 * environment pathname. As such, naming collisions are extremely
705 * unlikely. If a collision occurs, the results are unpredictable.
707 typedef struct MDB_txbody {
708 /** Stamp identifying this as an LMDB file. It must be set
711 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
713 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
714 char mtb_rmname[MNAME_LEN];
716 /** Mutex protecting access to this table.
717 * This is the reader table lock used with LOCK_MUTEX().
719 mdb_mutex_t mtb_rmutex;
721 /** The ID of the last transaction committed to the database.
722 * This is recorded here only for convenience; the value can always
723 * be determined by reading the main database meta pages.
725 volatile txnid_t mtb_txnid;
726 /** The number of slots that have been used in the reader table.
727 * This always records the maximum count, it is not decremented
728 * when readers release their slots.
730 volatile unsigned mtb_numreaders;
733 /** The actual reader table definition. */
734 typedef struct MDB_txninfo {
737 #define mti_magic mt1.mtb.mtb_magic
738 #define mti_format mt1.mtb.mtb_format
739 #define mti_rmutex mt1.mtb.mtb_rmutex
740 #define mti_rmname mt1.mtb.mtb_rmname
741 #define mti_txnid mt1.mtb.mtb_txnid
742 #define mti_numreaders mt1.mtb.mtb_numreaders
743 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
746 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
747 char mt2_wmname[MNAME_LEN];
748 #define mti_wmname mt2.mt2_wmname
750 mdb_mutex_t mt2_wmutex;
751 #define mti_wmutex mt2.mt2_wmutex
753 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
755 MDB_reader mti_readers[1];
758 /** Lockfile format signature: version, features and field layout */
759 #define MDB_LOCK_FORMAT \
761 ((MDB_LOCK_VERSION) \
762 /* Flags which describe functionality */ \
763 + (((MDB_PIDLOCK) != 0) << 16)))
766 /** Common header for all page types.
767 * Overflow records occupy a number of contiguous pages with no
768 * headers on any page after the first.
770 typedef struct MDB_page {
771 #define mp_pgno mp_p.p_pgno
772 #define mp_next mp_p.p_next
774 pgno_t p_pgno; /**< page number */
775 struct MDB_page *p_next; /**< for in-memory list of freed pages */
778 /** @defgroup mdb_page Page Flags
780 * Flags for the page headers.
783 #define P_BRANCH 0x01 /**< branch page */
784 #define P_LEAF 0x02 /**< leaf page */
785 #define P_OVERFLOW 0x04 /**< overflow page */
786 #define P_META 0x08 /**< meta page */
787 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
788 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
789 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
790 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
791 #define P_KEEP 0x8000 /**< leave this page alone during spill */
793 uint16_t mp_flags; /**< @ref mdb_page */
794 #define mp_lower mp_pb.pb.pb_lower
795 #define mp_upper mp_pb.pb.pb_upper
796 #define mp_pages mp_pb.pb_pages
799 indx_t pb_lower; /**< lower bound of free space */
800 indx_t pb_upper; /**< upper bound of free space */
802 uint32_t pb_pages; /**< number of overflow pages */
804 indx_t mp_ptrs[1]; /**< dynamic size */
807 /** Size of the page header, excluding dynamic data at the end */
808 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
810 /** Address of first usable data byte in a page, after the header */
811 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
813 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
814 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
816 /** Number of nodes on a page */
817 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
819 /** The amount of space remaining in the page */
820 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
822 /** The percentage of space used in the page, in tenths of a percent. */
823 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
824 ((env)->me_psize - PAGEHDRSZ))
825 /** The minimum page fill factor, in tenths of a percent.
826 * Pages emptier than this are candidates for merging.
828 #define FILL_THRESHOLD 250
830 /** Test if a page is a leaf page */
831 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
832 /** Test if a page is a LEAF2 page */
833 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
834 /** Test if a page is a branch page */
835 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
836 /** Test if a page is an overflow page */
837 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
838 /** Test if a page is a sub page */
839 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
841 /** The number of overflow pages needed to store the given size. */
842 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
844 /** Link in #MDB_txn.%mt_loose_pgs list */
845 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
847 /** Header for a single key/data pair within a page.
848 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
849 * We guarantee 2-byte alignment for 'MDB_node's.
851 typedef struct MDB_node {
852 /** lo and hi are used for data size on leaf nodes and for
853 * child pgno on branch nodes. On 64 bit platforms, flags
854 * is also used for pgno. (Branch nodes have no flags).
855 * They are in host byte order in case that lets some
856 * accesses be optimized into a 32-bit word access.
858 #if BYTE_ORDER == LITTLE_ENDIAN
859 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
861 unsigned short mn_hi, mn_lo;
863 /** @defgroup mdb_node Node Flags
865 * Flags for node headers.
868 #define F_BIGDATA 0x01 /**< data put on overflow page */
869 #define F_SUBDATA 0x02 /**< data is a sub-database */
870 #define F_DUPDATA 0x04 /**< data has duplicates */
872 /** valid flags for #mdb_node_add() */
873 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
876 unsigned short mn_flags; /**< @ref mdb_node */
877 unsigned short mn_ksize; /**< key size */
878 char mn_data[1]; /**< key and data are appended here */
881 /** Size of the node header, excluding dynamic data at the end */
882 #define NODESIZE offsetof(MDB_node, mn_data)
884 /** Bit position of top word in page number, for shifting mn_flags */
885 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
887 /** Size of a node in a branch page with a given key.
888 * This is just the node header plus the key, there is no data.
890 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
892 /** Size of a node in a leaf page with a given key and data.
893 * This is node header plus key plus data size.
895 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
897 /** Address of node \b i in page \b p */
898 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
900 /** Address of the key for the node */
901 #define NODEKEY(node) (void *)((node)->mn_data)
903 /** Address of the data for a node */
904 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
906 /** Get the page number pointed to by a branch node */
907 #define NODEPGNO(node) \
908 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
909 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
910 /** Set the page number in a branch node */
911 #define SETPGNO(node,pgno) do { \
912 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
913 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
915 /** Get the size of the data in a leaf node */
916 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
917 /** Set the size of the data for a leaf node */
918 #define SETDSZ(node,size) do { \
919 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
920 /** The size of a key in a node */
921 #define NODEKSZ(node) ((node)->mn_ksize)
923 /** Copy a page number from src to dst */
925 #define COPY_PGNO(dst,src) dst = src
927 #if SIZE_MAX > 4294967295UL
928 #define COPY_PGNO(dst,src) do { \
929 unsigned short *s, *d; \
930 s = (unsigned short *)&(src); \
931 d = (unsigned short *)&(dst); \
938 #define COPY_PGNO(dst,src) do { \
939 unsigned short *s, *d; \
940 s = (unsigned short *)&(src); \
941 d = (unsigned short *)&(dst); \
947 /** The address of a key in a LEAF2 page.
948 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
949 * There are no node headers, keys are stored contiguously.
951 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
953 /** Set the \b node's key into \b keyptr, if requested. */
954 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
955 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
957 /** Set the \b node's key into \b key. */
958 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
960 /** Information about a single database in the environment. */
961 typedef struct MDB_db {
962 uint32_t md_pad; /**< also ksize for LEAF2 pages */
963 uint16_t md_flags; /**< @ref mdb_dbi_open */
964 uint16_t md_depth; /**< depth of this tree */
965 pgno_t md_branch_pages; /**< number of internal pages */
966 pgno_t md_leaf_pages; /**< number of leaf pages */
967 pgno_t md_overflow_pages; /**< number of overflow pages */
968 size_t md_entries; /**< number of data items */
969 pgno_t md_root; /**< the root page of this tree */
972 /** mdb_dbi_open flags */
973 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
974 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
975 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
976 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
978 /** Handle for the DB used to track free pages. */
980 /** Handle for the default DB. */
982 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
985 /** Number of meta pages - also hardcoded elsewhere */
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[CORE_DBS]; /**< first is free space, 2nd is main db */
1001 /** The size of pages used in this DB */
1002 #define mm_psize mm_dbs[FREE_DBI].md_pad
1003 /** Any persistent environment flags. @ref mdb_env */
1004 #define mm_flags mm_dbs[FREE_DBI].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 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1041 pgno_t mt_next_pgno; /**< next unallocated page */
1042 /** The ID of this transaction. IDs are integers incrementing from 1.
1043 * Only committed write transactions increment the ID. If a transaction
1044 * aborts, the ID may be re-used by the next writer.
1047 MDB_env *mt_env; /**< the DB environment */
1048 /** The list of pages that became unused during this transaction.
1050 MDB_IDL mt_free_pgs;
1051 /** The list of loose pages that became unused and may be reused
1052 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1054 MDB_page *mt_loose_pgs;
1055 /* #Number of loose pages (#mt_loose_pgs) */
1057 /** The sorted list of dirty pages we temporarily wrote to disk
1058 * because the dirty list was full. page numbers in here are
1059 * shifted left by 1, deleted slots have the LSB set.
1061 MDB_IDL mt_spill_pgs;
1063 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1064 MDB_ID2L dirty_list;
1065 /** For read txns: This thread/txn's reader table slot, or NULL. */
1068 /** Array of records for each DB known in the environment. */
1070 /** Array of MDB_db records for each known DB */
1072 /** Array of sequence numbers for each DB handle */
1073 unsigned int *mt_dbiseqs;
1074 /** @defgroup mt_dbflag Transaction DB Flags
1078 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1079 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1080 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1081 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1082 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1084 /** In write txns, array of cursors for each DB */
1085 MDB_cursor **mt_cursors;
1086 /** Array of flags for each DB */
1087 unsigned char *mt_dbflags;
1088 /** Number of DB records in use, or 0 when the txn is finished.
1089 * This number only ever increments until the txn finishes; we
1090 * don't decrement it when individual DB handles are closed.
1094 /** @defgroup mdb_txn Transaction Flags
1098 /** #mdb_txn_begin() flags */
1099 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1100 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1101 /* internal txn flags */
1102 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1103 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1104 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1105 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1106 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1107 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1108 /** most operations on the txn are currently illegal */
1109 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1111 unsigned int mt_flags; /**< @ref mdb_txn */
1112 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1113 * Includes ancestor txns' dirty pages not hidden by other txns'
1114 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1115 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1117 unsigned int mt_dirty_room;
1120 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1121 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1122 * raise this on a 64 bit machine.
1124 #define CURSOR_STACK 32
1128 /** Cursors are used for all DB operations.
1129 * A cursor holds a path of (page pointer, key index) from the DB
1130 * root to a position in the DB, plus other state. #MDB_DUPSORT
1131 * cursors include an xcursor to the current data item. Write txns
1132 * track their cursors and keep them up to date when data moves.
1133 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1134 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1137 /** Next cursor on this DB in this txn */
1138 MDB_cursor *mc_next;
1139 /** Backup of the original cursor if this cursor is a shadow */
1140 MDB_cursor *mc_backup;
1141 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1142 struct MDB_xcursor *mc_xcursor;
1143 /** The transaction that owns this cursor */
1145 /** The database handle this cursor operates on */
1147 /** The database record for this cursor */
1149 /** The database auxiliary record for this cursor */
1151 /** The @ref mt_dbflag for this database */
1152 unsigned char *mc_dbflag;
1153 unsigned short mc_snum; /**< number of pushed pages */
1154 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1155 /** @defgroup mdb_cursor Cursor Flags
1157 * Cursor state flags.
1160 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1161 #define C_EOF 0x02 /**< No more data */
1162 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1163 #define C_DEL 0x08 /**< last op was a cursor_del */
1164 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1165 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1167 unsigned int mc_flags; /**< @ref mdb_cursor */
1168 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1169 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1172 /** Context for sorted-dup records.
1173 * We could have gone to a fully recursive design, with arbitrarily
1174 * deep nesting of sub-databases. But for now we only handle these
1175 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1177 typedef struct MDB_xcursor {
1178 /** A sub-cursor for traversing the Dup DB */
1179 MDB_cursor mx_cursor;
1180 /** The database record for this Dup DB */
1182 /** The auxiliary DB record for this Dup DB */
1184 /** The @ref mt_dbflag for this Dup DB */
1185 unsigned char mx_dbflag;
1188 /** State of FreeDB old pages, stored in the MDB_env */
1189 typedef struct MDB_pgstate {
1190 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1191 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1194 /** The database environment. */
1196 HANDLE me_fd; /**< The main data file */
1197 HANDLE me_lfd; /**< The lock file */
1198 HANDLE me_mfd; /**< just for writing the meta pages */
1199 /** Failed to update the meta page. Probably an I/O error. */
1200 #define MDB_FATAL_ERROR 0x80000000U
1201 /** Some fields are initialized. */
1202 #define MDB_ENV_ACTIVE 0x20000000U
1203 /** me_txkey is set */
1204 #define MDB_ENV_TXKEY 0x10000000U
1205 /** fdatasync is unreliable */
1206 #define MDB_FSYNCONLY 0x08000000U
1207 uint32_t me_flags; /**< @ref mdb_env */
1208 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1209 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1210 unsigned int me_maxreaders; /**< size of the reader table */
1211 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1212 volatile int me_close_readers;
1213 MDB_dbi me_numdbs; /**< number of DBs opened */
1214 MDB_dbi me_maxdbs; /**< size of the DB table */
1215 MDB_PID_T me_pid; /**< process ID of this env */
1216 char *me_path; /**< path to the DB files */
1217 char *me_map; /**< the memory map of the data file */
1218 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1219 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1220 void *me_pbuf; /**< scratch area for DUPSORT put() */
1221 MDB_txn *me_txn; /**< current write transaction */
1222 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1223 size_t me_mapsize; /**< size of the data memory map */
1224 off_t me_size; /**< current file size */
1225 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1226 MDB_dbx *me_dbxs; /**< array of static DB info */
1227 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1228 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1229 pthread_key_t me_txkey; /**< thread-key for readers */
1230 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1231 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1232 # define me_pglast me_pgstate.mf_pglast
1233 # define me_pghead me_pgstate.mf_pghead
1234 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1235 /** IDL of pages that became unused in a write txn */
1236 MDB_IDL me_free_pgs;
1237 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1238 MDB_ID2L me_dirty_list;
1239 /** Max number of freelist items that can fit in a single overflow page */
1241 /** Max size of a node on a page */
1242 unsigned int me_nodemax;
1243 #if !(MDB_MAXKEYSIZE)
1244 unsigned int me_maxkey; /**< max size of a key */
1246 int me_live_reader; /**< have liveness lock in reader table */
1248 int me_pidquery; /**< Used in OpenProcess */
1250 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1251 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1252 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1254 mdb_mutex_t me_rmutex;
1255 mdb_mutex_t me_wmutex;
1257 void *me_userctx; /**< User-settable context */
1258 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1261 /** Nested transaction */
1262 typedef struct MDB_ntxn {
1263 MDB_txn mnt_txn; /**< the transaction */
1264 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1267 /** max number of pages to commit in one writev() call */
1268 #define MDB_COMMIT_PAGES 64
1269 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1270 #undef MDB_COMMIT_PAGES
1271 #define MDB_COMMIT_PAGES IOV_MAX
1274 /** max bytes to write in one call */
1275 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1277 /** Check \b txn and \b dbi arguments to a function */
1278 #define TXN_DBI_EXIST(txn, dbi, validity) \
1279 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1281 /** Check for misused \b dbi handles */
1282 #define TXN_DBI_CHANGED(txn, dbi) \
1283 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1285 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1286 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1287 static int mdb_page_touch(MDB_cursor *mc);
1289 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1290 "reset-tmp", "fail-begin", "fail-beginchild"}
1292 /* mdb_txn_end operation number, for logging */
1293 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1294 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1296 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1297 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1298 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1299 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1300 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1302 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1303 static int mdb_page_search_root(MDB_cursor *mc,
1304 MDB_val *key, int modify);
1305 #define MDB_PS_MODIFY 1
1306 #define MDB_PS_ROOTONLY 2
1307 #define MDB_PS_FIRST 4
1308 #define MDB_PS_LAST 8
1309 static int mdb_page_search(MDB_cursor *mc,
1310 MDB_val *key, int flags);
1311 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1313 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1314 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1315 pgno_t newpgno, unsigned int nflags);
1317 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1318 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1319 static int mdb_env_write_meta(MDB_txn *txn);
1320 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1321 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1323 static void mdb_env_close0(MDB_env *env, int excl);
1325 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1326 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1327 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1328 static void mdb_node_del(MDB_cursor *mc, int ksize);
1329 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1330 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1331 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1332 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1333 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1335 static int mdb_rebalance(MDB_cursor *mc);
1336 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1338 static void mdb_cursor_pop(MDB_cursor *mc);
1339 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1341 static int mdb_cursor_del0(MDB_cursor *mc);
1342 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1343 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1344 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1345 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1346 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1348 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1349 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1351 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1352 static void mdb_xcursor_init0(MDB_cursor *mc);
1353 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1354 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1356 static int mdb_drop0(MDB_cursor *mc, int subs);
1357 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1358 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1361 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1364 /** Compare two items pointing at size_t's of unknown alignment. */
1365 #ifdef MISALIGNED_OK
1366 # define mdb_cmp_clong mdb_cmp_long
1368 # define mdb_cmp_clong mdb_cmp_cint
1372 static SECURITY_DESCRIPTOR mdb_null_sd;
1373 static SECURITY_ATTRIBUTES mdb_all_sa;
1374 static int mdb_sec_inited;
1377 /** Return the library version info. */
1379 mdb_version(int *major, int *minor, int *patch)
1381 if (major) *major = MDB_VERSION_MAJOR;
1382 if (minor) *minor = MDB_VERSION_MINOR;
1383 if (patch) *patch = MDB_VERSION_PATCH;
1384 return MDB_VERSION_STRING;
1387 /** Table of descriptions for LMDB @ref errors */
1388 static char *const mdb_errstr[] = {
1389 "MDB_KEYEXIST: Key/data pair already exists",
1390 "MDB_NOTFOUND: No matching key/data pair found",
1391 "MDB_PAGE_NOTFOUND: Requested page not found",
1392 "MDB_CORRUPTED: Located page was wrong type",
1393 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1394 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1395 "MDB_INVALID: File is not an LMDB file",
1396 "MDB_MAP_FULL: Environment mapsize limit reached",
1397 "MDB_DBS_FULL: Environment maxdbs limit reached",
1398 "MDB_READERS_FULL: Environment maxreaders limit reached",
1399 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1400 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1401 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1402 "MDB_PAGE_FULL: Internal error - page has no more space",
1403 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1404 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1405 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1406 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1407 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1408 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1412 mdb_strerror(int err)
1415 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1416 * This works as long as no function between the call to mdb_strerror
1417 * and the actual use of the message uses more than 4K of stack.
1420 char buf[1024], *ptr = buf;
1424 return ("Successful return: 0");
1426 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1427 i = err - MDB_KEYEXIST;
1428 return mdb_errstr[i];
1432 /* These are the C-runtime error codes we use. The comment indicates
1433 * their numeric value, and the Win32 error they would correspond to
1434 * if the error actually came from a Win32 API. A major mess, we should
1435 * have used LMDB-specific error codes for everything.
1438 case ENOENT: /* 2, FILE_NOT_FOUND */
1439 case EIO: /* 5, ACCESS_DENIED */
1440 case ENOMEM: /* 12, INVALID_ACCESS */
1441 case EACCES: /* 13, INVALID_DATA */
1442 case EBUSY: /* 16, CURRENT_DIRECTORY */
1443 case EINVAL: /* 22, BAD_COMMAND */
1444 case ENOSPC: /* 28, OUT_OF_PAPER */
1445 return strerror(err);
1450 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1451 FORMAT_MESSAGE_IGNORE_INSERTS,
1452 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1455 return strerror(err);
1459 /** assert(3) variant in cursor context */
1460 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1461 /** assert(3) variant in transaction context */
1462 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1463 /** assert(3) variant in environment context */
1464 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1467 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1468 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1471 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1472 const char *func, const char *file, int line)
1475 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1476 file, line, expr_txt, func);
1477 if (env->me_assert_func)
1478 env->me_assert_func(env, buf);
1479 fprintf(stderr, "%s\n", buf);
1483 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1487 /** Return the page number of \b mp which may be sub-page, for debug output */
1489 mdb_dbg_pgno(MDB_page *mp)
1492 COPY_PGNO(ret, mp->mp_pgno);
1496 /** Display a key in hexadecimal and return the address of the result.
1497 * @param[in] key the key to display
1498 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1499 * @return The key in hexadecimal form.
1502 mdb_dkey(MDB_val *key, char *buf)
1505 unsigned char *c = key->mv_data;
1511 if (key->mv_size > DKBUF_MAXKEYSIZE)
1512 return "MDB_MAXKEYSIZE";
1513 /* may want to make this a dynamic check: if the key is mostly
1514 * printable characters, print it as-is instead of converting to hex.
1518 for (i=0; i<key->mv_size; i++)
1519 ptr += sprintf(ptr, "%02x", *c++);
1521 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1527 mdb_leafnode_type(MDB_node *n)
1529 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1530 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1531 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1534 /** Display all the keys in the page. */
1536 mdb_page_list(MDB_page *mp)
1538 pgno_t pgno = mdb_dbg_pgno(mp);
1539 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1541 unsigned int i, nkeys, nsize, total = 0;
1545 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1546 case P_BRANCH: type = "Branch page"; break;
1547 case P_LEAF: type = "Leaf page"; break;
1548 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1549 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1550 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1552 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1553 pgno, mp->mp_pages, state);
1556 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1557 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1560 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1564 nkeys = NUMKEYS(mp);
1565 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1567 for (i=0; i<nkeys; i++) {
1568 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1569 key.mv_size = nsize = mp->mp_pad;
1570 key.mv_data = LEAF2KEY(mp, i, nsize);
1572 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1575 node = NODEPTR(mp, i);
1576 key.mv_size = node->mn_ksize;
1577 key.mv_data = node->mn_data;
1578 nsize = NODESIZE + key.mv_size;
1579 if (IS_BRANCH(mp)) {
1580 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1584 if (F_ISSET(node->mn_flags, F_BIGDATA))
1585 nsize += sizeof(pgno_t);
1587 nsize += NODEDSZ(node);
1589 nsize += sizeof(indx_t);
1590 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1591 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1593 total = EVEN(total);
1595 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1596 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1600 mdb_cursor_chk(MDB_cursor *mc)
1606 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1607 for (i=0; i<mc->mc_top; i++) {
1609 node = NODEPTR(mp, mc->mc_ki[i]);
1610 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1613 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1619 /** Count all the pages in each DB and in the freelist
1620 * and make sure it matches the actual number of pages
1622 * All named DBs must be open for a correct count.
1624 static void mdb_audit(MDB_txn *txn)
1628 MDB_ID freecount, count;
1633 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1634 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1635 freecount += *(MDB_ID *)data.mv_data;
1636 mdb_tassert(txn, rc == MDB_NOTFOUND);
1639 for (i = 0; i<txn->mt_numdbs; i++) {
1641 if (!(txn->mt_dbflags[i] & DB_VALID))
1643 mdb_cursor_init(&mc, txn, i, &mx);
1644 if (txn->mt_dbs[i].md_root == P_INVALID)
1646 count += txn->mt_dbs[i].md_branch_pages +
1647 txn->mt_dbs[i].md_leaf_pages +
1648 txn->mt_dbs[i].md_overflow_pages;
1649 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1650 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1651 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1654 mp = mc.mc_pg[mc.mc_top];
1655 for (j=0; j<NUMKEYS(mp); j++) {
1656 MDB_node *leaf = NODEPTR(mp, j);
1657 if (leaf->mn_flags & F_SUBDATA) {
1659 memcpy(&db, NODEDATA(leaf), sizeof(db));
1660 count += db.md_branch_pages + db.md_leaf_pages +
1661 db.md_overflow_pages;
1665 mdb_tassert(txn, rc == MDB_NOTFOUND);
1668 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1669 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1670 txn->mt_txnid, freecount, count+NUM_METAS,
1671 freecount+count+NUM_METAS, txn->mt_next_pgno);
1677 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1679 return txn->mt_dbxs[dbi].md_cmp(a, b);
1683 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1685 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1686 #if UINT_MAX < SIZE_MAX
1687 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1688 dcmp = mdb_cmp_clong;
1693 /** Allocate memory for a page.
1694 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1697 mdb_page_malloc(MDB_txn *txn, unsigned num)
1699 MDB_env *env = txn->mt_env;
1700 MDB_page *ret = env->me_dpages;
1701 size_t psize = env->me_psize, sz = psize, off;
1702 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1703 * For a single page alloc, we init everything after the page header.
1704 * For multi-page, we init the final page; if the caller needed that
1705 * many pages they will be filling in at least up to the last page.
1709 VGMEMP_ALLOC(env, ret, sz);
1710 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1711 env->me_dpages = ret->mp_next;
1714 psize -= off = PAGEHDRSZ;
1719 if ((ret = malloc(sz)) != NULL) {
1720 VGMEMP_ALLOC(env, ret, sz);
1721 if (!(env->me_flags & MDB_NOMEMINIT)) {
1722 memset((char *)ret + off, 0, psize);
1726 txn->mt_flags |= MDB_TXN_ERROR;
1730 /** Free a single page.
1731 * Saves single pages to a list, for future reuse.
1732 * (This is not used for multi-page overflow pages.)
1735 mdb_page_free(MDB_env *env, MDB_page *mp)
1737 mp->mp_next = env->me_dpages;
1738 VGMEMP_FREE(env, mp);
1739 env->me_dpages = mp;
1742 /** Free a dirty page */
1744 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1746 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1747 mdb_page_free(env, dp);
1749 /* large pages just get freed directly */
1750 VGMEMP_FREE(env, dp);
1755 /** Return all dirty pages to dpage list */
1757 mdb_dlist_free(MDB_txn *txn)
1759 MDB_env *env = txn->mt_env;
1760 MDB_ID2L dl = txn->mt_u.dirty_list;
1761 unsigned i, n = dl[0].mid;
1763 for (i = 1; i <= n; i++) {
1764 mdb_dpage_free(env, dl[i].mptr);
1769 /** Loosen or free a single page.
1770 * Saves single pages to a list for future reuse
1771 * in this same txn. It has been pulled from the freeDB
1772 * and already resides on the dirty list, but has been
1773 * deleted. Use these pages first before pulling again
1776 * If the page wasn't dirtied in this txn, just add it
1777 * to this txn's free list.
1780 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1783 pgno_t pgno = mp->mp_pgno;
1784 MDB_txn *txn = mc->mc_txn;
1786 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1787 if (txn->mt_parent) {
1788 MDB_ID2 *dl = txn->mt_u.dirty_list;
1789 /* If txn has a parent, make sure the page is in our
1793 unsigned x = mdb_mid2l_search(dl, pgno);
1794 if (x <= dl[0].mid && dl[x].mid == pgno) {
1795 if (mp != dl[x].mptr) { /* bad cursor? */
1796 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1797 txn->mt_flags |= MDB_TXN_ERROR;
1798 return MDB_CORRUPTED;
1805 /* no parent txn, so it's just ours */
1810 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1812 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1813 txn->mt_loose_pgs = mp;
1814 txn->mt_loose_count++;
1815 mp->mp_flags |= P_LOOSE;
1817 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1825 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1826 * @param[in] mc A cursor handle for the current operation.
1827 * @param[in] pflags Flags of the pages to update:
1828 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1829 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1830 * @return 0 on success, non-zero on failure.
1833 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1835 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1836 MDB_txn *txn = mc->mc_txn;
1842 int rc = MDB_SUCCESS, level;
1844 /* Mark pages seen by cursors */
1845 if (mc->mc_flags & C_UNTRACK)
1846 mc = NULL; /* will find mc in mt_cursors */
1847 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1848 for (; mc; mc=mc->mc_next) {
1849 if (!(mc->mc_flags & C_INITIALIZED))
1851 for (m3 = mc;; m3 = &mx->mx_cursor) {
1853 for (j=0; j<m3->mc_snum; j++) {
1855 if ((mp->mp_flags & Mask) == pflags)
1856 mp->mp_flags ^= P_KEEP;
1858 mx = m3->mc_xcursor;
1859 /* Proceed to mx if it is at a sub-database */
1860 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1862 if (! (mp && (mp->mp_flags & P_LEAF)))
1864 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1865 if (!(leaf->mn_flags & F_SUBDATA))
1874 /* Mark dirty root pages */
1875 for (i=0; i<txn->mt_numdbs; i++) {
1876 if (txn->mt_dbflags[i] & DB_DIRTY) {
1877 pgno_t pgno = txn->mt_dbs[i].md_root;
1878 if (pgno == P_INVALID)
1880 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1882 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1883 dp->mp_flags ^= P_KEEP;
1891 static int mdb_page_flush(MDB_txn *txn, int keep);
1893 /** Spill pages from the dirty list back to disk.
1894 * This is intended to prevent running into #MDB_TXN_FULL situations,
1895 * but note that they may still occur in a few cases:
1896 * 1) our estimate of the txn size could be too small. Currently this
1897 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1898 * 2) child txns may run out of space if their parents dirtied a
1899 * lot of pages and never spilled them. TODO: we probably should do
1900 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1901 * the parent's dirty_room is below a given threshold.
1903 * Otherwise, if not using nested txns, it is expected that apps will
1904 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1905 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1906 * If the txn never references them again, they can be left alone.
1907 * If the txn only reads them, they can be used without any fuss.
1908 * If the txn writes them again, they can be dirtied immediately without
1909 * going thru all of the work of #mdb_page_touch(). Such references are
1910 * handled by #mdb_page_unspill().
1912 * Also note, we never spill DB root pages, nor pages of active cursors,
1913 * because we'll need these back again soon anyway. And in nested txns,
1914 * we can't spill a page in a child txn if it was already spilled in a
1915 * parent txn. That would alter the parent txns' data even though
1916 * the child hasn't committed yet, and we'd have no way to undo it if
1917 * the child aborted.
1919 * @param[in] m0 cursor A cursor handle identifying the transaction and
1920 * database for which we are checking space.
1921 * @param[in] key For a put operation, the key being stored.
1922 * @param[in] data For a put operation, the data being stored.
1923 * @return 0 on success, non-zero on failure.
1926 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1928 MDB_txn *txn = m0->mc_txn;
1930 MDB_ID2L dl = txn->mt_u.dirty_list;
1931 unsigned int i, j, need;
1934 if (m0->mc_flags & C_SUB)
1937 /* Estimate how much space this op will take */
1938 i = m0->mc_db->md_depth;
1939 /* Named DBs also dirty the main DB */
1940 if (m0->mc_dbi >= CORE_DBS)
1941 i += txn->mt_dbs[MAIN_DBI].md_depth;
1942 /* For puts, roughly factor in the key+data size */
1944 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1945 i += i; /* double it for good measure */
1948 if (txn->mt_dirty_room > i)
1951 if (!txn->mt_spill_pgs) {
1952 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1953 if (!txn->mt_spill_pgs)
1956 /* purge deleted slots */
1957 MDB_IDL sl = txn->mt_spill_pgs;
1958 unsigned int num = sl[0];
1960 for (i=1; i<=num; i++) {
1967 /* Preserve pages which may soon be dirtied again */
1968 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1971 /* Less aggressive spill - we originally spilled the entire dirty list,
1972 * with a few exceptions for cursor pages and DB root pages. But this
1973 * turns out to be a lot of wasted effort because in a large txn many
1974 * of those pages will need to be used again. So now we spill only 1/8th
1975 * of the dirty pages. Testing revealed this to be a good tradeoff,
1976 * better than 1/2, 1/4, or 1/10.
1978 if (need < MDB_IDL_UM_MAX / 8)
1979 need = MDB_IDL_UM_MAX / 8;
1981 /* Save the page IDs of all the pages we're flushing */
1982 /* flush from the tail forward, this saves a lot of shifting later on. */
1983 for (i=dl[0].mid; i && need; i--) {
1984 MDB_ID pn = dl[i].mid << 1;
1986 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1988 /* Can't spill twice, make sure it's not already in a parent's
1991 if (txn->mt_parent) {
1993 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1994 if (tx2->mt_spill_pgs) {
1995 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1996 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1997 dp->mp_flags |= P_KEEP;
2005 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2009 mdb_midl_sort(txn->mt_spill_pgs);
2011 /* Flush the spilled part of dirty list */
2012 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2015 /* Reset any dirty pages we kept that page_flush didn't see */
2016 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2019 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2023 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2025 mdb_find_oldest(MDB_txn *txn)
2028 txnid_t mr, oldest = txn->mt_txnid - 1;
2029 if (txn->mt_env->me_txns) {
2030 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2031 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2042 /** Add a page to the txn's dirty list */
2044 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2047 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2049 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2050 insert = mdb_mid2l_append;
2052 insert = mdb_mid2l_insert;
2054 mid.mid = mp->mp_pgno;
2056 rc = insert(txn->mt_u.dirty_list, &mid);
2057 mdb_tassert(txn, rc == 0);
2058 txn->mt_dirty_room--;
2061 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2062 * me_pghead and mt_next_pgno.
2064 * If there are free pages available from older transactions, they
2065 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2066 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2067 * and move me_pglast to say which records were consumed. Only this
2068 * function can create me_pghead and move me_pglast/mt_next_pgno.
2069 * @param[in] mc cursor A cursor handle identifying the transaction and
2070 * database for which we are allocating.
2071 * @param[in] num the number of pages to allocate.
2072 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2073 * will always be satisfied by a single contiguous chunk of memory.
2074 * @return 0 on success, non-zero on failure.
2077 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2079 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2080 /* Get at most <Max_retries> more freeDB records once me_pghead
2081 * has enough pages. If not enough, use new pages from the map.
2082 * If <Paranoid> and mc is updating the freeDB, only get new
2083 * records if me_pghead is empty. Then the freelist cannot play
2084 * catch-up with itself by growing while trying to save it.
2086 enum { Paranoid = 1, Max_retries = 500 };
2088 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2090 int rc, retry = num * 60;
2091 MDB_txn *txn = mc->mc_txn;
2092 MDB_env *env = txn->mt_env;
2093 pgno_t pgno, *mop = env->me_pghead;
2094 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2096 txnid_t oldest = 0, last;
2101 /* If there are any loose pages, just use them */
2102 if (num == 1 && txn->mt_loose_pgs) {
2103 np = txn->mt_loose_pgs;
2104 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2105 txn->mt_loose_count--;
2106 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2114 /* If our dirty list is already full, we can't do anything */
2115 if (txn->mt_dirty_room == 0) {
2120 for (op = MDB_FIRST;; op = MDB_NEXT) {
2125 /* Seek a big enough contiguous page range. Prefer
2126 * pages at the tail, just truncating the list.
2132 if (mop[i-n2] == pgno+n2)
2139 if (op == MDB_FIRST) { /* 1st iteration */
2140 /* Prepare to fetch more and coalesce */
2141 last = env->me_pglast;
2142 oldest = env->me_pgoldest;
2143 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2146 key.mv_data = &last; /* will look up last+1 */
2147 key.mv_size = sizeof(last);
2149 if (Paranoid && mc->mc_dbi == FREE_DBI)
2152 if (Paranoid && retry < 0 && mop_len)
2156 /* Do not fetch more if the record will be too recent */
2157 if (oldest <= last) {
2159 oldest = mdb_find_oldest(txn);
2160 env->me_pgoldest = oldest;
2166 rc = mdb_cursor_get(&m2, &key, NULL, op);
2168 if (rc == MDB_NOTFOUND)
2172 last = *(txnid_t*)key.mv_data;
2173 if (oldest <= last) {
2175 oldest = mdb_find_oldest(txn);
2176 env->me_pgoldest = oldest;
2182 np = m2.mc_pg[m2.mc_top];
2183 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2184 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2187 idl = (MDB_ID *) data.mv_data;
2190 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2195 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2197 mop = env->me_pghead;
2199 env->me_pglast = last;
2201 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2202 last, txn->mt_dbs[FREE_DBI].md_root, i));
2204 DPRINTF(("IDL %"Z"u", idl[j]));
2206 /* Merge in descending sorted order */
2207 mdb_midl_xmerge(mop, idl);
2211 /* Use new pages from the map when nothing suitable in the freeDB */
2213 pgno = txn->mt_next_pgno;
2214 if (pgno + num >= env->me_maxpg) {
2215 DPUTS("DB size maxed out");
2221 if (env->me_flags & MDB_WRITEMAP) {
2222 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2224 if (!(np = mdb_page_malloc(txn, num))) {
2230 mop[0] = mop_len -= num;
2231 /* Move any stragglers down */
2232 for (j = i-num; j < mop_len; )
2233 mop[++j] = mop[++i];
2235 txn->mt_next_pgno = pgno + num;
2238 mdb_page_dirty(txn, np);
2244 txn->mt_flags |= MDB_TXN_ERROR;
2248 /** Copy the used portions of a non-overflow page.
2249 * @param[in] dst page to copy into
2250 * @param[in] src page to copy from
2251 * @param[in] psize size of a page
2254 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2256 enum { Align = sizeof(pgno_t) };
2257 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2259 /* If page isn't full, just copy the used portion. Adjust
2260 * alignment so memcpy may copy words instead of bytes.
2262 if ((unused &= -Align) && !IS_LEAF2(src)) {
2263 upper = (upper + PAGEBASE) & -Align;
2264 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2265 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2268 memcpy(dst, src, psize - unused);
2272 /** Pull a page off the txn's spill list, if present.
2273 * If a page being referenced was spilled to disk in this txn, bring
2274 * it back and make it dirty/writable again.
2275 * @param[in] txn the transaction handle.
2276 * @param[in] mp the page being referenced. It must not be dirty.
2277 * @param[out] ret the writable page, if any. ret is unchanged if
2278 * mp wasn't spilled.
2281 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2283 MDB_env *env = txn->mt_env;
2286 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2288 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2289 if (!tx2->mt_spill_pgs)
2291 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2292 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2295 if (txn->mt_dirty_room == 0)
2296 return MDB_TXN_FULL;
2297 if (IS_OVERFLOW(mp))
2301 if (env->me_flags & MDB_WRITEMAP) {
2304 np = mdb_page_malloc(txn, num);
2308 memcpy(np, mp, num * env->me_psize);
2310 mdb_page_copy(np, mp, env->me_psize);
2313 /* If in current txn, this page is no longer spilled.
2314 * If it happens to be the last page, truncate the spill list.
2315 * Otherwise mark it as deleted by setting the LSB.
2317 if (x == txn->mt_spill_pgs[0])
2318 txn->mt_spill_pgs[0]--;
2320 txn->mt_spill_pgs[x] |= 1;
2321 } /* otherwise, if belonging to a parent txn, the
2322 * page remains spilled until child commits
2325 mdb_page_dirty(txn, np);
2326 np->mp_flags |= P_DIRTY;
2334 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2335 * @param[in] mc cursor pointing to the page to be touched
2336 * @return 0 on success, non-zero on failure.
2339 mdb_page_touch(MDB_cursor *mc)
2341 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2342 MDB_txn *txn = mc->mc_txn;
2343 MDB_cursor *m2, *m3;
2347 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2348 if (txn->mt_flags & MDB_TXN_SPILLS) {
2350 rc = mdb_page_unspill(txn, mp, &np);
2356 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2357 (rc = mdb_page_alloc(mc, 1, &np)))
2360 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2361 mp->mp_pgno, pgno));
2362 mdb_cassert(mc, mp->mp_pgno != pgno);
2363 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2364 /* Update the parent page, if any, to point to the new page */
2366 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2367 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2368 SETPGNO(node, pgno);
2370 mc->mc_db->md_root = pgno;
2372 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2373 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2375 /* If txn has a parent, make sure the page is in our
2379 unsigned x = mdb_mid2l_search(dl, pgno);
2380 if (x <= dl[0].mid && dl[x].mid == pgno) {
2381 if (mp != dl[x].mptr) { /* bad cursor? */
2382 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2383 txn->mt_flags |= MDB_TXN_ERROR;
2384 return MDB_CORRUPTED;
2389 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2391 np = mdb_page_malloc(txn, 1);
2396 rc = mdb_mid2l_insert(dl, &mid);
2397 mdb_cassert(mc, rc == 0);
2402 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2404 np->mp_flags |= P_DIRTY;
2407 /* Adjust cursors pointing to mp */
2408 mc->mc_pg[mc->mc_top] = np;
2409 m2 = txn->mt_cursors[mc->mc_dbi];
2410 if (mc->mc_flags & C_SUB) {
2411 for (; m2; m2=m2->mc_next) {
2412 m3 = &m2->mc_xcursor->mx_cursor;
2413 if (m3->mc_snum < mc->mc_snum) continue;
2414 if (m3->mc_pg[mc->mc_top] == mp)
2415 m3->mc_pg[mc->mc_top] = np;
2418 for (; m2; m2=m2->mc_next) {
2419 if (m2->mc_snum < mc->mc_snum) continue;
2420 if (m2->mc_pg[mc->mc_top] == mp) {
2421 m2->mc_pg[mc->mc_top] = np;
2422 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2424 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2426 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2427 if (!(leaf->mn_flags & F_SUBDATA))
2428 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2436 txn->mt_flags |= MDB_TXN_ERROR;
2441 mdb_env_sync(MDB_env *env, int force)
2444 if (env->me_flags & MDB_RDONLY)
2446 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2447 if (env->me_flags & MDB_WRITEMAP) {
2448 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2449 ? MS_ASYNC : MS_SYNC;
2450 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2453 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2457 #ifdef BROKEN_FDATASYNC
2458 if (env->me_flags & MDB_FSYNCONLY) {
2459 if (fsync(env->me_fd))
2463 if (MDB_FDATASYNC(env->me_fd))
2470 /** Back up parent txn's cursors, then grab the originals for tracking */
2472 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2474 MDB_cursor *mc, *bk;
2479 for (i = src->mt_numdbs; --i >= 0; ) {
2480 if ((mc = src->mt_cursors[i]) != NULL) {
2481 size = sizeof(MDB_cursor);
2483 size += sizeof(MDB_xcursor);
2484 for (; mc; mc = bk->mc_next) {
2490 mc->mc_db = &dst->mt_dbs[i];
2491 /* Kill pointers into src - and dst to reduce abuse: The
2492 * user may not use mc until dst ends. Otherwise we'd...
2494 mc->mc_txn = NULL; /* ...set this to dst */
2495 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2496 if ((mx = mc->mc_xcursor) != NULL) {
2497 *(MDB_xcursor *)(bk+1) = *mx;
2498 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2500 mc->mc_next = dst->mt_cursors[i];
2501 dst->mt_cursors[i] = mc;
2508 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2509 * @param[in] txn the transaction handle.
2510 * @param[in] merge true to keep changes to parent cursors, false to revert.
2511 * @return 0 on success, non-zero on failure.
2514 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2516 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2520 for (i = txn->mt_numdbs; --i >= 0; ) {
2521 for (mc = cursors[i]; mc; mc = next) {
2523 if ((bk = mc->mc_backup) != NULL) {
2525 /* Commit changes to parent txn */
2526 mc->mc_next = bk->mc_next;
2527 mc->mc_backup = bk->mc_backup;
2528 mc->mc_txn = bk->mc_txn;
2529 mc->mc_db = bk->mc_db;
2530 mc->mc_dbflag = bk->mc_dbflag;
2531 if ((mx = mc->mc_xcursor) != NULL)
2532 mx->mx_cursor.mc_txn = bk->mc_txn;
2534 /* Abort nested txn */
2536 if ((mx = mc->mc_xcursor) != NULL)
2537 *mx = *(MDB_xcursor *)(bk+1);
2541 /* Only malloced cursors are permanently tracked. */
2548 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2554 Pidset = F_SETLK, Pidcheck = F_GETLK
2558 /** Set or check a pid lock. Set returns 0 on success.
2559 * Check returns 0 if the process is certainly dead, nonzero if it may
2560 * be alive (the lock exists or an error happened so we do not know).
2562 * On Windows Pidset is a no-op, we merely check for the existence
2563 * of the process with the given pid. On POSIX we use a single byte
2564 * lock on the lockfile, set at an offset equal to the pid.
2567 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2569 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2572 if (op == Pidcheck) {
2573 h = OpenProcess(env->me_pidquery, FALSE, pid);
2574 /* No documented "no such process" code, but other program use this: */
2576 return ErrCode() != ERROR_INVALID_PARAMETER;
2577 /* A process exists until all handles to it close. Has it exited? */
2578 ret = WaitForSingleObject(h, 0) != 0;
2585 struct flock lock_info;
2586 memset(&lock_info, 0, sizeof(lock_info));
2587 lock_info.l_type = F_WRLCK;
2588 lock_info.l_whence = SEEK_SET;
2589 lock_info.l_start = pid;
2590 lock_info.l_len = 1;
2591 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2592 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2594 } else if ((rc = ErrCode()) == EINTR) {
2602 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2603 * @param[in] txn the transaction handle to initialize
2604 * @return 0 on success, non-zero on failure.
2607 mdb_txn_renew0(MDB_txn *txn)
2609 MDB_env *env = txn->mt_env;
2610 MDB_txninfo *ti = env->me_txns;
2612 unsigned int i, nr, flags = txn->mt_flags;
2614 int rc, new_notls = 0;
2616 if ((flags &= MDB_TXN_RDONLY) != 0) {
2618 meta = mdb_env_pick_meta(env);
2619 txn->mt_txnid = meta->mm_txnid;
2620 txn->mt_u.reader = NULL;
2622 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2623 pthread_getspecific(env->me_txkey);
2625 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2626 return MDB_BAD_RSLOT;
2628 MDB_PID_T pid = env->me_pid;
2629 MDB_THR_T tid = pthread_self();
2630 mdb_mutexref_t rmutex = env->me_rmutex;
2632 if (!env->me_live_reader) {
2633 rc = mdb_reader_pid(env, Pidset, pid);
2636 env->me_live_reader = 1;
2639 if (LOCK_MUTEX(rc, env, rmutex))
2641 nr = ti->mti_numreaders;
2642 for (i=0; i<nr; i++)
2643 if (ti->mti_readers[i].mr_pid == 0)
2645 if (i == env->me_maxreaders) {
2646 UNLOCK_MUTEX(rmutex);
2647 return MDB_READERS_FULL;
2649 r = &ti->mti_readers[i];
2650 /* Claim the reader slot, carefully since other code
2651 * uses the reader table un-mutexed: First reset the
2652 * slot, next publish it in mti_numreaders. After
2653 * that, it is safe for mdb_env_close() to touch it.
2654 * When it will be closed, we can finally claim it.
2657 r->mr_txnid = (txnid_t)-1;
2660 ti->mti_numreaders = ++nr;
2661 env->me_close_readers = nr;
2663 UNLOCK_MUTEX(rmutex);
2665 new_notls = (env->me_flags & MDB_NOTLS);
2666 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2671 do /* LY: Retry on a race, ITS#7970. */
2672 r->mr_txnid = ti->mti_txnid;
2673 while(r->mr_txnid != ti->mti_txnid);
2674 txn->mt_txnid = r->mr_txnid;
2675 txn->mt_u.reader = r;
2676 meta = env->me_metas[txn->mt_txnid & 1];
2680 /* Not yet touching txn == env->me_txn0, it may be active */
2682 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2684 txn->mt_txnid = ti->mti_txnid;
2685 meta = env->me_metas[txn->mt_txnid & 1];
2687 meta = mdb_env_pick_meta(env);
2688 txn->mt_txnid = meta->mm_txnid;
2692 if (txn->mt_txnid == mdb_debug_start)
2695 txn->mt_child = NULL;
2696 txn->mt_loose_pgs = NULL;
2697 txn->mt_loose_count = 0;
2698 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2699 txn->mt_u.dirty_list = env->me_dirty_list;
2700 txn->mt_u.dirty_list[0].mid = 0;
2701 txn->mt_free_pgs = env->me_free_pgs;
2702 txn->mt_free_pgs[0] = 0;
2703 txn->mt_spill_pgs = NULL;
2705 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2708 /* Copy the DB info and flags */
2709 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2711 /* Moved to here to avoid a data race in read TXNs */
2712 txn->mt_next_pgno = meta->mm_last_pg+1;
2714 txn->mt_flags = flags;
2717 txn->mt_numdbs = env->me_numdbs;
2718 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2719 x = env->me_dbflags[i];
2720 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2721 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2723 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2724 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2726 if (env->me_flags & MDB_FATAL_ERROR) {
2727 DPUTS("environment had fatal error, must shutdown!");
2729 } else if (env->me_maxpg < txn->mt_next_pgno) {
2730 rc = MDB_MAP_RESIZED;
2734 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2739 mdb_txn_renew(MDB_txn *txn)
2743 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2746 rc = mdb_txn_renew0(txn);
2747 if (rc == MDB_SUCCESS) {
2748 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2749 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2750 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2756 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2760 int rc, size, tsize;
2762 flags &= MDB_TXN_BEGIN_FLAGS;
2763 flags |= env->me_flags & MDB_WRITEMAP;
2765 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2769 /* Nested transactions: Max 1 child, write txns only, no writemap */
2770 flags |= parent->mt_flags;
2771 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2772 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2774 /* Child txns save MDB_pgstate and use own copy of cursors */
2775 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2776 size += tsize = sizeof(MDB_ntxn);
2777 } else if (flags & MDB_RDONLY) {
2778 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2779 size += tsize = sizeof(MDB_txn);
2781 /* Reuse preallocated write txn. However, do not touch it until
2782 * mdb_txn_renew0() succeeds, since it currently may be active.
2787 if ((txn = calloc(1, size)) == NULL) {
2788 DPRINTF(("calloc: %s", strerror(errno)));
2791 txn->mt_dbxs = env->me_dbxs; /* static */
2792 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2793 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2794 txn->mt_flags = flags;
2799 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2800 txn->mt_dbiseqs = parent->mt_dbiseqs;
2801 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2802 if (!txn->mt_u.dirty_list ||
2803 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2805 free(txn->mt_u.dirty_list);
2809 txn->mt_txnid = parent->mt_txnid;
2810 txn->mt_dirty_room = parent->mt_dirty_room;
2811 txn->mt_u.dirty_list[0].mid = 0;
2812 txn->mt_spill_pgs = NULL;
2813 txn->mt_next_pgno = parent->mt_next_pgno;
2814 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2815 parent->mt_child = txn;
2816 txn->mt_parent = parent;
2817 txn->mt_numdbs = parent->mt_numdbs;
2818 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2819 /* Copy parent's mt_dbflags, but clear DB_NEW */
2820 for (i=0; i<txn->mt_numdbs; i++)
2821 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2823 ntxn = (MDB_ntxn *)txn;
2824 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2825 if (env->me_pghead) {
2826 size = MDB_IDL_SIZEOF(env->me_pghead);
2827 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2829 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2834 rc = mdb_cursor_shadow(parent, txn);
2836 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2837 } else { /* MDB_RDONLY */
2838 txn->mt_dbiseqs = env->me_dbiseqs;
2840 rc = mdb_txn_renew0(txn);
2843 if (txn != env->me_txn0)
2846 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2848 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2849 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2850 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2857 mdb_txn_env(MDB_txn *txn)
2859 if(!txn) return NULL;
2864 mdb_txn_id(MDB_txn *txn)
2867 return txn->mt_txnid;
2870 /** Export or close DBI handles opened in this txn. */
2872 mdb_dbis_update(MDB_txn *txn, int keep)
2875 MDB_dbi n = txn->mt_numdbs;
2876 MDB_env *env = txn->mt_env;
2877 unsigned char *tdbflags = txn->mt_dbflags;
2879 for (i = n; --i >= CORE_DBS;) {
2880 if (tdbflags[i] & DB_NEW) {
2882 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2884 char *ptr = env->me_dbxs[i].md_name.mv_data;
2886 env->me_dbxs[i].md_name.mv_data = NULL;
2887 env->me_dbxs[i].md_name.mv_size = 0;
2888 env->me_dbflags[i] = 0;
2889 env->me_dbiseqs[i]++;
2895 if (keep && env->me_numdbs < n)
2899 /** End a transaction, except successful commit of a nested transaction.
2900 * May be called twice for readonly txns: First reset it, then abort.
2901 * @param[in] txn the transaction handle to end
2902 * @param[in] mode why and how to end the transaction
2905 mdb_txn_end(MDB_txn *txn, unsigned mode)
2907 MDB_env *env = txn->mt_env;
2909 static const char *const names[] = MDB_END_NAMES;
2912 /* Export or close DBI handles opened in this txn */
2913 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2915 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2916 names[mode & MDB_END_OPMASK],
2917 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2918 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2920 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2921 if (txn->mt_u.reader) {
2922 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2923 if (!(env->me_flags & MDB_NOTLS)) {
2924 txn->mt_u.reader = NULL; /* txn does not own reader */
2925 } else if (mode & MDB_END_SLOT) {
2926 txn->mt_u.reader->mr_pid = 0;
2927 txn->mt_u.reader = NULL;
2928 } /* else txn owns the slot until it does MDB_END_SLOT */
2930 txn->mt_numdbs = 0; /* prevent further DBI activity */
2931 txn->mt_flags |= MDB_TXN_FINISHED;
2933 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2934 pgno_t *pghead = env->me_pghead;
2936 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2937 mdb_cursors_close(txn, 0);
2938 if (!(env->me_flags & MDB_WRITEMAP)) {
2939 mdb_dlist_free(txn);
2943 txn->mt_flags = MDB_TXN_FINISHED;
2945 if (!txn->mt_parent) {
2946 mdb_midl_shrink(&txn->mt_free_pgs);
2947 env->me_free_pgs = txn->mt_free_pgs;
2949 env->me_pghead = NULL;
2953 mode = 0; /* txn == env->me_txn0, do not free() it */
2955 /* The writer mutex was locked in mdb_txn_begin. */
2957 UNLOCK_MUTEX(env->me_wmutex);
2959 txn->mt_parent->mt_child = NULL;
2960 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2961 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2962 mdb_midl_free(txn->mt_free_pgs);
2963 mdb_midl_free(txn->mt_spill_pgs);
2964 free(txn->mt_u.dirty_list);
2967 mdb_midl_free(pghead);
2970 if (mode & MDB_END_FREE)
2975 mdb_txn_reset(MDB_txn *txn)
2980 /* This call is only valid for read-only txns */
2981 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2984 mdb_txn_end(txn, MDB_END_RESET);
2988 mdb_txn_abort(MDB_txn *txn)
2994 mdb_txn_abort(txn->mt_child);
2996 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
2999 /** Save the freelist as of this transaction to the freeDB.
3000 * This changes the freelist. Keep trying until it stabilizes.
3003 mdb_freelist_save(MDB_txn *txn)
3005 /* env->me_pghead[] can grow and shrink during this call.
3006 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3007 * Page numbers cannot disappear from txn->mt_free_pgs[].
3010 MDB_env *env = txn->mt_env;
3011 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3012 txnid_t pglast = 0, head_id = 0;
3013 pgno_t freecnt = 0, *free_pgs, *mop;
3014 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3016 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3018 if (env->me_pghead) {
3019 /* Make sure first page of freeDB is touched and on freelist */
3020 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3021 if (rc && rc != MDB_NOTFOUND)
3025 if (!env->me_pghead && txn->mt_loose_pgs) {
3026 /* Put loose page numbers in mt_free_pgs, since
3027 * we may be unable to return them to me_pghead.
3029 MDB_page *mp = txn->mt_loose_pgs;
3030 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3032 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3033 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3034 txn->mt_loose_pgs = NULL;
3035 txn->mt_loose_count = 0;
3038 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3039 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3040 ? SSIZE_MAX : maxfree_1pg;
3043 /* Come back here after each Put() in case freelist changed */
3048 /* If using records from freeDB which we have not yet
3049 * deleted, delete them and any we reserved for me_pghead.
3051 while (pglast < env->me_pglast) {
3052 rc = mdb_cursor_first(&mc, &key, NULL);
3055 pglast = head_id = *(txnid_t *)key.mv_data;
3056 total_room = head_room = 0;
3057 mdb_tassert(txn, pglast <= env->me_pglast);
3058 rc = mdb_cursor_del(&mc, 0);
3063 /* Save the IDL of pages freed by this txn, to a single record */
3064 if (freecnt < txn->mt_free_pgs[0]) {
3066 /* Make sure last page of freeDB is touched and on freelist */
3067 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3068 if (rc && rc != MDB_NOTFOUND)
3071 free_pgs = txn->mt_free_pgs;
3072 /* Write to last page of freeDB */
3073 key.mv_size = sizeof(txn->mt_txnid);
3074 key.mv_data = &txn->mt_txnid;
3076 freecnt = free_pgs[0];
3077 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3078 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3081 /* Retry if mt_free_pgs[] grew during the Put() */
3082 free_pgs = txn->mt_free_pgs;
3083 } while (freecnt < free_pgs[0]);
3084 mdb_midl_sort(free_pgs);
3085 memcpy(data.mv_data, free_pgs, data.mv_size);
3088 unsigned int i = free_pgs[0];
3089 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3090 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3092 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3098 mop = env->me_pghead;
3099 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3101 /* Reserve records for me_pghead[]. Split it if multi-page,
3102 * to avoid searching freeDB for a page range. Use keys in
3103 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3105 if (total_room >= mop_len) {
3106 if (total_room == mop_len || --more < 0)
3108 } else if (head_room >= maxfree_1pg && head_id > 1) {
3109 /* Keep current record (overflow page), add a new one */
3113 /* (Re)write {key = head_id, IDL length = head_room} */
3114 total_room -= head_room;
3115 head_room = mop_len - total_room;
3116 if (head_room > maxfree_1pg && head_id > 1) {
3117 /* Overflow multi-page for part of me_pghead */
3118 head_room /= head_id; /* amortize page sizes */
3119 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3120 } else if (head_room < 0) {
3121 /* Rare case, not bothering to delete this record */
3124 key.mv_size = sizeof(head_id);
3125 key.mv_data = &head_id;
3126 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3127 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3130 /* IDL is initially empty, zero out at least the length */
3131 pgs = (pgno_t *)data.mv_data;
3132 j = head_room > clean_limit ? head_room : 0;
3136 total_room += head_room;
3139 /* Return loose page numbers to me_pghead, though usually none are
3140 * left at this point. The pages themselves remain in dirty_list.
3142 if (txn->mt_loose_pgs) {
3143 MDB_page *mp = txn->mt_loose_pgs;
3144 unsigned count = txn->mt_loose_count;
3146 /* Room for loose pages + temp IDL with same */
3147 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3149 mop = env->me_pghead;
3150 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3151 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3152 loose[ ++count ] = mp->mp_pgno;
3154 mdb_midl_sort(loose);
3155 mdb_midl_xmerge(mop, loose);
3156 txn->mt_loose_pgs = NULL;
3157 txn->mt_loose_count = 0;
3161 /* Fill in the reserved me_pghead records */
3167 rc = mdb_cursor_first(&mc, &key, &data);
3168 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3169 txnid_t id = *(txnid_t *)key.mv_data;
3170 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3173 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3175 if (len > mop_len) {
3177 data.mv_size = (len + 1) * sizeof(MDB_ID);
3179 data.mv_data = mop -= len;
3182 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3184 if (rc || !(mop_len -= len))
3191 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3192 * @param[in] txn the transaction that's being committed
3193 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3194 * @return 0 on success, non-zero on failure.
3197 mdb_page_flush(MDB_txn *txn, int keep)
3199 MDB_env *env = txn->mt_env;
3200 MDB_ID2L dl = txn->mt_u.dirty_list;
3201 unsigned psize = env->me_psize, j;
3202 int i, pagecount = dl[0].mid, rc;
3203 size_t size = 0, pos = 0;
3205 MDB_page *dp = NULL;
3209 struct iovec iov[MDB_COMMIT_PAGES];
3210 ssize_t wpos = 0, wsize = 0, wres;
3211 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3217 if (env->me_flags & MDB_WRITEMAP) {
3218 /* Clear dirty flags */
3219 while (++i <= pagecount) {
3221 /* Don't flush this page yet */
3222 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3223 dp->mp_flags &= ~P_KEEP;
3227 dp->mp_flags &= ~P_DIRTY;
3232 /* Write the pages */
3234 if (++i <= pagecount) {
3236 /* Don't flush this page yet */
3237 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3238 dp->mp_flags &= ~P_KEEP;
3243 /* clear dirty flag */
3244 dp->mp_flags &= ~P_DIRTY;
3247 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3252 /* Windows actually supports scatter/gather I/O, but only on
3253 * unbuffered file handles. Since we're relying on the OS page
3254 * cache for all our data, that's self-defeating. So we just
3255 * write pages one at a time. We use the ov structure to set
3256 * the write offset, to at least save the overhead of a Seek
3259 DPRINTF(("committing page %"Z"u", pgno));
3260 memset(&ov, 0, sizeof(ov));
3261 ov.Offset = pos & 0xffffffff;
3262 ov.OffsetHigh = pos >> 16 >> 16;
3263 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3265 DPRINTF(("WriteFile: %d", rc));
3269 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3270 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3273 /* Write previous page(s) */
3274 #ifdef MDB_USE_PWRITEV
3275 wres = pwritev(env->me_fd, iov, n, wpos);
3278 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3281 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3285 DPRINTF(("lseek: %s", strerror(rc)));
3288 wres = writev(env->me_fd, iov, n);
3291 if (wres != wsize) {
3296 DPRINTF(("Write error: %s", strerror(rc)));
3298 rc = EIO; /* TODO: Use which error code? */
3299 DPUTS("short write, filesystem full?");
3310 DPRINTF(("committing page %"Z"u", pgno));
3311 next_pos = pos + size;
3312 iov[n].iov_len = size;
3313 iov[n].iov_base = (char *)dp;
3319 /* MIPS has cache coherency issues, this is a no-op everywhere else
3320 * Note: for any size >= on-chip cache size, entire on-chip cache is
3323 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3325 for (i = keep; ++i <= pagecount; ) {
3327 /* This is a page we skipped above */
3330 dl[j].mid = dp->mp_pgno;
3333 mdb_dpage_free(env, dp);
3338 txn->mt_dirty_room += i - j;
3344 mdb_txn_commit(MDB_txn *txn)
3347 unsigned int i, end_mode;
3353 /* mdb_txn_end() mode for a commit which writes nothing */
3354 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3356 if (txn->mt_child) {
3357 rc = mdb_txn_commit(txn->mt_child);
3364 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3368 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3369 DPUTS("txn has failed/finished, can't commit");
3371 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3376 if (txn->mt_parent) {
3377 MDB_txn *parent = txn->mt_parent;
3381 unsigned x, y, len, ps_len;
3383 /* Append our free list to parent's */
3384 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3387 mdb_midl_free(txn->mt_free_pgs);
3388 /* Failures after this must either undo the changes
3389 * to the parent or set MDB_TXN_ERROR in the parent.
3392 parent->mt_next_pgno = txn->mt_next_pgno;
3393 parent->mt_flags = txn->mt_flags;
3395 /* Merge our cursors into parent's and close them */
3396 mdb_cursors_close(txn, 1);
3398 /* Update parent's DB table. */
3399 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3400 parent->mt_numdbs = txn->mt_numdbs;
3401 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3402 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3403 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3404 /* preserve parent's DB_NEW status */
3405 x = parent->mt_dbflags[i] & DB_NEW;
3406 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3409 dst = parent->mt_u.dirty_list;
3410 src = txn->mt_u.dirty_list;
3411 /* Remove anything in our dirty list from parent's spill list */
3412 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3414 pspill[0] = (pgno_t)-1;
3415 /* Mark our dirty pages as deleted in parent spill list */
3416 for (i=0, len=src[0].mid; ++i <= len; ) {
3417 MDB_ID pn = src[i].mid << 1;
3418 while (pn > pspill[x])
3420 if (pn == pspill[x]) {
3425 /* Squash deleted pagenums if we deleted any */
3426 for (x=y; ++x <= ps_len; )
3427 if (!(pspill[x] & 1))
3428 pspill[++y] = pspill[x];
3432 /* Find len = length of merging our dirty list with parent's */
3434 dst[0].mid = 0; /* simplify loops */
3435 if (parent->mt_parent) {
3436 len = x + src[0].mid;
3437 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3438 for (i = x; y && i; y--) {
3439 pgno_t yp = src[y].mid;
3440 while (yp < dst[i].mid)
3442 if (yp == dst[i].mid) {
3447 } else { /* Simplify the above for single-ancestor case */
3448 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3450 /* Merge our dirty list with parent's */
3452 for (i = len; y; dst[i--] = src[y--]) {
3453 pgno_t yp = src[y].mid;
3454 while (yp < dst[x].mid)
3455 dst[i--] = dst[x--];
3456 if (yp == dst[x].mid)
3457 free(dst[x--].mptr);
3459 mdb_tassert(txn, i == x);
3461 free(txn->mt_u.dirty_list);
3462 parent->mt_dirty_room = txn->mt_dirty_room;
3463 if (txn->mt_spill_pgs) {
3464 if (parent->mt_spill_pgs) {
3465 /* TODO: Prevent failure here, so parent does not fail */
3466 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3468 parent->mt_flags |= MDB_TXN_ERROR;
3469 mdb_midl_free(txn->mt_spill_pgs);
3470 mdb_midl_sort(parent->mt_spill_pgs);
3472 parent->mt_spill_pgs = txn->mt_spill_pgs;
3476 /* Append our loose page list to parent's */
3477 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3479 *lp = txn->mt_loose_pgs;
3480 parent->mt_loose_count += txn->mt_loose_count;
3482 parent->mt_child = NULL;
3483 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3488 if (txn != env->me_txn) {
3489 DPUTS("attempt to commit unknown transaction");
3494 mdb_cursors_close(txn, 0);
3496 if (!txn->mt_u.dirty_list[0].mid &&
3497 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3500 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3501 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3503 /* Update DB root pointers */
3504 if (txn->mt_numdbs > CORE_DBS) {
3508 data.mv_size = sizeof(MDB_db);
3510 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3511 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3512 if (txn->mt_dbflags[i] & DB_DIRTY) {
3513 if (TXN_DBI_CHANGED(txn, i)) {
3517 data.mv_data = &txn->mt_dbs[i];
3518 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3526 rc = mdb_freelist_save(txn);
3530 mdb_midl_free(env->me_pghead);
3531 env->me_pghead = NULL;
3532 mdb_midl_shrink(&txn->mt_free_pgs);
3538 if ((rc = mdb_page_flush(txn, 0)) ||
3539 (rc = mdb_env_sync(env, 0)) ||
3540 (rc = mdb_env_write_meta(txn)))
3542 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3545 mdb_txn_end(txn, end_mode);
3553 /** Read the environment parameters of a DB environment before
3554 * mapping it into memory.
3555 * @param[in] env the environment handle
3556 * @param[out] meta address of where to store the meta information
3557 * @return 0 on success, non-zero on failure.
3560 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3566 enum { Size = sizeof(pbuf) };
3568 /* We don't know the page size yet, so use a minimum value.
3569 * Read both meta pages so we can use the latest one.
3572 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3576 memset(&ov, 0, sizeof(ov));
3578 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3579 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3582 rc = pread(env->me_fd, &pbuf, Size, off);
3585 if (rc == 0 && off == 0)
3587 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3588 DPRINTF(("read: %s", mdb_strerror(rc)));
3592 p = (MDB_page *)&pbuf;
3594 if (!F_ISSET(p->mp_flags, P_META)) {
3595 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3600 if (m->mm_magic != MDB_MAGIC) {
3601 DPUTS("meta has invalid magic");
3605 if (m->mm_version != MDB_DATA_VERSION) {
3606 DPRINTF(("database is version %u, expected version %u",
3607 m->mm_version, MDB_DATA_VERSION));
3608 return MDB_VERSION_MISMATCH;
3611 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3617 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3619 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3621 meta->mm_magic = MDB_MAGIC;
3622 meta->mm_version = MDB_DATA_VERSION;
3623 meta->mm_mapsize = env->me_mapsize;
3624 meta->mm_psize = env->me_psize;
3625 meta->mm_last_pg = NUM_METAS-1;
3626 meta->mm_flags = env->me_flags & 0xffff;
3627 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3628 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3629 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3632 /** Write the environment parameters of a freshly created DB environment.
3633 * @param[in] env the environment handle
3634 * @param[in] meta the #MDB_meta to write
3635 * @return 0 on success, non-zero on failure.
3638 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3646 memset(&ov, 0, sizeof(ov));
3647 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3649 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3652 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3653 len = pwrite(fd, ptr, size, pos); \
3654 if (len == -1 && ErrCode() == EINTR) continue; \
3655 rc = (len >= 0); break; } while(1)
3658 DPUTS("writing new meta page");
3660 psize = env->me_psize;
3662 p = calloc(NUM_METAS, psize);
3667 p->mp_flags = P_META;
3668 *(MDB_meta *)METADATA(p) = *meta;
3670 q = (MDB_page *)((char *)p + psize);
3672 q->mp_flags = P_META;
3673 *(MDB_meta *)METADATA(q) = *meta;
3675 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3678 else if ((unsigned) len == psize * NUM_METAS)
3686 /** Update the environment info to commit a transaction.
3687 * @param[in] txn the transaction that's being committed
3688 * @return 0 on success, non-zero on failure.
3691 mdb_env_write_meta(MDB_txn *txn)
3694 MDB_meta meta, metab, *mp;
3698 int rc, len, toggle;
3707 toggle = txn->mt_txnid & 1;
3708 DPRINTF(("writing meta page %d for root page %"Z"u",
3709 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3712 flags = env->me_flags;
3713 mp = env->me_metas[toggle];
3714 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3715 /* Persist any increases of mapsize config */
3716 if (mapsize < env->me_mapsize)
3717 mapsize = env->me_mapsize;
3719 if (flags & MDB_WRITEMAP) {
3720 mp->mm_mapsize = mapsize;
3721 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3722 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3723 mp->mm_last_pg = txn->mt_next_pgno - 1;
3724 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3725 !(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;
3733 ptr = (char *)mp - PAGEHDRSZ;
3734 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3735 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3739 if (MDB_MSYNC(ptr, meta_size, rc)) {
3746 metab.mm_txnid = mp->mm_txnid;
3747 metab.mm_last_pg = mp->mm_last_pg;
3749 meta.mm_mapsize = mapsize;
3750 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3751 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3752 meta.mm_last_pg = txn->mt_next_pgno - 1;
3753 meta.mm_txnid = txn->mt_txnid;
3755 off = offsetof(MDB_meta, mm_mapsize);
3756 ptr = (char *)&meta + off;
3757 len = sizeof(MDB_meta) - off;
3758 off += (char *)mp - env->me_map;
3760 /* Write to the SYNC fd */
3761 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3764 memset(&ov, 0, sizeof(ov));
3766 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3771 rc = pwrite(mfd, ptr, len, off);
3774 rc = rc < 0 ? ErrCode() : EIO;
3779 DPUTS("write failed, disk error?");
3780 /* On a failure, the pagecache still contains the new data.
3781 * Write some old data back, to prevent it from being used.
3782 * Use the non-SYNC fd; we know it will fail anyway.
3784 meta.mm_last_pg = metab.mm_last_pg;
3785 meta.mm_txnid = metab.mm_txnid;
3787 memset(&ov, 0, sizeof(ov));
3789 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3791 r2 = pwrite(env->me_fd, ptr, len, off);
3792 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3795 env->me_flags |= MDB_FATAL_ERROR;
3798 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3799 CACHEFLUSH(env->me_map + off, len, DCACHE);
3801 /* Memory ordering issues are irrelevant; since the entire writer
3802 * is wrapped by wmutex, all of these changes will become visible
3803 * after the wmutex is unlocked. Since the DB is multi-version,
3804 * readers will get consistent data regardless of how fresh or
3805 * how stale their view of these values is.
3808 env->me_txns->mti_txnid = txn->mt_txnid;
3813 /** Check both meta pages to see which one is newer.
3814 * @param[in] env the environment handle
3815 * @return newest #MDB_meta.
3818 mdb_env_pick_meta(const MDB_env *env)
3820 MDB_meta *const *metas = env->me_metas;
3821 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3825 mdb_env_create(MDB_env **env)
3829 e = calloc(1, sizeof(MDB_env));
3833 e->me_maxreaders = DEFAULT_READERS;
3834 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3835 e->me_fd = INVALID_HANDLE_VALUE;
3836 e->me_lfd = INVALID_HANDLE_VALUE;
3837 e->me_mfd = INVALID_HANDLE_VALUE;
3838 #ifdef MDB_USE_POSIX_SEM
3839 e->me_rmutex = SEM_FAILED;
3840 e->me_wmutex = SEM_FAILED;
3842 e->me_pid = getpid();
3843 GET_PAGESIZE(e->me_os_psize);
3844 VGMEMP_CREATE(e,0,0);
3850 mdb_env_map(MDB_env *env, void *addr)
3853 unsigned int flags = env->me_flags;
3857 LONG sizelo, sizehi;
3860 if (flags & MDB_RDONLY) {
3861 /* Don't set explicit map size, use whatever exists */
3866 msize = env->me_mapsize;
3867 sizelo = msize & 0xffffffff;
3868 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3870 /* Windows won't create mappings for zero length files.
3871 * and won't map more than the file size.
3872 * Just set the maxsize right now.
3874 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3875 || !SetEndOfFile(env->me_fd)
3876 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3880 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3881 PAGE_READWRITE : PAGE_READONLY,
3882 sizehi, sizelo, NULL);
3885 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3886 FILE_MAP_WRITE : FILE_MAP_READ,
3888 rc = env->me_map ? 0 : ErrCode();
3893 int prot = PROT_READ;
3894 if (flags & MDB_WRITEMAP) {
3896 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3899 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3901 if (env->me_map == MAP_FAILED) {
3906 if (flags & MDB_NORDAHEAD) {
3907 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3909 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3911 #ifdef POSIX_MADV_RANDOM
3912 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3913 #endif /* POSIX_MADV_RANDOM */
3914 #endif /* MADV_RANDOM */
3918 /* Can happen because the address argument to mmap() is just a
3919 * hint. mmap() can pick another, e.g. if the range is in use.
3920 * The MAP_FIXED flag would prevent that, but then mmap could
3921 * instead unmap existing pages to make room for the new map.
3923 if (addr && env->me_map != addr)
3924 return EBUSY; /* TODO: Make a new MDB_* error code? */
3926 p = (MDB_page *)env->me_map;
3927 env->me_metas[0] = METADATA(p);
3928 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3934 mdb_env_set_mapsize(MDB_env *env, size_t size)
3936 /* If env is already open, caller is responsible for making
3937 * sure there are no active txns.
3945 meta = mdb_env_pick_meta(env);
3947 size = meta->mm_mapsize;
3949 /* Silently round up to minimum if the size is too small */
3950 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3954 munmap(env->me_map, env->me_mapsize);
3955 env->me_mapsize = size;
3956 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3957 rc = mdb_env_map(env, old);
3961 env->me_mapsize = size;
3963 env->me_maxpg = env->me_mapsize / env->me_psize;
3968 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3972 env->me_maxdbs = dbs + CORE_DBS;
3977 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3979 if (env->me_map || readers < 1)
3981 env->me_maxreaders = readers;
3986 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3988 if (!env || !readers)
3990 *readers = env->me_maxreaders;
3995 mdb_fsize(HANDLE fd, size_t *size)
3998 LARGE_INTEGER fsize;
4000 if (!GetFileSizeEx(fd, &fsize))
4003 *size = fsize.QuadPart;
4015 #ifdef BROKEN_FDATASYNC
4016 #include <sys/utsname.h>
4017 #include <sys/vfs.h>
4020 /** Further setup required for opening an LMDB environment
4023 mdb_env_open2(MDB_env *env)
4025 unsigned int flags = env->me_flags;
4026 int i, newenv = 0, rc;
4030 /* See if we should use QueryLimited */
4032 if ((rc & 0xff) > 5)
4033 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4035 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4038 #ifdef BROKEN_FDATASYNC
4039 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4040 * https://lkml.org/lkml/2012/9/3/83
4041 * Kernels after 3.6-rc6 are known good.
4042 * https://lkml.org/lkml/2012/9/10/556
4043 * See if the DB is on ext3/ext4, then check for new enough kernel
4044 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4049 fstatfs(env->me_fd, &st);
4050 while (st.f_type == 0xEF53) {
4054 if (uts.release[0] < '3') {
4055 if (!strncmp(uts.release, "2.6.32.", 7)) {
4056 i = atoi(uts.release+7);
4058 break; /* 2.6.32.60 and newer is OK */
4059 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4060 i = atoi(uts.release+7);
4062 break; /* 2.6.34.15 and newer is OK */
4064 } else if (uts.release[0] == '3') {
4065 i = atoi(uts.release+2);
4067 break; /* 3.6 and newer is OK */
4069 i = atoi(uts.release+4);
4071 break; /* 3.5.4 and newer is OK */
4072 } else if (i == 2) {
4073 i = atoi(uts.release+4);
4075 break; /* 3.2.30 and newer is OK */
4077 } else { /* 4.x and newer is OK */
4080 env->me_flags |= MDB_FSYNCONLY;
4086 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4089 DPUTS("new mdbenv");
4091 env->me_psize = env->me_os_psize;
4092 if (env->me_psize > MAX_PAGESIZE)
4093 env->me_psize = MAX_PAGESIZE;
4094 memset(&meta, 0, sizeof(meta));
4095 mdb_env_init_meta0(env, &meta);
4096 meta.mm_mapsize = DEFAULT_MAPSIZE;
4098 env->me_psize = meta.mm_psize;
4101 /* Was a mapsize configured? */
4102 if (!env->me_mapsize) {
4103 env->me_mapsize = meta.mm_mapsize;
4106 /* Make sure mapsize >= committed data size. Even when using
4107 * mm_mapsize, which could be broken in old files (ITS#7789).
4109 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4110 if (env->me_mapsize < minsize)
4111 env->me_mapsize = minsize;
4113 meta.mm_mapsize = env->me_mapsize;
4115 if (newenv && !(flags & MDB_FIXEDMAP)) {
4116 /* mdb_env_map() may grow the datafile. Write the metapages
4117 * first, so the file will be valid if initialization fails.
4118 * Except with FIXEDMAP, since we do not yet know mm_address.
4119 * We could fill in mm_address later, but then a different
4120 * program might end up doing that - one with a memory layout
4121 * and map address which does not suit the main program.
4123 rc = mdb_env_init_meta(env, &meta);
4129 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4134 if (flags & MDB_FIXEDMAP)
4135 meta.mm_address = env->me_map;
4136 i = mdb_env_init_meta(env, &meta);
4137 if (i != MDB_SUCCESS) {
4142 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4143 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4145 #if !(MDB_MAXKEYSIZE)
4146 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4148 env->me_maxpg = env->me_mapsize / env->me_psize;
4152 MDB_meta *meta = mdb_env_pick_meta(env);
4153 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4155 DPRINTF(("opened database version %u, pagesize %u",
4156 meta->mm_version, env->me_psize));
4157 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4158 DPRINTF(("depth: %u", db->md_depth));
4159 DPRINTF(("entries: %"Z"u", db->md_entries));
4160 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4161 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4162 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4163 DPRINTF(("root: %"Z"u", db->md_root));
4171 /** Release a reader thread's slot in the reader lock table.
4172 * This function is called automatically when a thread exits.
4173 * @param[in] ptr This points to the slot in the reader lock table.
4176 mdb_env_reader_dest(void *ptr)
4178 MDB_reader *reader = ptr;
4184 /** Junk for arranging thread-specific callbacks on Windows. This is
4185 * necessarily platform and compiler-specific. Windows supports up
4186 * to 1088 keys. Let's assume nobody opens more than 64 environments
4187 * in a single process, for now. They can override this if needed.
4189 #ifndef MAX_TLS_KEYS
4190 #define MAX_TLS_KEYS 64
4192 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4193 static int mdb_tls_nkeys;
4195 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4199 case DLL_PROCESS_ATTACH: break;
4200 case DLL_THREAD_ATTACH: break;
4201 case DLL_THREAD_DETACH:
4202 for (i=0; i<mdb_tls_nkeys; i++) {
4203 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4205 mdb_env_reader_dest(r);
4209 case DLL_PROCESS_DETACH: break;
4214 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4216 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4220 /* Force some symbol references.
4221 * _tls_used forces the linker to create the TLS directory if not already done
4222 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4224 #pragma comment(linker, "/INCLUDE:_tls_used")
4225 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4226 #pragma const_seg(".CRT$XLB")
4227 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4228 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4231 #pragma comment(linker, "/INCLUDE:__tls_used")
4232 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4233 #pragma data_seg(".CRT$XLB")
4234 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4236 #endif /* WIN 32/64 */
4237 #endif /* !__GNUC__ */
4240 /** Downgrade the exclusive lock on the region back to shared */
4242 mdb_env_share_locks(MDB_env *env, int *excl)
4245 MDB_meta *meta = mdb_env_pick_meta(env);
4247 env->me_txns->mti_txnid = meta->mm_txnid;
4252 /* First acquire a shared lock. The Unlock will
4253 * then release the existing exclusive lock.
4255 memset(&ov, 0, sizeof(ov));
4256 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4259 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4265 struct flock lock_info;
4266 /* The shared lock replaces the existing lock */
4267 memset((void *)&lock_info, 0, sizeof(lock_info));
4268 lock_info.l_type = F_RDLCK;
4269 lock_info.l_whence = SEEK_SET;
4270 lock_info.l_start = 0;
4271 lock_info.l_len = 1;
4272 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4273 (rc = ErrCode()) == EINTR) ;
4274 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4281 /** Try to get exclusive lock, otherwise shared.
4282 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4285 mdb_env_excl_lock(MDB_env *env, int *excl)
4289 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4293 memset(&ov, 0, sizeof(ov));
4294 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4301 struct flock lock_info;
4302 memset((void *)&lock_info, 0, sizeof(lock_info));
4303 lock_info.l_type = F_WRLCK;
4304 lock_info.l_whence = SEEK_SET;
4305 lock_info.l_start = 0;
4306 lock_info.l_len = 1;
4307 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4308 (rc = ErrCode()) == EINTR) ;
4312 # ifndef MDB_USE_POSIX_MUTEX
4313 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4316 lock_info.l_type = F_RDLCK;
4317 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4318 (rc = ErrCode()) == EINTR) ;
4328 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4330 * @(#) $Revision: 5.1 $
4331 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4332 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4334 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4338 * Please do not copyright this code. This code is in the public domain.
4340 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4341 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4342 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4343 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4344 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4345 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4346 * PERFORMANCE OF THIS SOFTWARE.
4349 * chongo <Landon Curt Noll> /\oo/\
4350 * http://www.isthe.com/chongo/
4352 * Share and Enjoy! :-)
4355 typedef unsigned long long mdb_hash_t;
4356 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4358 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4359 * @param[in] val value to hash
4360 * @param[in] hval initial value for hash
4361 * @return 64 bit hash
4363 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4364 * hval arg on the first call.
4367 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4369 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4370 unsigned char *end = s + val->mv_size;
4372 * FNV-1a hash each octet of the string
4375 /* xor the bottom with the current octet */
4376 hval ^= (mdb_hash_t)*s++;
4378 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4379 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4380 (hval << 7) + (hval << 8) + (hval << 40);
4382 /* return our new hash value */
4386 /** Hash the string and output the encoded hash.
4387 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4388 * very short name limits. We don't care about the encoding being reversible,
4389 * we just want to preserve as many bits of the input as possible in a
4390 * small printable string.
4391 * @param[in] str string to hash
4392 * @param[out] encbuf an array of 11 chars to hold the hash
4394 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4397 mdb_pack85(unsigned long l, char *out)
4401 for (i=0; i<5; i++) {
4402 *out++ = mdb_a85[l % 85];
4408 mdb_hash_enc(MDB_val *val, char *encbuf)
4410 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4412 mdb_pack85(h, encbuf);
4413 mdb_pack85(h>>32, encbuf+5);
4418 /** Open and/or initialize the lock region for the environment.
4419 * @param[in] env The LMDB environment.
4420 * @param[in] lpath The pathname of the file used for the lock region.
4421 * @param[in] mode The Unix permissions for the file, if we create it.
4422 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4423 * @return 0 on success, non-zero on failure.
4426 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4429 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4431 # define MDB_ERRCODE_ROFS EROFS
4432 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4433 # define MDB_CLOEXEC O_CLOEXEC
4436 # define MDB_CLOEXEC 0
4443 env->me_lfd = CreateFileA(lpath, GENERIC_READ|GENERIC_WRITE,
4444 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4445 FILE_ATTRIBUTE_NORMAL, NULL);
4447 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4449 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4451 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4456 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4457 /* Lose record locks when exec*() */
4458 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4459 fcntl(env->me_lfd, F_SETFD, fdflags);
4462 if (!(env->me_flags & MDB_NOTLS)) {
4463 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4466 env->me_flags |= MDB_ENV_TXKEY;
4468 /* Windows TLS callbacks need help finding their TLS info. */
4469 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4473 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4477 /* Try to get exclusive lock. If we succeed, then
4478 * nobody is using the lock region and we should initialize it.
4480 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4483 size = GetFileSize(env->me_lfd, NULL);
4485 size = lseek(env->me_lfd, 0, SEEK_END);
4486 if (size == -1) goto fail_errno;
4488 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4489 if (size < rsize && *excl > 0) {
4491 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4492 || !SetEndOfFile(env->me_lfd))
4495 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4499 size = rsize - sizeof(MDB_txninfo);
4500 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4505 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4507 if (!mh) goto fail_errno;
4508 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4510 if (!env->me_txns) goto fail_errno;
4512 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4514 if (m == MAP_FAILED) goto fail_errno;
4520 BY_HANDLE_FILE_INFORMATION stbuf;
4529 if (!mdb_sec_inited) {
4530 InitializeSecurityDescriptor(&mdb_null_sd,
4531 SECURITY_DESCRIPTOR_REVISION);
4532 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4533 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4534 mdb_all_sa.bInheritHandle = FALSE;
4535 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4538 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4539 idbuf.volume = stbuf.dwVolumeSerialNumber;
4540 idbuf.nhigh = stbuf.nFileIndexHigh;
4541 idbuf.nlow = stbuf.nFileIndexLow;
4542 val.mv_data = &idbuf;
4543 val.mv_size = sizeof(idbuf);
4544 mdb_hash_enc(&val, encbuf);
4545 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4546 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4547 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4548 if (!env->me_rmutex) goto fail_errno;
4549 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4550 if (!env->me_wmutex) goto fail_errno;
4551 #elif defined(MDB_USE_POSIX_SEM)
4560 #if defined(__NetBSD__)
4561 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4563 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4564 idbuf.dev = stbuf.st_dev;
4565 idbuf.ino = stbuf.st_ino;
4566 val.mv_data = &idbuf;
4567 val.mv_size = sizeof(idbuf);
4568 mdb_hash_enc(&val, encbuf);
4569 #ifdef MDB_SHORT_SEMNAMES
4570 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4572 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4573 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4574 /* Clean up after a previous run, if needed: Try to
4575 * remove both semaphores before doing anything else.
4577 sem_unlink(env->me_txns->mti_rmname);
4578 sem_unlink(env->me_txns->mti_wmname);
4579 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4580 O_CREAT|O_EXCL, mode, 1);
4581 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4582 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4583 O_CREAT|O_EXCL, mode, 1);
4584 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4585 #else /* MDB_USE_POSIX_MUTEX: */
4586 pthread_mutexattr_t mattr;
4588 if ((rc = pthread_mutexattr_init(&mattr))
4589 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4590 #ifdef MDB_ROBUST_SUPPORTED
4591 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4593 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4594 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4596 pthread_mutexattr_destroy(&mattr);
4597 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4599 env->me_txns->mti_magic = MDB_MAGIC;
4600 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4601 env->me_txns->mti_txnid = 0;
4602 env->me_txns->mti_numreaders = 0;
4605 if (env->me_txns->mti_magic != MDB_MAGIC) {
4606 DPUTS("lock region has invalid magic");
4610 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4611 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4612 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4613 rc = MDB_VERSION_MISMATCH;
4617 if (rc && rc != EACCES && rc != EAGAIN) {
4621 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4622 if (!env->me_rmutex) goto fail_errno;
4623 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4624 if (!env->me_wmutex) goto fail_errno;
4625 #elif defined(MDB_USE_POSIX_SEM)
4626 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4627 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4628 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4629 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4640 /** The name of the lock file in the DB environment */
4641 #define LOCKNAME "/lock.mdb"
4642 /** The name of the data file in the DB environment */
4643 #define DATANAME "/data.mdb"
4644 /** The suffix of the lock file when no subdir is used */
4645 #define LOCKSUFF "-lock"
4646 /** Only a subset of the @ref mdb_env flags can be changed
4647 * at runtime. Changing other flags requires closing the
4648 * environment and re-opening it with the new flags.
4650 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4651 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4652 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4654 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4655 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4659 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4661 int oflags, rc, len, excl = -1;
4662 char *lpath, *dpath;
4664 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4668 if (flags & MDB_NOSUBDIR) {
4669 rc = len + sizeof(LOCKSUFF) + len + 1;
4671 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4676 if (flags & MDB_NOSUBDIR) {
4677 dpath = lpath + len + sizeof(LOCKSUFF);
4678 sprintf(lpath, "%s" LOCKSUFF, path);
4679 strcpy(dpath, path);
4681 dpath = lpath + len + sizeof(LOCKNAME);
4682 sprintf(lpath, "%s" LOCKNAME, path);
4683 sprintf(dpath, "%s" DATANAME, path);
4687 flags |= env->me_flags;
4688 if (flags & MDB_RDONLY) {
4689 /* silently ignore WRITEMAP when we're only getting read access */
4690 flags &= ~MDB_WRITEMAP;
4692 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4693 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4696 env->me_flags = flags |= MDB_ENV_ACTIVE;
4700 env->me_path = strdup(path);
4701 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4702 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4703 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4704 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4708 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4710 /* For RDONLY, get lockfile after we know datafile exists */
4711 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4712 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4718 if (F_ISSET(flags, MDB_RDONLY)) {
4719 oflags = GENERIC_READ;
4720 len = OPEN_EXISTING;
4722 oflags = GENERIC_READ|GENERIC_WRITE;
4725 mode = FILE_ATTRIBUTE_NORMAL;
4726 env->me_fd = CreateFileA(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4727 NULL, len, mode, NULL);
4729 if (F_ISSET(flags, MDB_RDONLY))
4732 oflags = O_RDWR | O_CREAT;
4734 env->me_fd = open(dpath, oflags, mode);
4736 if (env->me_fd == INVALID_HANDLE_VALUE) {
4741 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4742 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4747 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4748 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4749 env->me_mfd = env->me_fd;
4751 /* Synchronous fd for meta writes. Needed even with
4752 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4755 len = OPEN_EXISTING;
4756 env->me_mfd = CreateFileA(dpath, oflags,
4757 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4758 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4761 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4763 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4768 DPRINTF(("opened dbenv %p", (void *) env));
4770 rc = mdb_env_share_locks(env, &excl);
4774 if (!(flags & MDB_RDONLY)) {
4776 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4777 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4778 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4779 (txn = calloc(1, size)))
4781 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4782 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4783 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4784 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4786 txn->mt_dbxs = env->me_dbxs;
4787 txn->mt_flags = MDB_TXN_FINISHED;
4797 mdb_env_close0(env, excl);
4803 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4805 mdb_env_close0(MDB_env *env, int excl)
4809 if (!(env->me_flags & MDB_ENV_ACTIVE))
4812 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4814 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4815 free(env->me_dbxs[i].md_name.mv_data);
4820 free(env->me_dbiseqs);
4821 free(env->me_dbflags);
4823 free(env->me_dirty_list);
4825 mdb_midl_free(env->me_free_pgs);
4827 if (env->me_flags & MDB_ENV_TXKEY) {
4828 pthread_key_delete(env->me_txkey);
4830 /* Delete our key from the global list */
4831 for (i=0; i<mdb_tls_nkeys; i++)
4832 if (mdb_tls_keys[i] == env->me_txkey) {
4833 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4841 munmap(env->me_map, env->me_mapsize);
4843 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4844 (void) close(env->me_mfd);
4845 if (env->me_fd != INVALID_HANDLE_VALUE)
4846 (void) close(env->me_fd);
4848 MDB_PID_T pid = env->me_pid;
4849 /* Clearing readers is done in this function because
4850 * me_txkey with its destructor must be disabled first.
4852 * We skip the the reader mutex, so we touch only
4853 * data owned by this process (me_close_readers and
4854 * our readers), and clear each reader atomically.
4856 for (i = env->me_close_readers; --i >= 0; )
4857 if (env->me_txns->mti_readers[i].mr_pid == pid)
4858 env->me_txns->mti_readers[i].mr_pid = 0;
4860 if (env->me_rmutex) {
4861 CloseHandle(env->me_rmutex);
4862 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4864 /* Windows automatically destroys the mutexes when
4865 * the last handle closes.
4867 #elif defined(MDB_USE_POSIX_SEM)
4868 if (env->me_rmutex != SEM_FAILED) {
4869 sem_close(env->me_rmutex);
4870 if (env->me_wmutex != SEM_FAILED)
4871 sem_close(env->me_wmutex);
4872 /* If we have the filelock: If we are the
4873 * only remaining user, clean up semaphores.
4876 mdb_env_excl_lock(env, &excl);
4878 sem_unlink(env->me_txns->mti_rmname);
4879 sem_unlink(env->me_txns->mti_wmname);
4883 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4885 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4888 /* Unlock the lockfile. Windows would have unlocked it
4889 * after closing anyway, but not necessarily at once.
4891 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4894 (void) close(env->me_lfd);
4897 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4901 mdb_env_close(MDB_env *env)
4908 VGMEMP_DESTROY(env);
4909 while ((dp = env->me_dpages) != NULL) {
4910 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4911 env->me_dpages = dp->mp_next;
4915 mdb_env_close0(env, 0);
4919 /** Compare two items pointing at aligned size_t's */
4921 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4923 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4924 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4927 /** Compare two items pointing at aligned unsigned int's.
4929 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4930 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4933 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4935 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4936 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4939 /** Compare two items pointing at unsigned ints of unknown alignment.
4940 * Nodes and keys are guaranteed to be 2-byte aligned.
4943 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4945 #if BYTE_ORDER == LITTLE_ENDIAN
4946 unsigned short *u, *c;
4949 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4950 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4953 } while(!x && u > (unsigned short *)a->mv_data);
4956 unsigned short *u, *c, *end;
4959 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4960 u = (unsigned short *)a->mv_data;
4961 c = (unsigned short *)b->mv_data;
4964 } while(!x && u < end);
4969 /** Compare two items lexically */
4971 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4978 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4984 diff = memcmp(a->mv_data, b->mv_data, len);
4985 return diff ? diff : len_diff<0 ? -1 : len_diff;
4988 /** Compare two items in reverse byte order */
4990 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4992 const unsigned char *p1, *p2, *p1_lim;
4996 p1_lim = (const unsigned char *)a->mv_data;
4997 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4998 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5000 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5006 while (p1 > p1_lim) {
5007 diff = *--p1 - *--p2;
5011 return len_diff<0 ? -1 : len_diff;
5014 /** Search for key within a page, using binary search.
5015 * Returns the smallest entry larger or equal to the key.
5016 * If exactp is non-null, stores whether the found entry was an exact match
5017 * in *exactp (1 or 0).
5018 * Updates the cursor index with the index of the found entry.
5019 * If no entry larger or equal to the key is found, returns NULL.
5022 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5024 unsigned int i = 0, nkeys;
5027 MDB_page *mp = mc->mc_pg[mc->mc_top];
5028 MDB_node *node = NULL;
5033 nkeys = NUMKEYS(mp);
5035 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5036 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5039 low = IS_LEAF(mp) ? 0 : 1;
5041 cmp = mc->mc_dbx->md_cmp;
5043 /* Branch pages have no data, so if using integer keys,
5044 * alignment is guaranteed. Use faster mdb_cmp_int.
5046 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5047 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5054 nodekey.mv_size = mc->mc_db->md_pad;
5055 node = NODEPTR(mp, 0); /* fake */
5056 while (low <= high) {
5057 i = (low + high) >> 1;
5058 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5059 rc = cmp(key, &nodekey);
5060 DPRINTF(("found leaf index %u [%s], rc = %i",
5061 i, DKEY(&nodekey), rc));
5070 while (low <= high) {
5071 i = (low + high) >> 1;
5073 node = NODEPTR(mp, i);
5074 nodekey.mv_size = NODEKSZ(node);
5075 nodekey.mv_data = NODEKEY(node);
5077 rc = cmp(key, &nodekey);
5080 DPRINTF(("found leaf index %u [%s], rc = %i",
5081 i, DKEY(&nodekey), rc));
5083 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5084 i, DKEY(&nodekey), NODEPGNO(node), rc));
5095 if (rc > 0) { /* Found entry is less than the key. */
5096 i++; /* Skip to get the smallest entry larger than key. */
5098 node = NODEPTR(mp, i);
5101 *exactp = (rc == 0 && nkeys > 0);
5102 /* store the key index */
5103 mc->mc_ki[mc->mc_top] = i;
5105 /* There is no entry larger or equal to the key. */
5108 /* nodeptr is fake for LEAF2 */
5114 mdb_cursor_adjust(MDB_cursor *mc, func)
5118 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5119 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5126 /** Pop a page off the top of the cursor's stack. */
5128 mdb_cursor_pop(MDB_cursor *mc)
5131 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5132 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5140 /** Push a page onto the top of the cursor's stack. */
5142 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5144 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5145 DDBI(mc), (void *) mc));
5147 if (mc->mc_snum >= CURSOR_STACK) {
5148 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5149 return MDB_CURSOR_FULL;
5152 mc->mc_top = mc->mc_snum++;
5153 mc->mc_pg[mc->mc_top] = mp;
5154 mc->mc_ki[mc->mc_top] = 0;
5159 /** Find the address of the page corresponding to a given page number.
5160 * @param[in] txn the transaction for this access.
5161 * @param[in] pgno the page number for the page to retrieve.
5162 * @param[out] ret address of a pointer where the page's address will be stored.
5163 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5164 * @return 0 on success, non-zero on failure.
5167 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5169 MDB_env *env = txn->mt_env;
5173 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5177 MDB_ID2L dl = tx2->mt_u.dirty_list;
5179 /* Spilled pages were dirtied in this txn and flushed
5180 * because the dirty list got full. Bring this page
5181 * back in from the map (but don't unspill it here,
5182 * leave that unless page_touch happens again).
5184 if (tx2->mt_spill_pgs) {
5185 MDB_ID pn = pgno << 1;
5186 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5187 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5188 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5193 unsigned x = mdb_mid2l_search(dl, pgno);
5194 if (x <= dl[0].mid && dl[x].mid == pgno) {
5200 } while ((tx2 = tx2->mt_parent) != NULL);
5203 if (pgno < txn->mt_next_pgno) {
5205 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5207 DPRINTF(("page %"Z"u not found", pgno));
5208 txn->mt_flags |= MDB_TXN_ERROR;
5209 return MDB_PAGE_NOTFOUND;
5219 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5220 * The cursor is at the root page, set up the rest of it.
5223 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5225 MDB_page *mp = mc->mc_pg[mc->mc_top];
5229 while (IS_BRANCH(mp)) {
5233 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5234 mdb_cassert(mc, NUMKEYS(mp) > 1);
5235 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5237 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5239 if (flags & MDB_PS_LAST)
5240 i = NUMKEYS(mp) - 1;
5243 node = mdb_node_search(mc, key, &exact);
5245 i = NUMKEYS(mp) - 1;
5247 i = mc->mc_ki[mc->mc_top];
5249 mdb_cassert(mc, i > 0);
5253 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5256 mdb_cassert(mc, i < NUMKEYS(mp));
5257 node = NODEPTR(mp, i);
5259 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5262 mc->mc_ki[mc->mc_top] = i;
5263 if ((rc = mdb_cursor_push(mc, mp)))
5266 if (flags & MDB_PS_MODIFY) {
5267 if ((rc = mdb_page_touch(mc)) != 0)
5269 mp = mc->mc_pg[mc->mc_top];
5274 DPRINTF(("internal error, index points to a %02X page!?",
5276 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5277 return MDB_CORRUPTED;
5280 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5281 key ? DKEY(key) : "null"));
5282 mc->mc_flags |= C_INITIALIZED;
5283 mc->mc_flags &= ~C_EOF;
5288 /** Search for the lowest key under the current branch page.
5289 * This just bypasses a NUMKEYS check in the current page
5290 * before calling mdb_page_search_root(), because the callers
5291 * are all in situations where the current page is known to
5295 mdb_page_search_lowest(MDB_cursor *mc)
5297 MDB_page *mp = mc->mc_pg[mc->mc_top];
5298 MDB_node *node = NODEPTR(mp, 0);
5301 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5304 mc->mc_ki[mc->mc_top] = 0;
5305 if ((rc = mdb_cursor_push(mc, mp)))
5307 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5310 /** Search for the page a given key should be in.
5311 * Push it and its parent pages on the cursor stack.
5312 * @param[in,out] mc the cursor for this operation.
5313 * @param[in] key the key to search for, or NULL for first/last page.
5314 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5315 * are touched (updated with new page numbers).
5316 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5317 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5318 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5319 * @return 0 on success, non-zero on failure.
5322 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5327 /* Make sure the txn is still viable, then find the root from
5328 * the txn's db table and set it as the root of the cursor's stack.
5330 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5331 DPUTS("transaction may not be used now");
5334 /* Make sure we're using an up-to-date root */
5335 if (*mc->mc_dbflag & DB_STALE) {
5337 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5339 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5340 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5347 MDB_node *leaf = mdb_node_search(&mc2,
5348 &mc->mc_dbx->md_name, &exact);
5350 return MDB_NOTFOUND;
5351 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5352 return MDB_INCOMPATIBLE; /* not a named DB */
5353 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5356 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5358 /* The txn may not know this DBI, or another process may
5359 * have dropped and recreated the DB with other flags.
5361 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5362 return MDB_INCOMPATIBLE;
5363 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5365 *mc->mc_dbflag &= ~DB_STALE;
5367 root = mc->mc_db->md_root;
5369 if (root == P_INVALID) { /* Tree is empty. */
5370 DPUTS("tree is empty");
5371 return MDB_NOTFOUND;
5375 mdb_cassert(mc, root > 1);
5376 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5377 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5383 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5384 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5386 if (flags & MDB_PS_MODIFY) {
5387 if ((rc = mdb_page_touch(mc)))
5391 if (flags & MDB_PS_ROOTONLY)
5394 return mdb_page_search_root(mc, key, flags);
5398 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5400 MDB_txn *txn = mc->mc_txn;
5401 pgno_t pg = mp->mp_pgno;
5402 unsigned x = 0, ovpages = mp->mp_pages;
5403 MDB_env *env = txn->mt_env;
5404 MDB_IDL sl = txn->mt_spill_pgs;
5405 MDB_ID pn = pg << 1;
5408 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5409 /* If the page is dirty or on the spill list we just acquired it,
5410 * so we should give it back to our current free list, if any.
5411 * Otherwise put it onto the list of pages we freed in this txn.
5413 * Won't create me_pghead: me_pglast must be inited along with it.
5414 * Unsupported in nested txns: They would need to hide the page
5415 * range in ancestor txns' dirty and spilled lists.
5417 if (env->me_pghead &&
5419 ((mp->mp_flags & P_DIRTY) ||
5420 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5424 MDB_ID2 *dl, ix, iy;
5425 rc = mdb_midl_need(&env->me_pghead, ovpages);
5428 if (!(mp->mp_flags & P_DIRTY)) {
5429 /* This page is no longer spilled */
5436 /* Remove from dirty list */
5437 dl = txn->mt_u.dirty_list;
5439 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5445 mdb_cassert(mc, x > 1);
5447 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5448 txn->mt_flags |= MDB_TXN_ERROR;
5449 return MDB_CORRUPTED;
5452 if (!(env->me_flags & MDB_WRITEMAP))
5453 mdb_dpage_free(env, mp);
5455 /* Insert in me_pghead */
5456 mop = env->me_pghead;
5457 j = mop[0] + ovpages;
5458 for (i = mop[0]; i && mop[i] < pg; i--)
5464 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5468 mc->mc_db->md_overflow_pages -= ovpages;
5472 /** Return the data associated with a given node.
5473 * @param[in] txn The transaction for this operation.
5474 * @param[in] leaf The node being read.
5475 * @param[out] data Updated to point to the node's data.
5476 * @return 0 on success, non-zero on failure.
5479 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5481 MDB_page *omp; /* overflow page */
5485 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5486 data->mv_size = NODEDSZ(leaf);
5487 data->mv_data = NODEDATA(leaf);
5491 /* Read overflow data.
5493 data->mv_size = NODEDSZ(leaf);
5494 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5495 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5496 DPRINTF(("read overflow page %"Z"u failed", pgno));
5499 data->mv_data = METADATA(omp);
5505 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5506 MDB_val *key, MDB_val *data)
5513 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5515 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5518 if (txn->mt_flags & MDB_TXN_BLOCKED)
5521 mdb_cursor_init(&mc, txn, dbi, &mx);
5522 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5525 /** Find a sibling for a page.
5526 * Replaces the page at the top of the cursor's stack with the
5527 * specified sibling, if one exists.
5528 * @param[in] mc The cursor for this operation.
5529 * @param[in] move_right Non-zero if the right sibling is requested,
5530 * otherwise the left sibling.
5531 * @return 0 on success, non-zero on failure.
5534 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5540 if (mc->mc_snum < 2) {
5541 return MDB_NOTFOUND; /* root has no siblings */
5545 DPRINTF(("parent page is page %"Z"u, index %u",
5546 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5548 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5549 : (mc->mc_ki[mc->mc_top] == 0)) {
5550 DPRINTF(("no more keys left, moving to %s sibling",
5551 move_right ? "right" : "left"));
5552 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5553 /* undo cursor_pop before returning */
5560 mc->mc_ki[mc->mc_top]++;
5562 mc->mc_ki[mc->mc_top]--;
5563 DPRINTF(("just moving to %s index key %u",
5564 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5566 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5568 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5569 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5570 /* mc will be inconsistent if caller does mc_snum++ as above */
5571 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5575 mdb_cursor_push(mc, mp);
5577 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5582 /** Move the cursor to the next data item. */
5584 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5590 if (mc->mc_flags & C_EOF) {
5591 return MDB_NOTFOUND;
5594 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5596 mp = mc->mc_pg[mc->mc_top];
5598 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5599 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5600 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5601 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5602 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5603 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5604 if (rc == MDB_SUCCESS)
5605 MDB_GET_KEY(leaf, key);
5610 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5611 if (op == MDB_NEXT_DUP)
5612 return MDB_NOTFOUND;
5616 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5617 mdb_dbg_pgno(mp), (void *) mc));
5618 if (mc->mc_flags & C_DEL)
5621 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5622 DPUTS("=====> move to next sibling page");
5623 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5624 mc->mc_flags |= C_EOF;
5627 mp = mc->mc_pg[mc->mc_top];
5628 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5630 mc->mc_ki[mc->mc_top]++;
5633 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5634 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5637 key->mv_size = mc->mc_db->md_pad;
5638 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5642 mdb_cassert(mc, IS_LEAF(mp));
5643 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5645 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5646 mdb_xcursor_init1(mc, leaf);
5649 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5652 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5653 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5654 if (rc != MDB_SUCCESS)
5659 MDB_GET_KEY(leaf, key);
5663 /** Move the cursor to the previous data item. */
5665 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5671 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5673 mp = mc->mc_pg[mc->mc_top];
5675 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5676 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5677 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5678 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5679 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5680 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5681 if (rc == MDB_SUCCESS) {
5682 MDB_GET_KEY(leaf, key);
5683 mc->mc_flags &= ~C_EOF;
5689 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5690 if (op == MDB_PREV_DUP)
5691 return MDB_NOTFOUND;
5695 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5696 mdb_dbg_pgno(mp), (void *) mc));
5698 if (mc->mc_ki[mc->mc_top] == 0) {
5699 DPUTS("=====> move to prev sibling page");
5700 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5703 mp = mc->mc_pg[mc->mc_top];
5704 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5705 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5707 mc->mc_ki[mc->mc_top]--;
5709 mc->mc_flags &= ~C_EOF;
5711 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5712 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5715 key->mv_size = mc->mc_db->md_pad;
5716 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5720 mdb_cassert(mc, IS_LEAF(mp));
5721 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5723 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5724 mdb_xcursor_init1(mc, leaf);
5727 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5730 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5731 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5732 if (rc != MDB_SUCCESS)
5737 MDB_GET_KEY(leaf, key);
5741 /** Set the cursor on a specific data item. */
5743 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5744 MDB_cursor_op op, int *exactp)
5748 MDB_node *leaf = NULL;
5751 if (key->mv_size == 0)
5752 return MDB_BAD_VALSIZE;
5755 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5757 /* See if we're already on the right page */
5758 if (mc->mc_flags & C_INITIALIZED) {
5761 mp = mc->mc_pg[mc->mc_top];
5763 mc->mc_ki[mc->mc_top] = 0;
5764 return MDB_NOTFOUND;
5766 if (mp->mp_flags & P_LEAF2) {
5767 nodekey.mv_size = mc->mc_db->md_pad;
5768 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5770 leaf = NODEPTR(mp, 0);
5771 MDB_GET_KEY2(leaf, nodekey);
5773 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5775 /* Probably happens rarely, but first node on the page
5776 * was the one we wanted.
5778 mc->mc_ki[mc->mc_top] = 0;
5785 unsigned int nkeys = NUMKEYS(mp);
5787 if (mp->mp_flags & P_LEAF2) {
5788 nodekey.mv_data = LEAF2KEY(mp,
5789 nkeys-1, nodekey.mv_size);
5791 leaf = NODEPTR(mp, nkeys-1);
5792 MDB_GET_KEY2(leaf, nodekey);
5794 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5796 /* last node was the one we wanted */
5797 mc->mc_ki[mc->mc_top] = nkeys-1;
5803 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5804 /* This is definitely the right page, skip search_page */
5805 if (mp->mp_flags & P_LEAF2) {
5806 nodekey.mv_data = LEAF2KEY(mp,
5807 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5809 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5810 MDB_GET_KEY2(leaf, nodekey);
5812 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5814 /* current node was the one we wanted */
5824 /* If any parents have right-sibs, search.
5825 * Otherwise, there's nothing further.
5827 for (i=0; i<mc->mc_top; i++)
5829 NUMKEYS(mc->mc_pg[i])-1)
5831 if (i == mc->mc_top) {
5832 /* There are no other pages */
5833 mc->mc_ki[mc->mc_top] = nkeys;
5834 return MDB_NOTFOUND;
5838 /* There are no other pages */
5839 mc->mc_ki[mc->mc_top] = 0;
5840 if (op == MDB_SET_RANGE && !exactp) {
5844 return MDB_NOTFOUND;
5848 rc = mdb_page_search(mc, key, 0);
5849 if (rc != MDB_SUCCESS)
5852 mp = mc->mc_pg[mc->mc_top];
5853 mdb_cassert(mc, IS_LEAF(mp));
5856 leaf = mdb_node_search(mc, key, exactp);
5857 if (exactp != NULL && !*exactp) {
5858 /* MDB_SET specified and not an exact match. */
5859 return MDB_NOTFOUND;
5863 DPUTS("===> inexact leaf not found, goto sibling");
5864 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5865 mc->mc_flags |= C_EOF;
5866 return rc; /* no entries matched */
5868 mp = mc->mc_pg[mc->mc_top];
5869 mdb_cassert(mc, IS_LEAF(mp));
5870 leaf = NODEPTR(mp, 0);
5874 mc->mc_flags |= C_INITIALIZED;
5875 mc->mc_flags &= ~C_EOF;
5878 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5879 key->mv_size = mc->mc_db->md_pad;
5880 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5885 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5886 mdb_xcursor_init1(mc, leaf);
5889 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5890 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5891 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5894 if (op == MDB_GET_BOTH) {
5900 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5901 if (rc != MDB_SUCCESS)
5904 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5907 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5909 dcmp = mc->mc_dbx->md_dcmp;
5910 #if UINT_MAX < SIZE_MAX
5911 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5912 dcmp = mdb_cmp_clong;
5914 rc = dcmp(data, &olddata);
5916 if (op == MDB_GET_BOTH || rc > 0)
5917 return MDB_NOTFOUND;
5924 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5925 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5930 /* The key already matches in all other cases */
5931 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5932 MDB_GET_KEY(leaf, key);
5933 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5938 /** Move the cursor to the first item in the database. */
5940 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5946 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5948 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5949 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5950 if (rc != MDB_SUCCESS)
5953 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5955 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5956 mc->mc_flags |= C_INITIALIZED;
5957 mc->mc_flags &= ~C_EOF;
5959 mc->mc_ki[mc->mc_top] = 0;
5961 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5962 key->mv_size = mc->mc_db->md_pad;
5963 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5968 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5969 mdb_xcursor_init1(mc, leaf);
5970 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5974 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5978 MDB_GET_KEY(leaf, key);
5982 /** Move the cursor to the last item in the database. */
5984 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5990 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5992 if (!(mc->mc_flags & C_EOF)) {
5994 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5995 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5996 if (rc != MDB_SUCCESS)
5999 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6002 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6003 mc->mc_flags |= C_INITIALIZED|C_EOF;
6004 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6006 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6007 key->mv_size = mc->mc_db->md_pad;
6008 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6013 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6014 mdb_xcursor_init1(mc, leaf);
6015 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6019 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6024 MDB_GET_KEY(leaf, key);
6029 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6034 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6039 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6043 case MDB_GET_CURRENT:
6044 if (!(mc->mc_flags & C_INITIALIZED)) {
6047 MDB_page *mp = mc->mc_pg[mc->mc_top];
6048 int nkeys = NUMKEYS(mp);
6049 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6050 mc->mc_ki[mc->mc_top] = nkeys;
6056 key->mv_size = mc->mc_db->md_pad;
6057 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6059 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6060 MDB_GET_KEY(leaf, key);
6062 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6063 if (mc->mc_flags & C_DEL)
6064 mdb_xcursor_init1(mc, leaf);
6065 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6067 rc = mdb_node_read(mc->mc_txn, leaf, data);
6074 case MDB_GET_BOTH_RANGE:
6079 if (mc->mc_xcursor == NULL) {
6080 rc = MDB_INCOMPATIBLE;
6090 rc = mdb_cursor_set(mc, key, data, op,
6091 op == MDB_SET_RANGE ? NULL : &exact);
6094 case MDB_GET_MULTIPLE:
6095 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6099 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6100 rc = MDB_INCOMPATIBLE;
6104 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6105 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6108 case MDB_NEXT_MULTIPLE:
6113 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6114 rc = MDB_INCOMPATIBLE;
6117 if (!(mc->mc_flags & C_INITIALIZED))
6118 rc = mdb_cursor_first(mc, key, data);
6120 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6121 if (rc == MDB_SUCCESS) {
6122 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6125 mx = &mc->mc_xcursor->mx_cursor;
6126 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6128 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6129 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6137 case MDB_NEXT_NODUP:
6138 if (!(mc->mc_flags & C_INITIALIZED))
6139 rc = mdb_cursor_first(mc, key, data);
6141 rc = mdb_cursor_next(mc, key, data, op);
6145 case MDB_PREV_NODUP:
6146 if (!(mc->mc_flags & C_INITIALIZED)) {
6147 rc = mdb_cursor_last(mc, key, data);
6150 mc->mc_flags |= C_INITIALIZED;
6151 mc->mc_ki[mc->mc_top]++;
6153 rc = mdb_cursor_prev(mc, key, data, op);
6156 rc = mdb_cursor_first(mc, key, data);
6159 mfunc = mdb_cursor_first;
6161 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6165 if (mc->mc_xcursor == NULL) {
6166 rc = MDB_INCOMPATIBLE;
6170 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6171 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6172 MDB_GET_KEY(leaf, key);
6173 rc = mdb_node_read(mc->mc_txn, leaf, data);
6177 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6181 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6184 rc = mdb_cursor_last(mc, key, data);
6187 mfunc = mdb_cursor_last;
6190 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6195 if (mc->mc_flags & C_DEL)
6196 mc->mc_flags ^= C_DEL;
6201 /** Touch all the pages in the cursor stack. Set mc_top.
6202 * Makes sure all the pages are writable, before attempting a write operation.
6203 * @param[in] mc The cursor to operate on.
6206 mdb_cursor_touch(MDB_cursor *mc)
6208 int rc = MDB_SUCCESS;
6210 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6213 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6215 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6216 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6219 *mc->mc_dbflag |= DB_DIRTY;
6224 rc = mdb_page_touch(mc);
6225 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6226 mc->mc_top = mc->mc_snum-1;
6231 /** Do not spill pages to disk if txn is getting full, may fail instead */
6232 #define MDB_NOSPILL 0x8000
6235 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6239 MDB_node *leaf = NULL;
6240 MDB_page *fp, *mp, *sub_root = NULL;
6242 MDB_val xdata, *rdata, dkey, olddata;
6244 int do_sub = 0, insert_key, insert_data;
6245 unsigned int mcount = 0, dcount = 0, nospill;
6248 unsigned int nflags;
6251 if (mc == NULL || key == NULL)
6254 env = mc->mc_txn->mt_env;
6256 /* Check this first so counter will always be zero on any
6259 if (flags & MDB_MULTIPLE) {
6260 dcount = data[1].mv_size;
6261 data[1].mv_size = 0;
6262 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6263 return MDB_INCOMPATIBLE;
6266 nospill = flags & MDB_NOSPILL;
6267 flags &= ~MDB_NOSPILL;
6269 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6270 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6272 if (key->mv_size-1 >= ENV_MAXKEY(env))
6273 return MDB_BAD_VALSIZE;
6275 #if SIZE_MAX > MAXDATASIZE
6276 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6277 return MDB_BAD_VALSIZE;
6279 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6280 return MDB_BAD_VALSIZE;
6283 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6284 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6288 if (flags == MDB_CURRENT) {
6289 if (!(mc->mc_flags & C_INITIALIZED))
6292 } else if (mc->mc_db->md_root == P_INVALID) {
6293 /* new database, cursor has nothing to point to */
6296 mc->mc_flags &= ~C_INITIALIZED;
6301 if (flags & MDB_APPEND) {
6303 rc = mdb_cursor_last(mc, &k2, &d2);
6305 rc = mc->mc_dbx->md_cmp(key, &k2);
6308 mc->mc_ki[mc->mc_top]++;
6310 /* new key is <= last key */
6315 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6317 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6318 DPRINTF(("duplicate key [%s]", DKEY(key)));
6320 return MDB_KEYEXIST;
6322 if (rc && rc != MDB_NOTFOUND)
6326 if (mc->mc_flags & C_DEL)
6327 mc->mc_flags ^= C_DEL;
6329 /* Cursor is positioned, check for room in the dirty list */
6331 if (flags & MDB_MULTIPLE) {
6333 xdata.mv_size = data->mv_size * dcount;
6337 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6341 if (rc == MDB_NO_ROOT) {
6343 /* new database, write a root leaf page */
6344 DPUTS("allocating new root leaf page");
6345 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6348 mdb_cursor_push(mc, np);
6349 mc->mc_db->md_root = np->mp_pgno;
6350 mc->mc_db->md_depth++;
6351 *mc->mc_dbflag |= DB_DIRTY;
6352 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6354 np->mp_flags |= P_LEAF2;
6355 mc->mc_flags |= C_INITIALIZED;
6357 /* make sure all cursor pages are writable */
6358 rc2 = mdb_cursor_touch(mc);
6363 insert_key = insert_data = rc;
6365 /* The key does not exist */
6366 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6367 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6368 LEAFSIZE(key, data) > env->me_nodemax)
6370 /* Too big for a node, insert in sub-DB. Set up an empty
6371 * "old sub-page" for prep_subDB to expand to a full page.
6373 fp_flags = P_LEAF|P_DIRTY;
6375 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6376 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6377 olddata.mv_size = PAGEHDRSZ;
6381 /* there's only a key anyway, so this is a no-op */
6382 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6384 unsigned int ksize = mc->mc_db->md_pad;
6385 if (key->mv_size != ksize)
6386 return MDB_BAD_VALSIZE;
6387 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6388 memcpy(ptr, key->mv_data, ksize);
6390 /* if overwriting slot 0 of leaf, need to
6391 * update branch key if there is a parent page
6393 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6394 unsigned short dtop = 1;
6396 /* slot 0 is always an empty key, find real slot */
6397 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6401 if (mc->mc_ki[mc->mc_top])
6402 rc2 = mdb_update_key(mc, key);
6413 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6414 olddata.mv_size = NODEDSZ(leaf);
6415 olddata.mv_data = NODEDATA(leaf);
6418 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6419 /* Prepare (sub-)page/sub-DB to accept the new item,
6420 * if needed. fp: old sub-page or a header faking
6421 * it. mp: new (sub-)page. offset: growth in page
6422 * size. xdata: node data with new page or DB.
6424 unsigned i, offset = 0;
6425 mp = fp = xdata.mv_data = env->me_pbuf;
6426 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6428 /* Was a single item before, must convert now */
6429 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6431 /* Just overwrite the current item */
6432 if (flags == MDB_CURRENT)
6434 dcmp = mc->mc_dbx->md_dcmp;
6435 #if UINT_MAX < SIZE_MAX
6436 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6437 dcmp = mdb_cmp_clong;
6439 /* does data match? */
6440 if (!dcmp(data, &olddata)) {
6441 if (flags & MDB_NODUPDATA)
6442 return MDB_KEYEXIST;
6447 /* Back up original data item */
6448 dkey.mv_size = olddata.mv_size;
6449 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6451 /* Make sub-page header for the dup items, with dummy body */
6452 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6453 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6454 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6455 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6456 fp->mp_flags |= P_LEAF2;
6457 fp->mp_pad = data->mv_size;
6458 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6460 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6461 (dkey.mv_size & 1) + (data->mv_size & 1);
6463 fp->mp_upper = xdata.mv_size - PAGEBASE;
6464 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6465 } else if (leaf->mn_flags & F_SUBDATA) {
6466 /* Data is on sub-DB, just store it */
6467 flags |= F_DUPDATA|F_SUBDATA;
6470 /* Data is on sub-page */
6471 fp = olddata.mv_data;
6474 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6475 offset = EVEN(NODESIZE + sizeof(indx_t) +
6479 offset = fp->mp_pad;
6480 if (SIZELEFT(fp) < offset) {
6481 offset *= 4; /* space for 4 more */
6484 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6486 fp->mp_flags |= P_DIRTY;
6487 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6488 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6492 xdata.mv_size = olddata.mv_size + offset;
6495 fp_flags = fp->mp_flags;
6496 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6497 /* Too big for a sub-page, convert to sub-DB */
6498 fp_flags &= ~P_SUBP;
6500 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6501 fp_flags |= P_LEAF2;
6502 dummy.md_pad = fp->mp_pad;
6503 dummy.md_flags = MDB_DUPFIXED;
6504 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6505 dummy.md_flags |= MDB_INTEGERKEY;
6511 dummy.md_branch_pages = 0;
6512 dummy.md_leaf_pages = 1;
6513 dummy.md_overflow_pages = 0;
6514 dummy.md_entries = NUMKEYS(fp);
6515 xdata.mv_size = sizeof(MDB_db);
6516 xdata.mv_data = &dummy;
6517 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6519 offset = env->me_psize - olddata.mv_size;
6520 flags |= F_DUPDATA|F_SUBDATA;
6521 dummy.md_root = mp->mp_pgno;
6525 mp->mp_flags = fp_flags | P_DIRTY;
6526 mp->mp_pad = fp->mp_pad;
6527 mp->mp_lower = fp->mp_lower;
6528 mp->mp_upper = fp->mp_upper + offset;
6529 if (fp_flags & P_LEAF2) {
6530 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6532 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6533 olddata.mv_size - fp->mp_upper - PAGEBASE);
6534 for (i=0; i<NUMKEYS(fp); i++)
6535 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6543 mdb_node_del(mc, 0);
6547 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6548 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6549 return MDB_INCOMPATIBLE;
6550 /* overflow page overwrites need special handling */
6551 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6554 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6556 memcpy(&pg, olddata.mv_data, sizeof(pg));
6557 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6559 ovpages = omp->mp_pages;
6561 /* Is the ov page large enough? */
6562 if (ovpages >= dpages) {
6563 if (!(omp->mp_flags & P_DIRTY) &&
6564 (level || (env->me_flags & MDB_WRITEMAP)))
6566 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6569 level = 0; /* dirty in this txn or clean */
6572 if (omp->mp_flags & P_DIRTY) {
6573 /* yes, overwrite it. Note in this case we don't
6574 * bother to try shrinking the page if the new data
6575 * is smaller than the overflow threshold.
6578 /* It is writable only in a parent txn */
6579 size_t sz = (size_t) env->me_psize * ovpages, off;
6580 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6586 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6587 mdb_cassert(mc, rc2 == 0);
6588 if (!(flags & MDB_RESERVE)) {
6589 /* Copy end of page, adjusting alignment so
6590 * compiler may copy words instead of bytes.
6592 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6593 memcpy((size_t *)((char *)np + off),
6594 (size_t *)((char *)omp + off), sz - off);
6597 memcpy(np, omp, sz); /* Copy beginning of page */
6600 SETDSZ(leaf, data->mv_size);
6601 if (F_ISSET(flags, MDB_RESERVE))
6602 data->mv_data = METADATA(omp);
6604 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6608 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6610 } else if (data->mv_size == olddata.mv_size) {
6611 /* same size, just replace it. Note that we could
6612 * also reuse this node if the new data is smaller,
6613 * but instead we opt to shrink the node in that case.
6615 if (F_ISSET(flags, MDB_RESERVE))
6616 data->mv_data = olddata.mv_data;
6617 else if (!(mc->mc_flags & C_SUB))
6618 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6620 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6625 mdb_node_del(mc, 0);
6631 nflags = flags & NODE_ADD_FLAGS;
6632 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6633 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6634 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6635 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6637 nflags |= MDB_SPLIT_REPLACE;
6638 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6640 /* There is room already in this leaf page. */
6641 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6642 if (rc == 0 && insert_key) {
6643 /* Adjust other cursors pointing to mp */
6644 MDB_cursor *m2, *m3;
6645 MDB_dbi dbi = mc->mc_dbi;
6646 unsigned i = mc->mc_top;
6647 MDB_page *mp = mc->mc_pg[i];
6649 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6650 if (mc->mc_flags & C_SUB)
6651 m3 = &m2->mc_xcursor->mx_cursor;
6654 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6655 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6662 if (rc == MDB_SUCCESS) {
6663 /* Now store the actual data in the child DB. Note that we're
6664 * storing the user data in the keys field, so there are strict
6665 * size limits on dupdata. The actual data fields of the child
6666 * DB are all zero size.
6669 int xflags, new_dupdata;
6674 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6675 if (flags & MDB_CURRENT) {
6676 xflags = MDB_CURRENT|MDB_NOSPILL;
6678 mdb_xcursor_init1(mc, leaf);
6679 xflags = (flags & MDB_NODUPDATA) ?
6680 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6683 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6684 new_dupdata = (int)dkey.mv_size;
6685 /* converted, write the original data first */
6687 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6690 /* we've done our job */
6693 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6694 /* Adjust other cursors pointing to mp */
6696 MDB_xcursor *mx = mc->mc_xcursor;
6697 unsigned i = mc->mc_top;
6698 MDB_page *mp = mc->mc_pg[i];
6700 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6701 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6702 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6703 if (m2->mc_pg[i] == mp) {
6704 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6705 mdb_xcursor_init2(m2, mx, new_dupdata);
6706 } else if (!insert_key) {
6707 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6708 if (!(n2->mn_flags & F_SUBDATA))
6709 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6714 ecount = mc->mc_xcursor->mx_db.md_entries;
6715 if (flags & MDB_APPENDDUP)
6716 xflags |= MDB_APPEND;
6717 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6718 if (flags & F_SUBDATA) {
6719 void *db = NODEDATA(leaf);
6720 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6722 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6724 /* Increment count unless we just replaced an existing item. */
6726 mc->mc_db->md_entries++;
6728 /* Invalidate txn if we created an empty sub-DB */
6731 /* If we succeeded and the key didn't exist before,
6732 * make sure the cursor is marked valid.
6734 mc->mc_flags |= C_INITIALIZED;
6736 if (flags & MDB_MULTIPLE) {
6739 /* let caller know how many succeeded, if any */
6740 data[1].mv_size = mcount;
6741 if (mcount < dcount) {
6742 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6743 insert_key = insert_data = 0;
6750 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6753 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6758 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6764 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6765 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6767 if (!(mc->mc_flags & C_INITIALIZED))
6770 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6771 return MDB_NOTFOUND;
6773 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6776 rc = mdb_cursor_touch(mc);
6780 mp = mc->mc_pg[mc->mc_top];
6783 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6785 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6786 if (flags & MDB_NODUPDATA) {
6787 /* mdb_cursor_del0() will subtract the final entry */
6788 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6790 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6791 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6793 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6796 /* If sub-DB still has entries, we're done */
6797 if (mc->mc_xcursor->mx_db.md_entries) {
6798 if (leaf->mn_flags & F_SUBDATA) {
6799 /* update subDB info */
6800 void *db = NODEDATA(leaf);
6801 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6804 /* shrink fake page */
6805 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6806 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6807 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6808 /* fix other sub-DB cursors pointed at fake pages on this page */
6809 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6810 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6811 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6812 if (m2->mc_pg[mc->mc_top] == mp) {
6813 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6814 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6816 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6817 if (!(n2->mn_flags & F_SUBDATA))
6818 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6823 mc->mc_db->md_entries--;
6824 mc->mc_flags |= C_DEL;
6827 /* otherwise fall thru and delete the sub-DB */
6830 if (leaf->mn_flags & F_SUBDATA) {
6831 /* add all the child DB's pages to the free list */
6832 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6837 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6838 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6839 rc = MDB_INCOMPATIBLE;
6843 /* add overflow pages to free list */
6844 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6848 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6849 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6850 (rc = mdb_ovpage_free(mc, omp)))
6855 return mdb_cursor_del0(mc);
6858 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6862 /** Allocate and initialize new pages for a database.
6863 * @param[in] mc a cursor on the database being added to.
6864 * @param[in] flags flags defining what type of page is being allocated.
6865 * @param[in] num the number of pages to allocate. This is usually 1,
6866 * unless allocating overflow pages for a large record.
6867 * @param[out] mp Address of a page, or NULL on failure.
6868 * @return 0 on success, non-zero on failure.
6871 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6876 if ((rc = mdb_page_alloc(mc, num, &np)))
6878 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6879 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6880 np->mp_flags = flags | P_DIRTY;
6881 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6882 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6885 mc->mc_db->md_branch_pages++;
6886 else if (IS_LEAF(np))
6887 mc->mc_db->md_leaf_pages++;
6888 else if (IS_OVERFLOW(np)) {
6889 mc->mc_db->md_overflow_pages += num;
6897 /** Calculate the size of a leaf node.
6898 * The size depends on the environment's page size; if a data item
6899 * is too large it will be put onto an overflow page and the node
6900 * size will only include the key and not the data. Sizes are always
6901 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6902 * of the #MDB_node headers.
6903 * @param[in] env The environment handle.
6904 * @param[in] key The key for the node.
6905 * @param[in] data The data for the node.
6906 * @return The number of bytes needed to store the node.
6909 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6913 sz = LEAFSIZE(key, data);
6914 if (sz > env->me_nodemax) {
6915 /* put on overflow page */
6916 sz -= data->mv_size - sizeof(pgno_t);
6919 return EVEN(sz + sizeof(indx_t));
6922 /** Calculate the size of a branch node.
6923 * The size should depend on the environment's page size but since
6924 * we currently don't support spilling large keys onto overflow
6925 * pages, it's simply the size of the #MDB_node header plus the
6926 * size of the key. Sizes are always rounded up to an even number
6927 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6928 * @param[in] env The environment handle.
6929 * @param[in] key The key for the node.
6930 * @return The number of bytes needed to store the node.
6933 mdb_branch_size(MDB_env *env, MDB_val *key)
6938 if (sz > env->me_nodemax) {
6939 /* put on overflow page */
6940 /* not implemented */
6941 /* sz -= key->size - sizeof(pgno_t); */
6944 return sz + sizeof(indx_t);
6947 /** Add a node to the page pointed to by the cursor.
6948 * @param[in] mc The cursor for this operation.
6949 * @param[in] indx The index on the page where the new node should be added.
6950 * @param[in] key The key for the new node.
6951 * @param[in] data The data for the new node, if any.
6952 * @param[in] pgno The page number, if adding a branch node.
6953 * @param[in] flags Flags for the node.
6954 * @return 0 on success, non-zero on failure. Possible errors are:
6956 * <li>ENOMEM - failed to allocate overflow pages for the node.
6957 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6958 * should never happen since all callers already calculate the
6959 * page's free space before calling this function.
6963 mdb_node_add(MDB_cursor *mc, indx_t indx,
6964 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6967 size_t node_size = NODESIZE;
6971 MDB_page *mp = mc->mc_pg[mc->mc_top];
6972 MDB_page *ofp = NULL; /* overflow page */
6976 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6978 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6979 IS_LEAF(mp) ? "leaf" : "branch",
6980 IS_SUBP(mp) ? "sub-" : "",
6981 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6982 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6985 /* Move higher keys up one slot. */
6986 int ksize = mc->mc_db->md_pad, dif;
6987 char *ptr = LEAF2KEY(mp, indx, ksize);
6988 dif = NUMKEYS(mp) - indx;
6990 memmove(ptr+ksize, ptr, dif*ksize);
6991 /* insert new key */
6992 memcpy(ptr, key->mv_data, ksize);
6994 /* Just using these for counting */
6995 mp->mp_lower += sizeof(indx_t);
6996 mp->mp_upper -= ksize - sizeof(indx_t);
7000 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7002 node_size += key->mv_size;
7004 mdb_cassert(mc, key && data);
7005 if (F_ISSET(flags, F_BIGDATA)) {
7006 /* Data already on overflow page. */
7007 node_size += sizeof(pgno_t);
7008 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7009 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7011 /* Put data on overflow page. */
7012 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7013 data->mv_size, node_size+data->mv_size));
7014 node_size = EVEN(node_size + sizeof(pgno_t));
7015 if ((ssize_t)node_size > room)
7017 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7019 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7023 node_size += data->mv_size;
7026 node_size = EVEN(node_size);
7027 if ((ssize_t)node_size > room)
7031 /* Move higher pointers up one slot. */
7032 for (i = NUMKEYS(mp); i > indx; i--)
7033 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7035 /* Adjust free space offsets. */
7036 ofs = mp->mp_upper - node_size;
7037 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7038 mp->mp_ptrs[indx] = ofs;
7040 mp->mp_lower += sizeof(indx_t);
7042 /* Write the node data. */
7043 node = NODEPTR(mp, indx);
7044 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7045 node->mn_flags = flags;
7047 SETDSZ(node,data->mv_size);
7052 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7055 ndata = NODEDATA(node);
7057 if (F_ISSET(flags, F_BIGDATA))
7058 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7059 else if (F_ISSET(flags, MDB_RESERVE))
7060 data->mv_data = ndata;
7062 memcpy(ndata, data->mv_data, data->mv_size);
7064 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7065 ndata = METADATA(ofp);
7066 if (F_ISSET(flags, MDB_RESERVE))
7067 data->mv_data = ndata;
7069 memcpy(ndata, data->mv_data, data->mv_size);
7076 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7077 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7078 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7079 DPRINTF(("node size = %"Z"u", node_size));
7080 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7081 return MDB_PAGE_FULL;
7084 /** Delete the specified node from a page.
7085 * @param[in] mc Cursor pointing to the node to delete.
7086 * @param[in] ksize The size of a node. Only used if the page is
7087 * part of a #MDB_DUPFIXED database.
7090 mdb_node_del(MDB_cursor *mc, int ksize)
7092 MDB_page *mp = mc->mc_pg[mc->mc_top];
7093 indx_t indx = mc->mc_ki[mc->mc_top];
7095 indx_t i, j, numkeys, ptr;
7099 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7100 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7101 numkeys = NUMKEYS(mp);
7102 mdb_cassert(mc, indx < numkeys);
7105 int x = numkeys - 1 - indx;
7106 base = LEAF2KEY(mp, indx, ksize);
7108 memmove(base, base + ksize, x * ksize);
7109 mp->mp_lower -= sizeof(indx_t);
7110 mp->mp_upper += ksize - sizeof(indx_t);
7114 node = NODEPTR(mp, indx);
7115 sz = NODESIZE + node->mn_ksize;
7117 if (F_ISSET(node->mn_flags, F_BIGDATA))
7118 sz += sizeof(pgno_t);
7120 sz += NODEDSZ(node);
7124 ptr = mp->mp_ptrs[indx];
7125 for (i = j = 0; i < numkeys; i++) {
7127 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7128 if (mp->mp_ptrs[i] < ptr)
7129 mp->mp_ptrs[j] += sz;
7134 base = (char *)mp + mp->mp_upper + PAGEBASE;
7135 memmove(base + sz, base, ptr - mp->mp_upper);
7137 mp->mp_lower -= sizeof(indx_t);
7141 /** Compact the main page after deleting a node on a subpage.
7142 * @param[in] mp The main page to operate on.
7143 * @param[in] indx The index of the subpage on the main page.
7146 mdb_node_shrink(MDB_page *mp, indx_t indx)
7151 indx_t delta, nsize, len, ptr;
7154 node = NODEPTR(mp, indx);
7155 sp = (MDB_page *)NODEDATA(node);
7156 delta = SIZELEFT(sp);
7157 nsize = NODEDSZ(node) - delta;
7159 /* Prepare to shift upward, set len = length(subpage part to shift) */
7163 return; /* do not make the node uneven-sized */
7165 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7166 for (i = NUMKEYS(sp); --i >= 0; )
7167 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7170 sp->mp_upper = sp->mp_lower;
7171 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7172 SETDSZ(node, nsize);
7174 /* Shift <lower nodes...initial part of subpage> upward */
7175 base = (char *)mp + mp->mp_upper + PAGEBASE;
7176 memmove(base + delta, base, (char *)sp + len - base);
7178 ptr = mp->mp_ptrs[indx];
7179 for (i = NUMKEYS(mp); --i >= 0; ) {
7180 if (mp->mp_ptrs[i] <= ptr)
7181 mp->mp_ptrs[i] += delta;
7183 mp->mp_upper += delta;
7186 /** Initial setup of a sorted-dups cursor.
7187 * Sorted duplicates are implemented as a sub-database for the given key.
7188 * The duplicate data items are actually keys of the sub-database.
7189 * Operations on the duplicate data items are performed using a sub-cursor
7190 * initialized when the sub-database is first accessed. This function does
7191 * the preliminary setup of the sub-cursor, filling in the fields that
7192 * depend only on the parent DB.
7193 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7196 mdb_xcursor_init0(MDB_cursor *mc)
7198 MDB_xcursor *mx = mc->mc_xcursor;
7200 mx->mx_cursor.mc_xcursor = NULL;
7201 mx->mx_cursor.mc_txn = mc->mc_txn;
7202 mx->mx_cursor.mc_db = &mx->mx_db;
7203 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7204 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7205 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7206 mx->mx_cursor.mc_snum = 0;
7207 mx->mx_cursor.mc_top = 0;
7208 mx->mx_cursor.mc_flags = C_SUB;
7209 mx->mx_dbx.md_name.mv_size = 0;
7210 mx->mx_dbx.md_name.mv_data = NULL;
7211 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7212 mx->mx_dbx.md_dcmp = NULL;
7213 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7216 /** Final setup of a sorted-dups cursor.
7217 * Sets up the fields that depend on the data from the main cursor.
7218 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7219 * @param[in] node The data containing the #MDB_db record for the
7220 * sorted-dup database.
7223 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7225 MDB_xcursor *mx = mc->mc_xcursor;
7227 if (node->mn_flags & F_SUBDATA) {
7228 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7229 mx->mx_cursor.mc_pg[0] = 0;
7230 mx->mx_cursor.mc_snum = 0;
7231 mx->mx_cursor.mc_top = 0;
7232 mx->mx_cursor.mc_flags = C_SUB;
7234 MDB_page *fp = NODEDATA(node);
7235 mx->mx_db.md_pad = 0;
7236 mx->mx_db.md_flags = 0;
7237 mx->mx_db.md_depth = 1;
7238 mx->mx_db.md_branch_pages = 0;
7239 mx->mx_db.md_leaf_pages = 1;
7240 mx->mx_db.md_overflow_pages = 0;
7241 mx->mx_db.md_entries = NUMKEYS(fp);
7242 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7243 mx->mx_cursor.mc_snum = 1;
7244 mx->mx_cursor.mc_top = 0;
7245 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7246 mx->mx_cursor.mc_pg[0] = fp;
7247 mx->mx_cursor.mc_ki[0] = 0;
7248 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7249 mx->mx_db.md_flags = MDB_DUPFIXED;
7250 mx->mx_db.md_pad = fp->mp_pad;
7251 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7252 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7255 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7256 mx->mx_db.md_root));
7257 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7258 #if UINT_MAX < SIZE_MAX
7259 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7260 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7265 /** Fixup a sorted-dups cursor due to underlying update.
7266 * Sets up some fields that depend on the data from the main cursor.
7267 * Almost the same as init1, but skips initialization steps if the
7268 * xcursor had already been used.
7269 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7270 * @param[in] src_mx The xcursor of an up-to-date cursor.
7271 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7274 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7276 MDB_xcursor *mx = mc->mc_xcursor;
7279 mx->mx_cursor.mc_snum = 1;
7280 mx->mx_cursor.mc_top = 0;
7281 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7282 mx->mx_cursor.mc_ki[0] = 0;
7283 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7284 #if UINT_MAX < SIZE_MAX
7285 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7287 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7290 mx->mx_db = src_mx->mx_db;
7291 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7292 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7293 mx->mx_db.md_root));
7296 /** Initialize a cursor for a given transaction and database. */
7298 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7301 mc->mc_backup = NULL;
7304 mc->mc_db = &txn->mt_dbs[dbi];
7305 mc->mc_dbx = &txn->mt_dbxs[dbi];
7306 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7312 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7313 mdb_tassert(txn, mx != NULL);
7314 mc->mc_xcursor = mx;
7315 mdb_xcursor_init0(mc);
7317 mc->mc_xcursor = NULL;
7319 if (*mc->mc_dbflag & DB_STALE) {
7320 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7325 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7328 size_t size = sizeof(MDB_cursor);
7330 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7333 if (txn->mt_flags & MDB_TXN_BLOCKED)
7336 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7339 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7340 size += sizeof(MDB_xcursor);
7342 if ((mc = malloc(size)) != NULL) {
7343 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7344 if (txn->mt_cursors) {
7345 mc->mc_next = txn->mt_cursors[dbi];
7346 txn->mt_cursors[dbi] = mc;
7347 mc->mc_flags |= C_UNTRACK;
7359 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7361 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7364 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7367 if (txn->mt_flags & MDB_TXN_BLOCKED)
7370 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7374 /* Return the count of duplicate data items for the current key */
7376 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7380 if (mc == NULL || countp == NULL)
7383 if (mc->mc_xcursor == NULL)
7384 return MDB_INCOMPATIBLE;
7386 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7389 if (!(mc->mc_flags & C_INITIALIZED))
7392 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7393 return MDB_NOTFOUND;
7395 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7396 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7399 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7402 *countp = mc->mc_xcursor->mx_db.md_entries;
7408 mdb_cursor_close(MDB_cursor *mc)
7410 if (mc && !mc->mc_backup) {
7411 /* remove from txn, if tracked */
7412 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7413 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7414 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7416 *prev = mc->mc_next;
7423 mdb_cursor_txn(MDB_cursor *mc)
7425 if (!mc) return NULL;
7430 mdb_cursor_dbi(MDB_cursor *mc)
7435 /** Replace the key for a branch node with a new key.
7436 * @param[in] mc Cursor pointing to the node to operate on.
7437 * @param[in] key The new key to use.
7438 * @return 0 on success, non-zero on failure.
7441 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7447 int delta, ksize, oksize;
7448 indx_t ptr, i, numkeys, indx;
7451 indx = mc->mc_ki[mc->mc_top];
7452 mp = mc->mc_pg[mc->mc_top];
7453 node = NODEPTR(mp, indx);
7454 ptr = mp->mp_ptrs[indx];
7458 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7459 k2.mv_data = NODEKEY(node);
7460 k2.mv_size = node->mn_ksize;
7461 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7463 mdb_dkey(&k2, kbuf2),
7469 /* Sizes must be 2-byte aligned. */
7470 ksize = EVEN(key->mv_size);
7471 oksize = EVEN(node->mn_ksize);
7472 delta = ksize - oksize;
7474 /* Shift node contents if EVEN(key length) changed. */
7476 if (delta > 0 && SIZELEFT(mp) < delta) {
7478 /* not enough space left, do a delete and split */
7479 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7480 pgno = NODEPGNO(node);
7481 mdb_node_del(mc, 0);
7482 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7485 numkeys = NUMKEYS(mp);
7486 for (i = 0; i < numkeys; i++) {
7487 if (mp->mp_ptrs[i] <= ptr)
7488 mp->mp_ptrs[i] -= delta;
7491 base = (char *)mp + mp->mp_upper + PAGEBASE;
7492 len = ptr - mp->mp_upper + NODESIZE;
7493 memmove(base - delta, base, len);
7494 mp->mp_upper -= delta;
7496 node = NODEPTR(mp, indx);
7499 /* But even if no shift was needed, update ksize */
7500 if (node->mn_ksize != key->mv_size)
7501 node->mn_ksize = key->mv_size;
7504 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7510 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7512 /** Move a node from csrc to cdst.
7515 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7522 unsigned short flags;
7526 /* Mark src and dst as dirty. */
7527 if ((rc = mdb_page_touch(csrc)) ||
7528 (rc = mdb_page_touch(cdst)))
7531 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7532 key.mv_size = csrc->mc_db->md_pad;
7533 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7535 data.mv_data = NULL;
7539 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7540 mdb_cassert(csrc, !((size_t)srcnode & 1));
7541 srcpg = NODEPGNO(srcnode);
7542 flags = srcnode->mn_flags;
7543 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7544 unsigned int snum = csrc->mc_snum;
7546 /* must find the lowest key below src */
7547 rc = mdb_page_search_lowest(csrc);
7550 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7551 key.mv_size = csrc->mc_db->md_pad;
7552 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7554 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7555 key.mv_size = NODEKSZ(s2);
7556 key.mv_data = NODEKEY(s2);
7558 csrc->mc_snum = snum--;
7559 csrc->mc_top = snum;
7561 key.mv_size = NODEKSZ(srcnode);
7562 key.mv_data = NODEKEY(srcnode);
7564 data.mv_size = NODEDSZ(srcnode);
7565 data.mv_data = NODEDATA(srcnode);
7567 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7568 unsigned int snum = cdst->mc_snum;
7571 /* must find the lowest key below dst */
7572 mdb_cursor_copy(cdst, &mn);
7573 rc = mdb_page_search_lowest(&mn);
7576 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7577 bkey.mv_size = mn.mc_db->md_pad;
7578 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7580 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7581 bkey.mv_size = NODEKSZ(s2);
7582 bkey.mv_data = NODEKEY(s2);
7584 mn.mc_snum = snum--;
7587 rc = mdb_update_key(&mn, &bkey);
7592 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7593 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7594 csrc->mc_ki[csrc->mc_top],
7596 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7597 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7599 /* Add the node to the destination page.
7601 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7602 if (rc != MDB_SUCCESS)
7605 /* Delete the node from the source page.
7607 mdb_node_del(csrc, key.mv_size);
7610 /* Adjust other cursors pointing to mp */
7611 MDB_cursor *m2, *m3;
7612 MDB_dbi dbi = csrc->mc_dbi;
7613 MDB_page *mpd, *mps;
7615 mps = csrc->mc_pg[csrc->mc_top];
7616 /* If we're adding on the left, bump others up */
7617 if (!cdst->mc_ki[csrc->mc_top]) {
7618 mpd = cdst->mc_pg[csrc->mc_top];
7619 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7620 if (csrc->mc_flags & C_SUB)
7621 m3 = &m2->mc_xcursor->mx_cursor;
7625 m3->mc_pg[csrc->mc_top] == mpd &&
7626 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7627 m3->mc_ki[csrc->mc_top]++;
7630 m3->mc_pg[csrc->mc_top] == mps &&
7631 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7632 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7633 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7634 m3->mc_ki[csrc->mc_top-1]++;
7638 /* Adding on the right, bump others down */
7640 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7641 if (csrc->mc_flags & C_SUB)
7642 m3 = &m2->mc_xcursor->mx_cursor;
7645 if (m3 == csrc) continue;
7646 if (m3->mc_pg[csrc->mc_top] == mps) {
7647 if (!m3->mc_ki[csrc->mc_top]) {
7648 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7649 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7650 m3->mc_ki[csrc->mc_top-1]--;
7652 m3->mc_ki[csrc->mc_top]--;
7659 /* Update the parent separators.
7661 if (csrc->mc_ki[csrc->mc_top] == 0) {
7662 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7663 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7664 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7666 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7667 key.mv_size = NODEKSZ(srcnode);
7668 key.mv_data = NODEKEY(srcnode);
7670 DPRINTF(("update separator for source page %"Z"u to [%s]",
7671 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7672 mdb_cursor_copy(csrc, &mn);
7675 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7678 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7680 indx_t ix = csrc->mc_ki[csrc->mc_top];
7681 nullkey.mv_size = 0;
7682 csrc->mc_ki[csrc->mc_top] = 0;
7683 rc = mdb_update_key(csrc, &nullkey);
7684 csrc->mc_ki[csrc->mc_top] = ix;
7685 mdb_cassert(csrc, rc == MDB_SUCCESS);
7689 if (cdst->mc_ki[cdst->mc_top] == 0) {
7690 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7691 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7692 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7694 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7695 key.mv_size = NODEKSZ(srcnode);
7696 key.mv_data = NODEKEY(srcnode);
7698 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7699 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7700 mdb_cursor_copy(cdst, &mn);
7703 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7706 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7708 indx_t ix = cdst->mc_ki[cdst->mc_top];
7709 nullkey.mv_size = 0;
7710 cdst->mc_ki[cdst->mc_top] = 0;
7711 rc = mdb_update_key(cdst, &nullkey);
7712 cdst->mc_ki[cdst->mc_top] = ix;
7713 mdb_cassert(cdst, rc == MDB_SUCCESS);
7720 /** Merge one page into another.
7721 * The nodes from the page pointed to by \b csrc will
7722 * be copied to the page pointed to by \b cdst and then
7723 * the \b csrc page will be freed.
7724 * @param[in] csrc Cursor pointing to the source page.
7725 * @param[in] cdst Cursor pointing to the destination page.
7726 * @return 0 on success, non-zero on failure.
7729 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7731 MDB_page *psrc, *pdst;
7738 psrc = csrc->mc_pg[csrc->mc_top];
7739 pdst = cdst->mc_pg[cdst->mc_top];
7741 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7743 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7744 mdb_cassert(csrc, cdst->mc_snum > 1);
7746 /* Mark dst as dirty. */
7747 if ((rc = mdb_page_touch(cdst)))
7750 /* get dst page again now that we've touched it. */
7751 pdst = cdst->mc_pg[cdst->mc_top];
7753 /* Move all nodes from src to dst.
7755 j = nkeys = NUMKEYS(pdst);
7756 if (IS_LEAF2(psrc)) {
7757 key.mv_size = csrc->mc_db->md_pad;
7758 key.mv_data = METADATA(psrc);
7759 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7760 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7761 if (rc != MDB_SUCCESS)
7763 key.mv_data = (char *)key.mv_data + key.mv_size;
7766 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7767 srcnode = NODEPTR(psrc, i);
7768 if (i == 0 && IS_BRANCH(psrc)) {
7771 mdb_cursor_copy(csrc, &mn);
7772 /* must find the lowest key below src */
7773 rc = mdb_page_search_lowest(&mn);
7776 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7777 key.mv_size = mn.mc_db->md_pad;
7778 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7780 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7781 key.mv_size = NODEKSZ(s2);
7782 key.mv_data = NODEKEY(s2);
7785 key.mv_size = srcnode->mn_ksize;
7786 key.mv_data = NODEKEY(srcnode);
7789 data.mv_size = NODEDSZ(srcnode);
7790 data.mv_data = NODEDATA(srcnode);
7791 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7792 if (rc != MDB_SUCCESS)
7797 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7798 pdst->mp_pgno, NUMKEYS(pdst),
7799 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7801 /* Unlink the src page from parent and add to free list.
7804 mdb_node_del(csrc, 0);
7805 if (csrc->mc_ki[csrc->mc_top] == 0) {
7807 rc = mdb_update_key(csrc, &key);
7815 psrc = csrc->mc_pg[csrc->mc_top];
7816 /* If not operating on FreeDB, allow this page to be reused
7817 * in this txn. Otherwise just add to free list.
7819 rc = mdb_page_loose(csrc, psrc);
7823 csrc->mc_db->md_leaf_pages--;
7825 csrc->mc_db->md_branch_pages--;
7827 /* Adjust other cursors pointing to mp */
7828 MDB_cursor *m2, *m3;
7829 MDB_dbi dbi = csrc->mc_dbi;
7830 unsigned int top = csrc->mc_top;
7832 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7833 if (csrc->mc_flags & C_SUB)
7834 m3 = &m2->mc_xcursor->mx_cursor;
7837 if (m3 == csrc) continue;
7838 if (m3->mc_snum < csrc->mc_snum) continue;
7839 if (m3->mc_pg[top] == psrc) {
7840 m3->mc_pg[top] = pdst;
7841 m3->mc_ki[top] += nkeys;
7842 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
7847 unsigned int snum = cdst->mc_snum;
7848 uint16_t depth = cdst->mc_db->md_depth;
7849 mdb_cursor_pop(cdst);
7850 rc = mdb_rebalance(cdst);
7851 /* Did the tree height change? */
7852 if (depth != cdst->mc_db->md_depth)
7853 snum += cdst->mc_db->md_depth - depth;
7854 cdst->mc_snum = snum;
7855 cdst->mc_top = snum-1;
7860 /** Copy the contents of a cursor.
7861 * @param[in] csrc The cursor to copy from.
7862 * @param[out] cdst The cursor to copy to.
7865 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7869 cdst->mc_txn = csrc->mc_txn;
7870 cdst->mc_dbi = csrc->mc_dbi;
7871 cdst->mc_db = csrc->mc_db;
7872 cdst->mc_dbx = csrc->mc_dbx;
7873 cdst->mc_snum = csrc->mc_snum;
7874 cdst->mc_top = csrc->mc_top;
7875 cdst->mc_flags = csrc->mc_flags;
7877 for (i=0; i<csrc->mc_snum; i++) {
7878 cdst->mc_pg[i] = csrc->mc_pg[i];
7879 cdst->mc_ki[i] = csrc->mc_ki[i];
7883 /** Rebalance the tree after a delete operation.
7884 * @param[in] mc Cursor pointing to the page where rebalancing
7886 * @return 0 on success, non-zero on failure.
7889 mdb_rebalance(MDB_cursor *mc)
7893 unsigned int ptop, minkeys, thresh;
7897 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7902 thresh = FILL_THRESHOLD;
7904 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7905 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7906 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7907 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7909 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7910 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7911 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7912 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7916 if (mc->mc_snum < 2) {
7917 MDB_page *mp = mc->mc_pg[0];
7919 DPUTS("Can't rebalance a subpage, ignoring");
7922 if (NUMKEYS(mp) == 0) {
7923 DPUTS("tree is completely empty");
7924 mc->mc_db->md_root = P_INVALID;
7925 mc->mc_db->md_depth = 0;
7926 mc->mc_db->md_leaf_pages = 0;
7927 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7930 /* Adjust cursors pointing to mp */
7933 mc->mc_flags &= ~C_INITIALIZED;
7935 MDB_cursor *m2, *m3;
7936 MDB_dbi dbi = mc->mc_dbi;
7938 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7939 if (mc->mc_flags & C_SUB)
7940 m3 = &m2->mc_xcursor->mx_cursor;
7943 if (m3->mc_snum < mc->mc_snum) continue;
7944 if (m3->mc_pg[0] == mp) {
7947 m3->mc_flags &= ~C_INITIALIZED;
7951 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7953 DPUTS("collapsing root page!");
7954 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7957 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7958 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7961 mc->mc_db->md_depth--;
7962 mc->mc_db->md_branch_pages--;
7963 mc->mc_ki[0] = mc->mc_ki[1];
7964 for (i = 1; i<mc->mc_db->md_depth; i++) {
7965 mc->mc_pg[i] = mc->mc_pg[i+1];
7966 mc->mc_ki[i] = mc->mc_ki[i+1];
7969 /* Adjust other cursors pointing to mp */
7970 MDB_cursor *m2, *m3;
7971 MDB_dbi dbi = mc->mc_dbi;
7973 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7974 if (mc->mc_flags & C_SUB)
7975 m3 = &m2->mc_xcursor->mx_cursor;
7978 if (m3 == mc) continue;
7979 if (m3->mc_pg[0] == mp) {
7980 for (i=0; i<mc->mc_db->md_depth; i++) {
7981 m3->mc_pg[i] = m3->mc_pg[i+1];
7982 m3->mc_ki[i] = m3->mc_ki[i+1];
7990 DPUTS("root page doesn't need rebalancing");
7994 /* The parent (branch page) must have at least 2 pointers,
7995 * otherwise the tree is invalid.
7997 ptop = mc->mc_top-1;
7998 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8000 /* Leaf page fill factor is below the threshold.
8001 * Try to move keys from left or right neighbor, or
8002 * merge with a neighbor page.
8007 mdb_cursor_copy(mc, &mn);
8008 mn.mc_xcursor = NULL;
8010 oldki = mc->mc_ki[mc->mc_top];
8011 if (mc->mc_ki[ptop] == 0) {
8012 /* We're the leftmost leaf in our parent.
8014 DPUTS("reading right neighbor");
8016 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8017 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8020 mn.mc_ki[mn.mc_top] = 0;
8021 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8023 /* There is at least one neighbor to the left.
8025 DPUTS("reading left neighbor");
8027 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8028 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8031 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8032 mc->mc_ki[mc->mc_top] = 0;
8035 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8036 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8037 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8039 /* If the neighbor page is above threshold and has enough keys,
8040 * move one key from it. Otherwise we should try to merge them.
8041 * (A branch page must never have less than 2 keys.)
8043 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8044 rc = mdb_node_move(&mn, mc);
8045 if (!mc->mc_ki[mc->mc_top]) {
8046 /* if we inserted on left, bump position up */
8050 if (mc->mc_ki[ptop] == 0) {
8051 rc = mdb_page_merge(&mn, mc);
8054 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8055 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8056 /* We want mdb_rebalance to find mn when doing fixups */
8057 if (mc->mc_flags & C_SUB) {
8058 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8059 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8060 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8062 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8063 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8065 rc = mdb_page_merge(mc, &mn);
8066 if (mc->mc_flags & C_SUB)
8067 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8069 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8070 mdb_cursor_copy(&mn, mc);
8072 mc->mc_flags &= ~C_EOF;
8074 mc->mc_ki[mc->mc_top] = oldki;
8078 /** Complete a delete operation started by #mdb_cursor_del(). */
8080 mdb_cursor_del0(MDB_cursor *mc)
8086 MDB_cursor *m2, *m3;
8087 MDB_dbi dbi = mc->mc_dbi;
8089 ki = mc->mc_ki[mc->mc_top];
8090 mp = mc->mc_pg[mc->mc_top];
8091 mdb_node_del(mc, mc->mc_db->md_pad);
8092 mc->mc_db->md_entries--;
8094 /* Adjust other cursors pointing to mp */
8095 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8096 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8097 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8099 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8101 if (m3->mc_pg[mc->mc_top] == mp) {
8102 if (m3->mc_ki[mc->mc_top] >= ki) {
8103 m3->mc_flags |= C_DEL;
8104 if (m3->mc_ki[mc->mc_top] > ki)
8105 m3->mc_ki[mc->mc_top]--;
8106 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8107 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8112 rc = mdb_rebalance(mc);
8114 if (rc == MDB_SUCCESS) {
8115 /* DB is totally empty now, just bail out.
8116 * Other cursors adjustments were already done
8117 * by mdb_rebalance and aren't needed here.
8122 mp = mc->mc_pg[mc->mc_top];
8123 nkeys = NUMKEYS(mp);
8125 /* Adjust other cursors pointing to mp */
8126 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8127 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8128 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8130 if (m3->mc_snum < mc->mc_snum)
8132 if (m3->mc_pg[mc->mc_top] == mp) {
8133 /* if m3 points past last node in page, find next sibling */
8134 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8135 rc = mdb_cursor_sibling(m3, 1);
8136 if (rc == MDB_NOTFOUND) {
8137 m3->mc_flags |= C_EOF;
8143 mc->mc_flags |= C_DEL;
8147 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8152 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8153 MDB_val *key, MDB_val *data)
8155 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8158 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8159 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8161 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8162 /* must ignore any data */
8166 return mdb_del0(txn, dbi, key, data, 0);
8170 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8171 MDB_val *key, MDB_val *data, unsigned flags)
8176 MDB_val rdata, *xdata;
8180 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8182 mdb_cursor_init(&mc, txn, dbi, &mx);
8191 flags |= MDB_NODUPDATA;
8193 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8195 /* let mdb_page_split know about this cursor if needed:
8196 * delete will trigger a rebalance; if it needs to move
8197 * a node from one page to another, it will have to
8198 * update the parent's separator key(s). If the new sepkey
8199 * is larger than the current one, the parent page may
8200 * run out of space, triggering a split. We need this
8201 * cursor to be consistent until the end of the rebalance.
8203 mc.mc_flags |= C_UNTRACK;
8204 mc.mc_next = txn->mt_cursors[dbi];
8205 txn->mt_cursors[dbi] = &mc;
8206 rc = mdb_cursor_del(&mc, flags);
8207 txn->mt_cursors[dbi] = mc.mc_next;
8212 /** Split a page and insert a new node.
8213 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8214 * The cursor will be updated to point to the actual page and index where
8215 * the node got inserted after the split.
8216 * @param[in] newkey The key for the newly inserted node.
8217 * @param[in] newdata The data for the newly inserted node.
8218 * @param[in] newpgno The page number, if the new node is a branch node.
8219 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8220 * @return 0 on success, non-zero on failure.
8223 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8224 unsigned int nflags)
8227 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8230 int i, j, split_indx, nkeys, pmax;
8231 MDB_env *env = mc->mc_txn->mt_env;
8233 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8234 MDB_page *copy = NULL;
8235 MDB_page *mp, *rp, *pp;
8240 mp = mc->mc_pg[mc->mc_top];
8241 newindx = mc->mc_ki[mc->mc_top];
8242 nkeys = NUMKEYS(mp);
8244 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8245 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8246 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8248 /* Create a right sibling. */
8249 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8251 rp->mp_pad = mp->mp_pad;
8252 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8254 /* Usually when splitting the root page, the cursor
8255 * height is 1. But when called from mdb_update_key,
8256 * the cursor height may be greater because it walks
8257 * up the stack while finding the branch slot to update.
8259 if (mc->mc_top < 1) {
8260 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8262 /* shift current top to make room for new parent */
8263 for (i=mc->mc_snum; i>0; i--) {
8264 mc->mc_pg[i] = mc->mc_pg[i-1];
8265 mc->mc_ki[i] = mc->mc_ki[i-1];
8269 mc->mc_db->md_root = pp->mp_pgno;
8270 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8271 new_root = mc->mc_db->md_depth++;
8273 /* Add left (implicit) pointer. */
8274 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8275 /* undo the pre-push */
8276 mc->mc_pg[0] = mc->mc_pg[1];
8277 mc->mc_ki[0] = mc->mc_ki[1];
8278 mc->mc_db->md_root = mp->mp_pgno;
8279 mc->mc_db->md_depth--;
8286 ptop = mc->mc_top-1;
8287 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8290 mc->mc_flags |= C_SPLITTING;
8291 mdb_cursor_copy(mc, &mn);
8292 mn.mc_pg[mn.mc_top] = rp;
8293 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8295 if (nflags & MDB_APPEND) {
8296 mn.mc_ki[mn.mc_top] = 0;
8298 split_indx = newindx;
8302 split_indx = (nkeys+1) / 2;
8307 unsigned int lsize, rsize, ksize;
8308 /* Move half of the keys to the right sibling */
8309 x = mc->mc_ki[mc->mc_top] - split_indx;
8310 ksize = mc->mc_db->md_pad;
8311 split = LEAF2KEY(mp, split_indx, ksize);
8312 rsize = (nkeys - split_indx) * ksize;
8313 lsize = (nkeys - split_indx) * sizeof(indx_t);
8314 mp->mp_lower -= lsize;
8315 rp->mp_lower += lsize;
8316 mp->mp_upper += rsize - lsize;
8317 rp->mp_upper -= rsize - lsize;
8318 sepkey.mv_size = ksize;
8319 if (newindx == split_indx) {
8320 sepkey.mv_data = newkey->mv_data;
8322 sepkey.mv_data = split;
8325 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8326 memcpy(rp->mp_ptrs, split, rsize);
8327 sepkey.mv_data = rp->mp_ptrs;
8328 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8329 memcpy(ins, newkey->mv_data, ksize);
8330 mp->mp_lower += sizeof(indx_t);
8331 mp->mp_upper -= ksize - sizeof(indx_t);
8334 memcpy(rp->mp_ptrs, split, x * ksize);
8335 ins = LEAF2KEY(rp, x, ksize);
8336 memcpy(ins, newkey->mv_data, ksize);
8337 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8338 rp->mp_lower += sizeof(indx_t);
8339 rp->mp_upper -= ksize - sizeof(indx_t);
8340 mc->mc_ki[mc->mc_top] = x;
8341 mc->mc_pg[mc->mc_top] = rp;
8345 int psize, nsize, k;
8346 /* Maximum free space in an empty page */
8347 pmax = env->me_psize - PAGEHDRSZ;
8349 nsize = mdb_leaf_size(env, newkey, newdata);
8351 nsize = mdb_branch_size(env, newkey);
8352 nsize = EVEN(nsize);
8354 /* grab a page to hold a temporary copy */
8355 copy = mdb_page_malloc(mc->mc_txn, 1);
8360 copy->mp_pgno = mp->mp_pgno;
8361 copy->mp_flags = mp->mp_flags;
8362 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8363 copy->mp_upper = env->me_psize - PAGEBASE;
8365 /* prepare to insert */
8366 for (i=0, j=0; i<nkeys; i++) {
8368 copy->mp_ptrs[j++] = 0;
8370 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8373 /* When items are relatively large the split point needs
8374 * to be checked, because being off-by-one will make the
8375 * difference between success or failure in mdb_node_add.
8377 * It's also relevant if a page happens to be laid out
8378 * such that one half of its nodes are all "small" and
8379 * the other half of its nodes are "large." If the new
8380 * item is also "large" and falls on the half with
8381 * "large" nodes, it also may not fit.
8383 * As a final tweak, if the new item goes on the last
8384 * spot on the page (and thus, onto the new page), bias
8385 * the split so the new page is emptier than the old page.
8386 * This yields better packing during sequential inserts.
8388 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8389 /* Find split point */
8391 if (newindx <= split_indx || newindx >= nkeys) {
8393 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8398 for (; i!=k; i+=j) {
8403 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8404 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8406 if (F_ISSET(node->mn_flags, F_BIGDATA))
8407 psize += sizeof(pgno_t);
8409 psize += NODEDSZ(node);
8411 psize = EVEN(psize);
8413 if (psize > pmax || i == k-j) {
8414 split_indx = i + (j<0);
8419 if (split_indx == newindx) {
8420 sepkey.mv_size = newkey->mv_size;
8421 sepkey.mv_data = newkey->mv_data;
8423 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8424 sepkey.mv_size = node->mn_ksize;
8425 sepkey.mv_data = NODEKEY(node);
8430 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8432 /* Copy separator key to the parent.
8434 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8438 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8443 if (mn.mc_snum == mc->mc_snum) {
8444 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8445 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8446 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8447 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8452 /* Right page might now have changed parent.
8453 * Check if left page also changed parent.
8455 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8456 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8457 for (i=0; i<ptop; i++) {
8458 mc->mc_pg[i] = mn.mc_pg[i];
8459 mc->mc_ki[i] = mn.mc_ki[i];
8461 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8462 if (mn.mc_ki[ptop]) {
8463 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8465 /* find right page's left sibling */
8466 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8467 mdb_cursor_sibling(mc, 0);
8472 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8475 mc->mc_flags ^= C_SPLITTING;
8476 if (rc != MDB_SUCCESS) {
8479 if (nflags & MDB_APPEND) {
8480 mc->mc_pg[mc->mc_top] = rp;
8481 mc->mc_ki[mc->mc_top] = 0;
8482 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8485 for (i=0; i<mc->mc_top; i++)
8486 mc->mc_ki[i] = mn.mc_ki[i];
8487 } else if (!IS_LEAF2(mp)) {
8489 mc->mc_pg[mc->mc_top] = rp;
8494 rkey.mv_data = newkey->mv_data;
8495 rkey.mv_size = newkey->mv_size;
8501 /* Update index for the new key. */
8502 mc->mc_ki[mc->mc_top] = j;
8504 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8505 rkey.mv_data = NODEKEY(node);
8506 rkey.mv_size = node->mn_ksize;
8508 xdata.mv_data = NODEDATA(node);
8509 xdata.mv_size = NODEDSZ(node);
8512 pgno = NODEPGNO(node);
8513 flags = node->mn_flags;
8516 if (!IS_LEAF(mp) && j == 0) {
8517 /* First branch index doesn't need key data. */
8521 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8527 mc->mc_pg[mc->mc_top] = copy;
8532 } while (i != split_indx);
8534 nkeys = NUMKEYS(copy);
8535 for (i=0; i<nkeys; i++)
8536 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8537 mp->mp_lower = copy->mp_lower;
8538 mp->mp_upper = copy->mp_upper;
8539 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8540 env->me_psize - copy->mp_upper - PAGEBASE);
8542 /* reset back to original page */
8543 if (newindx < split_indx) {
8544 mc->mc_pg[mc->mc_top] = mp;
8545 if (nflags & MDB_RESERVE) {
8546 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8547 if (!(node->mn_flags & F_BIGDATA))
8548 newdata->mv_data = NODEDATA(node);
8551 mc->mc_pg[mc->mc_top] = rp;
8553 /* Make sure mc_ki is still valid.
8555 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8556 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8557 for (i=0; i<=ptop; i++) {
8558 mc->mc_pg[i] = mn.mc_pg[i];
8559 mc->mc_ki[i] = mn.mc_ki[i];
8566 /* Adjust other cursors pointing to mp */
8567 MDB_cursor *m2, *m3;
8568 MDB_dbi dbi = mc->mc_dbi;
8569 int fixup = NUMKEYS(mp);
8571 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8572 if (mc->mc_flags & C_SUB)
8573 m3 = &m2->mc_xcursor->mx_cursor;
8578 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8580 if (m3->mc_flags & C_SPLITTING)
8585 for (k=new_root; k>=0; k--) {
8586 m3->mc_ki[k+1] = m3->mc_ki[k];
8587 m3->mc_pg[k+1] = m3->mc_pg[k];
8589 if (m3->mc_ki[0] >= split_indx) {
8594 m3->mc_pg[0] = mc->mc_pg[0];
8598 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8599 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8600 m3->mc_ki[mc->mc_top]++;
8601 if (m3->mc_ki[mc->mc_top] >= fixup) {
8602 m3->mc_pg[mc->mc_top] = rp;
8603 m3->mc_ki[mc->mc_top] -= fixup;
8604 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8606 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8607 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8612 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8615 if (copy) /* tmp page */
8616 mdb_page_free(env, copy);
8618 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8623 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8624 MDB_val *key, MDB_val *data, unsigned int flags)
8629 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8632 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8635 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8636 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8638 mdb_cursor_init(&mc, txn, dbi, &mx);
8639 return mdb_cursor_put(&mc, key, data, flags);
8643 #define MDB_WBUF (1024*1024)
8646 /** State needed for a compacting copy. */
8647 typedef struct mdb_copy {
8648 pthread_mutex_t mc_mutex;
8649 pthread_cond_t mc_cond;
8656 pgno_t mc_next_pgno;
8659 volatile int mc_new;
8664 /** Dedicated writer thread for compacting copy. */
8665 static THREAD_RET ESECT CALL_CONV
8666 mdb_env_copythr(void *arg)
8670 int toggle = 0, wsize, rc;
8673 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8676 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8679 pthread_mutex_lock(&my->mc_mutex);
8681 pthread_cond_signal(&my->mc_cond);
8684 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8685 if (my->mc_new < 0) {
8690 wsize = my->mc_wlen[toggle];
8691 ptr = my->mc_wbuf[toggle];
8694 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8698 } else if (len > 0) {
8712 /* If there's an overflow page tail, write it too */
8713 if (my->mc_olen[toggle]) {
8714 wsize = my->mc_olen[toggle];
8715 ptr = my->mc_over[toggle];
8716 my->mc_olen[toggle] = 0;
8719 my->mc_wlen[toggle] = 0;
8721 pthread_cond_signal(&my->mc_cond);
8723 pthread_cond_signal(&my->mc_cond);
8724 pthread_mutex_unlock(&my->mc_mutex);
8725 return (THREAD_RET)0;
8729 /** Tell the writer thread there's a buffer ready to write */
8731 mdb_env_cthr_toggle(mdb_copy *my, int st)
8733 int toggle = my->mc_toggle ^ 1;
8734 pthread_mutex_lock(&my->mc_mutex);
8735 if (my->mc_status) {
8736 pthread_mutex_unlock(&my->mc_mutex);
8737 return my->mc_status;
8739 while (my->mc_new == 1)
8740 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8742 my->mc_toggle = toggle;
8743 pthread_cond_signal(&my->mc_cond);
8744 pthread_mutex_unlock(&my->mc_mutex);
8748 /** Depth-first tree traversal for compacting copy. */
8750 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8753 MDB_txn *txn = my->mc_txn;
8755 MDB_page *mo, *mp, *leaf;
8760 /* Empty DB, nothing to do */
8761 if (*pg == P_INVALID)
8768 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8771 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8775 /* Make cursor pages writable */
8776 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8780 for (i=0; i<mc.mc_top; i++) {
8781 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8782 mc.mc_pg[i] = (MDB_page *)ptr;
8783 ptr += my->mc_env->me_psize;
8786 /* This is writable space for a leaf page. Usually not needed. */
8787 leaf = (MDB_page *)ptr;
8789 toggle = my->mc_toggle;
8790 while (mc.mc_snum > 0) {
8792 mp = mc.mc_pg[mc.mc_top];
8796 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8797 for (i=0; i<n; i++) {
8798 ni = NODEPTR(mp, i);
8799 if (ni->mn_flags & F_BIGDATA) {
8803 /* Need writable leaf */
8805 mc.mc_pg[mc.mc_top] = leaf;
8806 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8808 ni = NODEPTR(mp, i);
8811 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8812 rc = mdb_page_get(txn, pg, &omp, NULL);
8815 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8816 rc = mdb_env_cthr_toggle(my, 1);
8819 toggle = my->mc_toggle;
8821 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8822 memcpy(mo, omp, my->mc_env->me_psize);
8823 mo->mp_pgno = my->mc_next_pgno;
8824 my->mc_next_pgno += omp->mp_pages;
8825 my->mc_wlen[toggle] += my->mc_env->me_psize;
8826 if (omp->mp_pages > 1) {
8827 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8828 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8829 rc = mdb_env_cthr_toggle(my, 1);
8832 toggle = my->mc_toggle;
8834 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8835 } else if (ni->mn_flags & F_SUBDATA) {
8838 /* Need writable leaf */
8840 mc.mc_pg[mc.mc_top] = leaf;
8841 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8843 ni = NODEPTR(mp, i);
8846 memcpy(&db, NODEDATA(ni), sizeof(db));
8847 my->mc_toggle = toggle;
8848 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8851 toggle = my->mc_toggle;
8852 memcpy(NODEDATA(ni), &db, sizeof(db));
8857 mc.mc_ki[mc.mc_top]++;
8858 if (mc.mc_ki[mc.mc_top] < n) {
8861 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8863 rc = mdb_page_get(txn, pg, &mp, NULL);
8868 mc.mc_ki[mc.mc_top] = 0;
8869 if (IS_BRANCH(mp)) {
8870 /* Whenever we advance to a sibling branch page,
8871 * we must proceed all the way down to its first leaf.
8873 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8876 mc.mc_pg[mc.mc_top] = mp;
8880 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8881 rc = mdb_env_cthr_toggle(my, 1);
8884 toggle = my->mc_toggle;
8886 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8887 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8888 mo->mp_pgno = my->mc_next_pgno++;
8889 my->mc_wlen[toggle] += my->mc_env->me_psize;
8891 /* Update parent if there is one */
8892 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8893 SETPGNO(ni, mo->mp_pgno);
8894 mdb_cursor_pop(&mc);
8896 /* Otherwise we're done */
8906 /** Copy environment with compaction. */
8908 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8913 MDB_txn *txn = NULL;
8918 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8919 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8920 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8921 if (my.mc_wbuf[0] == NULL)
8924 pthread_mutex_init(&my.mc_mutex, NULL);
8925 pthread_cond_init(&my.mc_cond, NULL);
8926 #ifdef HAVE_MEMALIGN
8927 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8928 if (my.mc_wbuf[0] == NULL)
8931 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8936 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8937 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8942 my.mc_next_pgno = NUM_METAS;
8948 THREAD_CREATE(thr, mdb_env_copythr, &my);
8950 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8954 mp = (MDB_page *)my.mc_wbuf[0];
8955 memset(mp, 0, NUM_METAS * env->me_psize);
8957 mp->mp_flags = P_META;
8958 mm = (MDB_meta *)METADATA(mp);
8959 mdb_env_init_meta0(env, mm);
8960 mm->mm_address = env->me_metas[0]->mm_address;
8962 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8964 mp->mp_flags = P_META;
8965 *(MDB_meta *)METADATA(mp) = *mm;
8966 mm = (MDB_meta *)METADATA(mp);
8968 /* Count the number of free pages, subtract from lastpg to find
8969 * number of active pages
8972 MDB_ID freecount = 0;
8975 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8976 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8977 freecount += *(MDB_ID *)data.mv_data;
8978 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
8979 txn->mt_dbs[FREE_DBI].md_leaf_pages +
8980 txn->mt_dbs[FREE_DBI].md_overflow_pages;
8982 /* Set metapage 1 */
8983 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8984 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
8985 if (mm->mm_last_pg > NUM_METAS-1) {
8986 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
8989 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
8992 my.mc_wlen[0] = env->me_psize * NUM_METAS;
8994 pthread_mutex_lock(&my.mc_mutex);
8996 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8997 pthread_mutex_unlock(&my.mc_mutex);
8998 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
8999 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
9000 rc = mdb_env_cthr_toggle(&my, 1);
9001 mdb_env_cthr_toggle(&my, -1);
9002 pthread_mutex_lock(&my.mc_mutex);
9004 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9005 pthread_mutex_unlock(&my.mc_mutex);
9010 CloseHandle(my.mc_cond);
9011 CloseHandle(my.mc_mutex);
9012 _aligned_free(my.mc_wbuf[0]);
9014 pthread_cond_destroy(&my.mc_cond);
9015 pthread_mutex_destroy(&my.mc_mutex);
9016 free(my.mc_wbuf[0]);
9021 /** Copy environment as-is. */
9023 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9025 MDB_txn *txn = NULL;
9026 mdb_mutexref_t wmutex = NULL;
9032 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9036 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9039 /* Do the lock/unlock of the reader mutex before starting the
9040 * write txn. Otherwise other read txns could block writers.
9042 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9047 /* We must start the actual read txn after blocking writers */
9048 mdb_txn_end(txn, MDB_END_RESET_TMP);
9050 /* Temporarily block writers until we snapshot the meta pages */
9051 wmutex = env->me_wmutex;
9052 if (LOCK_MUTEX(rc, env, wmutex))
9055 rc = mdb_txn_renew0(txn);
9057 UNLOCK_MUTEX(wmutex);
9062 wsize = env->me_psize * NUM_METAS;
9066 DO_WRITE(rc, fd, ptr, w2, len);
9070 } else if (len > 0) {
9076 /* Non-blocking or async handles are not supported */
9082 UNLOCK_MUTEX(wmutex);
9087 w2 = txn->mt_next_pgno * env->me_psize;
9090 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9097 if (wsize > MAX_WRITE)
9101 DO_WRITE(rc, fd, ptr, w2, len);
9105 } else if (len > 0) {
9122 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9124 if (flags & MDB_CP_COMPACT)
9125 return mdb_env_copyfd1(env, fd);
9127 return mdb_env_copyfd0(env, fd);
9131 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9133 return mdb_env_copyfd2(env, fd, 0);
9137 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9141 HANDLE newfd = INVALID_HANDLE_VALUE;
9143 if (env->me_flags & MDB_NOSUBDIR) {
9144 lpath = (char *)path;
9147 len += sizeof(DATANAME);
9148 lpath = malloc(len);
9151 sprintf(lpath, "%s" DATANAME, path);
9154 /* The destination path must exist, but the destination file must not.
9155 * We don't want the OS to cache the writes, since the source data is
9156 * already in the OS cache.
9159 newfd = CreateFileA(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9160 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9162 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9164 if (newfd == INVALID_HANDLE_VALUE) {
9169 if (env->me_psize >= env->me_os_psize) {
9171 /* Set O_DIRECT if the file system supports it */
9172 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9173 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9175 #ifdef F_NOCACHE /* __APPLE__ */
9176 rc = fcntl(newfd, F_NOCACHE, 1);
9184 rc = mdb_env_copyfd2(env, newfd, flags);
9187 if (!(env->me_flags & MDB_NOSUBDIR))
9189 if (newfd != INVALID_HANDLE_VALUE)
9190 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9197 mdb_env_copy(MDB_env *env, const char *path)
9199 return mdb_env_copy2(env, path, 0);
9203 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9205 if (flag & ~CHANGEABLE)
9208 env->me_flags |= flag;
9210 env->me_flags &= ~flag;
9215 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9220 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9225 mdb_env_set_userctx(MDB_env *env, void *ctx)
9229 env->me_userctx = ctx;
9234 mdb_env_get_userctx(MDB_env *env)
9236 return env ? env->me_userctx : NULL;
9240 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9245 env->me_assert_func = func;
9251 mdb_env_get_path(MDB_env *env, const char **arg)
9256 *arg = env->me_path;
9261 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9270 /** Common code for #mdb_stat() and #mdb_env_stat().
9271 * @param[in] env the environment to operate in.
9272 * @param[in] db the #MDB_db record containing the stats to return.
9273 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9274 * @return 0, this function always succeeds.
9277 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9279 arg->ms_psize = env->me_psize;
9280 arg->ms_depth = db->md_depth;
9281 arg->ms_branch_pages = db->md_branch_pages;
9282 arg->ms_leaf_pages = db->md_leaf_pages;
9283 arg->ms_overflow_pages = db->md_overflow_pages;
9284 arg->ms_entries = db->md_entries;
9290 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9294 if (env == NULL || arg == NULL)
9297 meta = mdb_env_pick_meta(env);
9299 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9303 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9307 if (env == NULL || arg == NULL)
9310 meta = mdb_env_pick_meta(env);
9311 arg->me_mapaddr = meta->mm_address;
9312 arg->me_last_pgno = meta->mm_last_pg;
9313 arg->me_last_txnid = meta->mm_txnid;
9315 arg->me_mapsize = env->me_mapsize;
9316 arg->me_maxreaders = env->me_maxreaders;
9317 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9321 /** Set the default comparison functions for a database.
9322 * Called immediately after a database is opened to set the defaults.
9323 * The user can then override them with #mdb_set_compare() or
9324 * #mdb_set_dupsort().
9325 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9326 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9329 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9331 uint16_t f = txn->mt_dbs[dbi].md_flags;
9333 txn->mt_dbxs[dbi].md_cmp =
9334 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9335 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9337 txn->mt_dbxs[dbi].md_dcmp =
9338 !(f & MDB_DUPSORT) ? 0 :
9339 ((f & MDB_INTEGERDUP)
9340 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9341 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9344 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9350 int rc, dbflag, exact;
9351 unsigned int unused = 0, seq;
9354 if (flags & ~VALID_FLAGS)
9356 if (txn->mt_flags & MDB_TXN_BLOCKED)
9362 if (flags & PERSISTENT_FLAGS) {
9363 uint16_t f2 = flags & PERSISTENT_FLAGS;
9364 /* make sure flag changes get committed */
9365 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9366 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9367 txn->mt_flags |= MDB_TXN_DIRTY;
9370 mdb_default_cmp(txn, MAIN_DBI);
9374 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9375 mdb_default_cmp(txn, MAIN_DBI);
9378 /* Is the DB already open? */
9380 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9381 if (!txn->mt_dbxs[i].md_name.mv_size) {
9382 /* Remember this free slot */
9383 if (!unused) unused = i;
9386 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9387 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9393 /* If no free slot and max hit, fail */
9394 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9395 return MDB_DBS_FULL;
9397 /* Cannot mix named databases with some mainDB flags */
9398 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9399 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9401 /* Find the DB info */
9402 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9405 key.mv_data = (void *)name;
9406 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9407 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9408 if (rc == MDB_SUCCESS) {
9409 /* make sure this is actually a DB */
9410 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9411 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9412 return MDB_INCOMPATIBLE;
9413 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9414 /* Create if requested */
9415 data.mv_size = sizeof(MDB_db);
9416 data.mv_data = &dummy;
9417 memset(&dummy, 0, sizeof(dummy));
9418 dummy.md_root = P_INVALID;
9419 dummy.md_flags = flags & PERSISTENT_FLAGS;
9420 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9424 /* OK, got info, add to table */
9425 if (rc == MDB_SUCCESS) {
9426 unsigned int slot = unused ? unused : txn->mt_numdbs;
9427 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9428 txn->mt_dbxs[slot].md_name.mv_size = len;
9429 txn->mt_dbxs[slot].md_rel = NULL;
9430 txn->mt_dbflags[slot] = dbflag;
9431 /* txn-> and env-> are the same in read txns, use
9432 * tmp variable to avoid undefined assignment
9434 seq = ++txn->mt_env->me_dbiseqs[slot];
9435 txn->mt_dbiseqs[slot] = seq;
9437 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9439 mdb_default_cmp(txn, slot);
9449 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9451 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9454 if (txn->mt_flags & MDB_TXN_BLOCKED)
9457 if (txn->mt_dbflags[dbi] & DB_STALE) {
9460 /* Stale, must read the DB's root. cursor_init does it for us. */
9461 mdb_cursor_init(&mc, txn, dbi, &mx);
9463 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9466 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9469 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9471 ptr = env->me_dbxs[dbi].md_name.mv_data;
9472 /* If there was no name, this was already closed */
9474 env->me_dbxs[dbi].md_name.mv_data = NULL;
9475 env->me_dbxs[dbi].md_name.mv_size = 0;
9476 env->me_dbflags[dbi] = 0;
9477 env->me_dbiseqs[dbi]++;
9482 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9484 /* We could return the flags for the FREE_DBI too but what's the point? */
9485 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9487 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9491 /** Add all the DB's pages to the free list.
9492 * @param[in] mc Cursor on the DB to free.
9493 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9494 * @return 0 on success, non-zero on failure.
9497 mdb_drop0(MDB_cursor *mc, int subs)
9501 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9502 if (rc == MDB_SUCCESS) {
9503 MDB_txn *txn = mc->mc_txn;
9508 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9509 * This also avoids any P_LEAF2 pages, which have no nodes.
9511 if (mc->mc_flags & C_SUB)
9514 mdb_cursor_copy(mc, &mx);
9515 while (mc->mc_snum > 0) {
9516 MDB_page *mp = mc->mc_pg[mc->mc_top];
9517 unsigned n = NUMKEYS(mp);
9519 for (i=0; i<n; i++) {
9520 ni = NODEPTR(mp, i);
9521 if (ni->mn_flags & F_BIGDATA) {
9524 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9525 rc = mdb_page_get(txn, pg, &omp, NULL);
9528 mdb_cassert(mc, IS_OVERFLOW(omp));
9529 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9533 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9534 mdb_xcursor_init1(mc, ni);
9535 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9541 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9543 for (i=0; i<n; i++) {
9545 ni = NODEPTR(mp, i);
9548 mdb_midl_xappend(txn->mt_free_pgs, pg);
9553 mc->mc_ki[mc->mc_top] = i;
9554 rc = mdb_cursor_sibling(mc, 1);
9556 if (rc != MDB_NOTFOUND)
9558 /* no more siblings, go back to beginning
9559 * of previous level.
9563 for (i=1; i<mc->mc_snum; i++) {
9565 mc->mc_pg[i] = mx.mc_pg[i];
9570 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9573 txn->mt_flags |= MDB_TXN_ERROR;
9574 } else if (rc == MDB_NOTFOUND) {
9580 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9582 MDB_cursor *mc, *m2;
9585 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9588 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9591 if (TXN_DBI_CHANGED(txn, dbi))
9594 rc = mdb_cursor_open(txn, dbi, &mc);
9598 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9599 /* Invalidate the dropped DB's cursors */
9600 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9601 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9605 /* Can't delete the main DB */
9606 if (del && dbi >= CORE_DBS) {
9607 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9609 txn->mt_dbflags[dbi] = DB_STALE;
9610 mdb_dbi_close(txn->mt_env, dbi);
9612 txn->mt_flags |= MDB_TXN_ERROR;
9615 /* reset the DB record, mark it dirty */
9616 txn->mt_dbflags[dbi] |= DB_DIRTY;
9617 txn->mt_dbs[dbi].md_depth = 0;
9618 txn->mt_dbs[dbi].md_branch_pages = 0;
9619 txn->mt_dbs[dbi].md_leaf_pages = 0;
9620 txn->mt_dbs[dbi].md_overflow_pages = 0;
9621 txn->mt_dbs[dbi].md_entries = 0;
9622 txn->mt_dbs[dbi].md_root = P_INVALID;
9624 txn->mt_flags |= MDB_TXN_DIRTY;
9627 mdb_cursor_close(mc);
9631 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9633 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9636 txn->mt_dbxs[dbi].md_cmp = cmp;
9640 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9642 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9645 txn->mt_dbxs[dbi].md_dcmp = cmp;
9649 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9651 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9654 txn->mt_dbxs[dbi].md_rel = rel;
9658 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9660 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9663 txn->mt_dbxs[dbi].md_relctx = ctx;
9668 mdb_env_get_maxkeysize(MDB_env *env)
9670 return ENV_MAXKEY(env);
9674 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9676 unsigned int i, rdrs;
9679 int rc = 0, first = 1;
9683 if (!env->me_txns) {
9684 return func("(no reader locks)\n", ctx);
9686 rdrs = env->me_txns->mti_numreaders;
9687 mr = env->me_txns->mti_readers;
9688 for (i=0; i<rdrs; i++) {
9690 txnid_t txnid = mr[i].mr_txnid;
9691 sprintf(buf, txnid == (txnid_t)-1 ?
9692 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9693 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9696 rc = func(" pid thread txnid\n", ctx);
9700 rc = func(buf, ctx);
9706 rc = func("(no active readers)\n", ctx);
9711 /** Insert pid into list if not already present.
9712 * return -1 if already present.
9715 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9717 /* binary search of pid in list */
9719 unsigned cursor = 1;
9721 unsigned n = ids[0];
9724 unsigned pivot = n >> 1;
9725 cursor = base + pivot + 1;
9726 val = pid - ids[cursor];
9731 } else if ( val > 0 ) {
9736 /* found, so it's a duplicate */
9745 for (n = ids[0]; n > cursor; n--)
9752 mdb_reader_check(MDB_env *env, int *dead)
9758 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9761 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9763 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9765 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9766 unsigned int i, j, rdrs;
9768 MDB_PID_T *pids, pid;
9769 int rc = MDB_SUCCESS, count = 0;
9771 rdrs = env->me_txns->mti_numreaders;
9772 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9776 mr = env->me_txns->mti_readers;
9777 for (i=0; i<rdrs; i++) {
9779 if (pid && pid != env->me_pid) {
9780 if (mdb_pid_insert(pids, pid) == 0) {
9781 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9782 /* Stale reader found */
9785 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9786 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9788 rdrs = 0; /* the above checked all readers */
9790 /* Recheck, a new process may have reused pid */
9791 if (mdb_reader_pid(env, Pidcheck, pid))
9796 if (mr[j].mr_pid == pid) {
9797 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9798 (unsigned) pid, mr[j].mr_txnid));
9803 UNLOCK_MUTEX(rmutex);
9814 #ifdef MDB_ROBUST_SUPPORTED
9815 /** Handle #LOCK_MUTEX0() failure.
9816 * Try to repair the lock file if the mutex owner died.
9817 * @param[in] env the environment handle
9818 * @param[in] mutex LOCK_MUTEX0() mutex
9819 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9820 * @return 0 on success with the mutex locked, or an error code on failure.
9823 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9828 if (rc == MDB_OWNERDEAD) {
9829 /* We own the mutex. Clean up after dead previous owner. */
9831 rlocked = (mutex == env->me_rmutex);
9833 /* Keep mti_txnid updated, otherwise next writer can
9834 * overwrite data which latest meta page refers to.
9836 meta = mdb_env_pick_meta(env);
9837 env->me_txns->mti_txnid = meta->mm_txnid;
9838 /* env is hosed if the dead thread was ours */
9840 env->me_flags |= MDB_FATAL_ERROR;
9845 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9846 (rc ? "this process' env is hosed" : "recovering")));
9847 rc2 = mdb_reader_check0(env, rlocked, NULL);
9849 rc2 = mdb_mutex_consistent(mutex);
9850 if (rc || (rc = rc2)) {
9851 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9852 UNLOCK_MUTEX(mutex);
9858 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9863 #endif /* MDB_ROBUST_SUPPORTED */