2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2015 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments. You can safely
85 * define MDB_FDATASYNC_WORKS if this code will only be run
86 * on kernels 3.6 and newer.
88 #define BROKEN_FDATASYNC
101 #if defined(__sun) || defined(ANDROID)
102 /* Most platforms have posix_memalign, older may only have memalign */
103 #define HAVE_MEMALIGN 1
107 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
108 #include <netinet/in.h>
109 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
112 #if defined(__APPLE__) || defined (BSD)
113 # define MDB_USE_POSIX_SEM 1
114 # define MDB_FDATASYNC fsync
115 #elif defined(ANDROID)
116 # define MDB_FDATASYNC fsync
121 #ifdef MDB_USE_POSIX_SEM
122 # define MDB_USE_HASH 1
123 #include <semaphore.h>
125 #define MDB_USE_POSIX_MUTEX 1
129 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \
130 + defined(MDB_USE_POSIX_MUTEX) != 1
131 # error "Ambiguous shared-lock implementation"
135 #include <valgrind/memcheck.h>
136 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
137 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
138 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
139 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
140 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
142 #define VGMEMP_CREATE(h,r,z)
143 #define VGMEMP_ALLOC(h,a,s)
144 #define VGMEMP_FREE(h,a)
145 #define VGMEMP_DESTROY(h)
146 #define VGMEMP_DEFINED(a,s)
150 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
151 /* Solaris just defines one or the other */
152 # define LITTLE_ENDIAN 1234
153 # define BIG_ENDIAN 4321
154 # ifdef _LITTLE_ENDIAN
155 # define BYTE_ORDER LITTLE_ENDIAN
157 # define BYTE_ORDER BIG_ENDIAN
160 # define BYTE_ORDER __BYTE_ORDER
164 #ifndef LITTLE_ENDIAN
165 #define LITTLE_ENDIAN __LITTLE_ENDIAN
168 #define BIG_ENDIAN __BIG_ENDIAN
171 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
172 #define MISALIGNED_OK 1
178 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
179 # error "Unknown or unsupported endianness (BYTE_ORDER)"
180 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
181 # error "Two's complement, reasonably sized integer types, please"
185 /** Put infrequently used env functions in separate section */
187 # define ESECT __attribute__ ((section("__TEXT,text_env")))
189 # define ESECT __attribute__ ((section("text_env")))
196 #define CALL_CONV WINAPI
201 /** @defgroup internal LMDB Internals
204 /** @defgroup compat Compatibility Macros
205 * A bunch of macros to minimize the amount of platform-specific ifdefs
206 * needed throughout the rest of the code. When the features this library
207 * needs are similar enough to POSIX to be hidden in a one-or-two line
208 * replacement, this macro approach is used.
212 /** Features under development */
217 /** Wrapper around __func__, which is a C99 feature */
218 #if __STDC_VERSION__ >= 199901L
219 # define mdb_func_ __func__
220 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
221 # define mdb_func_ __FUNCTION__
223 /* If a debug message says <mdb_unknown>(), update the #if statements above */
224 # define mdb_func_ "<mdb_unknown>"
227 /* Internal error codes, not exposed outside liblmdb */
228 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
230 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
231 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
232 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
236 #define MDB_ROBUST_SUPPORTED 1
240 #define MDB_USE_HASH 1
241 #define MDB_PIDLOCK 0
242 #define THREAD_RET DWORD
243 #define pthread_t HANDLE
244 #define pthread_mutex_t HANDLE
245 #define pthread_cond_t HANDLE
246 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
247 #define pthread_key_t DWORD
248 #define pthread_self() GetCurrentThreadId()
249 #define pthread_key_create(x,y) \
250 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
251 #define pthread_key_delete(x) TlsFree(x)
252 #define pthread_getspecific(x) TlsGetValue(x)
253 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
254 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
255 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
256 #define pthread_cond_signal(x) SetEvent(*x)
257 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
258 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
259 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
260 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
261 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
262 #define mdb_mutex_consistent(mutex) 0
263 #define getpid() GetCurrentProcessId()
264 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
265 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
266 #define ErrCode() GetLastError()
267 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
268 #define close(fd) (CloseHandle(fd) ? 0 : -1)
269 #define munmap(ptr,len) UnmapViewOfFile(ptr)
270 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
271 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
273 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
277 #define THREAD_RET void *
278 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
279 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
280 #define Z "z" /**< printf format modifier for size_t */
282 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
283 #define MDB_PIDLOCK 1
285 #ifdef MDB_USE_POSIX_SEM
287 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
288 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
289 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
292 mdb_sem_wait(sem_t *sem)
295 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
299 #else /* MDB_USE_POSIX_MUTEX: */
300 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
301 * local variables keep it (mdb_mutexref_t).
303 * When #mdb_mutexref_t is a pointer declaration and #mdb_mutex_t is
304 * not, then it is array[size 1] so it can be assigned to a pointer.
307 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
309 /** Lock the reader or writer mutex.
310 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
312 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
313 /** Unlock the reader or writer mutex.
315 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
316 /** Mark mutex-protected data as repaired, after death of previous owner.
318 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
319 #endif /* MDB_USE_POSIX_SEM */
321 /** Get the error code for the last failed system function.
323 #define ErrCode() errno
325 /** An abstraction for a file handle.
326 * On POSIX systems file handles are small integers. On Windows
327 * they're opaque pointers.
331 /** A value for an invalid file handle.
332 * Mainly used to initialize file variables and signify that they are
335 #define INVALID_HANDLE_VALUE (-1)
337 /** Get the size of a memory page for the system.
338 * This is the basic size that the platform's memory manager uses, and is
339 * fundamental to the use of memory-mapped files.
341 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
344 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
347 #define MNAME_LEN (sizeof(pthread_mutex_t))
352 #ifdef MDB_ROBUST_SUPPORTED
353 /** Lock mutex, handle any error, set rc = result.
354 * Return 0 on success, nonzero (not rc) on error.
356 #define LOCK_MUTEX(rc, env, mutex) \
357 (((rc) = LOCK_MUTEX0(mutex)) && \
358 ((rc) = mdb_mutex_failed(env, mutex, rc)))
359 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
361 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
362 #define mdb_mutex_failed(env, mutex, rc) (rc)
366 /** A flag for opening a file and requesting synchronous data writes.
367 * This is only used when writing a meta page. It's not strictly needed;
368 * we could just do a normal write and then immediately perform a flush.
369 * But if this flag is available it saves us an extra system call.
371 * @note If O_DSYNC is undefined but exists in /usr/include,
372 * preferably set some compiler flag to get the definition.
373 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
376 # define MDB_DSYNC O_DSYNC
380 /** Function for flushing the data of a file. Define this to fsync
381 * if fdatasync() is not supported.
383 #ifndef MDB_FDATASYNC
384 # define MDB_FDATASYNC fdatasync
388 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
399 /** A page number in the database.
400 * Note that 64 bit page numbers are overkill, since pages themselves
401 * already represent 12-13 bits of addressable memory, and the OS will
402 * always limit applications to a maximum of 63 bits of address space.
404 * @note In the #MDB_node structure, we only store 48 bits of this value,
405 * which thus limits us to only 60 bits of addressable data.
407 typedef MDB_ID pgno_t;
409 /** A transaction ID.
410 * See struct MDB_txn.mt_txnid for details.
412 typedef MDB_ID txnid_t;
414 /** @defgroup debug Debug Macros
418 /** Enable debug output. Needs variable argument macros (a C99 feature).
419 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
420 * read from and written to the database (used for free space management).
426 static int mdb_debug;
427 static txnid_t mdb_debug_start;
429 /** Print a debug message with printf formatting.
430 * Requires double parenthesis around 2 or more args.
432 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
433 # define DPRINTF0(fmt, ...) \
434 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
436 # define DPRINTF(args) ((void) 0)
438 /** Print a debug string.
439 * The string is printed literally, with no format processing.
441 #define DPUTS(arg) DPRINTF(("%s", arg))
442 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
444 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
447 /** @brief The maximum size of a database page.
449 * It is 32k or 64k, since value-PAGEBASE must fit in
450 * #MDB_page.%mp_upper.
452 * LMDB will use database pages < OS pages if needed.
453 * That causes more I/O in write transactions: The OS must
454 * know (read) the whole page before writing a partial page.
456 * Note that we don't currently support Huge pages. On Linux,
457 * regular data files cannot use Huge pages, and in general
458 * Huge pages aren't actually pageable. We rely on the OS
459 * demand-pager to read our data and page it out when memory
460 * pressure from other processes is high. So until OSs have
461 * actual paging support for Huge pages, they're not viable.
463 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
465 /** The minimum number of keys required in a database page.
466 * Setting this to a larger value will place a smaller bound on the
467 * maximum size of a data item. Data items larger than this size will
468 * be pushed into overflow pages instead of being stored directly in
469 * the B-tree node. This value used to default to 4. With a page size
470 * of 4096 bytes that meant that any item larger than 1024 bytes would
471 * go into an overflow page. That also meant that on average 2-3KB of
472 * each overflow page was wasted space. The value cannot be lower than
473 * 2 because then there would no longer be a tree structure. With this
474 * value, items larger than 2KB will go into overflow pages, and on
475 * average only 1KB will be wasted.
477 #define MDB_MINKEYS 2
479 /** A stamp that identifies a file as an LMDB file.
480 * There's nothing special about this value other than that it is easily
481 * recognizable, and it will reflect any byte order mismatches.
483 #define MDB_MAGIC 0xBEEFC0DE
485 /** The version number for a database's datafile format. */
486 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
487 /** The version number for a database's lockfile format. */
488 #define MDB_LOCK_VERSION 1
490 /** @brief The max size of a key we can write, or 0 for computed max.
492 * This macro should normally be left alone or set to 0.
493 * Note that a database with big keys or dupsort data cannot be
494 * reliably modified by a liblmdb which uses a smaller max.
495 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
497 * Other values are allowed, for backwards compat. However:
498 * A value bigger than the computed max can break if you do not
499 * know what you are doing, and liblmdb <= 0.9.10 can break when
500 * modifying a DB with keys/dupsort data bigger than its max.
502 * Data items in an #MDB_DUPSORT database are also limited to
503 * this size, since they're actually keys of a sub-DB. Keys and
504 * #MDB_DUPSORT data items must fit on a node in a regular page.
506 #ifndef MDB_MAXKEYSIZE
507 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
510 /** The maximum size of a key we can write to the environment. */
512 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
514 #define ENV_MAXKEY(env) ((env)->me_maxkey)
517 /** @brief The maximum size of a data item.
519 * We only store a 32 bit value for node sizes.
521 #define MAXDATASIZE 0xffffffffUL
524 /** Key size which fits in a #DKBUF.
527 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
530 * This is used for printing a hex dump of a key's contents.
532 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
533 /** Display a key in hex.
535 * Invoke a function to display a key in hex.
537 #define DKEY(x) mdb_dkey(x, kbuf)
543 /** An invalid page number.
544 * Mainly used to denote an empty tree.
546 #define P_INVALID (~(pgno_t)0)
548 /** Test if the flags \b f are set in a flag word \b w. */
549 #define F_ISSET(w, f) (((w) & (f)) == (f))
551 /** Round \b n up to an even number. */
552 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
554 /** Used for offsets within a single page.
555 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
558 typedef uint16_t indx_t;
560 /** Default size of memory map.
561 * This is certainly too small for any actual applications. Apps should always set
562 * the size explicitly using #mdb_env_set_mapsize().
564 #define DEFAULT_MAPSIZE 1048576
566 /** @defgroup readers Reader Lock Table
567 * Readers don't acquire any locks for their data access. Instead, they
568 * simply record their transaction ID in the reader table. The reader
569 * mutex is needed just to find an empty slot in the reader table. The
570 * slot's address is saved in thread-specific data so that subsequent read
571 * transactions started by the same thread need no further locking to proceed.
573 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
575 * No reader table is used if the database is on a read-only filesystem, or
576 * if #MDB_NOLOCK is set.
578 * Since the database uses multi-version concurrency control, readers don't
579 * actually need any locking. This table is used to keep track of which
580 * readers are using data from which old transactions, so that we'll know
581 * when a particular old transaction is no longer in use. Old transactions
582 * that have discarded any data pages can then have those pages reclaimed
583 * for use by a later write transaction.
585 * The lock table is constructed such that reader slots are aligned with the
586 * processor's cache line size. Any slot is only ever used by one thread.
587 * This alignment guarantees that there will be no contention or cache
588 * thrashing as threads update their own slot info, and also eliminates
589 * any need for locking when accessing a slot.
591 * A writer thread will scan every slot in the table to determine the oldest
592 * outstanding reader transaction. Any freed pages older than this will be
593 * reclaimed by the writer. The writer doesn't use any locks when scanning
594 * this table. This means that there's no guarantee that the writer will
595 * see the most up-to-date reader info, but that's not required for correct
596 * operation - all we need is to know the upper bound on the oldest reader,
597 * we don't care at all about the newest reader. So the only consequence of
598 * reading stale information here is that old pages might hang around a
599 * while longer before being reclaimed. That's actually good anyway, because
600 * the longer we delay reclaiming old pages, the more likely it is that a
601 * string of contiguous pages can be found after coalescing old pages from
602 * many old transactions together.
605 /** Number of slots in the reader table.
606 * This value was chosen somewhat arbitrarily. 126 readers plus a
607 * couple mutexes fit exactly into 8KB on my development machine.
608 * Applications should set the table size using #mdb_env_set_maxreaders().
610 #define DEFAULT_READERS 126
612 /** The size of a CPU cache line in bytes. We want our lock structures
613 * aligned to this size to avoid false cache line sharing in the
615 * This value works for most CPUs. For Itanium this should be 128.
621 /** The information we store in a single slot of the reader table.
622 * In addition to a transaction ID, we also record the process and
623 * thread ID that owns a slot, so that we can detect stale information,
624 * e.g. threads or processes that went away without cleaning up.
625 * @note We currently don't check for stale records. We simply re-init
626 * the table when we know that we're the only process opening the
629 typedef struct MDB_rxbody {
630 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
631 * Multiple readers that start at the same time will probably have the
632 * same ID here. Again, it's not important to exclude them from
633 * anything; all we need to know is which version of the DB they
634 * started from so we can avoid overwriting any data used in that
635 * particular version.
637 volatile txnid_t mrb_txnid;
638 /** The process ID of the process owning this reader txn. */
639 volatile MDB_PID_T mrb_pid;
640 /** The thread ID of the thread owning this txn. */
641 volatile MDB_THR_T mrb_tid;
644 /** The actual reader record, with cacheline padding. */
645 typedef struct MDB_reader {
648 /** shorthand for mrb_txnid */
649 #define mr_txnid mru.mrx.mrb_txnid
650 #define mr_pid mru.mrx.mrb_pid
651 #define mr_tid mru.mrx.mrb_tid
652 /** cache line alignment */
653 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
657 /** The header for the reader table.
658 * The table resides in a memory-mapped file. (This is a different file
659 * than is used for the main database.)
661 * For POSIX the actual mutexes reside in the shared memory of this
662 * mapped file. On Windows, mutexes are named objects allocated by the
663 * kernel; we store the mutex names in this mapped file so that other
664 * processes can grab them. This same approach is also used on
665 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
666 * process-shared POSIX mutexes. For these cases where a named object
667 * is used, the object name is derived from a 64 bit FNV hash of the
668 * environment pathname. As such, naming collisions are extremely
669 * unlikely. If a collision occurs, the results are unpredictable.
671 typedef struct MDB_txbody {
672 /** Stamp identifying this as an LMDB file. It must be set
675 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
677 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
678 char mtb_rmname[MNAME_LEN];
680 /** Mutex protecting access to this table.
681 * This is the reader table lock used with LOCK_MUTEX().
683 mdb_mutex_t mtb_rmutex;
685 /** The ID of the last transaction committed to the database.
686 * This is recorded here only for convenience; the value can always
687 * be determined by reading the main database meta pages.
689 volatile txnid_t mtb_txnid;
690 /** The number of slots that have been used in the reader table.
691 * This always records the maximum count, it is not decremented
692 * when readers release their slots.
694 volatile unsigned mtb_numreaders;
697 /** The actual reader table definition. */
698 typedef struct MDB_txninfo {
701 #define mti_magic mt1.mtb.mtb_magic
702 #define mti_format mt1.mtb.mtb_format
703 #define mti_rmutex mt1.mtb.mtb_rmutex
704 #define mti_rmname mt1.mtb.mtb_rmname
705 #define mti_txnid mt1.mtb.mtb_txnid
706 #define mti_numreaders mt1.mtb.mtb_numreaders
707 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
710 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
711 char mt2_wmname[MNAME_LEN];
712 #define mti_wmname mt2.mt2_wmname
714 mdb_mutex_t mt2_wmutex;
715 #define mti_wmutex mt2.mt2_wmutex
717 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
719 MDB_reader mti_readers[1];
722 /** Lockfile format signature: version, features and field layout */
723 #define MDB_LOCK_FORMAT \
725 ((MDB_LOCK_VERSION) \
726 /* Flags which describe functionality */ \
727 + (((MDB_PIDLOCK) != 0) << 16)))
730 /** Common header for all page types.
731 * Overflow records occupy a number of contiguous pages with no
732 * headers on any page after the first.
734 typedef struct MDB_page {
735 #define mp_pgno mp_p.p_pgno
736 #define mp_next mp_p.p_next
738 pgno_t p_pgno; /**< page number */
739 struct MDB_page *p_next; /**< for in-memory list of freed pages */
742 /** @defgroup mdb_page Page Flags
744 * Flags for the page headers.
747 #define P_BRANCH 0x01 /**< branch page */
748 #define P_LEAF 0x02 /**< leaf page */
749 #define P_OVERFLOW 0x04 /**< overflow page */
750 #define P_META 0x08 /**< meta page */
751 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
752 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
753 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
754 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
755 #define P_KEEP 0x8000 /**< leave this page alone during spill */
757 uint16_t mp_flags; /**< @ref mdb_page */
758 #define mp_lower mp_pb.pb.pb_lower
759 #define mp_upper mp_pb.pb.pb_upper
760 #define mp_pages mp_pb.pb_pages
763 indx_t pb_lower; /**< lower bound of free space */
764 indx_t pb_upper; /**< upper bound of free space */
766 uint32_t pb_pages; /**< number of overflow pages */
768 indx_t mp_ptrs[1]; /**< dynamic size */
771 /** Size of the page header, excluding dynamic data at the end */
772 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
774 /** Address of first usable data byte in a page, after the header */
775 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
777 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
778 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
780 /** Number of nodes on a page */
781 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
783 /** The amount of space remaining in the page */
784 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
786 /** The percentage of space used in the page, in tenths of a percent. */
787 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
788 ((env)->me_psize - PAGEHDRSZ))
789 /** The minimum page fill factor, in tenths of a percent.
790 * Pages emptier than this are candidates for merging.
792 #define FILL_THRESHOLD 250
794 /** Test if a page is a leaf page */
795 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
796 /** Test if a page is a LEAF2 page */
797 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
798 /** Test if a page is a branch page */
799 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
800 /** Test if a page is an overflow page */
801 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
802 /** Test if a page is a sub page */
803 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
805 /** The number of overflow pages needed to store the given size. */
806 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
808 /** Link in #MDB_txn.%mt_loose_pgs list */
809 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
811 /** Header for a single key/data pair within a page.
812 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
813 * We guarantee 2-byte alignment for 'MDB_node's.
815 typedef struct MDB_node {
816 /** lo and hi are used for data size on leaf nodes and for
817 * child pgno on branch nodes. On 64 bit platforms, flags
818 * is also used for pgno. (Branch nodes have no flags).
819 * They are in host byte order in case that lets some
820 * accesses be optimized into a 32-bit word access.
822 #if BYTE_ORDER == LITTLE_ENDIAN
823 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
825 unsigned short mn_hi, mn_lo;
827 /** @defgroup mdb_node Node Flags
829 * Flags for node headers.
832 #define F_BIGDATA 0x01 /**< data put on overflow page */
833 #define F_SUBDATA 0x02 /**< data is a sub-database */
834 #define F_DUPDATA 0x04 /**< data has duplicates */
836 /** valid flags for #mdb_node_add() */
837 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
840 unsigned short mn_flags; /**< @ref mdb_node */
841 unsigned short mn_ksize; /**< key size */
842 char mn_data[1]; /**< key and data are appended here */
845 /** Size of the node header, excluding dynamic data at the end */
846 #define NODESIZE offsetof(MDB_node, mn_data)
848 /** Bit position of top word in page number, for shifting mn_flags */
849 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
851 /** Size of a node in a branch page with a given key.
852 * This is just the node header plus the key, there is no data.
854 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
856 /** Size of a node in a leaf page with a given key and data.
857 * This is node header plus key plus data size.
859 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
861 /** Address of node \b i in page \b p */
862 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
864 /** Address of the key for the node */
865 #define NODEKEY(node) (void *)((node)->mn_data)
867 /** Address of the data for a node */
868 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
870 /** Get the page number pointed to by a branch node */
871 #define NODEPGNO(node) \
872 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
873 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
874 /** Set the page number in a branch node */
875 #define SETPGNO(node,pgno) do { \
876 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
877 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
879 /** Get the size of the data in a leaf node */
880 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
881 /** Set the size of the data for a leaf node */
882 #define SETDSZ(node,size) do { \
883 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
884 /** The size of a key in a node */
885 #define NODEKSZ(node) ((node)->mn_ksize)
887 /** Copy a page number from src to dst */
889 #define COPY_PGNO(dst,src) dst = src
891 #if SIZE_MAX > 4294967295UL
892 #define COPY_PGNO(dst,src) do { \
893 unsigned short *s, *d; \
894 s = (unsigned short *)&(src); \
895 d = (unsigned short *)&(dst); \
902 #define COPY_PGNO(dst,src) do { \
903 unsigned short *s, *d; \
904 s = (unsigned short *)&(src); \
905 d = (unsigned short *)&(dst); \
911 /** The address of a key in a LEAF2 page.
912 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
913 * There are no node headers, keys are stored contiguously.
915 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
917 /** Set the \b node's key into \b keyptr, if requested. */
918 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
919 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
921 /** Set the \b node's key into \b key. */
922 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
924 /** Information about a single database in the environment. */
925 typedef struct MDB_db {
926 uint32_t md_pad; /**< also ksize for LEAF2 pages */
927 uint16_t md_flags; /**< @ref mdb_dbi_open */
928 uint16_t md_depth; /**< depth of this tree */
929 pgno_t md_branch_pages; /**< number of internal pages */
930 pgno_t md_leaf_pages; /**< number of leaf pages */
931 pgno_t md_overflow_pages; /**< number of overflow pages */
932 size_t md_entries; /**< number of data items */
933 pgno_t md_root; /**< the root page of this tree */
936 /** mdb_dbi_open flags */
937 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
938 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
939 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
940 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
942 /** Handle for the DB used to track free pages. */
944 /** Handle for the default DB. */
946 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
949 /** Number of meta pages - also hardcoded elsewhere */
952 /** Meta page content.
953 * A meta page is the start point for accessing a database snapshot.
954 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
956 typedef struct MDB_meta {
957 /** Stamp identifying this as an LMDB file. It must be set
960 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
962 void *mm_address; /**< address for fixed mapping */
963 size_t mm_mapsize; /**< size of mmap region */
964 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
965 /** The size of pages used in this DB */
966 #define mm_psize mm_dbs[FREE_DBI].md_pad
967 /** Any persistent environment flags. @ref mdb_env */
968 #define mm_flags mm_dbs[FREE_DBI].md_flags
969 pgno_t mm_last_pg; /**< last used page in file */
970 volatile txnid_t mm_txnid; /**< txnid that committed this page */
973 /** Buffer for a stack-allocated meta page.
974 * The members define size and alignment, and silence type
975 * aliasing warnings. They are not used directly; that could
976 * mean incorrectly using several union members in parallel.
978 typedef union MDB_metabuf {
981 char mm_pad[PAGEHDRSZ];
986 /** Auxiliary DB info.
987 * The information here is mostly static/read-only. There is
988 * only a single copy of this record in the environment.
990 typedef struct MDB_dbx {
991 MDB_val md_name; /**< name of the database */
992 MDB_cmp_func *md_cmp; /**< function for comparing keys */
993 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
994 MDB_rel_func *md_rel; /**< user relocate function */
995 void *md_relctx; /**< user-provided context for md_rel */
998 /** A database transaction.
999 * Every operation requires a transaction handle.
1002 MDB_txn *mt_parent; /**< parent of a nested txn */
1003 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1005 pgno_t mt_next_pgno; /**< next unallocated page */
1006 /** The ID of this transaction. IDs are integers incrementing from 1.
1007 * Only committed write transactions increment the ID. If a transaction
1008 * aborts, the ID may be re-used by the next writer.
1011 MDB_env *mt_env; /**< the DB environment */
1012 /** The list of pages that became unused during this transaction.
1014 MDB_IDL mt_free_pgs;
1015 /** The list of loose pages that became unused and may be reused
1016 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1018 MDB_page *mt_loose_pgs;
1019 /* #Number of loose pages (#mt_loose_pgs) */
1021 /** The sorted list of dirty pages we temporarily wrote to disk
1022 * because the dirty list was full. page numbers in here are
1023 * shifted left by 1, deleted slots have the LSB set.
1025 MDB_IDL mt_spill_pgs;
1027 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1028 MDB_ID2L dirty_list;
1029 /** For read txns: This thread/txn's reader table slot, or NULL. */
1032 /** Array of records for each DB known in the environment. */
1034 /** Array of MDB_db records for each known DB */
1036 /** Array of sequence numbers for each DB handle */
1037 unsigned int *mt_dbiseqs;
1038 /** @defgroup mt_dbflag Transaction DB Flags
1042 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1043 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1044 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1045 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1046 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1048 /** In write txns, array of cursors for each DB */
1049 MDB_cursor **mt_cursors;
1050 /** Array of flags for each DB */
1051 unsigned char *mt_dbflags;
1052 /** Number of DB records in use, or 0 when the txn is finished.
1053 * This number only ever increments until the txn finishes; we
1054 * don't decrement it when individual DB handles are closed.
1058 /** @defgroup mdb_txn Transaction Flags
1062 /** #mdb_txn_begin() flags */
1063 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1064 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1065 /* internal txn flags */
1066 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1067 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1068 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1069 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1070 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1071 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1072 /** most operations on the txn are currently illegal */
1073 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1075 unsigned int mt_flags; /**< @ref mdb_txn */
1076 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1077 * Includes ancestor txns' dirty pages not hidden by other txns'
1078 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1079 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1081 unsigned int mt_dirty_room;
1084 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1085 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1086 * raise this on a 64 bit machine.
1088 #define CURSOR_STACK 32
1092 /** Cursors are used for all DB operations.
1093 * A cursor holds a path of (page pointer, key index) from the DB
1094 * root to a position in the DB, plus other state. #MDB_DUPSORT
1095 * cursors include an xcursor to the current data item. Write txns
1096 * track their cursors and keep them up to date when data moves.
1097 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1098 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1101 /** Next cursor on this DB in this txn */
1102 MDB_cursor *mc_next;
1103 /** Backup of the original cursor if this cursor is a shadow */
1104 MDB_cursor *mc_backup;
1105 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1106 struct MDB_xcursor *mc_xcursor;
1107 /** The transaction that owns this cursor */
1109 /** The database handle this cursor operates on */
1111 /** The database record for this cursor */
1113 /** The database auxiliary record for this cursor */
1115 /** The @ref mt_dbflag for this database */
1116 unsigned char *mc_dbflag;
1117 unsigned short mc_snum; /**< number of pushed pages */
1118 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1119 /** @defgroup mdb_cursor Cursor Flags
1121 * Cursor state flags.
1124 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1125 #define C_EOF 0x02 /**< No more data */
1126 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1127 #define C_DEL 0x08 /**< last op was a cursor_del */
1128 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1129 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1131 unsigned int mc_flags; /**< @ref mdb_cursor */
1132 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1133 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1136 /** Context for sorted-dup records.
1137 * We could have gone to a fully recursive design, with arbitrarily
1138 * deep nesting of sub-databases. But for now we only handle these
1139 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1141 typedef struct MDB_xcursor {
1142 /** A sub-cursor for traversing the Dup DB */
1143 MDB_cursor mx_cursor;
1144 /** The database record for this Dup DB */
1146 /** The auxiliary DB record for this Dup DB */
1148 /** The @ref mt_dbflag for this Dup DB */
1149 unsigned char mx_dbflag;
1152 /** State of FreeDB old pages, stored in the MDB_env */
1153 typedef struct MDB_pgstate {
1154 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1155 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1158 /** The database environment. */
1160 HANDLE me_fd; /**< The main data file */
1161 HANDLE me_lfd; /**< The lock file */
1162 HANDLE me_mfd; /**< just for writing the meta pages */
1163 /** Failed to update the meta page. Probably an I/O error. */
1164 #define MDB_FATAL_ERROR 0x80000000U
1165 /** Some fields are initialized. */
1166 #define MDB_ENV_ACTIVE 0x20000000U
1167 /** me_txkey is set */
1168 #define MDB_ENV_TXKEY 0x10000000U
1169 /** fdatasync is unreliable */
1170 #define MDB_FSYNCONLY 0x08000000U
1171 uint32_t me_flags; /**< @ref mdb_env */
1172 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1173 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1174 unsigned int me_maxreaders; /**< size of the reader table */
1175 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1176 volatile int me_close_readers;
1177 MDB_dbi me_numdbs; /**< number of DBs opened */
1178 MDB_dbi me_maxdbs; /**< size of the DB table */
1179 MDB_PID_T me_pid; /**< process ID of this env */
1180 char *me_path; /**< path to the DB files */
1181 char *me_map; /**< the memory map of the data file */
1182 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1183 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1184 void *me_pbuf; /**< scratch area for DUPSORT put() */
1185 MDB_txn *me_txn; /**< current write transaction */
1186 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1187 size_t me_mapsize; /**< size of the data memory map */
1188 off_t me_size; /**< current file size */
1189 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1190 MDB_dbx *me_dbxs; /**< array of static DB info */
1191 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1192 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1193 pthread_key_t me_txkey; /**< thread-key for readers */
1194 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1195 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1196 # define me_pglast me_pgstate.mf_pglast
1197 # define me_pghead me_pgstate.mf_pghead
1198 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1199 /** IDL of pages that became unused in a write txn */
1200 MDB_IDL me_free_pgs;
1201 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1202 MDB_ID2L me_dirty_list;
1203 /** Max number of freelist items that can fit in a single overflow page */
1205 /** Max size of a node on a page */
1206 unsigned int me_nodemax;
1207 #if !(MDB_MAXKEYSIZE)
1208 unsigned int me_maxkey; /**< max size of a key */
1210 int me_live_reader; /**< have liveness lock in reader table */
1212 int me_pidquery; /**< Used in OpenProcess */
1214 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1215 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1216 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1218 mdb_mutex_t me_rmutex;
1219 mdb_mutex_t me_wmutex;
1221 void *me_userctx; /**< User-settable context */
1222 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1225 /** Nested transaction */
1226 typedef struct MDB_ntxn {
1227 MDB_txn mnt_txn; /**< the transaction */
1228 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1231 /** max number of pages to commit in one writev() call */
1232 #define MDB_COMMIT_PAGES 64
1233 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1234 #undef MDB_COMMIT_PAGES
1235 #define MDB_COMMIT_PAGES IOV_MAX
1238 /** max bytes to write in one call */
1239 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1241 /** Check \b txn and \b dbi arguments to a function */
1242 #define TXN_DBI_EXIST(txn, dbi, validity) \
1243 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1245 /** Check for misused \b dbi handles */
1246 #define TXN_DBI_CHANGED(txn, dbi) \
1247 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1249 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1250 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1251 static int mdb_page_touch(MDB_cursor *mc);
1253 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1254 "reset-tmp", "fail-begin", "fail-beginchild"}
1256 /* mdb_txn_end operation number, for logging */
1257 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1258 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1260 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1261 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1262 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1263 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1264 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1266 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1267 static int mdb_page_search_root(MDB_cursor *mc,
1268 MDB_val *key, int modify);
1269 #define MDB_PS_MODIFY 1
1270 #define MDB_PS_ROOTONLY 2
1271 #define MDB_PS_FIRST 4
1272 #define MDB_PS_LAST 8
1273 static int mdb_page_search(MDB_cursor *mc,
1274 MDB_val *key, int flags);
1275 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1277 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1278 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1279 pgno_t newpgno, unsigned int nflags);
1281 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1282 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1283 static int mdb_env_write_meta(MDB_txn *txn);
1284 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1285 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1287 static void mdb_env_close0(MDB_env *env, int excl);
1289 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1290 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1291 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1292 static void mdb_node_del(MDB_cursor *mc, int ksize);
1293 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1294 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1295 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1296 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1297 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1299 static int mdb_rebalance(MDB_cursor *mc);
1300 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1302 static void mdb_cursor_pop(MDB_cursor *mc);
1303 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1305 static int mdb_cursor_del0(MDB_cursor *mc);
1306 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1307 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1308 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1309 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1310 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1312 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1313 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1315 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1316 static void mdb_xcursor_init0(MDB_cursor *mc);
1317 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1318 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1320 static int mdb_drop0(MDB_cursor *mc, int subs);
1321 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1322 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1325 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1328 /** Compare two items pointing at size_t's of unknown alignment. */
1329 #ifdef MISALIGNED_OK
1330 # define mdb_cmp_clong mdb_cmp_long
1332 # define mdb_cmp_clong mdb_cmp_cint
1336 static SECURITY_DESCRIPTOR mdb_null_sd;
1337 static SECURITY_ATTRIBUTES mdb_all_sa;
1338 static int mdb_sec_inited;
1341 /** Return the library version info. */
1343 mdb_version(int *major, int *minor, int *patch)
1345 if (major) *major = MDB_VERSION_MAJOR;
1346 if (minor) *minor = MDB_VERSION_MINOR;
1347 if (patch) *patch = MDB_VERSION_PATCH;
1348 return MDB_VERSION_STRING;
1351 /** Table of descriptions for LMDB @ref errors */
1352 static char *const mdb_errstr[] = {
1353 "MDB_KEYEXIST: Key/data pair already exists",
1354 "MDB_NOTFOUND: No matching key/data pair found",
1355 "MDB_PAGE_NOTFOUND: Requested page not found",
1356 "MDB_CORRUPTED: Located page was wrong type",
1357 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1358 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1359 "MDB_INVALID: File is not an LMDB file",
1360 "MDB_MAP_FULL: Environment mapsize limit reached",
1361 "MDB_DBS_FULL: Environment maxdbs limit reached",
1362 "MDB_READERS_FULL: Environment maxreaders limit reached",
1363 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1364 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1365 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1366 "MDB_PAGE_FULL: Internal error - page has no more space",
1367 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1368 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1369 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1370 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1371 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1372 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1376 mdb_strerror(int err)
1379 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1380 * This works as long as no function between the call to mdb_strerror
1381 * and the actual use of the message uses more than 4K of stack.
1384 char buf[1024], *ptr = buf;
1388 return ("Successful return: 0");
1390 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1391 i = err - MDB_KEYEXIST;
1392 return mdb_errstr[i];
1396 /* These are the C-runtime error codes we use. The comment indicates
1397 * their numeric value, and the Win32 error they would correspond to
1398 * if the error actually came from a Win32 API. A major mess, we should
1399 * have used LMDB-specific error codes for everything.
1402 case ENOENT: /* 2, FILE_NOT_FOUND */
1403 case EIO: /* 5, ACCESS_DENIED */
1404 case ENOMEM: /* 12, INVALID_ACCESS */
1405 case EACCES: /* 13, INVALID_DATA */
1406 case EBUSY: /* 16, CURRENT_DIRECTORY */
1407 case EINVAL: /* 22, BAD_COMMAND */
1408 case ENOSPC: /* 28, OUT_OF_PAPER */
1409 return strerror(err);
1414 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1415 FORMAT_MESSAGE_IGNORE_INSERTS,
1416 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1419 return strerror(err);
1423 /** assert(3) variant in cursor context */
1424 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1425 /** assert(3) variant in transaction context */
1426 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1427 /** assert(3) variant in environment context */
1428 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1431 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1432 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1435 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1436 const char *func, const char *file, int line)
1439 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1440 file, line, expr_txt, func);
1441 if (env->me_assert_func)
1442 env->me_assert_func(env, buf);
1443 fprintf(stderr, "%s\n", buf);
1447 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1451 /** Return the page number of \b mp which may be sub-page, for debug output */
1453 mdb_dbg_pgno(MDB_page *mp)
1456 COPY_PGNO(ret, mp->mp_pgno);
1460 /** Display a key in hexadecimal and return the address of the result.
1461 * @param[in] key the key to display
1462 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1463 * @return The key in hexadecimal form.
1466 mdb_dkey(MDB_val *key, char *buf)
1469 unsigned char *c = key->mv_data;
1475 if (key->mv_size > DKBUF_MAXKEYSIZE)
1476 return "MDB_MAXKEYSIZE";
1477 /* may want to make this a dynamic check: if the key is mostly
1478 * printable characters, print it as-is instead of converting to hex.
1482 for (i=0; i<key->mv_size; i++)
1483 ptr += sprintf(ptr, "%02x", *c++);
1485 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1491 mdb_leafnode_type(MDB_node *n)
1493 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1494 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1495 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1498 /** Display all the keys in the page. */
1500 mdb_page_list(MDB_page *mp)
1502 pgno_t pgno = mdb_dbg_pgno(mp);
1503 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1505 unsigned int i, nkeys, nsize, total = 0;
1509 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1510 case P_BRANCH: type = "Branch page"; break;
1511 case P_LEAF: type = "Leaf page"; break;
1512 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1513 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1514 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1516 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1517 pgno, mp->mp_pages, state);
1520 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1521 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1524 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1528 nkeys = NUMKEYS(mp);
1529 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1531 for (i=0; i<nkeys; i++) {
1532 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1533 key.mv_size = nsize = mp->mp_pad;
1534 key.mv_data = LEAF2KEY(mp, i, nsize);
1536 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1539 node = NODEPTR(mp, i);
1540 key.mv_size = node->mn_ksize;
1541 key.mv_data = node->mn_data;
1542 nsize = NODESIZE + key.mv_size;
1543 if (IS_BRANCH(mp)) {
1544 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1548 if (F_ISSET(node->mn_flags, F_BIGDATA))
1549 nsize += sizeof(pgno_t);
1551 nsize += NODEDSZ(node);
1553 nsize += sizeof(indx_t);
1554 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1555 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1557 total = EVEN(total);
1559 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1560 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1564 mdb_cursor_chk(MDB_cursor *mc)
1570 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1571 for (i=0; i<mc->mc_top; i++) {
1573 node = NODEPTR(mp, mc->mc_ki[i]);
1574 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1577 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1583 /** Count all the pages in each DB and in the freelist
1584 * and make sure it matches the actual number of pages
1586 * All named DBs must be open for a correct count.
1588 static void mdb_audit(MDB_txn *txn)
1592 MDB_ID freecount, count;
1597 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1598 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1599 freecount += *(MDB_ID *)data.mv_data;
1600 mdb_tassert(txn, rc == MDB_NOTFOUND);
1603 for (i = 0; i<txn->mt_numdbs; i++) {
1605 if (!(txn->mt_dbflags[i] & DB_VALID))
1607 mdb_cursor_init(&mc, txn, i, &mx);
1608 if (txn->mt_dbs[i].md_root == P_INVALID)
1610 count += txn->mt_dbs[i].md_branch_pages +
1611 txn->mt_dbs[i].md_leaf_pages +
1612 txn->mt_dbs[i].md_overflow_pages;
1613 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1614 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1615 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1618 mp = mc.mc_pg[mc.mc_top];
1619 for (j=0; j<NUMKEYS(mp); j++) {
1620 MDB_node *leaf = NODEPTR(mp, j);
1621 if (leaf->mn_flags & F_SUBDATA) {
1623 memcpy(&db, NODEDATA(leaf), sizeof(db));
1624 count += db.md_branch_pages + db.md_leaf_pages +
1625 db.md_overflow_pages;
1629 mdb_tassert(txn, rc == MDB_NOTFOUND);
1632 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1633 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1634 txn->mt_txnid, freecount, count+NUM_METAS,
1635 freecount+count+NUM_METAS, txn->mt_next_pgno);
1641 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1643 return txn->mt_dbxs[dbi].md_cmp(a, b);
1647 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1649 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1650 #if UINT_MAX < SIZE_MAX
1651 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1652 dcmp = mdb_cmp_clong;
1657 /** Allocate memory for a page.
1658 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1661 mdb_page_malloc(MDB_txn *txn, unsigned num)
1663 MDB_env *env = txn->mt_env;
1664 MDB_page *ret = env->me_dpages;
1665 size_t psize = env->me_psize, sz = psize, off;
1666 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1667 * For a single page alloc, we init everything after the page header.
1668 * For multi-page, we init the final page; if the caller needed that
1669 * many pages they will be filling in at least up to the last page.
1673 VGMEMP_ALLOC(env, ret, sz);
1674 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1675 env->me_dpages = ret->mp_next;
1678 psize -= off = PAGEHDRSZ;
1683 if ((ret = malloc(sz)) != NULL) {
1684 VGMEMP_ALLOC(env, ret, sz);
1685 if (!(env->me_flags & MDB_NOMEMINIT)) {
1686 memset((char *)ret + off, 0, psize);
1690 txn->mt_flags |= MDB_TXN_ERROR;
1694 /** Free a single page.
1695 * Saves single pages to a list, for future reuse.
1696 * (This is not used for multi-page overflow pages.)
1699 mdb_page_free(MDB_env *env, MDB_page *mp)
1701 mp->mp_next = env->me_dpages;
1702 VGMEMP_FREE(env, mp);
1703 env->me_dpages = mp;
1706 /** Free a dirty page */
1708 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1710 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1711 mdb_page_free(env, dp);
1713 /* large pages just get freed directly */
1714 VGMEMP_FREE(env, dp);
1719 /** Return all dirty pages to dpage list */
1721 mdb_dlist_free(MDB_txn *txn)
1723 MDB_env *env = txn->mt_env;
1724 MDB_ID2L dl = txn->mt_u.dirty_list;
1725 unsigned i, n = dl[0].mid;
1727 for (i = 1; i <= n; i++) {
1728 mdb_dpage_free(env, dl[i].mptr);
1733 /** Loosen or free a single page.
1734 * Saves single pages to a list for future reuse
1735 * in this same txn. It has been pulled from the freeDB
1736 * and already resides on the dirty list, but has been
1737 * deleted. Use these pages first before pulling again
1740 * If the page wasn't dirtied in this txn, just add it
1741 * to this txn's free list.
1744 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1747 pgno_t pgno = mp->mp_pgno;
1748 MDB_txn *txn = mc->mc_txn;
1750 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1751 if (txn->mt_parent) {
1752 MDB_ID2 *dl = txn->mt_u.dirty_list;
1753 /* If txn has a parent, make sure the page is in our
1757 unsigned x = mdb_mid2l_search(dl, pgno);
1758 if (x <= dl[0].mid && dl[x].mid == pgno) {
1759 if (mp != dl[x].mptr) { /* bad cursor? */
1760 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1761 txn->mt_flags |= MDB_TXN_ERROR;
1762 return MDB_CORRUPTED;
1769 /* no parent txn, so it's just ours */
1774 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1776 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1777 txn->mt_loose_pgs = mp;
1778 txn->mt_loose_count++;
1779 mp->mp_flags |= P_LOOSE;
1781 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1789 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1790 * @param[in] mc A cursor handle for the current operation.
1791 * @param[in] pflags Flags of the pages to update:
1792 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1793 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1794 * @return 0 on success, non-zero on failure.
1797 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1799 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1800 MDB_txn *txn = mc->mc_txn;
1806 int rc = MDB_SUCCESS, level;
1808 /* Mark pages seen by cursors */
1809 if (mc->mc_flags & C_UNTRACK)
1810 mc = NULL; /* will find mc in mt_cursors */
1811 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1812 for (; mc; mc=mc->mc_next) {
1813 if (!(mc->mc_flags & C_INITIALIZED))
1815 for (m3 = mc;; m3 = &mx->mx_cursor) {
1817 for (j=0; j<m3->mc_snum; j++) {
1819 if ((mp->mp_flags & Mask) == pflags)
1820 mp->mp_flags ^= P_KEEP;
1822 mx = m3->mc_xcursor;
1823 /* Proceed to mx if it is at a sub-database */
1824 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1826 if (! (mp && (mp->mp_flags & P_LEAF)))
1828 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1829 if (!(leaf->mn_flags & F_SUBDATA))
1838 /* Mark dirty root pages */
1839 for (i=0; i<txn->mt_numdbs; i++) {
1840 if (txn->mt_dbflags[i] & DB_DIRTY) {
1841 pgno_t pgno = txn->mt_dbs[i].md_root;
1842 if (pgno == P_INVALID)
1844 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1846 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1847 dp->mp_flags ^= P_KEEP;
1855 static int mdb_page_flush(MDB_txn *txn, int keep);
1857 /** Spill pages from the dirty list back to disk.
1858 * This is intended to prevent running into #MDB_TXN_FULL situations,
1859 * but note that they may still occur in a few cases:
1860 * 1) our estimate of the txn size could be too small. Currently this
1861 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1862 * 2) child txns may run out of space if their parents dirtied a
1863 * lot of pages and never spilled them. TODO: we probably should do
1864 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1865 * the parent's dirty_room is below a given threshold.
1867 * Otherwise, if not using nested txns, it is expected that apps will
1868 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1869 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1870 * If the txn never references them again, they can be left alone.
1871 * If the txn only reads them, they can be used without any fuss.
1872 * If the txn writes them again, they can be dirtied immediately without
1873 * going thru all of the work of #mdb_page_touch(). Such references are
1874 * handled by #mdb_page_unspill().
1876 * Also note, we never spill DB root pages, nor pages of active cursors,
1877 * because we'll need these back again soon anyway. And in nested txns,
1878 * we can't spill a page in a child txn if it was already spilled in a
1879 * parent txn. That would alter the parent txns' data even though
1880 * the child hasn't committed yet, and we'd have no way to undo it if
1881 * the child aborted.
1883 * @param[in] m0 cursor A cursor handle identifying the transaction and
1884 * database for which we are checking space.
1885 * @param[in] key For a put operation, the key being stored.
1886 * @param[in] data For a put operation, the data being stored.
1887 * @return 0 on success, non-zero on failure.
1890 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1892 MDB_txn *txn = m0->mc_txn;
1894 MDB_ID2L dl = txn->mt_u.dirty_list;
1895 unsigned int i, j, need;
1898 if (m0->mc_flags & C_SUB)
1901 /* Estimate how much space this op will take */
1902 i = m0->mc_db->md_depth;
1903 /* Named DBs also dirty the main DB */
1904 if (m0->mc_dbi >= CORE_DBS)
1905 i += txn->mt_dbs[MAIN_DBI].md_depth;
1906 /* For puts, roughly factor in the key+data size */
1908 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1909 i += i; /* double it for good measure */
1912 if (txn->mt_dirty_room > i)
1915 if (!txn->mt_spill_pgs) {
1916 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1917 if (!txn->mt_spill_pgs)
1920 /* purge deleted slots */
1921 MDB_IDL sl = txn->mt_spill_pgs;
1922 unsigned int num = sl[0];
1924 for (i=1; i<=num; i++) {
1931 /* Preserve pages which may soon be dirtied again */
1932 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1935 /* Less aggressive spill - we originally spilled the entire dirty list,
1936 * with a few exceptions for cursor pages and DB root pages. But this
1937 * turns out to be a lot of wasted effort because in a large txn many
1938 * of those pages will need to be used again. So now we spill only 1/8th
1939 * of the dirty pages. Testing revealed this to be a good tradeoff,
1940 * better than 1/2, 1/4, or 1/10.
1942 if (need < MDB_IDL_UM_MAX / 8)
1943 need = MDB_IDL_UM_MAX / 8;
1945 /* Save the page IDs of all the pages we're flushing */
1946 /* flush from the tail forward, this saves a lot of shifting later on. */
1947 for (i=dl[0].mid; i && need; i--) {
1948 MDB_ID pn = dl[i].mid << 1;
1950 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1952 /* Can't spill twice, make sure it's not already in a parent's
1955 if (txn->mt_parent) {
1957 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1958 if (tx2->mt_spill_pgs) {
1959 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1960 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1961 dp->mp_flags |= P_KEEP;
1969 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1973 mdb_midl_sort(txn->mt_spill_pgs);
1975 /* Flush the spilled part of dirty list */
1976 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1979 /* Reset any dirty pages we kept that page_flush didn't see */
1980 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1983 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1987 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1989 mdb_find_oldest(MDB_txn *txn)
1992 txnid_t mr, oldest = txn->mt_txnid - 1;
1993 if (txn->mt_env->me_txns) {
1994 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1995 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2006 /** Add a page to the txn's dirty list */
2008 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2011 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2013 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2014 insert = mdb_mid2l_append;
2016 insert = mdb_mid2l_insert;
2018 mid.mid = mp->mp_pgno;
2020 rc = insert(txn->mt_u.dirty_list, &mid);
2021 mdb_tassert(txn, rc == 0);
2022 txn->mt_dirty_room--;
2025 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2026 * me_pghead and mt_next_pgno.
2028 * If there are free pages available from older transactions, they
2029 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2030 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2031 * and move me_pglast to say which records were consumed. Only this
2032 * function can create me_pghead and move me_pglast/mt_next_pgno.
2033 * @param[in] mc cursor A cursor handle identifying the transaction and
2034 * database for which we are allocating.
2035 * @param[in] num the number of pages to allocate.
2036 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2037 * will always be satisfied by a single contiguous chunk of memory.
2038 * @return 0 on success, non-zero on failure.
2041 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2043 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2044 /* Get at most <Max_retries> more freeDB records once me_pghead
2045 * has enough pages. If not enough, use new pages from the map.
2046 * If <Paranoid> and mc is updating the freeDB, only get new
2047 * records if me_pghead is empty. Then the freelist cannot play
2048 * catch-up with itself by growing while trying to save it.
2050 enum { Paranoid = 1, Max_retries = 500 };
2052 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2054 int rc, retry = num * 60;
2055 MDB_txn *txn = mc->mc_txn;
2056 MDB_env *env = txn->mt_env;
2057 pgno_t pgno, *mop = env->me_pghead;
2058 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2060 txnid_t oldest = 0, last;
2065 /* If there are any loose pages, just use them */
2066 if (num == 1 && txn->mt_loose_pgs) {
2067 np = txn->mt_loose_pgs;
2068 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2069 txn->mt_loose_count--;
2070 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2078 /* If our dirty list is already full, we can't do anything */
2079 if (txn->mt_dirty_room == 0) {
2084 for (op = MDB_FIRST;; op = MDB_NEXT) {
2089 /* Seek a big enough contiguous page range. Prefer
2090 * pages at the tail, just truncating the list.
2096 if (mop[i-n2] == pgno+n2)
2103 if (op == MDB_FIRST) { /* 1st iteration */
2104 /* Prepare to fetch more and coalesce */
2105 last = env->me_pglast;
2106 oldest = env->me_pgoldest;
2107 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2110 key.mv_data = &last; /* will look up last+1 */
2111 key.mv_size = sizeof(last);
2113 if (Paranoid && mc->mc_dbi == FREE_DBI)
2116 if (Paranoid && retry < 0 && mop_len)
2120 /* Do not fetch more if the record will be too recent */
2121 if (oldest <= last) {
2123 oldest = mdb_find_oldest(txn);
2124 env->me_pgoldest = oldest;
2130 rc = mdb_cursor_get(&m2, &key, NULL, op);
2132 if (rc == MDB_NOTFOUND)
2136 last = *(txnid_t*)key.mv_data;
2137 if (oldest <= last) {
2139 oldest = mdb_find_oldest(txn);
2140 env->me_pgoldest = oldest;
2146 np = m2.mc_pg[m2.mc_top];
2147 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2148 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2151 idl = (MDB_ID *) data.mv_data;
2154 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2159 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2161 mop = env->me_pghead;
2163 env->me_pglast = last;
2165 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2166 last, txn->mt_dbs[FREE_DBI].md_root, i));
2168 DPRINTF(("IDL %"Z"u", idl[j]));
2170 /* Merge in descending sorted order */
2171 mdb_midl_xmerge(mop, idl);
2175 /* Use new pages from the map when nothing suitable in the freeDB */
2177 pgno = txn->mt_next_pgno;
2178 if (pgno + num >= env->me_maxpg) {
2179 DPUTS("DB size maxed out");
2185 if (env->me_flags & MDB_WRITEMAP) {
2186 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2188 if (!(np = mdb_page_malloc(txn, num))) {
2194 mop[0] = mop_len -= num;
2195 /* Move any stragglers down */
2196 for (j = i-num; j < mop_len; )
2197 mop[++j] = mop[++i];
2199 txn->mt_next_pgno = pgno + num;
2202 mdb_page_dirty(txn, np);
2208 txn->mt_flags |= MDB_TXN_ERROR;
2212 /** Copy the used portions of a non-overflow page.
2213 * @param[in] dst page to copy into
2214 * @param[in] src page to copy from
2215 * @param[in] psize size of a page
2218 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2220 enum { Align = sizeof(pgno_t) };
2221 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2223 /* If page isn't full, just copy the used portion. Adjust
2224 * alignment so memcpy may copy words instead of bytes.
2226 if ((unused &= -Align) && !IS_LEAF2(src)) {
2227 upper = (upper + PAGEBASE) & -Align;
2228 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2229 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2232 memcpy(dst, src, psize - unused);
2236 /** Pull a page off the txn's spill list, if present.
2237 * If a page being referenced was spilled to disk in this txn, bring
2238 * it back and make it dirty/writable again.
2239 * @param[in] txn the transaction handle.
2240 * @param[in] mp the page being referenced. It must not be dirty.
2241 * @param[out] ret the writable page, if any. ret is unchanged if
2242 * mp wasn't spilled.
2245 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2247 MDB_env *env = txn->mt_env;
2250 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2252 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2253 if (!tx2->mt_spill_pgs)
2255 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2256 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2259 if (txn->mt_dirty_room == 0)
2260 return MDB_TXN_FULL;
2261 if (IS_OVERFLOW(mp))
2265 if (env->me_flags & MDB_WRITEMAP) {
2268 np = mdb_page_malloc(txn, num);
2272 memcpy(np, mp, num * env->me_psize);
2274 mdb_page_copy(np, mp, env->me_psize);
2277 /* If in current txn, this page is no longer spilled.
2278 * If it happens to be the last page, truncate the spill list.
2279 * Otherwise mark it as deleted by setting the LSB.
2281 if (x == txn->mt_spill_pgs[0])
2282 txn->mt_spill_pgs[0]--;
2284 txn->mt_spill_pgs[x] |= 1;
2285 } /* otherwise, if belonging to a parent txn, the
2286 * page remains spilled until child commits
2289 mdb_page_dirty(txn, np);
2290 np->mp_flags |= P_DIRTY;
2298 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2299 * @param[in] mc cursor pointing to the page to be touched
2300 * @return 0 on success, non-zero on failure.
2303 mdb_page_touch(MDB_cursor *mc)
2305 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2306 MDB_txn *txn = mc->mc_txn;
2307 MDB_cursor *m2, *m3;
2311 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2312 if (txn->mt_flags & MDB_TXN_SPILLS) {
2314 rc = mdb_page_unspill(txn, mp, &np);
2320 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2321 (rc = mdb_page_alloc(mc, 1, &np)))
2324 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2325 mp->mp_pgno, pgno));
2326 mdb_cassert(mc, mp->mp_pgno != pgno);
2327 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2328 /* Update the parent page, if any, to point to the new page */
2330 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2331 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2332 SETPGNO(node, pgno);
2334 mc->mc_db->md_root = pgno;
2336 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2337 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2339 /* If txn has a parent, make sure the page is in our
2343 unsigned x = mdb_mid2l_search(dl, pgno);
2344 if (x <= dl[0].mid && dl[x].mid == pgno) {
2345 if (mp != dl[x].mptr) { /* bad cursor? */
2346 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2347 txn->mt_flags |= MDB_TXN_ERROR;
2348 return MDB_CORRUPTED;
2353 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2355 np = mdb_page_malloc(txn, 1);
2360 rc = mdb_mid2l_insert(dl, &mid);
2361 mdb_cassert(mc, rc == 0);
2366 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2368 np->mp_flags |= P_DIRTY;
2371 /* Adjust cursors pointing to mp */
2372 mc->mc_pg[mc->mc_top] = np;
2373 m2 = txn->mt_cursors[mc->mc_dbi];
2374 if (mc->mc_flags & C_SUB) {
2375 for (; m2; m2=m2->mc_next) {
2376 m3 = &m2->mc_xcursor->mx_cursor;
2377 if (m3->mc_snum < mc->mc_snum) continue;
2378 if (m3->mc_pg[mc->mc_top] == mp)
2379 m3->mc_pg[mc->mc_top] = np;
2382 for (; m2; m2=m2->mc_next) {
2383 if (m2->mc_snum < mc->mc_snum) continue;
2384 if (m2->mc_pg[mc->mc_top] == mp) {
2385 m2->mc_pg[mc->mc_top] = np;
2386 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2388 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2390 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2391 if (!(leaf->mn_flags & F_SUBDATA))
2392 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2400 txn->mt_flags |= MDB_TXN_ERROR;
2405 mdb_env_sync(MDB_env *env, int force)
2408 if (env->me_flags & MDB_RDONLY)
2410 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2411 if (env->me_flags & MDB_WRITEMAP) {
2412 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2413 ? MS_ASYNC : MS_SYNC;
2414 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2417 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2421 #ifdef BROKEN_FDATASYNC
2422 if (env->me_flags & MDB_FSYNCONLY) {
2423 if (fsync(env->me_fd))
2427 if (MDB_FDATASYNC(env->me_fd))
2434 /** Back up parent txn's cursors, then grab the originals for tracking */
2436 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2438 MDB_cursor *mc, *bk;
2443 for (i = src->mt_numdbs; --i >= 0; ) {
2444 if ((mc = src->mt_cursors[i]) != NULL) {
2445 size = sizeof(MDB_cursor);
2447 size += sizeof(MDB_xcursor);
2448 for (; mc; mc = bk->mc_next) {
2454 mc->mc_db = &dst->mt_dbs[i];
2455 /* Kill pointers into src - and dst to reduce abuse: The
2456 * user may not use mc until dst ends. Otherwise we'd...
2458 mc->mc_txn = NULL; /* ...set this to dst */
2459 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2460 if ((mx = mc->mc_xcursor) != NULL) {
2461 *(MDB_xcursor *)(bk+1) = *mx;
2462 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2464 mc->mc_next = dst->mt_cursors[i];
2465 dst->mt_cursors[i] = mc;
2472 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2473 * @param[in] txn the transaction handle.
2474 * @param[in] merge true to keep changes to parent cursors, false to revert.
2475 * @return 0 on success, non-zero on failure.
2478 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2480 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2484 for (i = txn->mt_numdbs; --i >= 0; ) {
2485 for (mc = cursors[i]; mc; mc = next) {
2487 if ((bk = mc->mc_backup) != NULL) {
2489 /* Commit changes to parent txn */
2490 mc->mc_next = bk->mc_next;
2491 mc->mc_backup = bk->mc_backup;
2492 mc->mc_txn = bk->mc_txn;
2493 mc->mc_db = bk->mc_db;
2494 mc->mc_dbflag = bk->mc_dbflag;
2495 if ((mx = mc->mc_xcursor) != NULL)
2496 mx->mx_cursor.mc_txn = bk->mc_txn;
2498 /* Abort nested txn */
2500 if ((mx = mc->mc_xcursor) != NULL)
2501 *mx = *(MDB_xcursor *)(bk+1);
2505 /* Only malloced cursors are permanently tracked. */
2512 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2518 Pidset = F_SETLK, Pidcheck = F_GETLK
2522 /** Set or check a pid lock. Set returns 0 on success.
2523 * Check returns 0 if the process is certainly dead, nonzero if it may
2524 * be alive (the lock exists or an error happened so we do not know).
2526 * On Windows Pidset is a no-op, we merely check for the existence
2527 * of the process with the given pid. On POSIX we use a single byte
2528 * lock on the lockfile, set at an offset equal to the pid.
2531 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2533 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2536 if (op == Pidcheck) {
2537 h = OpenProcess(env->me_pidquery, FALSE, pid);
2538 /* No documented "no such process" code, but other program use this: */
2540 return ErrCode() != ERROR_INVALID_PARAMETER;
2541 /* A process exists until all handles to it close. Has it exited? */
2542 ret = WaitForSingleObject(h, 0) != 0;
2549 struct flock lock_info;
2550 memset(&lock_info, 0, sizeof(lock_info));
2551 lock_info.l_type = F_WRLCK;
2552 lock_info.l_whence = SEEK_SET;
2553 lock_info.l_start = pid;
2554 lock_info.l_len = 1;
2555 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2556 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2558 } else if ((rc = ErrCode()) == EINTR) {
2566 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2567 * @param[in] txn the transaction handle to initialize
2568 * @return 0 on success, non-zero on failure.
2571 mdb_txn_renew0(MDB_txn *txn)
2573 MDB_env *env = txn->mt_env;
2574 MDB_txninfo *ti = env->me_txns;
2576 unsigned int i, nr, flags = txn->mt_flags;
2578 int rc, new_notls = 0;
2580 if ((flags &= MDB_TXN_RDONLY) != 0) {
2582 meta = mdb_env_pick_meta(env);
2583 txn->mt_txnid = meta->mm_txnid;
2584 txn->mt_u.reader = NULL;
2586 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2587 pthread_getspecific(env->me_txkey);
2589 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2590 return MDB_BAD_RSLOT;
2592 MDB_PID_T pid = env->me_pid;
2593 MDB_THR_T tid = pthread_self();
2594 mdb_mutexref_t rmutex = env->me_rmutex;
2596 if (!env->me_live_reader) {
2597 rc = mdb_reader_pid(env, Pidset, pid);
2600 env->me_live_reader = 1;
2603 if (LOCK_MUTEX(rc, env, rmutex))
2605 nr = ti->mti_numreaders;
2606 for (i=0; i<nr; i++)
2607 if (ti->mti_readers[i].mr_pid == 0)
2609 if (i == env->me_maxreaders) {
2610 UNLOCK_MUTEX(rmutex);
2611 return MDB_READERS_FULL;
2613 r = &ti->mti_readers[i];
2614 /* Claim the reader slot, carefully since other code
2615 * uses the reader table un-mutexed: First reset the
2616 * slot, next publish it in mti_numreaders. After
2617 * that, it is safe for mdb_env_close() to touch it.
2618 * When it will be closed, we can finally claim it.
2621 r->mr_txnid = (txnid_t)-1;
2624 ti->mti_numreaders = ++nr;
2625 env->me_close_readers = nr;
2627 UNLOCK_MUTEX(rmutex);
2629 new_notls = (env->me_flags & MDB_NOTLS);
2630 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2635 do /* LY: Retry on a race, ITS#7970. */
2636 r->mr_txnid = ti->mti_txnid;
2637 while(r->mr_txnid != ti->mti_txnid);
2638 txn->mt_txnid = r->mr_txnid;
2639 txn->mt_u.reader = r;
2640 meta = env->me_metas[txn->mt_txnid & 1];
2644 /* Not yet touching txn == env->me_txn0, it may be active */
2646 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2648 txn->mt_txnid = ti->mti_txnid;
2649 meta = env->me_metas[txn->mt_txnid & 1];
2651 meta = mdb_env_pick_meta(env);
2652 txn->mt_txnid = meta->mm_txnid;
2656 if (txn->mt_txnid == mdb_debug_start)
2659 txn->mt_child = NULL;
2660 txn->mt_loose_pgs = NULL;
2661 txn->mt_loose_count = 0;
2662 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2663 txn->mt_u.dirty_list = env->me_dirty_list;
2664 txn->mt_u.dirty_list[0].mid = 0;
2665 txn->mt_free_pgs = env->me_free_pgs;
2666 txn->mt_free_pgs[0] = 0;
2667 txn->mt_spill_pgs = NULL;
2669 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2672 /* Copy the DB info and flags */
2673 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2675 /* Moved to here to avoid a data race in read TXNs */
2676 txn->mt_next_pgno = meta->mm_last_pg+1;
2678 txn->mt_flags = flags;
2681 txn->mt_numdbs = env->me_numdbs;
2682 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2683 x = env->me_dbflags[i];
2684 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2685 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2687 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2688 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2690 if (env->me_flags & MDB_FATAL_ERROR) {
2691 DPUTS("environment had fatal error, must shutdown!");
2693 } else if (env->me_maxpg < txn->mt_next_pgno) {
2694 rc = MDB_MAP_RESIZED;
2698 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2703 mdb_txn_renew(MDB_txn *txn)
2707 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2710 rc = mdb_txn_renew0(txn);
2711 if (rc == MDB_SUCCESS) {
2712 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2713 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2714 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2720 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2724 int rc, size, tsize;
2726 flags &= MDB_TXN_BEGIN_FLAGS;
2727 flags |= env->me_flags & MDB_WRITEMAP;
2729 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2733 /* Nested transactions: Max 1 child, write txns only, no writemap */
2734 flags |= parent->mt_flags;
2735 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2736 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2738 /* Child txns save MDB_pgstate and use own copy of cursors */
2739 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2740 size += tsize = sizeof(MDB_ntxn);
2741 } else if (flags & MDB_RDONLY) {
2742 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2743 size += tsize = sizeof(MDB_txn);
2745 /* Reuse preallocated write txn. However, do not touch it until
2746 * mdb_txn_renew0() succeeds, since it currently may be active.
2751 if ((txn = calloc(1, size)) == NULL) {
2752 DPRINTF(("calloc: %s", strerror(errno)));
2755 txn->mt_dbxs = env->me_dbxs; /* static */
2756 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2757 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2758 txn->mt_flags = flags;
2763 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2764 txn->mt_dbiseqs = parent->mt_dbiseqs;
2765 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2766 if (!txn->mt_u.dirty_list ||
2767 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2769 free(txn->mt_u.dirty_list);
2773 txn->mt_txnid = parent->mt_txnid;
2774 txn->mt_dirty_room = parent->mt_dirty_room;
2775 txn->mt_u.dirty_list[0].mid = 0;
2776 txn->mt_spill_pgs = NULL;
2777 txn->mt_next_pgno = parent->mt_next_pgno;
2778 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2779 parent->mt_child = txn;
2780 txn->mt_parent = parent;
2781 txn->mt_numdbs = parent->mt_numdbs;
2782 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2783 /* Copy parent's mt_dbflags, but clear DB_NEW */
2784 for (i=0; i<txn->mt_numdbs; i++)
2785 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2787 ntxn = (MDB_ntxn *)txn;
2788 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2789 if (env->me_pghead) {
2790 size = MDB_IDL_SIZEOF(env->me_pghead);
2791 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2793 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2798 rc = mdb_cursor_shadow(parent, txn);
2800 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2801 } else { /* MDB_RDONLY */
2802 txn->mt_dbiseqs = env->me_dbiseqs;
2804 rc = mdb_txn_renew0(txn);
2807 if (txn != env->me_txn0)
2810 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2812 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2813 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2814 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2821 mdb_txn_env(MDB_txn *txn)
2823 if(!txn) return NULL;
2828 mdb_txn_id(MDB_txn *txn)
2831 return txn->mt_txnid;
2834 /** Export or close DBI handles opened in this txn. */
2836 mdb_dbis_update(MDB_txn *txn, int keep)
2839 MDB_dbi n = txn->mt_numdbs;
2840 MDB_env *env = txn->mt_env;
2841 unsigned char *tdbflags = txn->mt_dbflags;
2843 for (i = n; --i >= CORE_DBS;) {
2844 if (tdbflags[i] & DB_NEW) {
2846 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2848 char *ptr = env->me_dbxs[i].md_name.mv_data;
2850 env->me_dbxs[i].md_name.mv_data = NULL;
2851 env->me_dbxs[i].md_name.mv_size = 0;
2852 env->me_dbflags[i] = 0;
2853 env->me_dbiseqs[i]++;
2859 if (keep && env->me_numdbs < n)
2863 /** End a transaction, except successful commit of a nested transaction.
2864 * May be called twice for readonly txns: First reset it, then abort.
2865 * @param[in] txn the transaction handle to end
2866 * @param[in] mode why and how to end the transaction
2869 mdb_txn_end(MDB_txn *txn, unsigned mode)
2871 MDB_env *env = txn->mt_env;
2873 static const char *const names[] = MDB_END_NAMES;
2876 /* Export or close DBI handles opened in this txn */
2877 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2879 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2880 names[mode & MDB_END_OPMASK],
2881 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2882 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2884 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2885 if (txn->mt_u.reader) {
2886 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2887 if (!(env->me_flags & MDB_NOTLS)) {
2888 txn->mt_u.reader = NULL; /* txn does not own reader */
2889 } else if (mode & MDB_END_SLOT) {
2890 txn->mt_u.reader->mr_pid = 0;
2891 txn->mt_u.reader = NULL;
2892 } /* else txn owns the slot until it does MDB_END_SLOT */
2894 txn->mt_numdbs = 0; /* prevent further DBI activity */
2895 txn->mt_flags |= MDB_TXN_FINISHED;
2897 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2898 pgno_t *pghead = env->me_pghead;
2900 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2901 mdb_cursors_close(txn, 0);
2902 if (!(env->me_flags & MDB_WRITEMAP)) {
2903 mdb_dlist_free(txn);
2907 txn->mt_flags = MDB_TXN_FINISHED;
2909 if (!txn->mt_parent) {
2910 mdb_midl_shrink(&txn->mt_free_pgs);
2911 env->me_free_pgs = txn->mt_free_pgs;
2913 env->me_pghead = NULL;
2917 mode = 0; /* txn == env->me_txn0, do not free() it */
2919 /* The writer mutex was locked in mdb_txn_begin. */
2921 UNLOCK_MUTEX(env->me_wmutex);
2923 txn->mt_parent->mt_child = NULL;
2924 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2925 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2926 mdb_midl_free(txn->mt_free_pgs);
2927 mdb_midl_free(txn->mt_spill_pgs);
2928 free(txn->mt_u.dirty_list);
2931 mdb_midl_free(pghead);
2934 if (mode & MDB_END_FREE)
2939 mdb_txn_reset(MDB_txn *txn)
2944 /* This call is only valid for read-only txns */
2945 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2948 mdb_txn_end(txn, MDB_END_RESET);
2952 mdb_txn_abort(MDB_txn *txn)
2958 mdb_txn_abort(txn->mt_child);
2960 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
2963 /** Save the freelist as of this transaction to the freeDB.
2964 * This changes the freelist. Keep trying until it stabilizes.
2967 mdb_freelist_save(MDB_txn *txn)
2969 /* env->me_pghead[] can grow and shrink during this call.
2970 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2971 * Page numbers cannot disappear from txn->mt_free_pgs[].
2974 MDB_env *env = txn->mt_env;
2975 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2976 txnid_t pglast = 0, head_id = 0;
2977 pgno_t freecnt = 0, *free_pgs, *mop;
2978 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2980 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2982 if (env->me_pghead) {
2983 /* Make sure first page of freeDB is touched and on freelist */
2984 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2985 if (rc && rc != MDB_NOTFOUND)
2989 if (!env->me_pghead && txn->mt_loose_pgs) {
2990 /* Put loose page numbers in mt_free_pgs, since
2991 * we may be unable to return them to me_pghead.
2993 MDB_page *mp = txn->mt_loose_pgs;
2994 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2996 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2997 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2998 txn->mt_loose_pgs = NULL;
2999 txn->mt_loose_count = 0;
3002 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3003 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3004 ? SSIZE_MAX : maxfree_1pg;
3007 /* Come back here after each Put() in case freelist changed */
3012 /* If using records from freeDB which we have not yet
3013 * deleted, delete them and any we reserved for me_pghead.
3015 while (pglast < env->me_pglast) {
3016 rc = mdb_cursor_first(&mc, &key, NULL);
3019 pglast = head_id = *(txnid_t *)key.mv_data;
3020 total_room = head_room = 0;
3021 mdb_tassert(txn, pglast <= env->me_pglast);
3022 rc = mdb_cursor_del(&mc, 0);
3027 /* Save the IDL of pages freed by this txn, to a single record */
3028 if (freecnt < txn->mt_free_pgs[0]) {
3030 /* Make sure last page of freeDB is touched and on freelist */
3031 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3032 if (rc && rc != MDB_NOTFOUND)
3035 free_pgs = txn->mt_free_pgs;
3036 /* Write to last page of freeDB */
3037 key.mv_size = sizeof(txn->mt_txnid);
3038 key.mv_data = &txn->mt_txnid;
3040 freecnt = free_pgs[0];
3041 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3042 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3045 /* Retry if mt_free_pgs[] grew during the Put() */
3046 free_pgs = txn->mt_free_pgs;
3047 } while (freecnt < free_pgs[0]);
3048 mdb_midl_sort(free_pgs);
3049 memcpy(data.mv_data, free_pgs, data.mv_size);
3052 unsigned int i = free_pgs[0];
3053 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3054 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3056 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3062 mop = env->me_pghead;
3063 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3065 /* Reserve records for me_pghead[]. Split it if multi-page,
3066 * to avoid searching freeDB for a page range. Use keys in
3067 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3069 if (total_room >= mop_len) {
3070 if (total_room == mop_len || --more < 0)
3072 } else if (head_room >= maxfree_1pg && head_id > 1) {
3073 /* Keep current record (overflow page), add a new one */
3077 /* (Re)write {key = head_id, IDL length = head_room} */
3078 total_room -= head_room;
3079 head_room = mop_len - total_room;
3080 if (head_room > maxfree_1pg && head_id > 1) {
3081 /* Overflow multi-page for part of me_pghead */
3082 head_room /= head_id; /* amortize page sizes */
3083 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3084 } else if (head_room < 0) {
3085 /* Rare case, not bothering to delete this record */
3088 key.mv_size = sizeof(head_id);
3089 key.mv_data = &head_id;
3090 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3091 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3094 /* IDL is initially empty, zero out at least the length */
3095 pgs = (pgno_t *)data.mv_data;
3096 j = head_room > clean_limit ? head_room : 0;
3100 total_room += head_room;
3103 /* Return loose page numbers to me_pghead, though usually none are
3104 * left at this point. The pages themselves remain in dirty_list.
3106 if (txn->mt_loose_pgs) {
3107 MDB_page *mp = txn->mt_loose_pgs;
3108 unsigned count = txn->mt_loose_count;
3110 /* Room for loose pages + temp IDL with same */
3111 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3113 mop = env->me_pghead;
3114 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3115 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3116 loose[ ++count ] = mp->mp_pgno;
3118 mdb_midl_sort(loose);
3119 mdb_midl_xmerge(mop, loose);
3120 txn->mt_loose_pgs = NULL;
3121 txn->mt_loose_count = 0;
3125 /* Fill in the reserved me_pghead records */
3131 rc = mdb_cursor_first(&mc, &key, &data);
3132 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3133 txnid_t id = *(txnid_t *)key.mv_data;
3134 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3137 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3139 if (len > mop_len) {
3141 data.mv_size = (len + 1) * sizeof(MDB_ID);
3143 data.mv_data = mop -= len;
3146 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3148 if (rc || !(mop_len -= len))
3155 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3156 * @param[in] txn the transaction that's being committed
3157 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3158 * @return 0 on success, non-zero on failure.
3161 mdb_page_flush(MDB_txn *txn, int keep)
3163 MDB_env *env = txn->mt_env;
3164 MDB_ID2L dl = txn->mt_u.dirty_list;
3165 unsigned psize = env->me_psize, j;
3166 int i, pagecount = dl[0].mid, rc;
3167 size_t size = 0, pos = 0;
3169 MDB_page *dp = NULL;
3173 struct iovec iov[MDB_COMMIT_PAGES];
3174 ssize_t wpos = 0, wsize = 0, wres;
3175 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3181 if (env->me_flags & MDB_WRITEMAP) {
3182 /* Clear dirty flags */
3183 while (++i <= pagecount) {
3185 /* Don't flush this page yet */
3186 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3187 dp->mp_flags &= ~P_KEEP;
3191 dp->mp_flags &= ~P_DIRTY;
3196 /* Write the pages */
3198 if (++i <= pagecount) {
3200 /* Don't flush this page yet */
3201 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3202 dp->mp_flags &= ~P_KEEP;
3207 /* clear dirty flag */
3208 dp->mp_flags &= ~P_DIRTY;
3211 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3216 /* Windows actually supports scatter/gather I/O, but only on
3217 * unbuffered file handles. Since we're relying on the OS page
3218 * cache for all our data, that's self-defeating. So we just
3219 * write pages one at a time. We use the ov structure to set
3220 * the write offset, to at least save the overhead of a Seek
3223 DPRINTF(("committing page %"Z"u", pgno));
3224 memset(&ov, 0, sizeof(ov));
3225 ov.Offset = pos & 0xffffffff;
3226 ov.OffsetHigh = pos >> 16 >> 16;
3227 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3229 DPRINTF(("WriteFile: %d", rc));
3233 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3234 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3237 /* Write previous page(s) */
3238 #ifdef MDB_USE_PWRITEV
3239 wres = pwritev(env->me_fd, iov, n, wpos);
3242 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3245 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3249 DPRINTF(("lseek: %s", strerror(rc)));
3252 wres = writev(env->me_fd, iov, n);
3255 if (wres != wsize) {
3260 DPRINTF(("Write error: %s", strerror(rc)));
3262 rc = EIO; /* TODO: Use which error code? */
3263 DPUTS("short write, filesystem full?");
3274 DPRINTF(("committing page %"Z"u", pgno));
3275 next_pos = pos + size;
3276 iov[n].iov_len = size;
3277 iov[n].iov_base = (char *)dp;
3283 /* MIPS has cache coherency issues, this is a no-op everywhere else
3284 * Note: for any size >= on-chip cache size, entire on-chip cache is
3287 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3289 for (i = keep; ++i <= pagecount; ) {
3291 /* This is a page we skipped above */
3294 dl[j].mid = dp->mp_pgno;
3297 mdb_dpage_free(env, dp);
3302 txn->mt_dirty_room += i - j;
3308 mdb_txn_commit(MDB_txn *txn)
3311 unsigned int i, end_mode;
3317 /* mdb_txn_end() mode for a commit which writes nothing */
3318 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3320 if (txn->mt_child) {
3321 rc = mdb_txn_commit(txn->mt_child);
3328 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3332 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3333 DPUTS("txn has failed/finished, can't commit");
3335 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3340 if (txn->mt_parent) {
3341 MDB_txn *parent = txn->mt_parent;
3345 unsigned x, y, len, ps_len;
3347 /* Append our free list to parent's */
3348 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3351 mdb_midl_free(txn->mt_free_pgs);
3352 /* Failures after this must either undo the changes
3353 * to the parent or set MDB_TXN_ERROR in the parent.
3356 parent->mt_next_pgno = txn->mt_next_pgno;
3357 parent->mt_flags = txn->mt_flags;
3359 /* Merge our cursors into parent's and close them */
3360 mdb_cursors_close(txn, 1);
3362 /* Update parent's DB table. */
3363 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3364 parent->mt_numdbs = txn->mt_numdbs;
3365 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3366 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3367 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3368 /* preserve parent's DB_NEW status */
3369 x = parent->mt_dbflags[i] & DB_NEW;
3370 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3373 dst = parent->mt_u.dirty_list;
3374 src = txn->mt_u.dirty_list;
3375 /* Remove anything in our dirty list from parent's spill list */
3376 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3378 pspill[0] = (pgno_t)-1;
3379 /* Mark our dirty pages as deleted in parent spill list */
3380 for (i=0, len=src[0].mid; ++i <= len; ) {
3381 MDB_ID pn = src[i].mid << 1;
3382 while (pn > pspill[x])
3384 if (pn == pspill[x]) {
3389 /* Squash deleted pagenums if we deleted any */
3390 for (x=y; ++x <= ps_len; )
3391 if (!(pspill[x] & 1))
3392 pspill[++y] = pspill[x];
3396 /* Find len = length of merging our dirty list with parent's */
3398 dst[0].mid = 0; /* simplify loops */
3399 if (parent->mt_parent) {
3400 len = x + src[0].mid;
3401 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3402 for (i = x; y && i; y--) {
3403 pgno_t yp = src[y].mid;
3404 while (yp < dst[i].mid)
3406 if (yp == dst[i].mid) {
3411 } else { /* Simplify the above for single-ancestor case */
3412 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3414 /* Merge our dirty list with parent's */
3416 for (i = len; y; dst[i--] = src[y--]) {
3417 pgno_t yp = src[y].mid;
3418 while (yp < dst[x].mid)
3419 dst[i--] = dst[x--];
3420 if (yp == dst[x].mid)
3421 free(dst[x--].mptr);
3423 mdb_tassert(txn, i == x);
3425 free(txn->mt_u.dirty_list);
3426 parent->mt_dirty_room = txn->mt_dirty_room;
3427 if (txn->mt_spill_pgs) {
3428 if (parent->mt_spill_pgs) {
3429 /* TODO: Prevent failure here, so parent does not fail */
3430 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3432 parent->mt_flags |= MDB_TXN_ERROR;
3433 mdb_midl_free(txn->mt_spill_pgs);
3434 mdb_midl_sort(parent->mt_spill_pgs);
3436 parent->mt_spill_pgs = txn->mt_spill_pgs;
3440 /* Append our loose page list to parent's */
3441 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3443 *lp = txn->mt_loose_pgs;
3444 parent->mt_loose_count += txn->mt_loose_count;
3446 parent->mt_child = NULL;
3447 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3452 if (txn != env->me_txn) {
3453 DPUTS("attempt to commit unknown transaction");
3458 mdb_cursors_close(txn, 0);
3460 if (!txn->mt_u.dirty_list[0].mid &&
3461 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3464 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3465 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3467 /* Update DB root pointers */
3468 if (txn->mt_numdbs > CORE_DBS) {
3472 data.mv_size = sizeof(MDB_db);
3474 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3475 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3476 if (txn->mt_dbflags[i] & DB_DIRTY) {
3477 if (TXN_DBI_CHANGED(txn, i)) {
3481 data.mv_data = &txn->mt_dbs[i];
3482 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3490 rc = mdb_freelist_save(txn);
3494 mdb_midl_free(env->me_pghead);
3495 env->me_pghead = NULL;
3496 mdb_midl_shrink(&txn->mt_free_pgs);
3502 if ((rc = mdb_page_flush(txn, 0)) ||
3503 (rc = mdb_env_sync(env, 0)) ||
3504 (rc = mdb_env_write_meta(txn)))
3506 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3509 mdb_txn_end(txn, end_mode);
3517 /** Read the environment parameters of a DB environment before
3518 * mapping it into memory.
3519 * @param[in] env the environment handle
3520 * @param[out] meta address of where to store the meta information
3521 * @return 0 on success, non-zero on failure.
3524 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3530 enum { Size = sizeof(pbuf) };
3532 /* We don't know the page size yet, so use a minimum value.
3533 * Read both meta pages so we can use the latest one.
3536 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3540 memset(&ov, 0, sizeof(ov));
3542 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3543 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3546 rc = pread(env->me_fd, &pbuf, Size, off);
3549 if (rc == 0 && off == 0)
3551 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3552 DPRINTF(("read: %s", mdb_strerror(rc)));
3556 p = (MDB_page *)&pbuf;
3558 if (!F_ISSET(p->mp_flags, P_META)) {
3559 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3564 if (m->mm_magic != MDB_MAGIC) {
3565 DPUTS("meta has invalid magic");
3569 if (m->mm_version != MDB_DATA_VERSION) {
3570 DPRINTF(("database is version %u, expected version %u",
3571 m->mm_version, MDB_DATA_VERSION));
3572 return MDB_VERSION_MISMATCH;
3575 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3581 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3583 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3585 meta->mm_magic = MDB_MAGIC;
3586 meta->mm_version = MDB_DATA_VERSION;
3587 meta->mm_mapsize = env->me_mapsize;
3588 meta->mm_psize = env->me_psize;
3589 meta->mm_last_pg = NUM_METAS-1;
3590 meta->mm_flags = env->me_flags & 0xffff;
3591 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3592 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3593 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3596 /** Write the environment parameters of a freshly created DB environment.
3597 * @param[in] env the environment handle
3598 * @param[in] meta the #MDB_meta to write
3599 * @return 0 on success, non-zero on failure.
3602 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3610 memset(&ov, 0, sizeof(ov));
3611 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3613 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3616 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3617 len = pwrite(fd, ptr, size, pos); \
3618 if (len == -1 && ErrCode() == EINTR) continue; \
3619 rc = (len >= 0); break; } while(1)
3622 DPUTS("writing new meta page");
3624 psize = env->me_psize;
3626 p = calloc(NUM_METAS, psize);
3631 p->mp_flags = P_META;
3632 *(MDB_meta *)METADATA(p) = *meta;
3634 q = (MDB_page *)((char *)p + psize);
3636 q->mp_flags = P_META;
3637 *(MDB_meta *)METADATA(q) = *meta;
3639 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3642 else if ((unsigned) len == psize * NUM_METAS)
3650 /** Update the environment info to commit a transaction.
3651 * @param[in] txn the transaction that's being committed
3652 * @return 0 on success, non-zero on failure.
3655 mdb_env_write_meta(MDB_txn *txn)
3658 MDB_meta meta, metab, *mp;
3662 int rc, len, toggle;
3671 toggle = txn->mt_txnid & 1;
3672 DPRINTF(("writing meta page %d for root page %"Z"u",
3673 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3676 flags = env->me_flags;
3677 mp = env->me_metas[toggle];
3678 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3679 /* Persist any increases of mapsize config */
3680 if (mapsize < env->me_mapsize)
3681 mapsize = env->me_mapsize;
3683 if (flags & MDB_WRITEMAP) {
3684 mp->mm_mapsize = mapsize;
3685 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3686 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3687 mp->mm_last_pg = txn->mt_next_pgno - 1;
3688 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3689 !(defined(__i386__) || defined(__x86_64__))
3690 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3691 __sync_synchronize();
3693 mp->mm_txnid = txn->mt_txnid;
3694 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3695 unsigned meta_size = env->me_psize;
3696 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3699 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3700 if (meta_size < env->me_os_psize)
3701 meta_size += meta_size;
3706 if (MDB_MSYNC(ptr, meta_size, rc)) {
3713 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3714 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3716 meta.mm_mapsize = mapsize;
3717 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3718 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3719 meta.mm_last_pg = txn->mt_next_pgno - 1;
3720 meta.mm_txnid = txn->mt_txnid;
3722 off = offsetof(MDB_meta, mm_mapsize);
3723 ptr = (char *)&meta + off;
3724 len = sizeof(MDB_meta) - off;
3726 off += env->me_psize;
3729 /* Write to the SYNC fd */
3730 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3733 memset(&ov, 0, sizeof(ov));
3735 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3740 rc = pwrite(mfd, ptr, len, off);
3743 rc = rc < 0 ? ErrCode() : EIO;
3748 DPUTS("write failed, disk error?");
3749 /* On a failure, the pagecache still contains the new data.
3750 * Write some old data back, to prevent it from being used.
3751 * Use the non-SYNC fd; we know it will fail anyway.
3753 meta.mm_last_pg = metab.mm_last_pg;
3754 meta.mm_txnid = metab.mm_txnid;
3756 memset(&ov, 0, sizeof(ov));
3758 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3760 r2 = pwrite(env->me_fd, ptr, len, off);
3761 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3764 env->me_flags |= MDB_FATAL_ERROR;
3767 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3768 CACHEFLUSH(env->me_map + off, len, DCACHE);
3770 /* Memory ordering issues are irrelevant; since the entire writer
3771 * is wrapped by wmutex, all of these changes will become visible
3772 * after the wmutex is unlocked. Since the DB is multi-version,
3773 * readers will get consistent data regardless of how fresh or
3774 * how stale their view of these values is.
3777 env->me_txns->mti_txnid = txn->mt_txnid;
3782 /** Check both meta pages to see which one is newer.
3783 * @param[in] env the environment handle
3784 * @return newest #MDB_meta.
3787 mdb_env_pick_meta(const MDB_env *env)
3789 MDB_meta *const *metas = env->me_metas;
3790 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3794 mdb_env_create(MDB_env **env)
3798 e = calloc(1, sizeof(MDB_env));
3802 e->me_maxreaders = DEFAULT_READERS;
3803 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3804 e->me_fd = INVALID_HANDLE_VALUE;
3805 e->me_lfd = INVALID_HANDLE_VALUE;
3806 e->me_mfd = INVALID_HANDLE_VALUE;
3807 #ifdef MDB_USE_POSIX_SEM
3808 e->me_rmutex = SEM_FAILED;
3809 e->me_wmutex = SEM_FAILED;
3811 e->me_pid = getpid();
3812 GET_PAGESIZE(e->me_os_psize);
3813 VGMEMP_CREATE(e,0,0);
3819 mdb_env_map(MDB_env *env, void *addr)
3822 unsigned int flags = env->me_flags;
3826 LONG sizelo, sizehi;
3829 if (flags & MDB_RDONLY) {
3830 /* Don't set explicit map size, use whatever exists */
3835 msize = env->me_mapsize;
3836 sizelo = msize & 0xffffffff;
3837 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3839 /* Windows won't create mappings for zero length files.
3840 * and won't map more than the file size.
3841 * Just set the maxsize right now.
3843 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3844 || !SetEndOfFile(env->me_fd)
3845 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3849 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3850 PAGE_READWRITE : PAGE_READONLY,
3851 sizehi, sizelo, NULL);
3854 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3855 FILE_MAP_WRITE : FILE_MAP_READ,
3857 rc = env->me_map ? 0 : ErrCode();
3862 int prot = PROT_READ;
3863 if (flags & MDB_WRITEMAP) {
3865 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3868 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3870 if (env->me_map == MAP_FAILED) {
3875 if (flags & MDB_NORDAHEAD) {
3876 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3878 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3880 #ifdef POSIX_MADV_RANDOM
3881 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3882 #endif /* POSIX_MADV_RANDOM */
3883 #endif /* MADV_RANDOM */
3887 /* Can happen because the address argument to mmap() is just a
3888 * hint. mmap() can pick another, e.g. if the range is in use.
3889 * The MAP_FIXED flag would prevent that, but then mmap could
3890 * instead unmap existing pages to make room for the new map.
3892 if (addr && env->me_map != addr)
3893 return EBUSY; /* TODO: Make a new MDB_* error code? */
3895 p = (MDB_page *)env->me_map;
3896 env->me_metas[0] = METADATA(p);
3897 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3903 mdb_env_set_mapsize(MDB_env *env, size_t size)
3905 /* If env is already open, caller is responsible for making
3906 * sure there are no active txns.
3914 meta = mdb_env_pick_meta(env);
3916 size = meta->mm_mapsize;
3918 /* Silently round up to minimum if the size is too small */
3919 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3923 munmap(env->me_map, env->me_mapsize);
3924 env->me_mapsize = size;
3925 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3926 rc = mdb_env_map(env, old);
3930 env->me_mapsize = size;
3932 env->me_maxpg = env->me_mapsize / env->me_psize;
3937 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3941 env->me_maxdbs = dbs + CORE_DBS;
3946 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3948 if (env->me_map || readers < 1)
3950 env->me_maxreaders = readers;
3955 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3957 if (!env || !readers)
3959 *readers = env->me_maxreaders;
3964 mdb_fsize(HANDLE fd, size_t *size)
3967 LARGE_INTEGER fsize;
3969 if (!GetFileSizeEx(fd, &fsize))
3972 *size = fsize.QuadPart;
3984 #ifdef BROKEN_FDATASYNC
3985 #include <sys/utsname.h>
3986 #include <sys/vfs.h>
3989 /** Further setup required for opening an LMDB environment
3992 mdb_env_open2(MDB_env *env)
3994 unsigned int flags = env->me_flags;
3995 int i, newenv = 0, rc;
3999 /* See if we should use QueryLimited */
4001 if ((rc & 0xff) > 5)
4002 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4004 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4007 #ifdef BROKEN_FDATASYNC
4008 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4009 * https://lkml.org/lkml/2012/9/3/83
4010 * Kernels after 3.6-rc6 are known good.
4011 * https://lkml.org/lkml/2012/9/10/556
4012 * See if the DB is on ext3/ext4, then check for new enough kernel
4013 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4018 fstatfs(env->me_fd, &st);
4019 while (st.f_type == 0xEF53) {
4023 if (uts.release[0] < '3') {
4024 if (!strncmp(uts.release, "2.6.32.", 7)) {
4025 i = atoi(uts.release+7);
4027 break; /* 2.6.32.60 and newer is OK */
4028 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4029 i = atoi(uts.release+7);
4031 break; /* 2.6.34.15 and newer is OK */
4033 } else if (uts.release[0] == '3') {
4034 i = atoi(uts.release+2);
4036 break; /* 3.6 and newer is OK */
4038 i = atoi(uts.release+4);
4040 break; /* 3.5.4 and newer is OK */
4041 } else if (i == 2) {
4042 i = atoi(uts.release+4);
4044 break; /* 3.2.30 and newer is OK */
4046 } else { /* 4.x and newer is OK */
4049 env->me_flags |= MDB_FSYNCONLY;
4055 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4058 DPUTS("new mdbenv");
4060 env->me_psize = env->me_os_psize;
4061 if (env->me_psize > MAX_PAGESIZE)
4062 env->me_psize = MAX_PAGESIZE;
4063 memset(&meta, 0, sizeof(meta));
4064 mdb_env_init_meta0(env, &meta);
4065 meta.mm_mapsize = DEFAULT_MAPSIZE;
4067 env->me_psize = meta.mm_psize;
4070 /* Was a mapsize configured? */
4071 if (!env->me_mapsize) {
4072 env->me_mapsize = meta.mm_mapsize;
4075 /* Make sure mapsize >= committed data size. Even when using
4076 * mm_mapsize, which could be broken in old files (ITS#7789).
4078 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4079 if (env->me_mapsize < minsize)
4080 env->me_mapsize = minsize;
4082 meta.mm_mapsize = env->me_mapsize;
4084 if (newenv && !(flags & MDB_FIXEDMAP)) {
4085 /* mdb_env_map() may grow the datafile. Write the metapages
4086 * first, so the file will be valid if initialization fails.
4087 * Except with FIXEDMAP, since we do not yet know mm_address.
4088 * We could fill in mm_address later, but then a different
4089 * program might end up doing that - one with a memory layout
4090 * and map address which does not suit the main program.
4092 rc = mdb_env_init_meta(env, &meta);
4098 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4103 if (flags & MDB_FIXEDMAP)
4104 meta.mm_address = env->me_map;
4105 i = mdb_env_init_meta(env, &meta);
4106 if (i != MDB_SUCCESS) {
4111 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4112 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4114 #if !(MDB_MAXKEYSIZE)
4115 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4117 env->me_maxpg = env->me_mapsize / env->me_psize;
4121 MDB_meta *meta = mdb_env_pick_meta(env);
4122 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4124 DPRINTF(("opened database version %u, pagesize %u",
4125 meta->mm_version, env->me_psize));
4126 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4127 DPRINTF(("depth: %u", db->md_depth));
4128 DPRINTF(("entries: %"Z"u", db->md_entries));
4129 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4130 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4131 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4132 DPRINTF(("root: %"Z"u", db->md_root));
4140 /** Release a reader thread's slot in the reader lock table.
4141 * This function is called automatically when a thread exits.
4142 * @param[in] ptr This points to the slot in the reader lock table.
4145 mdb_env_reader_dest(void *ptr)
4147 MDB_reader *reader = ptr;
4153 /** Junk for arranging thread-specific callbacks on Windows. This is
4154 * necessarily platform and compiler-specific. Windows supports up
4155 * to 1088 keys. Let's assume nobody opens more than 64 environments
4156 * in a single process, for now. They can override this if needed.
4158 #ifndef MAX_TLS_KEYS
4159 #define MAX_TLS_KEYS 64
4161 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4162 static int mdb_tls_nkeys;
4164 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4168 case DLL_PROCESS_ATTACH: break;
4169 case DLL_THREAD_ATTACH: break;
4170 case DLL_THREAD_DETACH:
4171 for (i=0; i<mdb_tls_nkeys; i++) {
4172 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4174 mdb_env_reader_dest(r);
4178 case DLL_PROCESS_DETACH: break;
4183 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4185 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4189 /* Force some symbol references.
4190 * _tls_used forces the linker to create the TLS directory if not already done
4191 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4193 #pragma comment(linker, "/INCLUDE:_tls_used")
4194 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4195 #pragma const_seg(".CRT$XLB")
4196 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4197 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4200 #pragma comment(linker, "/INCLUDE:__tls_used")
4201 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4202 #pragma data_seg(".CRT$XLB")
4203 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4205 #endif /* WIN 32/64 */
4206 #endif /* !__GNUC__ */
4209 /** Downgrade the exclusive lock on the region back to shared */
4211 mdb_env_share_locks(MDB_env *env, int *excl)
4214 MDB_meta *meta = mdb_env_pick_meta(env);
4216 env->me_txns->mti_txnid = meta->mm_txnid;
4221 /* First acquire a shared lock. The Unlock will
4222 * then release the existing exclusive lock.
4224 memset(&ov, 0, sizeof(ov));
4225 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4228 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4234 struct flock lock_info;
4235 /* The shared lock replaces the existing lock */
4236 memset((void *)&lock_info, 0, sizeof(lock_info));
4237 lock_info.l_type = F_RDLCK;
4238 lock_info.l_whence = SEEK_SET;
4239 lock_info.l_start = 0;
4240 lock_info.l_len = 1;
4241 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4242 (rc = ErrCode()) == EINTR) ;
4243 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4250 /** Try to get exclusive lock, otherwise shared.
4251 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4254 mdb_env_excl_lock(MDB_env *env, int *excl)
4258 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4262 memset(&ov, 0, sizeof(ov));
4263 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4270 struct flock lock_info;
4271 memset((void *)&lock_info, 0, sizeof(lock_info));
4272 lock_info.l_type = F_WRLCK;
4273 lock_info.l_whence = SEEK_SET;
4274 lock_info.l_start = 0;
4275 lock_info.l_len = 1;
4276 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4277 (rc = ErrCode()) == EINTR) ;
4281 # ifndef MDB_USE_POSIX_MUTEX
4282 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4285 lock_info.l_type = F_RDLCK;
4286 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4287 (rc = ErrCode()) == EINTR) ;
4297 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4299 * @(#) $Revision: 5.1 $
4300 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4301 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4303 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4307 * Please do not copyright this code. This code is in the public domain.
4309 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4310 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4311 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4312 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4313 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4314 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4315 * PERFORMANCE OF THIS SOFTWARE.
4318 * chongo <Landon Curt Noll> /\oo/\
4319 * http://www.isthe.com/chongo/
4321 * Share and Enjoy! :-)
4324 typedef unsigned long long mdb_hash_t;
4325 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4327 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4328 * @param[in] val value to hash
4329 * @param[in] hval initial value for hash
4330 * @return 64 bit hash
4332 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4333 * hval arg on the first call.
4336 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4338 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4339 unsigned char *end = s + val->mv_size;
4341 * FNV-1a hash each octet of the string
4344 /* xor the bottom with the current octet */
4345 hval ^= (mdb_hash_t)*s++;
4347 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4348 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4349 (hval << 7) + (hval << 8) + (hval << 40);
4351 /* return our new hash value */
4355 /** Hash the string and output the encoded hash.
4356 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4357 * very short name limits. We don't care about the encoding being reversible,
4358 * we just want to preserve as many bits of the input as possible in a
4359 * small printable string.
4360 * @param[in] str string to hash
4361 * @param[out] encbuf an array of 11 chars to hold the hash
4363 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4366 mdb_pack85(unsigned long l, char *out)
4370 for (i=0; i<5; i++) {
4371 *out++ = mdb_a85[l % 85];
4377 mdb_hash_enc(MDB_val *val, char *encbuf)
4379 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4381 mdb_pack85(h, encbuf);
4382 mdb_pack85(h>>32, encbuf+5);
4387 /** Open and/or initialize the lock region for the environment.
4388 * @param[in] env The LMDB environment.
4389 * @param[in] lpath The pathname of the file used for the lock region.
4390 * @param[in] mode The Unix permissions for the file, if we create it.
4391 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4392 * @return 0 on success, non-zero on failure.
4395 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4398 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4400 # define MDB_ERRCODE_ROFS EROFS
4401 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4402 # define MDB_CLOEXEC O_CLOEXEC
4405 # define MDB_CLOEXEC 0
4412 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4413 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4414 FILE_ATTRIBUTE_NORMAL, NULL);
4416 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4418 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4420 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4425 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4426 /* Lose record locks when exec*() */
4427 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4428 fcntl(env->me_lfd, F_SETFD, fdflags);
4431 if (!(env->me_flags & MDB_NOTLS)) {
4432 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4435 env->me_flags |= MDB_ENV_TXKEY;
4437 /* Windows TLS callbacks need help finding their TLS info. */
4438 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4442 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4446 /* Try to get exclusive lock. If we succeed, then
4447 * nobody is using the lock region and we should initialize it.
4449 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4452 size = GetFileSize(env->me_lfd, NULL);
4454 size = lseek(env->me_lfd, 0, SEEK_END);
4455 if (size == -1) goto fail_errno;
4457 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4458 if (size < rsize && *excl > 0) {
4460 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4461 || !SetEndOfFile(env->me_lfd))
4464 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4468 size = rsize - sizeof(MDB_txninfo);
4469 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4474 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4476 if (!mh) goto fail_errno;
4477 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4479 if (!env->me_txns) goto fail_errno;
4481 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4483 if (m == MAP_FAILED) goto fail_errno;
4489 BY_HANDLE_FILE_INFORMATION stbuf;
4498 if (!mdb_sec_inited) {
4499 InitializeSecurityDescriptor(&mdb_null_sd,
4500 SECURITY_DESCRIPTOR_REVISION);
4501 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4502 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4503 mdb_all_sa.bInheritHandle = FALSE;
4504 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4507 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4508 idbuf.volume = stbuf.dwVolumeSerialNumber;
4509 idbuf.nhigh = stbuf.nFileIndexHigh;
4510 idbuf.nlow = stbuf.nFileIndexLow;
4511 val.mv_data = &idbuf;
4512 val.mv_size = sizeof(idbuf);
4513 mdb_hash_enc(&val, encbuf);
4514 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4515 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4516 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4517 if (!env->me_rmutex) goto fail_errno;
4518 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4519 if (!env->me_wmutex) goto fail_errno;
4520 #elif defined(MDB_USE_POSIX_SEM)
4529 #if defined(__NetBSD__)
4530 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4532 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4533 idbuf.dev = stbuf.st_dev;
4534 idbuf.ino = stbuf.st_ino;
4535 val.mv_data = &idbuf;
4536 val.mv_size = sizeof(idbuf);
4537 mdb_hash_enc(&val, encbuf);
4538 #ifdef MDB_SHORT_SEMNAMES
4539 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4541 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4542 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4543 /* Clean up after a previous run, if needed: Try to
4544 * remove both semaphores before doing anything else.
4546 sem_unlink(env->me_txns->mti_rmname);
4547 sem_unlink(env->me_txns->mti_wmname);
4548 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4549 O_CREAT|O_EXCL, mode, 1);
4550 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4551 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4552 O_CREAT|O_EXCL, mode, 1);
4553 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4554 #else /* MDB_USE_POSIX_MUTEX: */
4555 pthread_mutexattr_t mattr;
4557 if ((rc = pthread_mutexattr_init(&mattr))
4558 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4559 #ifdef MDB_ROBUST_SUPPORTED
4560 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4562 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4563 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4565 pthread_mutexattr_destroy(&mattr);
4566 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4568 env->me_txns->mti_magic = MDB_MAGIC;
4569 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4570 env->me_txns->mti_txnid = 0;
4571 env->me_txns->mti_numreaders = 0;
4574 if (env->me_txns->mti_magic != MDB_MAGIC) {
4575 DPUTS("lock region has invalid magic");
4579 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4580 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4581 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4582 rc = MDB_VERSION_MISMATCH;
4586 if (rc && rc != EACCES && rc != EAGAIN) {
4590 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4591 if (!env->me_rmutex) goto fail_errno;
4592 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4593 if (!env->me_wmutex) goto fail_errno;
4594 #elif defined(MDB_USE_POSIX_SEM)
4595 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4596 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4597 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4598 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4609 /** The name of the lock file in the DB environment */
4610 #define LOCKNAME "/lock.mdb"
4611 /** The name of the data file in the DB environment */
4612 #define DATANAME "/data.mdb"
4613 /** The suffix of the lock file when no subdir is used */
4614 #define LOCKSUFF "-lock"
4615 /** Only a subset of the @ref mdb_env flags can be changed
4616 * at runtime. Changing other flags requires closing the
4617 * environment and re-opening it with the new flags.
4619 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4620 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4621 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4623 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4624 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4628 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4630 int oflags, rc, len, excl = -1;
4631 char *lpath, *dpath;
4633 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4637 if (flags & MDB_NOSUBDIR) {
4638 rc = len + sizeof(LOCKSUFF) + len + 1;
4640 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4645 if (flags & MDB_NOSUBDIR) {
4646 dpath = lpath + len + sizeof(LOCKSUFF);
4647 sprintf(lpath, "%s" LOCKSUFF, path);
4648 strcpy(dpath, path);
4650 dpath = lpath + len + sizeof(LOCKNAME);
4651 sprintf(lpath, "%s" LOCKNAME, path);
4652 sprintf(dpath, "%s" DATANAME, path);
4656 flags |= env->me_flags;
4657 if (flags & MDB_RDONLY) {
4658 /* silently ignore WRITEMAP when we're only getting read access */
4659 flags &= ~MDB_WRITEMAP;
4661 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4662 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4665 env->me_flags = flags |= MDB_ENV_ACTIVE;
4669 env->me_path = strdup(path);
4670 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4671 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4672 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4673 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4677 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4679 /* For RDONLY, get lockfile after we know datafile exists */
4680 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4681 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4687 if (F_ISSET(flags, MDB_RDONLY)) {
4688 oflags = GENERIC_READ;
4689 len = OPEN_EXISTING;
4691 oflags = GENERIC_READ|GENERIC_WRITE;
4694 mode = FILE_ATTRIBUTE_NORMAL;
4695 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4696 NULL, len, mode, NULL);
4698 if (F_ISSET(flags, MDB_RDONLY))
4701 oflags = O_RDWR | O_CREAT;
4703 env->me_fd = open(dpath, oflags, mode);
4705 if (env->me_fd == INVALID_HANDLE_VALUE) {
4710 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4711 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4716 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4717 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4718 env->me_mfd = env->me_fd;
4720 /* Synchronous fd for meta writes. Needed even with
4721 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4724 len = OPEN_EXISTING;
4725 env->me_mfd = CreateFile(dpath, oflags,
4726 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4727 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4730 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4732 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4737 DPRINTF(("opened dbenv %p", (void *) env));
4739 rc = mdb_env_share_locks(env, &excl);
4743 if (!(flags & MDB_RDONLY)) {
4745 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4746 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4747 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4748 (txn = calloc(1, size)))
4750 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4751 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4752 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4753 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4755 txn->mt_dbxs = env->me_dbxs;
4756 txn->mt_flags = MDB_TXN_FINISHED;
4766 mdb_env_close0(env, excl);
4772 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4774 mdb_env_close0(MDB_env *env, int excl)
4778 if (!(env->me_flags & MDB_ENV_ACTIVE))
4781 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4783 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4784 free(env->me_dbxs[i].md_name.mv_data);
4789 free(env->me_dbiseqs);
4790 free(env->me_dbflags);
4792 free(env->me_dirty_list);
4794 mdb_midl_free(env->me_free_pgs);
4796 if (env->me_flags & MDB_ENV_TXKEY) {
4797 pthread_key_delete(env->me_txkey);
4799 /* Delete our key from the global list */
4800 for (i=0; i<mdb_tls_nkeys; i++)
4801 if (mdb_tls_keys[i] == env->me_txkey) {
4802 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4810 munmap(env->me_map, env->me_mapsize);
4812 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4813 (void) close(env->me_mfd);
4814 if (env->me_fd != INVALID_HANDLE_VALUE)
4815 (void) close(env->me_fd);
4817 MDB_PID_T pid = env->me_pid;
4818 /* Clearing readers is done in this function because
4819 * me_txkey with its destructor must be disabled first.
4821 * We skip the the reader mutex, so we touch only
4822 * data owned by this process (me_close_readers and
4823 * our readers), and clear each reader atomically.
4825 for (i = env->me_close_readers; --i >= 0; )
4826 if (env->me_txns->mti_readers[i].mr_pid == pid)
4827 env->me_txns->mti_readers[i].mr_pid = 0;
4829 if (env->me_rmutex) {
4830 CloseHandle(env->me_rmutex);
4831 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4833 /* Windows automatically destroys the mutexes when
4834 * the last handle closes.
4836 #elif defined(MDB_USE_POSIX_SEM)
4837 if (env->me_rmutex != SEM_FAILED) {
4838 sem_close(env->me_rmutex);
4839 if (env->me_wmutex != SEM_FAILED)
4840 sem_close(env->me_wmutex);
4841 /* If we have the filelock: If we are the
4842 * only remaining user, clean up semaphores.
4845 mdb_env_excl_lock(env, &excl);
4847 sem_unlink(env->me_txns->mti_rmname);
4848 sem_unlink(env->me_txns->mti_wmname);
4852 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4854 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4857 /* Unlock the lockfile. Windows would have unlocked it
4858 * after closing anyway, but not necessarily at once.
4860 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4863 (void) close(env->me_lfd);
4866 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4870 mdb_env_close(MDB_env *env)
4877 VGMEMP_DESTROY(env);
4878 while ((dp = env->me_dpages) != NULL) {
4879 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4880 env->me_dpages = dp->mp_next;
4884 mdb_env_close0(env, 0);
4888 /** Compare two items pointing at aligned size_t's */
4890 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4892 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4893 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4896 /** Compare two items pointing at aligned unsigned int's.
4898 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4899 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4902 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4904 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4905 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4908 /** Compare two items pointing at unsigned ints of unknown alignment.
4909 * Nodes and keys are guaranteed to be 2-byte aligned.
4912 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4914 #if BYTE_ORDER == LITTLE_ENDIAN
4915 unsigned short *u, *c;
4918 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4919 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4922 } while(!x && u > (unsigned short *)a->mv_data);
4925 unsigned short *u, *c, *end;
4928 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4929 u = (unsigned short *)a->mv_data;
4930 c = (unsigned short *)b->mv_data;
4933 } while(!x && u < end);
4938 /** Compare two items lexically */
4940 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4947 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4953 diff = memcmp(a->mv_data, b->mv_data, len);
4954 return diff ? diff : len_diff<0 ? -1 : len_diff;
4957 /** Compare two items in reverse byte order */
4959 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4961 const unsigned char *p1, *p2, *p1_lim;
4965 p1_lim = (const unsigned char *)a->mv_data;
4966 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4967 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4969 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4975 while (p1 > p1_lim) {
4976 diff = *--p1 - *--p2;
4980 return len_diff<0 ? -1 : len_diff;
4983 /** Search for key within a page, using binary search.
4984 * Returns the smallest entry larger or equal to the key.
4985 * If exactp is non-null, stores whether the found entry was an exact match
4986 * in *exactp (1 or 0).
4987 * Updates the cursor index with the index of the found entry.
4988 * If no entry larger or equal to the key is found, returns NULL.
4991 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4993 unsigned int i = 0, nkeys;
4996 MDB_page *mp = mc->mc_pg[mc->mc_top];
4997 MDB_node *node = NULL;
5002 nkeys = NUMKEYS(mp);
5004 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5005 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5008 low = IS_LEAF(mp) ? 0 : 1;
5010 cmp = mc->mc_dbx->md_cmp;
5012 /* Branch pages have no data, so if using integer keys,
5013 * alignment is guaranteed. Use faster mdb_cmp_int.
5015 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5016 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5023 nodekey.mv_size = mc->mc_db->md_pad;
5024 node = NODEPTR(mp, 0); /* fake */
5025 while (low <= high) {
5026 i = (low + high) >> 1;
5027 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5028 rc = cmp(key, &nodekey);
5029 DPRINTF(("found leaf index %u [%s], rc = %i",
5030 i, DKEY(&nodekey), rc));
5039 while (low <= high) {
5040 i = (low + high) >> 1;
5042 node = NODEPTR(mp, i);
5043 nodekey.mv_size = NODEKSZ(node);
5044 nodekey.mv_data = NODEKEY(node);
5046 rc = cmp(key, &nodekey);
5049 DPRINTF(("found leaf index %u [%s], rc = %i",
5050 i, DKEY(&nodekey), rc));
5052 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5053 i, DKEY(&nodekey), NODEPGNO(node), rc));
5064 if (rc > 0) { /* Found entry is less than the key. */
5065 i++; /* Skip to get the smallest entry larger than key. */
5067 node = NODEPTR(mp, i);
5070 *exactp = (rc == 0 && nkeys > 0);
5071 /* store the key index */
5072 mc->mc_ki[mc->mc_top] = i;
5074 /* There is no entry larger or equal to the key. */
5077 /* nodeptr is fake for LEAF2 */
5083 mdb_cursor_adjust(MDB_cursor *mc, func)
5087 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5088 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5095 /** Pop a page off the top of the cursor's stack. */
5097 mdb_cursor_pop(MDB_cursor *mc)
5100 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5101 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5109 /** Push a page onto the top of the cursor's stack. */
5111 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5113 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5114 DDBI(mc), (void *) mc));
5116 if (mc->mc_snum >= CURSOR_STACK) {
5117 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5118 return MDB_CURSOR_FULL;
5121 mc->mc_top = mc->mc_snum++;
5122 mc->mc_pg[mc->mc_top] = mp;
5123 mc->mc_ki[mc->mc_top] = 0;
5128 /** Find the address of the page corresponding to a given page number.
5129 * @param[in] txn the transaction for this access.
5130 * @param[in] pgno the page number for the page to retrieve.
5131 * @param[out] ret address of a pointer where the page's address will be stored.
5132 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5133 * @return 0 on success, non-zero on failure.
5136 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5138 MDB_env *env = txn->mt_env;
5142 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5146 MDB_ID2L dl = tx2->mt_u.dirty_list;
5148 /* Spilled pages were dirtied in this txn and flushed
5149 * because the dirty list got full. Bring this page
5150 * back in from the map (but don't unspill it here,
5151 * leave that unless page_touch happens again).
5153 if (tx2->mt_spill_pgs) {
5154 MDB_ID pn = pgno << 1;
5155 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5156 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5157 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5162 unsigned x = mdb_mid2l_search(dl, pgno);
5163 if (x <= dl[0].mid && dl[x].mid == pgno) {
5169 } while ((tx2 = tx2->mt_parent) != NULL);
5172 if (pgno < txn->mt_next_pgno) {
5174 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5176 DPRINTF(("page %"Z"u not found", pgno));
5177 txn->mt_flags |= MDB_TXN_ERROR;
5178 return MDB_PAGE_NOTFOUND;
5188 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5189 * The cursor is at the root page, set up the rest of it.
5192 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5194 MDB_page *mp = mc->mc_pg[mc->mc_top];
5198 while (IS_BRANCH(mp)) {
5202 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5203 mdb_cassert(mc, NUMKEYS(mp) > 1);
5204 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5206 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5208 if (flags & MDB_PS_LAST)
5209 i = NUMKEYS(mp) - 1;
5212 node = mdb_node_search(mc, key, &exact);
5214 i = NUMKEYS(mp) - 1;
5216 i = mc->mc_ki[mc->mc_top];
5218 mdb_cassert(mc, i > 0);
5222 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5225 mdb_cassert(mc, i < NUMKEYS(mp));
5226 node = NODEPTR(mp, i);
5228 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5231 mc->mc_ki[mc->mc_top] = i;
5232 if ((rc = mdb_cursor_push(mc, mp)))
5235 if (flags & MDB_PS_MODIFY) {
5236 if ((rc = mdb_page_touch(mc)) != 0)
5238 mp = mc->mc_pg[mc->mc_top];
5243 DPRINTF(("internal error, index points to a %02X page!?",
5245 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5246 return MDB_CORRUPTED;
5249 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5250 key ? DKEY(key) : "null"));
5251 mc->mc_flags |= C_INITIALIZED;
5252 mc->mc_flags &= ~C_EOF;
5257 /** Search for the lowest key under the current branch page.
5258 * This just bypasses a NUMKEYS check in the current page
5259 * before calling mdb_page_search_root(), because the callers
5260 * are all in situations where the current page is known to
5264 mdb_page_search_lowest(MDB_cursor *mc)
5266 MDB_page *mp = mc->mc_pg[mc->mc_top];
5267 MDB_node *node = NODEPTR(mp, 0);
5270 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5273 mc->mc_ki[mc->mc_top] = 0;
5274 if ((rc = mdb_cursor_push(mc, mp)))
5276 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5279 /** Search for the page a given key should be in.
5280 * Push it and its parent pages on the cursor stack.
5281 * @param[in,out] mc the cursor for this operation.
5282 * @param[in] key the key to search for, or NULL for first/last page.
5283 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5284 * are touched (updated with new page numbers).
5285 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5286 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5287 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5288 * @return 0 on success, non-zero on failure.
5291 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5296 /* Make sure the txn is still viable, then find the root from
5297 * the txn's db table and set it as the root of the cursor's stack.
5299 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5300 DPUTS("transaction may not be used now");
5303 /* Make sure we're using an up-to-date root */
5304 if (*mc->mc_dbflag & DB_STALE) {
5306 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5308 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5309 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5316 MDB_node *leaf = mdb_node_search(&mc2,
5317 &mc->mc_dbx->md_name, &exact);
5319 return MDB_NOTFOUND;
5320 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5321 return MDB_INCOMPATIBLE; /* not a named DB */
5322 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5325 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5327 /* The txn may not know this DBI, or another process may
5328 * have dropped and recreated the DB with other flags.
5330 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5331 return MDB_INCOMPATIBLE;
5332 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5334 *mc->mc_dbflag &= ~DB_STALE;
5336 root = mc->mc_db->md_root;
5338 if (root == P_INVALID) { /* Tree is empty. */
5339 DPUTS("tree is empty");
5340 return MDB_NOTFOUND;
5344 mdb_cassert(mc, root > 1);
5345 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5346 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5352 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5353 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5355 if (flags & MDB_PS_MODIFY) {
5356 if ((rc = mdb_page_touch(mc)))
5360 if (flags & MDB_PS_ROOTONLY)
5363 return mdb_page_search_root(mc, key, flags);
5367 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5369 MDB_txn *txn = mc->mc_txn;
5370 pgno_t pg = mp->mp_pgno;
5371 unsigned x = 0, ovpages = mp->mp_pages;
5372 MDB_env *env = txn->mt_env;
5373 MDB_IDL sl = txn->mt_spill_pgs;
5374 MDB_ID pn = pg << 1;
5377 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5378 /* If the page is dirty or on the spill list we just acquired it,
5379 * so we should give it back to our current free list, if any.
5380 * Otherwise put it onto the list of pages we freed in this txn.
5382 * Won't create me_pghead: me_pglast must be inited along with it.
5383 * Unsupported in nested txns: They would need to hide the page
5384 * range in ancestor txns' dirty and spilled lists.
5386 if (env->me_pghead &&
5388 ((mp->mp_flags & P_DIRTY) ||
5389 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5393 MDB_ID2 *dl, ix, iy;
5394 rc = mdb_midl_need(&env->me_pghead, ovpages);
5397 if (!(mp->mp_flags & P_DIRTY)) {
5398 /* This page is no longer spilled */
5405 /* Remove from dirty list */
5406 dl = txn->mt_u.dirty_list;
5408 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5414 mdb_cassert(mc, x > 1);
5416 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5417 txn->mt_flags |= MDB_TXN_ERROR;
5418 return MDB_CORRUPTED;
5421 if (!(env->me_flags & MDB_WRITEMAP))
5422 mdb_dpage_free(env, mp);
5424 /* Insert in me_pghead */
5425 mop = env->me_pghead;
5426 j = mop[0] + ovpages;
5427 for (i = mop[0]; i && mop[i] < pg; i--)
5433 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5437 mc->mc_db->md_overflow_pages -= ovpages;
5441 /** Return the data associated with a given node.
5442 * @param[in] txn The transaction for this operation.
5443 * @param[in] leaf The node being read.
5444 * @param[out] data Updated to point to the node's data.
5445 * @return 0 on success, non-zero on failure.
5448 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5450 MDB_page *omp; /* overflow page */
5454 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5455 data->mv_size = NODEDSZ(leaf);
5456 data->mv_data = NODEDATA(leaf);
5460 /* Read overflow data.
5462 data->mv_size = NODEDSZ(leaf);
5463 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5464 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5465 DPRINTF(("read overflow page %"Z"u failed", pgno));
5468 data->mv_data = METADATA(omp);
5474 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5475 MDB_val *key, MDB_val *data)
5482 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5484 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5487 if (txn->mt_flags & MDB_TXN_BLOCKED)
5490 mdb_cursor_init(&mc, txn, dbi, &mx);
5491 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5494 /** Find a sibling for a page.
5495 * Replaces the page at the top of the cursor's stack with the
5496 * specified sibling, if one exists.
5497 * @param[in] mc The cursor for this operation.
5498 * @param[in] move_right Non-zero if the right sibling is requested,
5499 * otherwise the left sibling.
5500 * @return 0 on success, non-zero on failure.
5503 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5509 if (mc->mc_snum < 2) {
5510 return MDB_NOTFOUND; /* root has no siblings */
5514 DPRINTF(("parent page is page %"Z"u, index %u",
5515 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5517 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5518 : (mc->mc_ki[mc->mc_top] == 0)) {
5519 DPRINTF(("no more keys left, moving to %s sibling",
5520 move_right ? "right" : "left"));
5521 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5522 /* undo cursor_pop before returning */
5529 mc->mc_ki[mc->mc_top]++;
5531 mc->mc_ki[mc->mc_top]--;
5532 DPRINTF(("just moving to %s index key %u",
5533 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5535 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5537 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5538 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5539 /* mc will be inconsistent if caller does mc_snum++ as above */
5540 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5544 mdb_cursor_push(mc, mp);
5546 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5551 /** Move the cursor to the next data item. */
5553 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5559 if (mc->mc_flags & C_EOF) {
5560 return MDB_NOTFOUND;
5563 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5565 mp = mc->mc_pg[mc->mc_top];
5567 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5568 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5569 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5570 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5571 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5572 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5573 if (rc == MDB_SUCCESS)
5574 MDB_GET_KEY(leaf, key);
5579 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5580 if (op == MDB_NEXT_DUP)
5581 return MDB_NOTFOUND;
5585 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5586 mdb_dbg_pgno(mp), (void *) mc));
5587 if (mc->mc_flags & C_DEL)
5590 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5591 DPUTS("=====> move to next sibling page");
5592 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5593 mc->mc_flags |= C_EOF;
5596 mp = mc->mc_pg[mc->mc_top];
5597 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5599 mc->mc_ki[mc->mc_top]++;
5602 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5603 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5606 key->mv_size = mc->mc_db->md_pad;
5607 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5611 mdb_cassert(mc, IS_LEAF(mp));
5612 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5614 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5615 mdb_xcursor_init1(mc, leaf);
5618 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5621 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5622 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5623 if (rc != MDB_SUCCESS)
5628 MDB_GET_KEY(leaf, key);
5632 /** Move the cursor to the previous data item. */
5634 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5640 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5642 mp = mc->mc_pg[mc->mc_top];
5644 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5645 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5646 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5647 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5648 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5649 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5650 if (rc == MDB_SUCCESS) {
5651 MDB_GET_KEY(leaf, key);
5652 mc->mc_flags &= ~C_EOF;
5658 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5659 if (op == MDB_PREV_DUP)
5660 return MDB_NOTFOUND;
5664 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5665 mdb_dbg_pgno(mp), (void *) mc));
5667 if (mc->mc_ki[mc->mc_top] == 0) {
5668 DPUTS("=====> move to prev sibling page");
5669 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5672 mp = mc->mc_pg[mc->mc_top];
5673 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5674 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5676 mc->mc_ki[mc->mc_top]--;
5678 mc->mc_flags &= ~C_EOF;
5680 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5681 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5684 key->mv_size = mc->mc_db->md_pad;
5685 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5689 mdb_cassert(mc, IS_LEAF(mp));
5690 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5692 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5693 mdb_xcursor_init1(mc, leaf);
5696 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5699 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5700 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5701 if (rc != MDB_SUCCESS)
5706 MDB_GET_KEY(leaf, key);
5710 /** Set the cursor on a specific data item. */
5712 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5713 MDB_cursor_op op, int *exactp)
5717 MDB_node *leaf = NULL;
5720 if (key->mv_size == 0)
5721 return MDB_BAD_VALSIZE;
5724 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5726 /* See if we're already on the right page */
5727 if (mc->mc_flags & C_INITIALIZED) {
5730 mp = mc->mc_pg[mc->mc_top];
5732 mc->mc_ki[mc->mc_top] = 0;
5733 return MDB_NOTFOUND;
5735 if (mp->mp_flags & P_LEAF2) {
5736 nodekey.mv_size = mc->mc_db->md_pad;
5737 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5739 leaf = NODEPTR(mp, 0);
5740 MDB_GET_KEY2(leaf, nodekey);
5742 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5744 /* Probably happens rarely, but first node on the page
5745 * was the one we wanted.
5747 mc->mc_ki[mc->mc_top] = 0;
5754 unsigned int nkeys = NUMKEYS(mp);
5756 if (mp->mp_flags & P_LEAF2) {
5757 nodekey.mv_data = LEAF2KEY(mp,
5758 nkeys-1, nodekey.mv_size);
5760 leaf = NODEPTR(mp, nkeys-1);
5761 MDB_GET_KEY2(leaf, nodekey);
5763 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5765 /* last node was the one we wanted */
5766 mc->mc_ki[mc->mc_top] = nkeys-1;
5772 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5773 /* This is definitely the right page, skip search_page */
5774 if (mp->mp_flags & P_LEAF2) {
5775 nodekey.mv_data = LEAF2KEY(mp,
5776 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5778 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5779 MDB_GET_KEY2(leaf, nodekey);
5781 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5783 /* current node was the one we wanted */
5793 /* If any parents have right-sibs, search.
5794 * Otherwise, there's nothing further.
5796 for (i=0; i<mc->mc_top; i++)
5798 NUMKEYS(mc->mc_pg[i])-1)
5800 if (i == mc->mc_top) {
5801 /* There are no other pages */
5802 mc->mc_ki[mc->mc_top] = nkeys;
5803 return MDB_NOTFOUND;
5807 /* There are no other pages */
5808 mc->mc_ki[mc->mc_top] = 0;
5809 if (op == MDB_SET_RANGE && !exactp) {
5813 return MDB_NOTFOUND;
5817 rc = mdb_page_search(mc, key, 0);
5818 if (rc != MDB_SUCCESS)
5821 mp = mc->mc_pg[mc->mc_top];
5822 mdb_cassert(mc, IS_LEAF(mp));
5825 leaf = mdb_node_search(mc, key, exactp);
5826 if (exactp != NULL && !*exactp) {
5827 /* MDB_SET specified and not an exact match. */
5828 return MDB_NOTFOUND;
5832 DPUTS("===> inexact leaf not found, goto sibling");
5833 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5834 mc->mc_flags |= C_EOF;
5835 return rc; /* no entries matched */
5837 mp = mc->mc_pg[mc->mc_top];
5838 mdb_cassert(mc, IS_LEAF(mp));
5839 leaf = NODEPTR(mp, 0);
5843 mc->mc_flags |= C_INITIALIZED;
5844 mc->mc_flags &= ~C_EOF;
5847 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5848 key->mv_size = mc->mc_db->md_pad;
5849 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5854 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5855 mdb_xcursor_init1(mc, leaf);
5858 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5859 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5860 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5863 if (op == MDB_GET_BOTH) {
5869 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5870 if (rc != MDB_SUCCESS)
5873 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5876 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5878 dcmp = mc->mc_dbx->md_dcmp;
5879 #if UINT_MAX < SIZE_MAX
5880 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5881 dcmp = mdb_cmp_clong;
5883 rc = dcmp(data, &olddata);
5885 if (op == MDB_GET_BOTH || rc > 0)
5886 return MDB_NOTFOUND;
5893 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5894 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5899 /* The key already matches in all other cases */
5900 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5901 MDB_GET_KEY(leaf, key);
5902 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5907 /** Move the cursor to the first item in the database. */
5909 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5915 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5917 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5918 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5919 if (rc != MDB_SUCCESS)
5922 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5924 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5925 mc->mc_flags |= C_INITIALIZED;
5926 mc->mc_flags &= ~C_EOF;
5928 mc->mc_ki[mc->mc_top] = 0;
5930 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5931 key->mv_size = mc->mc_db->md_pad;
5932 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5937 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5938 mdb_xcursor_init1(mc, leaf);
5939 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5943 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5947 MDB_GET_KEY(leaf, key);
5951 /** Move the cursor to the last item in the database. */
5953 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5959 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5961 if (!(mc->mc_flags & C_EOF)) {
5963 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5964 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5965 if (rc != MDB_SUCCESS)
5968 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5971 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5972 mc->mc_flags |= C_INITIALIZED|C_EOF;
5973 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5975 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5976 key->mv_size = mc->mc_db->md_pad;
5977 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5982 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5983 mdb_xcursor_init1(mc, leaf);
5984 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5988 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5993 MDB_GET_KEY(leaf, key);
5998 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6003 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6008 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6012 case MDB_GET_CURRENT:
6013 if (!(mc->mc_flags & C_INITIALIZED)) {
6016 MDB_page *mp = mc->mc_pg[mc->mc_top];
6017 int nkeys = NUMKEYS(mp);
6018 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6019 mc->mc_ki[mc->mc_top] = nkeys;
6025 key->mv_size = mc->mc_db->md_pad;
6026 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6028 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6029 MDB_GET_KEY(leaf, key);
6031 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6032 if (mc->mc_flags & C_DEL)
6033 mdb_xcursor_init1(mc, leaf);
6034 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6036 rc = mdb_node_read(mc->mc_txn, leaf, data);
6043 case MDB_GET_BOTH_RANGE:
6048 if (mc->mc_xcursor == NULL) {
6049 rc = MDB_INCOMPATIBLE;
6059 rc = mdb_cursor_set(mc, key, data, op,
6060 op == MDB_SET_RANGE ? NULL : &exact);
6063 case MDB_GET_MULTIPLE:
6064 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6068 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6069 rc = MDB_INCOMPATIBLE;
6073 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6074 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6077 case MDB_NEXT_MULTIPLE:
6082 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6083 rc = MDB_INCOMPATIBLE;
6086 if (!(mc->mc_flags & C_INITIALIZED))
6087 rc = mdb_cursor_first(mc, key, data);
6089 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6090 if (rc == MDB_SUCCESS) {
6091 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6094 mx = &mc->mc_xcursor->mx_cursor;
6095 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6097 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6098 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6106 case MDB_NEXT_NODUP:
6107 if (!(mc->mc_flags & C_INITIALIZED))
6108 rc = mdb_cursor_first(mc, key, data);
6110 rc = mdb_cursor_next(mc, key, data, op);
6114 case MDB_PREV_NODUP:
6115 if (!(mc->mc_flags & C_INITIALIZED)) {
6116 rc = mdb_cursor_last(mc, key, data);
6119 mc->mc_flags |= C_INITIALIZED;
6120 mc->mc_ki[mc->mc_top]++;
6122 rc = mdb_cursor_prev(mc, key, data, op);
6125 rc = mdb_cursor_first(mc, key, data);
6128 mfunc = mdb_cursor_first;
6130 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6134 if (mc->mc_xcursor == NULL) {
6135 rc = MDB_INCOMPATIBLE;
6139 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6140 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6141 MDB_GET_KEY(leaf, key);
6142 rc = mdb_node_read(mc->mc_txn, leaf, data);
6146 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6150 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6153 rc = mdb_cursor_last(mc, key, data);
6156 mfunc = mdb_cursor_last;
6159 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6164 if (mc->mc_flags & C_DEL)
6165 mc->mc_flags ^= C_DEL;
6170 /** Touch all the pages in the cursor stack. Set mc_top.
6171 * Makes sure all the pages are writable, before attempting a write operation.
6172 * @param[in] mc The cursor to operate on.
6175 mdb_cursor_touch(MDB_cursor *mc)
6177 int rc = MDB_SUCCESS;
6179 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6182 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6184 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6185 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6188 *mc->mc_dbflag |= DB_DIRTY;
6193 rc = mdb_page_touch(mc);
6194 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6195 mc->mc_top = mc->mc_snum-1;
6200 /** Do not spill pages to disk if txn is getting full, may fail instead */
6201 #define MDB_NOSPILL 0x8000
6204 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6208 MDB_node *leaf = NULL;
6209 MDB_page *fp, *mp, *sub_root = NULL;
6211 MDB_val xdata, *rdata, dkey, olddata;
6213 int do_sub = 0, insert_key, insert_data;
6214 unsigned int mcount = 0, dcount = 0, nospill;
6217 unsigned int nflags;
6220 if (mc == NULL || key == NULL)
6223 env = mc->mc_txn->mt_env;
6225 /* Check this first so counter will always be zero on any
6228 if (flags & MDB_MULTIPLE) {
6229 dcount = data[1].mv_size;
6230 data[1].mv_size = 0;
6231 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6232 return MDB_INCOMPATIBLE;
6235 nospill = flags & MDB_NOSPILL;
6236 flags &= ~MDB_NOSPILL;
6238 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6239 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6241 if (key->mv_size-1 >= ENV_MAXKEY(env))
6242 return MDB_BAD_VALSIZE;
6244 #if SIZE_MAX > MAXDATASIZE
6245 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6246 return MDB_BAD_VALSIZE;
6248 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6249 return MDB_BAD_VALSIZE;
6252 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6253 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6257 if (flags == MDB_CURRENT) {
6258 if (!(mc->mc_flags & C_INITIALIZED))
6261 } else if (mc->mc_db->md_root == P_INVALID) {
6262 /* new database, cursor has nothing to point to */
6265 mc->mc_flags &= ~C_INITIALIZED;
6270 if (flags & MDB_APPEND) {
6272 rc = mdb_cursor_last(mc, &k2, &d2);
6274 rc = mc->mc_dbx->md_cmp(key, &k2);
6277 mc->mc_ki[mc->mc_top]++;
6279 /* new key is <= last key */
6284 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6286 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6287 DPRINTF(("duplicate key [%s]", DKEY(key)));
6289 return MDB_KEYEXIST;
6291 if (rc && rc != MDB_NOTFOUND)
6295 if (mc->mc_flags & C_DEL)
6296 mc->mc_flags ^= C_DEL;
6298 /* Cursor is positioned, check for room in the dirty list */
6300 if (flags & MDB_MULTIPLE) {
6302 xdata.mv_size = data->mv_size * dcount;
6306 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6310 if (rc == MDB_NO_ROOT) {
6312 /* new database, write a root leaf page */
6313 DPUTS("allocating new root leaf page");
6314 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6317 mdb_cursor_push(mc, np);
6318 mc->mc_db->md_root = np->mp_pgno;
6319 mc->mc_db->md_depth++;
6320 *mc->mc_dbflag |= DB_DIRTY;
6321 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6323 np->mp_flags |= P_LEAF2;
6324 mc->mc_flags |= C_INITIALIZED;
6326 /* make sure all cursor pages are writable */
6327 rc2 = mdb_cursor_touch(mc);
6332 insert_key = insert_data = rc;
6334 /* The key does not exist */
6335 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6336 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6337 LEAFSIZE(key, data) > env->me_nodemax)
6339 /* Too big for a node, insert in sub-DB. Set up an empty
6340 * "old sub-page" for prep_subDB to expand to a full page.
6342 fp_flags = P_LEAF|P_DIRTY;
6344 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6345 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6346 olddata.mv_size = PAGEHDRSZ;
6350 /* there's only a key anyway, so this is a no-op */
6351 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6353 unsigned int ksize = mc->mc_db->md_pad;
6354 if (key->mv_size != ksize)
6355 return MDB_BAD_VALSIZE;
6356 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6357 memcpy(ptr, key->mv_data, ksize);
6359 /* if overwriting slot 0 of leaf, need to
6360 * update branch key if there is a parent page
6362 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6363 unsigned short top = mc->mc_top;
6365 /* slot 0 is always an empty key, find real slot */
6366 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6368 if (mc->mc_ki[mc->mc_top])
6369 rc2 = mdb_update_key(mc, key);
6380 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6381 olddata.mv_size = NODEDSZ(leaf);
6382 olddata.mv_data = NODEDATA(leaf);
6385 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6386 /* Prepare (sub-)page/sub-DB to accept the new item,
6387 * if needed. fp: old sub-page or a header faking
6388 * it. mp: new (sub-)page. offset: growth in page
6389 * size. xdata: node data with new page or DB.
6391 unsigned i, offset = 0;
6392 mp = fp = xdata.mv_data = env->me_pbuf;
6393 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6395 /* Was a single item before, must convert now */
6396 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6398 /* Just overwrite the current item */
6399 if (flags == MDB_CURRENT)
6401 dcmp = mc->mc_dbx->md_dcmp;
6402 #if UINT_MAX < SIZE_MAX
6403 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6404 dcmp = mdb_cmp_clong;
6406 /* does data match? */
6407 if (!dcmp(data, &olddata)) {
6408 if (flags & MDB_NODUPDATA)
6409 return MDB_KEYEXIST;
6414 /* Back up original data item */
6415 dkey.mv_size = olddata.mv_size;
6416 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6418 /* Make sub-page header for the dup items, with dummy body */
6419 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6420 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6421 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6422 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6423 fp->mp_flags |= P_LEAF2;
6424 fp->mp_pad = data->mv_size;
6425 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6427 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6428 (dkey.mv_size & 1) + (data->mv_size & 1);
6430 fp->mp_upper = xdata.mv_size - PAGEBASE;
6431 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6432 } else if (leaf->mn_flags & F_SUBDATA) {
6433 /* Data is on sub-DB, just store it */
6434 flags |= F_DUPDATA|F_SUBDATA;
6437 /* Data is on sub-page */
6438 fp = olddata.mv_data;
6441 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6442 offset = EVEN(NODESIZE + sizeof(indx_t) +
6446 offset = fp->mp_pad;
6447 if (SIZELEFT(fp) < offset) {
6448 offset *= 4; /* space for 4 more */
6451 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6453 fp->mp_flags |= P_DIRTY;
6454 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6455 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6459 xdata.mv_size = olddata.mv_size + offset;
6462 fp_flags = fp->mp_flags;
6463 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6464 /* Too big for a sub-page, convert to sub-DB */
6465 fp_flags &= ~P_SUBP;
6467 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6468 fp_flags |= P_LEAF2;
6469 dummy.md_pad = fp->mp_pad;
6470 dummy.md_flags = MDB_DUPFIXED;
6471 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6472 dummy.md_flags |= MDB_INTEGERKEY;
6478 dummy.md_branch_pages = 0;
6479 dummy.md_leaf_pages = 1;
6480 dummy.md_overflow_pages = 0;
6481 dummy.md_entries = NUMKEYS(fp);
6482 xdata.mv_size = sizeof(MDB_db);
6483 xdata.mv_data = &dummy;
6484 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6486 offset = env->me_psize - olddata.mv_size;
6487 flags |= F_DUPDATA|F_SUBDATA;
6488 dummy.md_root = mp->mp_pgno;
6492 mp->mp_flags = fp_flags | P_DIRTY;
6493 mp->mp_pad = fp->mp_pad;
6494 mp->mp_lower = fp->mp_lower;
6495 mp->mp_upper = fp->mp_upper + offset;
6496 if (fp_flags & P_LEAF2) {
6497 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6499 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6500 olddata.mv_size - fp->mp_upper - PAGEBASE);
6501 for (i=0; i<NUMKEYS(fp); i++)
6502 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6510 mdb_node_del(mc, 0);
6514 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6515 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6516 return MDB_INCOMPATIBLE;
6517 /* overflow page overwrites need special handling */
6518 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6521 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6523 memcpy(&pg, olddata.mv_data, sizeof(pg));
6524 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6526 ovpages = omp->mp_pages;
6528 /* Is the ov page large enough? */
6529 if (ovpages >= dpages) {
6530 if (!(omp->mp_flags & P_DIRTY) &&
6531 (level || (env->me_flags & MDB_WRITEMAP)))
6533 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6536 level = 0; /* dirty in this txn or clean */
6539 if (omp->mp_flags & P_DIRTY) {
6540 /* yes, overwrite it. Note in this case we don't
6541 * bother to try shrinking the page if the new data
6542 * is smaller than the overflow threshold.
6545 /* It is writable only in a parent txn */
6546 size_t sz = (size_t) env->me_psize * ovpages, off;
6547 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6553 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6554 mdb_cassert(mc, rc2 == 0);
6555 if (!(flags & MDB_RESERVE)) {
6556 /* Copy end of page, adjusting alignment so
6557 * compiler may copy words instead of bytes.
6559 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6560 memcpy((size_t *)((char *)np + off),
6561 (size_t *)((char *)omp + off), sz - off);
6564 memcpy(np, omp, sz); /* Copy beginning of page */
6567 SETDSZ(leaf, data->mv_size);
6568 if (F_ISSET(flags, MDB_RESERVE))
6569 data->mv_data = METADATA(omp);
6571 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6575 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6577 } else if (data->mv_size == olddata.mv_size) {
6578 /* same size, just replace it. Note that we could
6579 * also reuse this node if the new data is smaller,
6580 * but instead we opt to shrink the node in that case.
6582 if (F_ISSET(flags, MDB_RESERVE))
6583 data->mv_data = olddata.mv_data;
6584 else if (!(mc->mc_flags & C_SUB))
6585 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6587 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6592 mdb_node_del(mc, 0);
6598 nflags = flags & NODE_ADD_FLAGS;
6599 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6600 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6601 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6602 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6604 nflags |= MDB_SPLIT_REPLACE;
6605 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6607 /* There is room already in this leaf page. */
6608 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6609 if (rc == 0 && insert_key) {
6610 /* Adjust other cursors pointing to mp */
6611 MDB_cursor *m2, *m3;
6612 MDB_dbi dbi = mc->mc_dbi;
6613 unsigned i = mc->mc_top;
6614 MDB_page *mp = mc->mc_pg[i];
6616 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6617 if (mc->mc_flags & C_SUB)
6618 m3 = &m2->mc_xcursor->mx_cursor;
6621 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6622 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6629 if (rc == MDB_SUCCESS) {
6630 /* Now store the actual data in the child DB. Note that we're
6631 * storing the user data in the keys field, so there are strict
6632 * size limits on dupdata. The actual data fields of the child
6633 * DB are all zero size.
6636 int xflags, new_dupdata;
6641 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6642 if (flags & MDB_CURRENT) {
6643 xflags = MDB_CURRENT|MDB_NOSPILL;
6645 mdb_xcursor_init1(mc, leaf);
6646 xflags = (flags & MDB_NODUPDATA) ?
6647 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6650 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6651 new_dupdata = (int)dkey.mv_size;
6652 /* converted, write the original data first */
6654 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6657 /* we've done our job */
6660 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6661 /* Adjust other cursors pointing to mp */
6663 MDB_xcursor *mx = mc->mc_xcursor;
6664 unsigned i = mc->mc_top;
6665 MDB_page *mp = mc->mc_pg[i];
6667 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6668 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6669 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6670 if (m2->mc_pg[i] == mp) {
6671 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6672 mdb_xcursor_init2(m2, mx, new_dupdata);
6673 } else if (!insert_key) {
6674 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6675 if (!(n2->mn_flags & F_SUBDATA))
6676 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6681 ecount = mc->mc_xcursor->mx_db.md_entries;
6682 if (flags & MDB_APPENDDUP)
6683 xflags |= MDB_APPEND;
6684 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6685 if (flags & F_SUBDATA) {
6686 void *db = NODEDATA(leaf);
6687 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6689 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6691 /* Increment count unless we just replaced an existing item. */
6693 mc->mc_db->md_entries++;
6695 /* Invalidate txn if we created an empty sub-DB */
6698 /* If we succeeded and the key didn't exist before,
6699 * make sure the cursor is marked valid.
6701 mc->mc_flags |= C_INITIALIZED;
6703 if (flags & MDB_MULTIPLE) {
6706 /* let caller know how many succeeded, if any */
6707 data[1].mv_size = mcount;
6708 if (mcount < dcount) {
6709 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6710 insert_key = insert_data = 0;
6717 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6720 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6725 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6731 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6732 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6734 if (!(mc->mc_flags & C_INITIALIZED))
6737 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6738 return MDB_NOTFOUND;
6740 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6743 rc = mdb_cursor_touch(mc);
6747 mp = mc->mc_pg[mc->mc_top];
6750 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6752 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6753 if (flags & MDB_NODUPDATA) {
6754 /* mdb_cursor_del0() will subtract the final entry */
6755 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6757 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6758 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6760 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6763 /* If sub-DB still has entries, we're done */
6764 if (mc->mc_xcursor->mx_db.md_entries) {
6765 if (leaf->mn_flags & F_SUBDATA) {
6766 /* update subDB info */
6767 void *db = NODEDATA(leaf);
6768 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6771 /* shrink fake page */
6772 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6773 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6774 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6775 /* fix other sub-DB cursors pointed at fake pages on this page */
6776 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6777 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6778 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6779 if (m2->mc_pg[mc->mc_top] == mp) {
6780 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6781 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6783 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6784 if (!(n2->mn_flags & F_SUBDATA))
6785 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6790 mc->mc_db->md_entries--;
6791 mc->mc_flags |= C_DEL;
6794 /* otherwise fall thru and delete the sub-DB */
6797 if (leaf->mn_flags & F_SUBDATA) {
6798 /* add all the child DB's pages to the free list */
6799 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6804 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6805 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6806 rc = MDB_INCOMPATIBLE;
6810 /* add overflow pages to free list */
6811 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6815 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6816 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6817 (rc = mdb_ovpage_free(mc, omp)))
6822 return mdb_cursor_del0(mc);
6825 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6829 /** Allocate and initialize new pages for a database.
6830 * @param[in] mc a cursor on the database being added to.
6831 * @param[in] flags flags defining what type of page is being allocated.
6832 * @param[in] num the number of pages to allocate. This is usually 1,
6833 * unless allocating overflow pages for a large record.
6834 * @param[out] mp Address of a page, or NULL on failure.
6835 * @return 0 on success, non-zero on failure.
6838 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6843 if ((rc = mdb_page_alloc(mc, num, &np)))
6845 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6846 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6847 np->mp_flags = flags | P_DIRTY;
6848 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6849 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6852 mc->mc_db->md_branch_pages++;
6853 else if (IS_LEAF(np))
6854 mc->mc_db->md_leaf_pages++;
6855 else if (IS_OVERFLOW(np)) {
6856 mc->mc_db->md_overflow_pages += num;
6864 /** Calculate the size of a leaf node.
6865 * The size depends on the environment's page size; if a data item
6866 * is too large it will be put onto an overflow page and the node
6867 * size will only include the key and not the data. Sizes are always
6868 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6869 * of the #MDB_node headers.
6870 * @param[in] env The environment handle.
6871 * @param[in] key The key for the node.
6872 * @param[in] data The data for the node.
6873 * @return The number of bytes needed to store the node.
6876 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6880 sz = LEAFSIZE(key, data);
6881 if (sz > env->me_nodemax) {
6882 /* put on overflow page */
6883 sz -= data->mv_size - sizeof(pgno_t);
6886 return EVEN(sz + sizeof(indx_t));
6889 /** Calculate the size of a branch node.
6890 * The size should depend on the environment's page size but since
6891 * we currently don't support spilling large keys onto overflow
6892 * pages, it's simply the size of the #MDB_node header plus the
6893 * size of the key. Sizes are always rounded up to an even number
6894 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6895 * @param[in] env The environment handle.
6896 * @param[in] key The key for the node.
6897 * @return The number of bytes needed to store the node.
6900 mdb_branch_size(MDB_env *env, MDB_val *key)
6905 if (sz > env->me_nodemax) {
6906 /* put on overflow page */
6907 /* not implemented */
6908 /* sz -= key->size - sizeof(pgno_t); */
6911 return sz + sizeof(indx_t);
6914 /** Add a node to the page pointed to by the cursor.
6915 * @param[in] mc The cursor for this operation.
6916 * @param[in] indx The index on the page where the new node should be added.
6917 * @param[in] key The key for the new node.
6918 * @param[in] data The data for the new node, if any.
6919 * @param[in] pgno The page number, if adding a branch node.
6920 * @param[in] flags Flags for the node.
6921 * @return 0 on success, non-zero on failure. Possible errors are:
6923 * <li>ENOMEM - failed to allocate overflow pages for the node.
6924 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6925 * should never happen since all callers already calculate the
6926 * page's free space before calling this function.
6930 mdb_node_add(MDB_cursor *mc, indx_t indx,
6931 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6934 size_t node_size = NODESIZE;
6938 MDB_page *mp = mc->mc_pg[mc->mc_top];
6939 MDB_page *ofp = NULL; /* overflow page */
6943 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6945 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6946 IS_LEAF(mp) ? "leaf" : "branch",
6947 IS_SUBP(mp) ? "sub-" : "",
6948 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6949 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6952 /* Move higher keys up one slot. */
6953 int ksize = mc->mc_db->md_pad, dif;
6954 char *ptr = LEAF2KEY(mp, indx, ksize);
6955 dif = NUMKEYS(mp) - indx;
6957 memmove(ptr+ksize, ptr, dif*ksize);
6958 /* insert new key */
6959 memcpy(ptr, key->mv_data, ksize);
6961 /* Just using these for counting */
6962 mp->mp_lower += sizeof(indx_t);
6963 mp->mp_upper -= ksize - sizeof(indx_t);
6967 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6969 node_size += key->mv_size;
6971 mdb_cassert(mc, key && data);
6972 if (F_ISSET(flags, F_BIGDATA)) {
6973 /* Data already on overflow page. */
6974 node_size += sizeof(pgno_t);
6975 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6976 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6978 /* Put data on overflow page. */
6979 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6980 data->mv_size, node_size+data->mv_size));
6981 node_size = EVEN(node_size + sizeof(pgno_t));
6982 if ((ssize_t)node_size > room)
6984 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6986 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6990 node_size += data->mv_size;
6993 node_size = EVEN(node_size);
6994 if ((ssize_t)node_size > room)
6998 /* Move higher pointers up one slot. */
6999 for (i = NUMKEYS(mp); i > indx; i--)
7000 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7002 /* Adjust free space offsets. */
7003 ofs = mp->mp_upper - node_size;
7004 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7005 mp->mp_ptrs[indx] = ofs;
7007 mp->mp_lower += sizeof(indx_t);
7009 /* Write the node data. */
7010 node = NODEPTR(mp, indx);
7011 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7012 node->mn_flags = flags;
7014 SETDSZ(node,data->mv_size);
7019 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7022 ndata = NODEDATA(node);
7024 if (F_ISSET(flags, F_BIGDATA))
7025 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7026 else if (F_ISSET(flags, MDB_RESERVE))
7027 data->mv_data = ndata;
7029 memcpy(ndata, data->mv_data, data->mv_size);
7031 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7032 ndata = METADATA(ofp);
7033 if (F_ISSET(flags, MDB_RESERVE))
7034 data->mv_data = ndata;
7036 memcpy(ndata, data->mv_data, data->mv_size);
7043 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7044 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7045 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7046 DPRINTF(("node size = %"Z"u", node_size));
7047 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7048 return MDB_PAGE_FULL;
7051 /** Delete the specified node from a page.
7052 * @param[in] mc Cursor pointing to the node to delete.
7053 * @param[in] ksize The size of a node. Only used if the page is
7054 * part of a #MDB_DUPFIXED database.
7057 mdb_node_del(MDB_cursor *mc, int ksize)
7059 MDB_page *mp = mc->mc_pg[mc->mc_top];
7060 indx_t indx = mc->mc_ki[mc->mc_top];
7062 indx_t i, j, numkeys, ptr;
7066 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7067 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7068 numkeys = NUMKEYS(mp);
7069 mdb_cassert(mc, indx < numkeys);
7072 int x = numkeys - 1 - indx;
7073 base = LEAF2KEY(mp, indx, ksize);
7075 memmove(base, base + ksize, x * ksize);
7076 mp->mp_lower -= sizeof(indx_t);
7077 mp->mp_upper += ksize - sizeof(indx_t);
7081 node = NODEPTR(mp, indx);
7082 sz = NODESIZE + node->mn_ksize;
7084 if (F_ISSET(node->mn_flags, F_BIGDATA))
7085 sz += sizeof(pgno_t);
7087 sz += NODEDSZ(node);
7091 ptr = mp->mp_ptrs[indx];
7092 for (i = j = 0; i < numkeys; i++) {
7094 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7095 if (mp->mp_ptrs[i] < ptr)
7096 mp->mp_ptrs[j] += sz;
7101 base = (char *)mp + mp->mp_upper + PAGEBASE;
7102 memmove(base + sz, base, ptr - mp->mp_upper);
7104 mp->mp_lower -= sizeof(indx_t);
7108 /** Compact the main page after deleting a node on a subpage.
7109 * @param[in] mp The main page to operate on.
7110 * @param[in] indx The index of the subpage on the main page.
7113 mdb_node_shrink(MDB_page *mp, indx_t indx)
7118 indx_t delta, nsize, len, ptr;
7121 node = NODEPTR(mp, indx);
7122 sp = (MDB_page *)NODEDATA(node);
7123 delta = SIZELEFT(sp);
7124 nsize = NODEDSZ(node) - delta;
7126 /* Prepare to shift upward, set len = length(subpage part to shift) */
7130 return; /* do not make the node uneven-sized */
7132 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7133 for (i = NUMKEYS(sp); --i >= 0; )
7134 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7137 sp->mp_upper = sp->mp_lower;
7138 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7139 SETDSZ(node, nsize);
7141 /* Shift <lower nodes...initial part of subpage> upward */
7142 base = (char *)mp + mp->mp_upper + PAGEBASE;
7143 memmove(base + delta, base, (char *)sp + len - base);
7145 ptr = mp->mp_ptrs[indx];
7146 for (i = NUMKEYS(mp); --i >= 0; ) {
7147 if (mp->mp_ptrs[i] <= ptr)
7148 mp->mp_ptrs[i] += delta;
7150 mp->mp_upper += delta;
7153 /** Initial setup of a sorted-dups cursor.
7154 * Sorted duplicates are implemented as a sub-database for the given key.
7155 * The duplicate data items are actually keys of the sub-database.
7156 * Operations on the duplicate data items are performed using a sub-cursor
7157 * initialized when the sub-database is first accessed. This function does
7158 * the preliminary setup of the sub-cursor, filling in the fields that
7159 * depend only on the parent DB.
7160 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7163 mdb_xcursor_init0(MDB_cursor *mc)
7165 MDB_xcursor *mx = mc->mc_xcursor;
7167 mx->mx_cursor.mc_xcursor = NULL;
7168 mx->mx_cursor.mc_txn = mc->mc_txn;
7169 mx->mx_cursor.mc_db = &mx->mx_db;
7170 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7171 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7172 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7173 mx->mx_cursor.mc_snum = 0;
7174 mx->mx_cursor.mc_top = 0;
7175 mx->mx_cursor.mc_flags = C_SUB;
7176 mx->mx_dbx.md_name.mv_size = 0;
7177 mx->mx_dbx.md_name.mv_data = NULL;
7178 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7179 mx->mx_dbx.md_dcmp = NULL;
7180 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7183 /** Final setup of a sorted-dups cursor.
7184 * Sets up the fields that depend on the data from the main cursor.
7185 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7186 * @param[in] node The data containing the #MDB_db record for the
7187 * sorted-dup database.
7190 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7192 MDB_xcursor *mx = mc->mc_xcursor;
7194 if (node->mn_flags & F_SUBDATA) {
7195 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7196 mx->mx_cursor.mc_pg[0] = 0;
7197 mx->mx_cursor.mc_snum = 0;
7198 mx->mx_cursor.mc_top = 0;
7199 mx->mx_cursor.mc_flags = C_SUB;
7201 MDB_page *fp = NODEDATA(node);
7202 mx->mx_db.md_pad = 0;
7203 mx->mx_db.md_flags = 0;
7204 mx->mx_db.md_depth = 1;
7205 mx->mx_db.md_branch_pages = 0;
7206 mx->mx_db.md_leaf_pages = 1;
7207 mx->mx_db.md_overflow_pages = 0;
7208 mx->mx_db.md_entries = NUMKEYS(fp);
7209 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7210 mx->mx_cursor.mc_snum = 1;
7211 mx->mx_cursor.mc_top = 0;
7212 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7213 mx->mx_cursor.mc_pg[0] = fp;
7214 mx->mx_cursor.mc_ki[0] = 0;
7215 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7216 mx->mx_db.md_flags = MDB_DUPFIXED;
7217 mx->mx_db.md_pad = fp->mp_pad;
7218 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7219 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7222 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7223 mx->mx_db.md_root));
7224 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7225 #if UINT_MAX < SIZE_MAX
7226 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7227 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7232 /** Fixup a sorted-dups cursor due to underlying update.
7233 * Sets up some fields that depend on the data from the main cursor.
7234 * Almost the same as init1, but skips initialization steps if the
7235 * xcursor had already been used.
7236 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7237 * @param[in] src_mx The xcursor of an up-to-date cursor.
7238 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7241 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7243 MDB_xcursor *mx = mc->mc_xcursor;
7246 mx->mx_cursor.mc_snum = 1;
7247 mx->mx_cursor.mc_top = 0;
7248 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7249 mx->mx_cursor.mc_ki[0] = 0;
7250 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7251 #if UINT_MAX < SIZE_MAX
7252 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7254 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7257 mx->mx_db = src_mx->mx_db;
7258 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7259 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7260 mx->mx_db.md_root));
7263 /** Initialize a cursor for a given transaction and database. */
7265 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7268 mc->mc_backup = NULL;
7271 mc->mc_db = &txn->mt_dbs[dbi];
7272 mc->mc_dbx = &txn->mt_dbxs[dbi];
7273 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7279 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7280 mdb_tassert(txn, mx != NULL);
7281 mc->mc_xcursor = mx;
7282 mdb_xcursor_init0(mc);
7284 mc->mc_xcursor = NULL;
7286 if (*mc->mc_dbflag & DB_STALE) {
7287 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7292 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7295 size_t size = sizeof(MDB_cursor);
7297 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7300 if (txn->mt_flags & MDB_TXN_BLOCKED)
7303 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7306 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7307 size += sizeof(MDB_xcursor);
7309 if ((mc = malloc(size)) != NULL) {
7310 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7311 if (txn->mt_cursors) {
7312 mc->mc_next = txn->mt_cursors[dbi];
7313 txn->mt_cursors[dbi] = mc;
7314 mc->mc_flags |= C_UNTRACK;
7326 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7328 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7331 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7334 if (txn->mt_flags & MDB_TXN_BLOCKED)
7337 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7341 /* Return the count of duplicate data items for the current key */
7343 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7347 if (mc == NULL || countp == NULL)
7350 if (mc->mc_xcursor == NULL)
7351 return MDB_INCOMPATIBLE;
7353 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7356 if (!(mc->mc_flags & C_INITIALIZED))
7359 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7360 return MDB_NOTFOUND;
7362 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7363 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7366 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7369 *countp = mc->mc_xcursor->mx_db.md_entries;
7375 mdb_cursor_close(MDB_cursor *mc)
7377 if (mc && !mc->mc_backup) {
7378 /* remove from txn, if tracked */
7379 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7380 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7381 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7383 *prev = mc->mc_next;
7390 mdb_cursor_txn(MDB_cursor *mc)
7392 if (!mc) return NULL;
7397 mdb_cursor_dbi(MDB_cursor *mc)
7402 /** Replace the key for a branch node with a new key.
7403 * @param[in] mc Cursor pointing to the node to operate on.
7404 * @param[in] key The new key to use.
7405 * @return 0 on success, non-zero on failure.
7408 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7414 int delta, ksize, oksize;
7415 indx_t ptr, i, numkeys, indx;
7418 indx = mc->mc_ki[mc->mc_top];
7419 mp = mc->mc_pg[mc->mc_top];
7420 node = NODEPTR(mp, indx);
7421 ptr = mp->mp_ptrs[indx];
7425 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7426 k2.mv_data = NODEKEY(node);
7427 k2.mv_size = node->mn_ksize;
7428 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7430 mdb_dkey(&k2, kbuf2),
7436 /* Sizes must be 2-byte aligned. */
7437 ksize = EVEN(key->mv_size);
7438 oksize = EVEN(node->mn_ksize);
7439 delta = ksize - oksize;
7441 /* Shift node contents if EVEN(key length) changed. */
7443 if (delta > 0 && SIZELEFT(mp) < delta) {
7445 /* not enough space left, do a delete and split */
7446 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7447 pgno = NODEPGNO(node);
7448 mdb_node_del(mc, 0);
7449 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7452 numkeys = NUMKEYS(mp);
7453 for (i = 0; i < numkeys; i++) {
7454 if (mp->mp_ptrs[i] <= ptr)
7455 mp->mp_ptrs[i] -= delta;
7458 base = (char *)mp + mp->mp_upper + PAGEBASE;
7459 len = ptr - mp->mp_upper + NODESIZE;
7460 memmove(base - delta, base, len);
7461 mp->mp_upper -= delta;
7463 node = NODEPTR(mp, indx);
7466 /* But even if no shift was needed, update ksize */
7467 if (node->mn_ksize != key->mv_size)
7468 node->mn_ksize = key->mv_size;
7471 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7477 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7479 /** Move a node from csrc to cdst.
7482 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7489 unsigned short flags;
7493 /* Mark src and dst as dirty. */
7494 if ((rc = mdb_page_touch(csrc)) ||
7495 (rc = mdb_page_touch(cdst)))
7498 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7499 key.mv_size = csrc->mc_db->md_pad;
7500 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7502 data.mv_data = NULL;
7506 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7507 mdb_cassert(csrc, !((size_t)srcnode & 1));
7508 srcpg = NODEPGNO(srcnode);
7509 flags = srcnode->mn_flags;
7510 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7511 unsigned int snum = csrc->mc_snum;
7513 /* must find the lowest key below src */
7514 rc = mdb_page_search_lowest(csrc);
7517 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7518 key.mv_size = csrc->mc_db->md_pad;
7519 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7521 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7522 key.mv_size = NODEKSZ(s2);
7523 key.mv_data = NODEKEY(s2);
7525 csrc->mc_snum = snum--;
7526 csrc->mc_top = snum;
7528 key.mv_size = NODEKSZ(srcnode);
7529 key.mv_data = NODEKEY(srcnode);
7531 data.mv_size = NODEDSZ(srcnode);
7532 data.mv_data = NODEDATA(srcnode);
7534 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7535 unsigned int snum = cdst->mc_snum;
7538 /* must find the lowest key below dst */
7539 mdb_cursor_copy(cdst, &mn);
7540 rc = mdb_page_search_lowest(&mn);
7543 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7544 bkey.mv_size = mn.mc_db->md_pad;
7545 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7547 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7548 bkey.mv_size = NODEKSZ(s2);
7549 bkey.mv_data = NODEKEY(s2);
7551 mn.mc_snum = snum--;
7554 rc = mdb_update_key(&mn, &bkey);
7559 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7560 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7561 csrc->mc_ki[csrc->mc_top],
7563 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7564 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7566 /* Add the node to the destination page.
7568 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7569 if (rc != MDB_SUCCESS)
7572 /* Delete the node from the source page.
7574 mdb_node_del(csrc, key.mv_size);
7577 /* Adjust other cursors pointing to mp */
7578 MDB_cursor *m2, *m3;
7579 MDB_dbi dbi = csrc->mc_dbi;
7582 mp = cdst->mc_pg[csrc->mc_top];
7583 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7584 if (csrc->mc_flags & C_SUB)
7585 m3 = &m2->mc_xcursor->mx_cursor;
7588 if (m3 == cdst) continue;
7589 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] >=
7590 cdst->mc_ki[csrc->mc_top]) {
7591 m3->mc_ki[csrc->mc_top]++;
7595 mp = csrc->mc_pg[csrc->mc_top];
7596 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7597 if (csrc->mc_flags & C_SUB)
7598 m3 = &m2->mc_xcursor->mx_cursor;
7601 if (m3 == csrc) continue;
7602 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7603 csrc->mc_ki[csrc->mc_top]) {
7604 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7605 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7610 /* Update the parent separators.
7612 if (csrc->mc_ki[csrc->mc_top] == 0) {
7613 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7614 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7615 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7617 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7618 key.mv_size = NODEKSZ(srcnode);
7619 key.mv_data = NODEKEY(srcnode);
7621 DPRINTF(("update separator for source page %"Z"u to [%s]",
7622 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7623 mdb_cursor_copy(csrc, &mn);
7626 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7629 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7631 indx_t ix = csrc->mc_ki[csrc->mc_top];
7632 nullkey.mv_size = 0;
7633 csrc->mc_ki[csrc->mc_top] = 0;
7634 rc = mdb_update_key(csrc, &nullkey);
7635 csrc->mc_ki[csrc->mc_top] = ix;
7636 mdb_cassert(csrc, rc == MDB_SUCCESS);
7640 if (cdst->mc_ki[cdst->mc_top] == 0) {
7641 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7642 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7643 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7645 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7646 key.mv_size = NODEKSZ(srcnode);
7647 key.mv_data = NODEKEY(srcnode);
7649 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7650 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7651 mdb_cursor_copy(cdst, &mn);
7654 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7657 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7659 indx_t ix = cdst->mc_ki[cdst->mc_top];
7660 nullkey.mv_size = 0;
7661 cdst->mc_ki[cdst->mc_top] = 0;
7662 rc = mdb_update_key(cdst, &nullkey);
7663 cdst->mc_ki[cdst->mc_top] = ix;
7664 mdb_cassert(cdst, rc == MDB_SUCCESS);
7671 /** Merge one page into another.
7672 * The nodes from the page pointed to by \b csrc will
7673 * be copied to the page pointed to by \b cdst and then
7674 * the \b csrc page will be freed.
7675 * @param[in] csrc Cursor pointing to the source page.
7676 * @param[in] cdst Cursor pointing to the destination page.
7677 * @return 0 on success, non-zero on failure.
7680 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7682 MDB_page *psrc, *pdst;
7689 psrc = csrc->mc_pg[csrc->mc_top];
7690 pdst = cdst->mc_pg[cdst->mc_top];
7692 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7694 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7695 mdb_cassert(csrc, cdst->mc_snum > 1);
7697 /* Mark dst as dirty. */
7698 if ((rc = mdb_page_touch(cdst)))
7701 /* Move all nodes from src to dst.
7703 j = nkeys = NUMKEYS(pdst);
7704 if (IS_LEAF2(psrc)) {
7705 key.mv_size = csrc->mc_db->md_pad;
7706 key.mv_data = METADATA(psrc);
7707 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7708 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7709 if (rc != MDB_SUCCESS)
7711 key.mv_data = (char *)key.mv_data + key.mv_size;
7714 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7715 srcnode = NODEPTR(psrc, i);
7716 if (i == 0 && IS_BRANCH(psrc)) {
7719 mdb_cursor_copy(csrc, &mn);
7720 /* must find the lowest key below src */
7721 rc = mdb_page_search_lowest(&mn);
7724 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7725 key.mv_size = mn.mc_db->md_pad;
7726 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7728 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7729 key.mv_size = NODEKSZ(s2);
7730 key.mv_data = NODEKEY(s2);
7733 key.mv_size = srcnode->mn_ksize;
7734 key.mv_data = NODEKEY(srcnode);
7737 data.mv_size = NODEDSZ(srcnode);
7738 data.mv_data = NODEDATA(srcnode);
7739 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7740 if (rc != MDB_SUCCESS)
7745 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7746 pdst->mp_pgno, NUMKEYS(pdst),
7747 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7749 /* Unlink the src page from parent and add to free list.
7752 mdb_node_del(csrc, 0);
7753 if (csrc->mc_ki[csrc->mc_top] == 0) {
7755 rc = mdb_update_key(csrc, &key);
7763 psrc = csrc->mc_pg[csrc->mc_top];
7764 /* If not operating on FreeDB, allow this page to be reused
7765 * in this txn. Otherwise just add to free list.
7767 rc = mdb_page_loose(csrc, psrc);
7771 csrc->mc_db->md_leaf_pages--;
7773 csrc->mc_db->md_branch_pages--;
7775 /* Adjust other cursors pointing to mp */
7776 MDB_cursor *m2, *m3;
7777 MDB_dbi dbi = csrc->mc_dbi;
7779 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7780 if (csrc->mc_flags & C_SUB)
7781 m3 = &m2->mc_xcursor->mx_cursor;
7784 if (m3 == csrc) continue;
7785 if (m3->mc_snum < csrc->mc_snum) continue;
7786 if (m3->mc_pg[csrc->mc_top] == psrc) {
7787 m3->mc_pg[csrc->mc_top] = pdst;
7788 m3->mc_ki[csrc->mc_top] += nkeys;
7793 unsigned int snum = cdst->mc_snum;
7794 uint16_t depth = cdst->mc_db->md_depth;
7795 mdb_cursor_pop(cdst);
7796 rc = mdb_rebalance(cdst);
7797 /* Did the tree height change? */
7798 if (depth != cdst->mc_db->md_depth)
7799 snum += cdst->mc_db->md_depth - depth;
7800 cdst->mc_snum = snum;
7801 cdst->mc_top = snum-1;
7806 /** Copy the contents of a cursor.
7807 * @param[in] csrc The cursor to copy from.
7808 * @param[out] cdst The cursor to copy to.
7811 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7815 cdst->mc_txn = csrc->mc_txn;
7816 cdst->mc_dbi = csrc->mc_dbi;
7817 cdst->mc_db = csrc->mc_db;
7818 cdst->mc_dbx = csrc->mc_dbx;
7819 cdst->mc_snum = csrc->mc_snum;
7820 cdst->mc_top = csrc->mc_top;
7821 cdst->mc_flags = csrc->mc_flags;
7823 for (i=0; i<csrc->mc_snum; i++) {
7824 cdst->mc_pg[i] = csrc->mc_pg[i];
7825 cdst->mc_ki[i] = csrc->mc_ki[i];
7829 /** Rebalance the tree after a delete operation.
7830 * @param[in] mc Cursor pointing to the page where rebalancing
7832 * @return 0 on success, non-zero on failure.
7835 mdb_rebalance(MDB_cursor *mc)
7839 unsigned int ptop, minkeys, thresh;
7843 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7848 thresh = FILL_THRESHOLD;
7850 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7851 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7852 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7853 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7855 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7856 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7857 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7858 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7862 if (mc->mc_snum < 2) {
7863 MDB_page *mp = mc->mc_pg[0];
7865 DPUTS("Can't rebalance a subpage, ignoring");
7868 if (NUMKEYS(mp) == 0) {
7869 DPUTS("tree is completely empty");
7870 mc->mc_db->md_root = P_INVALID;
7871 mc->mc_db->md_depth = 0;
7872 mc->mc_db->md_leaf_pages = 0;
7873 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7876 /* Adjust cursors pointing to mp */
7879 mc->mc_flags &= ~C_INITIALIZED;
7881 MDB_cursor *m2, *m3;
7882 MDB_dbi dbi = mc->mc_dbi;
7884 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7885 if (mc->mc_flags & C_SUB)
7886 m3 = &m2->mc_xcursor->mx_cursor;
7889 if (m3->mc_snum < mc->mc_snum) continue;
7890 if (m3->mc_pg[0] == mp) {
7893 m3->mc_flags &= ~C_INITIALIZED;
7897 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7899 DPUTS("collapsing root page!");
7900 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7903 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7904 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7907 mc->mc_db->md_depth--;
7908 mc->mc_db->md_branch_pages--;
7909 mc->mc_ki[0] = mc->mc_ki[1];
7910 for (i = 1; i<mc->mc_db->md_depth; i++) {
7911 mc->mc_pg[i] = mc->mc_pg[i+1];
7912 mc->mc_ki[i] = mc->mc_ki[i+1];
7915 /* Adjust other cursors pointing to mp */
7916 MDB_cursor *m2, *m3;
7917 MDB_dbi dbi = mc->mc_dbi;
7919 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7920 if (mc->mc_flags & C_SUB)
7921 m3 = &m2->mc_xcursor->mx_cursor;
7924 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7925 if (m3->mc_pg[0] == mp) {
7926 for (i=0; i<m3->mc_snum; i++) {
7927 m3->mc_pg[i] = m3->mc_pg[i+1];
7928 m3->mc_ki[i] = m3->mc_ki[i+1];
7936 DPUTS("root page doesn't need rebalancing");
7940 /* The parent (branch page) must have at least 2 pointers,
7941 * otherwise the tree is invalid.
7943 ptop = mc->mc_top-1;
7944 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7946 /* Leaf page fill factor is below the threshold.
7947 * Try to move keys from left or right neighbor, or
7948 * merge with a neighbor page.
7953 mdb_cursor_copy(mc, &mn);
7954 mn.mc_xcursor = NULL;
7956 oldki = mc->mc_ki[mc->mc_top];
7957 if (mc->mc_ki[ptop] == 0) {
7958 /* We're the leftmost leaf in our parent.
7960 DPUTS("reading right neighbor");
7962 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7963 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7966 mn.mc_ki[mn.mc_top] = 0;
7967 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7969 /* There is at least one neighbor to the left.
7971 DPUTS("reading left neighbor");
7973 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7974 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7977 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7978 mc->mc_ki[mc->mc_top] = 0;
7981 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7982 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7983 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7985 /* If the neighbor page is above threshold and has enough keys,
7986 * move one key from it. Otherwise we should try to merge them.
7987 * (A branch page must never have less than 2 keys.)
7989 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7990 rc = mdb_node_move(&mn, mc);
7991 if (mc->mc_ki[mc->mc_top-1]) {
7995 if (mc->mc_ki[ptop] == 0) {
7996 rc = mdb_page_merge(&mn, mc);
7999 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8000 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8001 /* We want mdb_rebalance to find mn when doing fixups */
8002 if (mc->mc_flags & C_SUB) {
8003 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8004 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8005 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8007 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8008 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8010 rc = mdb_page_merge(mc, &mn);
8011 if (mc->mc_flags & C_SUB)
8012 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8014 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8015 mdb_cursor_copy(&mn, mc);
8017 mc->mc_flags &= ~C_EOF;
8019 mc->mc_ki[mc->mc_top] = oldki;
8023 /** Complete a delete operation started by #mdb_cursor_del(). */
8025 mdb_cursor_del0(MDB_cursor *mc)
8031 MDB_cursor *m2, *m3;
8032 MDB_dbi dbi = mc->mc_dbi;
8034 ki = mc->mc_ki[mc->mc_top];
8035 mp = mc->mc_pg[mc->mc_top];
8036 mdb_node_del(mc, mc->mc_db->md_pad);
8037 mc->mc_db->md_entries--;
8039 /* Adjust other cursors pointing to mp */
8040 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8041 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8042 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8044 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8046 if (m3->mc_pg[mc->mc_top] == mp) {
8047 if (m3->mc_ki[mc->mc_top] >= ki) {
8048 m3->mc_flags |= C_DEL;
8049 if (m3->mc_ki[mc->mc_top] > ki)
8050 m3->mc_ki[mc->mc_top]--;
8051 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8052 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8057 rc = mdb_rebalance(mc);
8059 if (rc == MDB_SUCCESS) {
8060 /* DB is totally empty now, just bail out.
8061 * Other cursors adjustments were already done
8062 * by mdb_rebalance and aren't needed here.
8067 mp = mc->mc_pg[mc->mc_top];
8068 nkeys = NUMKEYS(mp);
8070 /* Adjust other cursors pointing to mp */
8071 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8072 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8073 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8075 if (m3->mc_snum < mc->mc_snum)
8077 if (m3->mc_pg[mc->mc_top] == mp) {
8078 /* if m3 points past last node in page, find next sibling */
8079 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8080 rc = mdb_cursor_sibling(m3, 1);
8081 if (rc == MDB_NOTFOUND) {
8082 m3->mc_flags |= C_EOF;
8088 mc->mc_flags |= C_DEL;
8092 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8097 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8098 MDB_val *key, MDB_val *data)
8100 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8103 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8104 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8106 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8107 /* must ignore any data */
8111 return mdb_del0(txn, dbi, key, data, 0);
8115 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8116 MDB_val *key, MDB_val *data, unsigned flags)
8121 MDB_val rdata, *xdata;
8125 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8127 mdb_cursor_init(&mc, txn, dbi, &mx);
8136 flags |= MDB_NODUPDATA;
8138 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8140 /* let mdb_page_split know about this cursor if needed:
8141 * delete will trigger a rebalance; if it needs to move
8142 * a node from one page to another, it will have to
8143 * update the parent's separator key(s). If the new sepkey
8144 * is larger than the current one, the parent page may
8145 * run out of space, triggering a split. We need this
8146 * cursor to be consistent until the end of the rebalance.
8148 mc.mc_flags |= C_UNTRACK;
8149 mc.mc_next = txn->mt_cursors[dbi];
8150 txn->mt_cursors[dbi] = &mc;
8151 rc = mdb_cursor_del(&mc, flags);
8152 txn->mt_cursors[dbi] = mc.mc_next;
8157 /** Split a page and insert a new node.
8158 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8159 * The cursor will be updated to point to the actual page and index where
8160 * the node got inserted after the split.
8161 * @param[in] newkey The key for the newly inserted node.
8162 * @param[in] newdata The data for the newly inserted node.
8163 * @param[in] newpgno The page number, if the new node is a branch node.
8164 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8165 * @return 0 on success, non-zero on failure.
8168 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8169 unsigned int nflags)
8172 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8175 int i, j, split_indx, nkeys, pmax;
8176 MDB_env *env = mc->mc_txn->mt_env;
8178 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8179 MDB_page *copy = NULL;
8180 MDB_page *mp, *rp, *pp;
8185 mp = mc->mc_pg[mc->mc_top];
8186 newindx = mc->mc_ki[mc->mc_top];
8187 nkeys = NUMKEYS(mp);
8189 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8190 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8191 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8193 /* Create a right sibling. */
8194 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8196 rp->mp_pad = mp->mp_pad;
8197 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8199 if (mc->mc_snum < 2) {
8200 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8202 /* shift current top to make room for new parent */
8203 mc->mc_pg[1] = mc->mc_pg[0];
8204 mc->mc_ki[1] = mc->mc_ki[0];
8207 mc->mc_db->md_root = pp->mp_pgno;
8208 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8209 new_root = mc->mc_db->md_depth++;
8211 /* Add left (implicit) pointer. */
8212 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8213 /* undo the pre-push */
8214 mc->mc_pg[0] = mc->mc_pg[1];
8215 mc->mc_ki[0] = mc->mc_ki[1];
8216 mc->mc_db->md_root = mp->mp_pgno;
8217 mc->mc_db->md_depth--;
8224 ptop = mc->mc_top-1;
8225 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8228 mc->mc_flags |= C_SPLITTING;
8229 mdb_cursor_copy(mc, &mn);
8230 mn.mc_pg[mn.mc_top] = rp;
8231 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8233 if (nflags & MDB_APPEND) {
8234 mn.mc_ki[mn.mc_top] = 0;
8236 split_indx = newindx;
8240 split_indx = (nkeys+1) / 2;
8245 unsigned int lsize, rsize, ksize;
8246 /* Move half of the keys to the right sibling */
8247 x = mc->mc_ki[mc->mc_top] - split_indx;
8248 ksize = mc->mc_db->md_pad;
8249 split = LEAF2KEY(mp, split_indx, ksize);
8250 rsize = (nkeys - split_indx) * ksize;
8251 lsize = (nkeys - split_indx) * sizeof(indx_t);
8252 mp->mp_lower -= lsize;
8253 rp->mp_lower += lsize;
8254 mp->mp_upper += rsize - lsize;
8255 rp->mp_upper -= rsize - lsize;
8256 sepkey.mv_size = ksize;
8257 if (newindx == split_indx) {
8258 sepkey.mv_data = newkey->mv_data;
8260 sepkey.mv_data = split;
8263 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8264 memcpy(rp->mp_ptrs, split, rsize);
8265 sepkey.mv_data = rp->mp_ptrs;
8266 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8267 memcpy(ins, newkey->mv_data, ksize);
8268 mp->mp_lower += sizeof(indx_t);
8269 mp->mp_upper -= ksize - sizeof(indx_t);
8272 memcpy(rp->mp_ptrs, split, x * ksize);
8273 ins = LEAF2KEY(rp, x, ksize);
8274 memcpy(ins, newkey->mv_data, ksize);
8275 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8276 rp->mp_lower += sizeof(indx_t);
8277 rp->mp_upper -= ksize - sizeof(indx_t);
8278 mc->mc_ki[mc->mc_top] = x;
8279 mc->mc_pg[mc->mc_top] = rp;
8282 int psize, nsize, k;
8283 /* Maximum free space in an empty page */
8284 pmax = env->me_psize - PAGEHDRSZ;
8286 nsize = mdb_leaf_size(env, newkey, newdata);
8288 nsize = mdb_branch_size(env, newkey);
8289 nsize = EVEN(nsize);
8291 /* grab a page to hold a temporary copy */
8292 copy = mdb_page_malloc(mc->mc_txn, 1);
8297 copy->mp_pgno = mp->mp_pgno;
8298 copy->mp_flags = mp->mp_flags;
8299 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8300 copy->mp_upper = env->me_psize - PAGEBASE;
8302 /* prepare to insert */
8303 for (i=0, j=0; i<nkeys; i++) {
8305 copy->mp_ptrs[j++] = 0;
8307 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8310 /* When items are relatively large the split point needs
8311 * to be checked, because being off-by-one will make the
8312 * difference between success or failure in mdb_node_add.
8314 * It's also relevant if a page happens to be laid out
8315 * such that one half of its nodes are all "small" and
8316 * the other half of its nodes are "large." If the new
8317 * item is also "large" and falls on the half with
8318 * "large" nodes, it also may not fit.
8320 * As a final tweak, if the new item goes on the last
8321 * spot on the page (and thus, onto the new page), bias
8322 * the split so the new page is emptier than the old page.
8323 * This yields better packing during sequential inserts.
8325 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8326 /* Find split point */
8328 if (newindx <= split_indx || newindx >= nkeys) {
8330 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8335 for (; i!=k; i+=j) {
8340 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8341 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8343 if (F_ISSET(node->mn_flags, F_BIGDATA))
8344 psize += sizeof(pgno_t);
8346 psize += NODEDSZ(node);
8348 psize = EVEN(psize);
8350 if (psize > pmax || i == k-j) {
8351 split_indx = i + (j<0);
8356 if (split_indx == newindx) {
8357 sepkey.mv_size = newkey->mv_size;
8358 sepkey.mv_data = newkey->mv_data;
8360 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8361 sepkey.mv_size = node->mn_ksize;
8362 sepkey.mv_data = NODEKEY(node);
8367 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8369 /* Copy separator key to the parent.
8371 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8375 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8380 if (mn.mc_snum == mc->mc_snum) {
8381 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8382 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8383 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8384 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8389 /* Right page might now have changed parent.
8390 * Check if left page also changed parent.
8392 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8393 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8394 for (i=0; i<ptop; i++) {
8395 mc->mc_pg[i] = mn.mc_pg[i];
8396 mc->mc_ki[i] = mn.mc_ki[i];
8398 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8399 if (mn.mc_ki[ptop]) {
8400 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8402 /* find right page's left sibling */
8403 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8404 mdb_cursor_sibling(mc, 0);
8409 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8412 mc->mc_flags ^= C_SPLITTING;
8413 if (rc != MDB_SUCCESS) {
8416 if (nflags & MDB_APPEND) {
8417 mc->mc_pg[mc->mc_top] = rp;
8418 mc->mc_ki[mc->mc_top] = 0;
8419 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8422 for (i=0; i<mc->mc_top; i++)
8423 mc->mc_ki[i] = mn.mc_ki[i];
8424 } else if (!IS_LEAF2(mp)) {
8426 mc->mc_pg[mc->mc_top] = rp;
8431 rkey.mv_data = newkey->mv_data;
8432 rkey.mv_size = newkey->mv_size;
8438 /* Update index for the new key. */
8439 mc->mc_ki[mc->mc_top] = j;
8441 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8442 rkey.mv_data = NODEKEY(node);
8443 rkey.mv_size = node->mn_ksize;
8445 xdata.mv_data = NODEDATA(node);
8446 xdata.mv_size = NODEDSZ(node);
8449 pgno = NODEPGNO(node);
8450 flags = node->mn_flags;
8453 if (!IS_LEAF(mp) && j == 0) {
8454 /* First branch index doesn't need key data. */
8458 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8464 mc->mc_pg[mc->mc_top] = copy;
8469 } while (i != split_indx);
8471 nkeys = NUMKEYS(copy);
8472 for (i=0; i<nkeys; i++)
8473 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8474 mp->mp_lower = copy->mp_lower;
8475 mp->mp_upper = copy->mp_upper;
8476 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8477 env->me_psize - copy->mp_upper - PAGEBASE);
8479 /* reset back to original page */
8480 if (newindx < split_indx) {
8481 mc->mc_pg[mc->mc_top] = mp;
8482 if (nflags & MDB_RESERVE) {
8483 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8484 if (!(node->mn_flags & F_BIGDATA))
8485 newdata->mv_data = NODEDATA(node);
8488 mc->mc_pg[mc->mc_top] = rp;
8490 /* Make sure mc_ki is still valid.
8492 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8493 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8494 for (i=0; i<=ptop; i++) {
8495 mc->mc_pg[i] = mn.mc_pg[i];
8496 mc->mc_ki[i] = mn.mc_ki[i];
8503 /* Adjust other cursors pointing to mp */
8504 MDB_cursor *m2, *m3;
8505 MDB_dbi dbi = mc->mc_dbi;
8506 int fixup = NUMKEYS(mp);
8508 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8509 if (mc->mc_flags & C_SUB)
8510 m3 = &m2->mc_xcursor->mx_cursor;
8515 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8517 if (m3->mc_flags & C_SPLITTING)
8522 for (k=new_root; k>=0; k--) {
8523 m3->mc_ki[k+1] = m3->mc_ki[k];
8524 m3->mc_pg[k+1] = m3->mc_pg[k];
8526 if (m3->mc_ki[0] >= split_indx) {
8531 m3->mc_pg[0] = mc->mc_pg[0];
8535 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8536 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8537 m3->mc_ki[mc->mc_top]++;
8538 if (m3->mc_ki[mc->mc_top] >= fixup) {
8539 m3->mc_pg[mc->mc_top] = rp;
8540 m3->mc_ki[mc->mc_top] -= fixup;
8541 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8543 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8544 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8549 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8552 if (copy) /* tmp page */
8553 mdb_page_free(env, copy);
8555 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8560 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8561 MDB_val *key, MDB_val *data, unsigned int flags)
8566 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8569 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8572 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8573 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8575 mdb_cursor_init(&mc, txn, dbi, &mx);
8576 return mdb_cursor_put(&mc, key, data, flags);
8580 #define MDB_WBUF (1024*1024)
8583 /** State needed for a compacting copy. */
8584 typedef struct mdb_copy {
8585 pthread_mutex_t mc_mutex;
8586 pthread_cond_t mc_cond;
8593 pgno_t mc_next_pgno;
8596 volatile int mc_new;
8601 /** Dedicated writer thread for compacting copy. */
8602 static THREAD_RET ESECT CALL_CONV
8603 mdb_env_copythr(void *arg)
8607 int toggle = 0, wsize, rc;
8610 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8613 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8616 pthread_mutex_lock(&my->mc_mutex);
8618 pthread_cond_signal(&my->mc_cond);
8621 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8622 if (my->mc_new < 0) {
8627 wsize = my->mc_wlen[toggle];
8628 ptr = my->mc_wbuf[toggle];
8631 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8635 } else if (len > 0) {
8649 /* If there's an overflow page tail, write it too */
8650 if (my->mc_olen[toggle]) {
8651 wsize = my->mc_olen[toggle];
8652 ptr = my->mc_over[toggle];
8653 my->mc_olen[toggle] = 0;
8656 my->mc_wlen[toggle] = 0;
8658 pthread_cond_signal(&my->mc_cond);
8660 pthread_cond_signal(&my->mc_cond);
8661 pthread_mutex_unlock(&my->mc_mutex);
8662 return (THREAD_RET)0;
8666 /** Tell the writer thread there's a buffer ready to write */
8668 mdb_env_cthr_toggle(mdb_copy *my, int st)
8670 int toggle = my->mc_toggle ^ 1;
8671 pthread_mutex_lock(&my->mc_mutex);
8672 if (my->mc_status) {
8673 pthread_mutex_unlock(&my->mc_mutex);
8674 return my->mc_status;
8676 while (my->mc_new == 1)
8677 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8679 my->mc_toggle = toggle;
8680 pthread_cond_signal(&my->mc_cond);
8681 pthread_mutex_unlock(&my->mc_mutex);
8685 /** Depth-first tree traversal for compacting copy. */
8687 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8690 MDB_txn *txn = my->mc_txn;
8692 MDB_page *mo, *mp, *leaf;
8697 /* Empty DB, nothing to do */
8698 if (*pg == P_INVALID)
8705 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8708 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8712 /* Make cursor pages writable */
8713 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8717 for (i=0; i<mc.mc_top; i++) {
8718 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8719 mc.mc_pg[i] = (MDB_page *)ptr;
8720 ptr += my->mc_env->me_psize;
8723 /* This is writable space for a leaf page. Usually not needed. */
8724 leaf = (MDB_page *)ptr;
8726 toggle = my->mc_toggle;
8727 while (mc.mc_snum > 0) {
8729 mp = mc.mc_pg[mc.mc_top];
8733 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8734 for (i=0; i<n; i++) {
8735 ni = NODEPTR(mp, i);
8736 if (ni->mn_flags & F_BIGDATA) {
8740 /* Need writable leaf */
8742 mc.mc_pg[mc.mc_top] = leaf;
8743 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8745 ni = NODEPTR(mp, i);
8748 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8749 rc = mdb_page_get(txn, pg, &omp, NULL);
8752 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8753 rc = mdb_env_cthr_toggle(my, 1);
8756 toggle = my->mc_toggle;
8758 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8759 memcpy(mo, omp, my->mc_env->me_psize);
8760 mo->mp_pgno = my->mc_next_pgno;
8761 my->mc_next_pgno += omp->mp_pages;
8762 my->mc_wlen[toggle] += my->mc_env->me_psize;
8763 if (omp->mp_pages > 1) {
8764 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8765 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8766 rc = mdb_env_cthr_toggle(my, 1);
8769 toggle = my->mc_toggle;
8771 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8772 } else if (ni->mn_flags & F_SUBDATA) {
8775 /* Need writable leaf */
8777 mc.mc_pg[mc.mc_top] = leaf;
8778 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8780 ni = NODEPTR(mp, i);
8783 memcpy(&db, NODEDATA(ni), sizeof(db));
8784 my->mc_toggle = toggle;
8785 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8788 toggle = my->mc_toggle;
8789 memcpy(NODEDATA(ni), &db, sizeof(db));
8794 mc.mc_ki[mc.mc_top]++;
8795 if (mc.mc_ki[mc.mc_top] < n) {
8798 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8800 rc = mdb_page_get(txn, pg, &mp, NULL);
8805 mc.mc_ki[mc.mc_top] = 0;
8806 if (IS_BRANCH(mp)) {
8807 /* Whenever we advance to a sibling branch page,
8808 * we must proceed all the way down to its first leaf.
8810 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8813 mc.mc_pg[mc.mc_top] = mp;
8817 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8818 rc = mdb_env_cthr_toggle(my, 1);
8821 toggle = my->mc_toggle;
8823 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8824 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8825 mo->mp_pgno = my->mc_next_pgno++;
8826 my->mc_wlen[toggle] += my->mc_env->me_psize;
8828 /* Update parent if there is one */
8829 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8830 SETPGNO(ni, mo->mp_pgno);
8831 mdb_cursor_pop(&mc);
8833 /* Otherwise we're done */
8843 /** Copy environment with compaction. */
8845 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8850 MDB_txn *txn = NULL;
8855 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8856 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8857 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8858 if (my.mc_wbuf[0] == NULL)
8861 pthread_mutex_init(&my.mc_mutex, NULL);
8862 pthread_cond_init(&my.mc_cond, NULL);
8863 #ifdef HAVE_MEMALIGN
8864 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8865 if (my.mc_wbuf[0] == NULL)
8868 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8873 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8874 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8879 my.mc_next_pgno = NUM_METAS;
8885 THREAD_CREATE(thr, mdb_env_copythr, &my);
8887 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8891 mp = (MDB_page *)my.mc_wbuf[0];
8892 memset(mp, 0, NUM_METAS * env->me_psize);
8894 mp->mp_flags = P_META;
8895 mm = (MDB_meta *)METADATA(mp);
8896 mdb_env_init_meta0(env, mm);
8897 mm->mm_address = env->me_metas[0]->mm_address;
8899 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8901 mp->mp_flags = P_META;
8902 *(MDB_meta *)METADATA(mp) = *mm;
8903 mm = (MDB_meta *)METADATA(mp);
8905 /* Count the number of free pages, subtract from lastpg to find
8906 * number of active pages
8909 MDB_ID freecount = 0;
8912 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8913 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8914 freecount += *(MDB_ID *)data.mv_data;
8915 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
8916 txn->mt_dbs[FREE_DBI].md_leaf_pages +
8917 txn->mt_dbs[FREE_DBI].md_overflow_pages;
8919 /* Set metapage 1 */
8920 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8921 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
8922 if (mm->mm_last_pg > NUM_METAS-1) {
8923 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
8926 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
8929 my.mc_wlen[0] = env->me_psize * NUM_METAS;
8931 pthread_mutex_lock(&my.mc_mutex);
8933 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8934 pthread_mutex_unlock(&my.mc_mutex);
8935 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
8936 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8937 rc = mdb_env_cthr_toggle(&my, 1);
8938 mdb_env_cthr_toggle(&my, -1);
8939 pthread_mutex_lock(&my.mc_mutex);
8941 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8942 pthread_mutex_unlock(&my.mc_mutex);
8947 CloseHandle(my.mc_cond);
8948 CloseHandle(my.mc_mutex);
8949 _aligned_free(my.mc_wbuf[0]);
8951 pthread_cond_destroy(&my.mc_cond);
8952 pthread_mutex_destroy(&my.mc_mutex);
8953 free(my.mc_wbuf[0]);
8958 /** Copy environment as-is. */
8960 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8962 MDB_txn *txn = NULL;
8963 mdb_mutexref_t wmutex = NULL;
8969 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8973 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8976 /* Do the lock/unlock of the reader mutex before starting the
8977 * write txn. Otherwise other read txns could block writers.
8979 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8984 /* We must start the actual read txn after blocking writers */
8985 mdb_txn_end(txn, MDB_END_RESET_TMP);
8987 /* Temporarily block writers until we snapshot the meta pages */
8988 wmutex = env->me_wmutex;
8989 if (LOCK_MUTEX(rc, env, wmutex))
8992 rc = mdb_txn_renew0(txn);
8994 UNLOCK_MUTEX(wmutex);
8999 wsize = env->me_psize * NUM_METAS;
9003 DO_WRITE(rc, fd, ptr, w2, len);
9007 } else if (len > 0) {
9013 /* Non-blocking or async handles are not supported */
9019 UNLOCK_MUTEX(wmutex);
9024 w2 = txn->mt_next_pgno * env->me_psize;
9027 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9034 if (wsize > MAX_WRITE)
9038 DO_WRITE(rc, fd, ptr, w2, len);
9042 } else if (len > 0) {
9059 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9061 if (flags & MDB_CP_COMPACT)
9062 return mdb_env_copyfd1(env, fd);
9064 return mdb_env_copyfd0(env, fd);
9068 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9070 return mdb_env_copyfd2(env, fd, 0);
9074 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9078 HANDLE newfd = INVALID_HANDLE_VALUE;
9080 if (env->me_flags & MDB_NOSUBDIR) {
9081 lpath = (char *)path;
9084 len += sizeof(DATANAME);
9085 lpath = malloc(len);
9088 sprintf(lpath, "%s" DATANAME, path);
9091 /* The destination path must exist, but the destination file must not.
9092 * We don't want the OS to cache the writes, since the source data is
9093 * already in the OS cache.
9096 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9097 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9099 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9101 if (newfd == INVALID_HANDLE_VALUE) {
9106 if (env->me_psize >= env->me_os_psize) {
9108 /* Set O_DIRECT if the file system supports it */
9109 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9110 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9112 #ifdef F_NOCACHE /* __APPLE__ */
9113 rc = fcntl(newfd, F_NOCACHE, 1);
9121 rc = mdb_env_copyfd2(env, newfd, flags);
9124 if (!(env->me_flags & MDB_NOSUBDIR))
9126 if (newfd != INVALID_HANDLE_VALUE)
9127 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9134 mdb_env_copy(MDB_env *env, const char *path)
9136 return mdb_env_copy2(env, path, 0);
9140 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9142 if (flag & ~CHANGEABLE)
9145 env->me_flags |= flag;
9147 env->me_flags &= ~flag;
9152 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9157 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9162 mdb_env_set_userctx(MDB_env *env, void *ctx)
9166 env->me_userctx = ctx;
9171 mdb_env_get_userctx(MDB_env *env)
9173 return env ? env->me_userctx : NULL;
9177 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9182 env->me_assert_func = func;
9188 mdb_env_get_path(MDB_env *env, const char **arg)
9193 *arg = env->me_path;
9198 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9207 /** Common code for #mdb_stat() and #mdb_env_stat().
9208 * @param[in] env the environment to operate in.
9209 * @param[in] db the #MDB_db record containing the stats to return.
9210 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9211 * @return 0, this function always succeeds.
9214 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9216 arg->ms_psize = env->me_psize;
9217 arg->ms_depth = db->md_depth;
9218 arg->ms_branch_pages = db->md_branch_pages;
9219 arg->ms_leaf_pages = db->md_leaf_pages;
9220 arg->ms_overflow_pages = db->md_overflow_pages;
9221 arg->ms_entries = db->md_entries;
9227 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9231 if (env == NULL || arg == NULL)
9234 meta = mdb_env_pick_meta(env);
9236 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9240 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9244 if (env == NULL || arg == NULL)
9247 meta = mdb_env_pick_meta(env);
9248 arg->me_mapaddr = meta->mm_address;
9249 arg->me_last_pgno = meta->mm_last_pg;
9250 arg->me_last_txnid = meta->mm_txnid;
9252 arg->me_mapsize = env->me_mapsize;
9253 arg->me_maxreaders = env->me_maxreaders;
9254 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9258 /** Set the default comparison functions for a database.
9259 * Called immediately after a database is opened to set the defaults.
9260 * The user can then override them with #mdb_set_compare() or
9261 * #mdb_set_dupsort().
9262 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9263 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9266 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9268 uint16_t f = txn->mt_dbs[dbi].md_flags;
9270 txn->mt_dbxs[dbi].md_cmp =
9271 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9272 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9274 txn->mt_dbxs[dbi].md_dcmp =
9275 !(f & MDB_DUPSORT) ? 0 :
9276 ((f & MDB_INTEGERDUP)
9277 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9278 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9281 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9287 int rc, dbflag, exact;
9288 unsigned int unused = 0, seq;
9291 if (flags & ~VALID_FLAGS)
9293 if (txn->mt_flags & MDB_TXN_BLOCKED)
9299 if (flags & PERSISTENT_FLAGS) {
9300 uint16_t f2 = flags & PERSISTENT_FLAGS;
9301 /* make sure flag changes get committed */
9302 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9303 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9304 txn->mt_flags |= MDB_TXN_DIRTY;
9307 mdb_default_cmp(txn, MAIN_DBI);
9311 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9312 mdb_default_cmp(txn, MAIN_DBI);
9315 /* Is the DB already open? */
9317 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9318 if (!txn->mt_dbxs[i].md_name.mv_size) {
9319 /* Remember this free slot */
9320 if (!unused) unused = i;
9323 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9324 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9330 /* If no free slot and max hit, fail */
9331 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9332 return MDB_DBS_FULL;
9334 /* Cannot mix named databases with some mainDB flags */
9335 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9336 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9338 /* Find the DB info */
9339 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9342 key.mv_data = (void *)name;
9343 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9344 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9345 if (rc == MDB_SUCCESS) {
9346 /* make sure this is actually a DB */
9347 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9348 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9349 return MDB_INCOMPATIBLE;
9350 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9351 /* Create if requested */
9352 data.mv_size = sizeof(MDB_db);
9353 data.mv_data = &dummy;
9354 memset(&dummy, 0, sizeof(dummy));
9355 dummy.md_root = P_INVALID;
9356 dummy.md_flags = flags & PERSISTENT_FLAGS;
9357 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9361 /* OK, got info, add to table */
9362 if (rc == MDB_SUCCESS) {
9363 unsigned int slot = unused ? unused : txn->mt_numdbs;
9364 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9365 txn->mt_dbxs[slot].md_name.mv_size = len;
9366 txn->mt_dbxs[slot].md_rel = NULL;
9367 txn->mt_dbflags[slot] = dbflag;
9368 /* txn-> and env-> are the same in read txns, use
9369 * tmp variable to avoid undefined assignment
9371 seq = ++txn->mt_env->me_dbiseqs[slot];
9372 txn->mt_dbiseqs[slot] = seq;
9374 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9376 mdb_default_cmp(txn, slot);
9386 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9388 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9391 if (txn->mt_flags & MDB_TXN_BLOCKED)
9394 if (txn->mt_dbflags[dbi] & DB_STALE) {
9397 /* Stale, must read the DB's root. cursor_init does it for us. */
9398 mdb_cursor_init(&mc, txn, dbi, &mx);
9400 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9403 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9406 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9408 ptr = env->me_dbxs[dbi].md_name.mv_data;
9409 /* If there was no name, this was already closed */
9411 env->me_dbxs[dbi].md_name.mv_data = NULL;
9412 env->me_dbxs[dbi].md_name.mv_size = 0;
9413 env->me_dbflags[dbi] = 0;
9414 env->me_dbiseqs[dbi]++;
9419 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9421 /* We could return the flags for the FREE_DBI too but what's the point? */
9422 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9424 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9428 /** Add all the DB's pages to the free list.
9429 * @param[in] mc Cursor on the DB to free.
9430 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9431 * @return 0 on success, non-zero on failure.
9434 mdb_drop0(MDB_cursor *mc, int subs)
9438 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9439 if (rc == MDB_SUCCESS) {
9440 MDB_txn *txn = mc->mc_txn;
9445 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9446 * This also avoids any P_LEAF2 pages, which have no nodes.
9448 if (mc->mc_flags & C_SUB)
9451 mdb_cursor_copy(mc, &mx);
9452 while (mc->mc_snum > 0) {
9453 MDB_page *mp = mc->mc_pg[mc->mc_top];
9454 unsigned n = NUMKEYS(mp);
9456 for (i=0; i<n; i++) {
9457 ni = NODEPTR(mp, i);
9458 if (ni->mn_flags & F_BIGDATA) {
9461 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9462 rc = mdb_page_get(txn, pg, &omp, NULL);
9465 mdb_cassert(mc, IS_OVERFLOW(omp));
9466 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9470 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9471 mdb_xcursor_init1(mc, ni);
9472 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9478 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9480 for (i=0; i<n; i++) {
9482 ni = NODEPTR(mp, i);
9485 mdb_midl_xappend(txn->mt_free_pgs, pg);
9490 mc->mc_ki[mc->mc_top] = i;
9491 rc = mdb_cursor_sibling(mc, 1);
9493 if (rc != MDB_NOTFOUND)
9495 /* no more siblings, go back to beginning
9496 * of previous level.
9500 for (i=1; i<mc->mc_snum; i++) {
9502 mc->mc_pg[i] = mx.mc_pg[i];
9507 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9510 txn->mt_flags |= MDB_TXN_ERROR;
9511 } else if (rc == MDB_NOTFOUND) {
9517 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9519 MDB_cursor *mc, *m2;
9522 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9525 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9528 if (TXN_DBI_CHANGED(txn, dbi))
9531 rc = mdb_cursor_open(txn, dbi, &mc);
9535 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9536 /* Invalidate the dropped DB's cursors */
9537 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9538 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9542 /* Can't delete the main DB */
9543 if (del && dbi >= CORE_DBS) {
9544 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9546 txn->mt_dbflags[dbi] = DB_STALE;
9547 mdb_dbi_close(txn->mt_env, dbi);
9549 txn->mt_flags |= MDB_TXN_ERROR;
9552 /* reset the DB record, mark it dirty */
9553 txn->mt_dbflags[dbi] |= DB_DIRTY;
9554 txn->mt_dbs[dbi].md_depth = 0;
9555 txn->mt_dbs[dbi].md_branch_pages = 0;
9556 txn->mt_dbs[dbi].md_leaf_pages = 0;
9557 txn->mt_dbs[dbi].md_overflow_pages = 0;
9558 txn->mt_dbs[dbi].md_entries = 0;
9559 txn->mt_dbs[dbi].md_root = P_INVALID;
9561 txn->mt_flags |= MDB_TXN_DIRTY;
9564 mdb_cursor_close(mc);
9568 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9570 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9573 txn->mt_dbxs[dbi].md_cmp = cmp;
9577 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9579 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9582 txn->mt_dbxs[dbi].md_dcmp = cmp;
9586 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9588 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9591 txn->mt_dbxs[dbi].md_rel = rel;
9595 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9597 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9600 txn->mt_dbxs[dbi].md_relctx = ctx;
9605 mdb_env_get_maxkeysize(MDB_env *env)
9607 return ENV_MAXKEY(env);
9611 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9613 unsigned int i, rdrs;
9616 int rc = 0, first = 1;
9620 if (!env->me_txns) {
9621 return func("(no reader locks)\n", ctx);
9623 rdrs = env->me_txns->mti_numreaders;
9624 mr = env->me_txns->mti_readers;
9625 for (i=0; i<rdrs; i++) {
9627 txnid_t txnid = mr[i].mr_txnid;
9628 sprintf(buf, txnid == (txnid_t)-1 ?
9629 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9630 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9633 rc = func(" pid thread txnid\n", ctx);
9637 rc = func(buf, ctx);
9643 rc = func("(no active readers)\n", ctx);
9648 /** Insert pid into list if not already present.
9649 * return -1 if already present.
9652 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9654 /* binary search of pid in list */
9656 unsigned cursor = 1;
9658 unsigned n = ids[0];
9661 unsigned pivot = n >> 1;
9662 cursor = base + pivot + 1;
9663 val = pid - ids[cursor];
9668 } else if ( val > 0 ) {
9673 /* found, so it's a duplicate */
9682 for (n = ids[0]; n > cursor; n--)
9689 mdb_reader_check(MDB_env *env, int *dead)
9695 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9698 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9700 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9702 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9703 unsigned int i, j, rdrs;
9705 MDB_PID_T *pids, pid;
9706 int rc = MDB_SUCCESS, count = 0;
9708 rdrs = env->me_txns->mti_numreaders;
9709 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9713 mr = env->me_txns->mti_readers;
9714 for (i=0; i<rdrs; i++) {
9716 if (pid && pid != env->me_pid) {
9717 if (mdb_pid_insert(pids, pid) == 0) {
9718 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9719 /* Stale reader found */
9722 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9723 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9725 rdrs = 0; /* the above checked all readers */
9727 /* Recheck, a new process may have reused pid */
9728 if (mdb_reader_pid(env, Pidcheck, pid))
9733 if (mr[j].mr_pid == pid) {
9734 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9735 (unsigned) pid, mr[j].mr_txnid));
9740 UNLOCK_MUTEX(rmutex);
9751 #ifdef MDB_ROBUST_SUPPORTED
9752 /** Handle #LOCK_MUTEX0() failure.
9753 * Try to repair the lock file if the mutex owner died.
9754 * @param[in] env the environment handle
9755 * @param[in] mutex LOCK_MUTEX0() mutex
9756 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9757 * @return 0 on success with the mutex locked, or an error code on failure.
9760 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9765 if (rc == MDB_OWNERDEAD) {
9766 /* We own the mutex. Clean up after dead previous owner. */
9768 rlocked = (mutex == env->me_rmutex);
9770 /* Keep mti_txnid updated, otherwise next writer can
9771 * overwrite data which latest meta page refers to.
9773 meta = mdb_env_pick_meta(env);
9774 env->me_txns->mti_txnid = meta->mm_txnid;
9775 /* env is hosed if the dead thread was ours */
9777 env->me_flags |= MDB_FATAL_ERROR;
9782 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9783 (rc ? "this process' env is hosed" : "recovering")));
9784 rc2 = mdb_reader_check0(env, rlocked, NULL);
9786 rc2 = mdb_mutex_consistent(mutex);
9787 if (rc || (rc = rc2)) {
9788 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9789 UNLOCK_MUTEX(mutex);
9795 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9800 #endif /* MDB_ROBUST_SUPPORTED */