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. */
947 /** Meta page content.
948 * A meta page is the start point for accessing a database snapshot.
949 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
951 typedef struct MDB_meta {
952 /** Stamp identifying this as an LMDB file. It must be set
955 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
957 void *mm_address; /**< address for fixed mapping */
958 size_t mm_mapsize; /**< size of mmap region */
959 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
960 /** The size of pages used in this DB */
961 #define mm_psize mm_dbs[0].md_pad
962 /** Any persistent environment flags. @ref mdb_env */
963 #define mm_flags mm_dbs[0].md_flags
964 pgno_t mm_last_pg; /**< last used page in file */
965 volatile txnid_t mm_txnid; /**< txnid that committed this page */
968 /** Buffer for a stack-allocated meta page.
969 * The members define size and alignment, and silence type
970 * aliasing warnings. They are not used directly; that could
971 * mean incorrectly using several union members in parallel.
973 typedef union MDB_metabuf {
976 char mm_pad[PAGEHDRSZ];
981 /** Auxiliary DB info.
982 * The information here is mostly static/read-only. There is
983 * only a single copy of this record in the environment.
985 typedef struct MDB_dbx {
986 MDB_val md_name; /**< name of the database */
987 MDB_cmp_func *md_cmp; /**< function for comparing keys */
988 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
989 MDB_rel_func *md_rel; /**< user relocate function */
990 void *md_relctx; /**< user-provided context for md_rel */
993 /** A database transaction.
994 * Every operation requires a transaction handle.
997 MDB_txn *mt_parent; /**< parent of a nested txn */
998 MDB_txn *mt_child; /**< nested txn under this txn */
999 pgno_t mt_next_pgno; /**< next unallocated page */
1000 /** The ID of this transaction. IDs are integers incrementing from 1.
1001 * Only committed write transactions increment the ID. If a transaction
1002 * aborts, the ID may be re-used by the next writer.
1005 MDB_env *mt_env; /**< the DB environment */
1006 /** The list of pages that became unused during this transaction.
1008 MDB_IDL mt_free_pgs;
1009 /** The list of loose pages that became unused and may be reused
1010 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1012 MDB_page *mt_loose_pgs;
1013 /* #Number of loose pages (#mt_loose_pgs) */
1015 /** The sorted list of dirty pages we temporarily wrote to disk
1016 * because the dirty list was full. page numbers in here are
1017 * shifted left by 1, deleted slots have the LSB set.
1019 MDB_IDL mt_spill_pgs;
1021 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1022 MDB_ID2L dirty_list;
1023 /** For read txns: This thread/txn's reader table slot, or NULL. */
1026 /** Array of records for each DB known in the environment. */
1028 /** Array of MDB_db records for each known DB */
1030 /** Array of sequence numbers for each DB handle */
1031 unsigned int *mt_dbiseqs;
1032 /** @defgroup mt_dbflag Transaction DB Flags
1036 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1037 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1038 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1039 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1041 /** In write txns, array of cursors for each DB */
1042 MDB_cursor **mt_cursors;
1043 /** Array of flags for each DB */
1044 unsigned char *mt_dbflags;
1045 /** Number of DB records in use. This number only ever increments;
1046 * we don't decrement it when individual DB handles are closed.
1050 /** @defgroup mdb_txn Transaction Flags
1054 /** #mdb_txn_begin() flags */
1055 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1056 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1057 /* internal txn flags */
1058 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1059 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1060 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1061 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1063 unsigned int mt_flags; /**< @ref mdb_txn */
1064 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1065 * Includes ancestor txns' dirty pages not hidden by other txns'
1066 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1067 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1069 unsigned int mt_dirty_room;
1072 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1073 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1074 * raise this on a 64 bit machine.
1076 #define CURSOR_STACK 32
1080 /** Cursors are used for all DB operations.
1081 * A cursor holds a path of (page pointer, key index) from the DB
1082 * root to a position in the DB, plus other state. #MDB_DUPSORT
1083 * cursors include an xcursor to the current data item. Write txns
1084 * track their cursors and keep them up to date when data moves.
1085 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1086 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1089 /** Next cursor on this DB in this txn */
1090 MDB_cursor *mc_next;
1091 /** Backup of the original cursor if this cursor is a shadow */
1092 MDB_cursor *mc_backup;
1093 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1094 struct MDB_xcursor *mc_xcursor;
1095 /** The transaction that owns this cursor */
1097 /** The database handle this cursor operates on */
1099 /** The database record for this cursor */
1101 /** The database auxiliary record for this cursor */
1103 /** The @ref mt_dbflag for this database */
1104 unsigned char *mc_dbflag;
1105 unsigned short mc_snum; /**< number of pushed pages */
1106 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1107 /** @defgroup mdb_cursor Cursor Flags
1109 * Cursor state flags.
1112 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1113 #define C_EOF 0x02 /**< No more data */
1114 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1115 #define C_DEL 0x08 /**< last op was a cursor_del */
1116 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1117 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1119 unsigned int mc_flags; /**< @ref mdb_cursor */
1120 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1121 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1124 /** Context for sorted-dup records.
1125 * We could have gone to a fully recursive design, with arbitrarily
1126 * deep nesting of sub-databases. But for now we only handle these
1127 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1129 typedef struct MDB_xcursor {
1130 /** A sub-cursor for traversing the Dup DB */
1131 MDB_cursor mx_cursor;
1132 /** The database record for this Dup DB */
1134 /** The auxiliary DB record for this Dup DB */
1136 /** The @ref mt_dbflag for this Dup DB */
1137 unsigned char mx_dbflag;
1140 /** State of FreeDB old pages, stored in the MDB_env */
1141 typedef struct MDB_pgstate {
1142 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1143 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1146 /** The database environment. */
1148 HANDLE me_fd; /**< The main data file */
1149 HANDLE me_lfd; /**< The lock file */
1150 HANDLE me_mfd; /**< just for writing the meta pages */
1151 /** Failed to update the meta page. Probably an I/O error. */
1152 #define MDB_FATAL_ERROR 0x80000000U
1153 /** Some fields are initialized. */
1154 #define MDB_ENV_ACTIVE 0x20000000U
1155 /** me_txkey is set */
1156 #define MDB_ENV_TXKEY 0x10000000U
1157 /** fdatasync is unreliable */
1158 #define MDB_FSYNCONLY 0x08000000U
1159 uint32_t me_flags; /**< @ref mdb_env */
1160 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1161 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1162 unsigned int me_maxreaders; /**< size of the reader table */
1163 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1164 volatile int me_close_readers;
1165 MDB_dbi me_numdbs; /**< number of DBs opened */
1166 MDB_dbi me_maxdbs; /**< size of the DB table */
1167 MDB_PID_T me_pid; /**< process ID of this env */
1168 char *me_path; /**< path to the DB files */
1169 char *me_map; /**< the memory map of the data file */
1170 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1171 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1172 void *me_pbuf; /**< scratch area for DUPSORT put() */
1173 MDB_txn *me_txn; /**< current write transaction */
1174 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1175 size_t me_mapsize; /**< size of the data memory map */
1176 off_t me_size; /**< current file size */
1177 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1178 MDB_dbx *me_dbxs; /**< array of static DB info */
1179 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1180 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1181 pthread_key_t me_txkey; /**< thread-key for readers */
1182 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1183 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1184 # define me_pglast me_pgstate.mf_pglast
1185 # define me_pghead me_pgstate.mf_pghead
1186 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1187 /** IDL of pages that became unused in a write txn */
1188 MDB_IDL me_free_pgs;
1189 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1190 MDB_ID2L me_dirty_list;
1191 /** Max number of freelist items that can fit in a single overflow page */
1193 /** Max size of a node on a page */
1194 unsigned int me_nodemax;
1195 #if !(MDB_MAXKEYSIZE)
1196 unsigned int me_maxkey; /**< max size of a key */
1198 int me_live_reader; /**< have liveness lock in reader table */
1200 int me_pidquery; /**< Used in OpenProcess */
1202 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1203 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1204 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1206 mdb_mutex_t me_rmutex;
1207 mdb_mutex_t me_wmutex;
1209 void *me_userctx; /**< User-settable context */
1210 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1213 /** Nested transaction */
1214 typedef struct MDB_ntxn {
1215 MDB_txn mnt_txn; /**< the transaction */
1216 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1219 /** max number of pages to commit in one writev() call */
1220 #define MDB_COMMIT_PAGES 64
1221 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1222 #undef MDB_COMMIT_PAGES
1223 #define MDB_COMMIT_PAGES IOV_MAX
1226 /** max bytes to write in one call */
1227 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1229 /** Check \b txn and \b dbi arguments to a function */
1230 #define TXN_DBI_EXIST(txn, dbi) \
1231 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1233 /** Check for misused \b dbi handles */
1234 #define TXN_DBI_CHANGED(txn, dbi) \
1235 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1237 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1238 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1239 static int mdb_page_touch(MDB_cursor *mc);
1241 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1242 static int mdb_page_search_root(MDB_cursor *mc,
1243 MDB_val *key, int modify);
1244 #define MDB_PS_MODIFY 1
1245 #define MDB_PS_ROOTONLY 2
1246 #define MDB_PS_FIRST 4
1247 #define MDB_PS_LAST 8
1248 static int mdb_page_search(MDB_cursor *mc,
1249 MDB_val *key, int flags);
1250 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1252 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1253 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1254 pgno_t newpgno, unsigned int nflags);
1256 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1257 static int mdb_env_pick_meta(const MDB_env *env);
1258 static int mdb_env_write_meta(MDB_txn *txn);
1259 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1260 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1262 static void mdb_env_close0(MDB_env *env, int excl);
1264 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1265 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1266 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1267 static void mdb_node_del(MDB_cursor *mc, int ksize);
1268 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1269 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1270 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1271 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1272 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1274 static int mdb_rebalance(MDB_cursor *mc);
1275 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1277 static void mdb_cursor_pop(MDB_cursor *mc);
1278 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1280 static int mdb_cursor_del0(MDB_cursor *mc);
1281 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1282 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1283 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1284 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1285 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1287 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1288 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1290 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1291 static void mdb_xcursor_init0(MDB_cursor *mc);
1292 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1293 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1295 static int mdb_drop0(MDB_cursor *mc, int subs);
1296 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1297 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1300 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1303 /** Compare two items pointing at size_t's of unknown alignment. */
1304 #ifdef MISALIGNED_OK
1305 # define mdb_cmp_clong mdb_cmp_long
1307 # define mdb_cmp_clong mdb_cmp_cint
1311 static SECURITY_DESCRIPTOR mdb_null_sd;
1312 static SECURITY_ATTRIBUTES mdb_all_sa;
1313 static int mdb_sec_inited;
1316 /** Return the library version info. */
1318 mdb_version(int *major, int *minor, int *patch)
1320 if (major) *major = MDB_VERSION_MAJOR;
1321 if (minor) *minor = MDB_VERSION_MINOR;
1322 if (patch) *patch = MDB_VERSION_PATCH;
1323 return MDB_VERSION_STRING;
1326 /** Table of descriptions for LMDB @ref errors */
1327 static char *const mdb_errstr[] = {
1328 "MDB_KEYEXIST: Key/data pair already exists",
1329 "MDB_NOTFOUND: No matching key/data pair found",
1330 "MDB_PAGE_NOTFOUND: Requested page not found",
1331 "MDB_CORRUPTED: Located page was wrong type",
1332 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1333 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1334 "MDB_INVALID: File is not an LMDB file",
1335 "MDB_MAP_FULL: Environment mapsize limit reached",
1336 "MDB_DBS_FULL: Environment maxdbs limit reached",
1337 "MDB_READERS_FULL: Environment maxreaders limit reached",
1338 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1339 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1340 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1341 "MDB_PAGE_FULL: Internal error - page has no more space",
1342 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1343 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1344 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1345 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1346 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1347 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1351 mdb_strerror(int err)
1354 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1355 * This works as long as no function between the call to mdb_strerror
1356 * and the actual use of the message uses more than 4K of stack.
1359 char buf[1024], *ptr = buf;
1363 return ("Successful return: 0");
1365 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1366 i = err - MDB_KEYEXIST;
1367 return mdb_errstr[i];
1371 /* These are the C-runtime error codes we use. The comment indicates
1372 * their numeric value, and the Win32 error they would correspond to
1373 * if the error actually came from a Win32 API. A major mess, we should
1374 * have used LMDB-specific error codes for everything.
1377 case ENOENT: /* 2, FILE_NOT_FOUND */
1378 case EIO: /* 5, ACCESS_DENIED */
1379 case ENOMEM: /* 12, INVALID_ACCESS */
1380 case EACCES: /* 13, INVALID_DATA */
1381 case EBUSY: /* 16, CURRENT_DIRECTORY */
1382 case EINVAL: /* 22, BAD_COMMAND */
1383 case ENOSPC: /* 28, OUT_OF_PAPER */
1384 return strerror(err);
1389 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1390 FORMAT_MESSAGE_IGNORE_INSERTS,
1391 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1394 return strerror(err);
1398 /** assert(3) variant in cursor context */
1399 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1400 /** assert(3) variant in transaction context */
1401 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1402 /** assert(3) variant in environment context */
1403 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1406 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1407 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1410 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1411 const char *func, const char *file, int line)
1414 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1415 file, line, expr_txt, func);
1416 if (env->me_assert_func)
1417 env->me_assert_func(env, buf);
1418 fprintf(stderr, "%s\n", buf);
1422 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1426 /** Return the page number of \b mp which may be sub-page, for debug output */
1428 mdb_dbg_pgno(MDB_page *mp)
1431 COPY_PGNO(ret, mp->mp_pgno);
1435 /** Display a key in hexadecimal and return the address of the result.
1436 * @param[in] key the key to display
1437 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1438 * @return The key in hexadecimal form.
1441 mdb_dkey(MDB_val *key, char *buf)
1444 unsigned char *c = key->mv_data;
1450 if (key->mv_size > DKBUF_MAXKEYSIZE)
1451 return "MDB_MAXKEYSIZE";
1452 /* may want to make this a dynamic check: if the key is mostly
1453 * printable characters, print it as-is instead of converting to hex.
1457 for (i=0; i<key->mv_size; i++)
1458 ptr += sprintf(ptr, "%02x", *c++);
1460 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1466 mdb_leafnode_type(MDB_node *n)
1468 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1469 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1470 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1473 /** Display all the keys in the page. */
1475 mdb_page_list(MDB_page *mp)
1477 pgno_t pgno = mdb_dbg_pgno(mp);
1478 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1480 unsigned int i, nkeys, nsize, total = 0;
1484 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1485 case P_BRANCH: type = "Branch page"; break;
1486 case P_LEAF: type = "Leaf page"; break;
1487 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1488 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1489 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1491 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1492 pgno, mp->mp_pages, state);
1495 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1496 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1499 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1503 nkeys = NUMKEYS(mp);
1504 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1506 for (i=0; i<nkeys; i++) {
1507 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1508 key.mv_size = nsize = mp->mp_pad;
1509 key.mv_data = LEAF2KEY(mp, i, nsize);
1511 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1514 node = NODEPTR(mp, i);
1515 key.mv_size = node->mn_ksize;
1516 key.mv_data = node->mn_data;
1517 nsize = NODESIZE + key.mv_size;
1518 if (IS_BRANCH(mp)) {
1519 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1523 if (F_ISSET(node->mn_flags, F_BIGDATA))
1524 nsize += sizeof(pgno_t);
1526 nsize += NODEDSZ(node);
1528 nsize += sizeof(indx_t);
1529 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1530 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1532 total = EVEN(total);
1534 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1535 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1539 mdb_cursor_chk(MDB_cursor *mc)
1545 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1546 for (i=0; i<mc->mc_top; i++) {
1548 node = NODEPTR(mp, mc->mc_ki[i]);
1549 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1552 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1558 /** Count all the pages in each DB and in the freelist
1559 * and make sure it matches the actual number of pages
1561 * All named DBs must be open for a correct count.
1563 static void mdb_audit(MDB_txn *txn)
1567 MDB_ID freecount, count;
1572 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1573 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1574 freecount += *(MDB_ID *)data.mv_data;
1575 mdb_tassert(txn, rc == MDB_NOTFOUND);
1578 for (i = 0; i<txn->mt_numdbs; i++) {
1580 if (!(txn->mt_dbflags[i] & DB_VALID))
1582 mdb_cursor_init(&mc, txn, i, &mx);
1583 if (txn->mt_dbs[i].md_root == P_INVALID)
1585 count += txn->mt_dbs[i].md_branch_pages +
1586 txn->mt_dbs[i].md_leaf_pages +
1587 txn->mt_dbs[i].md_overflow_pages;
1588 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1589 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1590 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1593 mp = mc.mc_pg[mc.mc_top];
1594 for (j=0; j<NUMKEYS(mp); j++) {
1595 MDB_node *leaf = NODEPTR(mp, j);
1596 if (leaf->mn_flags & F_SUBDATA) {
1598 memcpy(&db, NODEDATA(leaf), sizeof(db));
1599 count += db.md_branch_pages + db.md_leaf_pages +
1600 db.md_overflow_pages;
1604 mdb_tassert(txn, rc == MDB_NOTFOUND);
1607 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1608 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1609 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1615 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1617 return txn->mt_dbxs[dbi].md_cmp(a, b);
1621 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1623 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1624 #if UINT_MAX < SIZE_MAX
1625 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1626 dcmp = mdb_cmp_clong;
1631 /** Allocate memory for a page.
1632 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1635 mdb_page_malloc(MDB_txn *txn, unsigned num)
1637 MDB_env *env = txn->mt_env;
1638 MDB_page *ret = env->me_dpages;
1639 size_t psize = env->me_psize, sz = psize, off;
1640 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1641 * For a single page alloc, we init everything after the page header.
1642 * For multi-page, we init the final page; if the caller needed that
1643 * many pages they will be filling in at least up to the last page.
1647 VGMEMP_ALLOC(env, ret, sz);
1648 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1649 env->me_dpages = ret->mp_next;
1652 psize -= off = PAGEHDRSZ;
1657 if ((ret = malloc(sz)) != NULL) {
1658 VGMEMP_ALLOC(env, ret, sz);
1659 if (!(env->me_flags & MDB_NOMEMINIT)) {
1660 memset((char *)ret + off, 0, psize);
1664 txn->mt_flags |= MDB_TXN_ERROR;
1668 /** Free a single page.
1669 * Saves single pages to a list, for future reuse.
1670 * (This is not used for multi-page overflow pages.)
1673 mdb_page_free(MDB_env *env, MDB_page *mp)
1675 mp->mp_next = env->me_dpages;
1676 VGMEMP_FREE(env, mp);
1677 env->me_dpages = mp;
1680 /** Free a dirty page */
1682 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1684 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1685 mdb_page_free(env, dp);
1687 /* large pages just get freed directly */
1688 VGMEMP_FREE(env, dp);
1693 /** Return all dirty pages to dpage list */
1695 mdb_dlist_free(MDB_txn *txn)
1697 MDB_env *env = txn->mt_env;
1698 MDB_ID2L dl = txn->mt_u.dirty_list;
1699 unsigned i, n = dl[0].mid;
1701 for (i = 1; i <= n; i++) {
1702 mdb_dpage_free(env, dl[i].mptr);
1707 /** Loosen or free a single page.
1708 * Saves single pages to a list for future reuse
1709 * in this same txn. It has been pulled from the freeDB
1710 * and already resides on the dirty list, but has been
1711 * deleted. Use these pages first before pulling again
1714 * If the page wasn't dirtied in this txn, just add it
1715 * to this txn's free list.
1718 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1721 pgno_t pgno = mp->mp_pgno;
1722 MDB_txn *txn = mc->mc_txn;
1724 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1725 if (txn->mt_parent) {
1726 MDB_ID2 *dl = txn->mt_u.dirty_list;
1727 /* If txn has a parent, make sure the page is in our
1731 unsigned x = mdb_mid2l_search(dl, pgno);
1732 if (x <= dl[0].mid && dl[x].mid == pgno) {
1733 if (mp != dl[x].mptr) { /* bad cursor? */
1734 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1735 txn->mt_flags |= MDB_TXN_ERROR;
1736 return MDB_CORRUPTED;
1743 /* no parent txn, so it's just ours */
1748 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1750 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1751 txn->mt_loose_pgs = mp;
1752 txn->mt_loose_count++;
1753 mp->mp_flags |= P_LOOSE;
1755 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1763 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1764 * @param[in] mc A cursor handle for the current operation.
1765 * @param[in] pflags Flags of the pages to update:
1766 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1767 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1768 * @return 0 on success, non-zero on failure.
1771 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1773 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1774 MDB_txn *txn = mc->mc_txn;
1780 int rc = MDB_SUCCESS, level;
1782 /* Mark pages seen by cursors */
1783 if (mc->mc_flags & C_UNTRACK)
1784 mc = NULL; /* will find mc in mt_cursors */
1785 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1786 for (; mc; mc=mc->mc_next) {
1787 if (!(mc->mc_flags & C_INITIALIZED))
1789 for (m3 = mc;; m3 = &mx->mx_cursor) {
1791 for (j=0; j<m3->mc_snum; j++) {
1793 if ((mp->mp_flags & Mask) == pflags)
1794 mp->mp_flags ^= P_KEEP;
1796 mx = m3->mc_xcursor;
1797 /* Proceed to mx if it is at a sub-database */
1798 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1800 if (! (mp && (mp->mp_flags & P_LEAF)))
1802 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1803 if (!(leaf->mn_flags & F_SUBDATA))
1812 /* Mark dirty root pages */
1813 for (i=0; i<txn->mt_numdbs; i++) {
1814 if (txn->mt_dbflags[i] & DB_DIRTY) {
1815 pgno_t pgno = txn->mt_dbs[i].md_root;
1816 if (pgno == P_INVALID)
1818 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1820 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1821 dp->mp_flags ^= P_KEEP;
1829 static int mdb_page_flush(MDB_txn *txn, int keep);
1831 /** Spill pages from the dirty list back to disk.
1832 * This is intended to prevent running into #MDB_TXN_FULL situations,
1833 * but note that they may still occur in a few cases:
1834 * 1) our estimate of the txn size could be too small. Currently this
1835 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1836 * 2) child txns may run out of space if their parents dirtied a
1837 * lot of pages and never spilled them. TODO: we probably should do
1838 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1839 * the parent's dirty_room is below a given threshold.
1841 * Otherwise, if not using nested txns, it is expected that apps will
1842 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1843 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1844 * If the txn never references them again, they can be left alone.
1845 * If the txn only reads them, they can be used without any fuss.
1846 * If the txn writes them again, they can be dirtied immediately without
1847 * going thru all of the work of #mdb_page_touch(). Such references are
1848 * handled by #mdb_page_unspill().
1850 * Also note, we never spill DB root pages, nor pages of active cursors,
1851 * because we'll need these back again soon anyway. And in nested txns,
1852 * we can't spill a page in a child txn if it was already spilled in a
1853 * parent txn. That would alter the parent txns' data even though
1854 * the child hasn't committed yet, and we'd have no way to undo it if
1855 * the child aborted.
1857 * @param[in] m0 cursor A cursor handle identifying the transaction and
1858 * database for which we are checking space.
1859 * @param[in] key For a put operation, the key being stored.
1860 * @param[in] data For a put operation, the data being stored.
1861 * @return 0 on success, non-zero on failure.
1864 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1866 MDB_txn *txn = m0->mc_txn;
1868 MDB_ID2L dl = txn->mt_u.dirty_list;
1869 unsigned int i, j, need;
1872 if (m0->mc_flags & C_SUB)
1875 /* Estimate how much space this op will take */
1876 i = m0->mc_db->md_depth;
1877 /* Named DBs also dirty the main DB */
1878 if (m0->mc_dbi > MAIN_DBI)
1879 i += txn->mt_dbs[MAIN_DBI].md_depth;
1880 /* For puts, roughly factor in the key+data size */
1882 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1883 i += i; /* double it for good measure */
1886 if (txn->mt_dirty_room > i)
1889 if (!txn->mt_spill_pgs) {
1890 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1891 if (!txn->mt_spill_pgs)
1894 /* purge deleted slots */
1895 MDB_IDL sl = txn->mt_spill_pgs;
1896 unsigned int num = sl[0];
1898 for (i=1; i<=num; i++) {
1905 /* Preserve pages which may soon be dirtied again */
1906 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1909 /* Less aggressive spill - we originally spilled the entire dirty list,
1910 * with a few exceptions for cursor pages and DB root pages. But this
1911 * turns out to be a lot of wasted effort because in a large txn many
1912 * of those pages will need to be used again. So now we spill only 1/8th
1913 * of the dirty pages. Testing revealed this to be a good tradeoff,
1914 * better than 1/2, 1/4, or 1/10.
1916 if (need < MDB_IDL_UM_MAX / 8)
1917 need = MDB_IDL_UM_MAX / 8;
1919 /* Save the page IDs of all the pages we're flushing */
1920 /* flush from the tail forward, this saves a lot of shifting later on. */
1921 for (i=dl[0].mid; i && need; i--) {
1922 MDB_ID pn = dl[i].mid << 1;
1924 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1926 /* Can't spill twice, make sure it's not already in a parent's
1929 if (txn->mt_parent) {
1931 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1932 if (tx2->mt_spill_pgs) {
1933 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1934 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1935 dp->mp_flags |= P_KEEP;
1943 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1947 mdb_midl_sort(txn->mt_spill_pgs);
1949 /* Flush the spilled part of dirty list */
1950 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1953 /* Reset any dirty pages we kept that page_flush didn't see */
1954 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1957 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1961 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1963 mdb_find_oldest(MDB_txn *txn)
1966 txnid_t mr, oldest = txn->mt_txnid - 1;
1967 if (txn->mt_env->me_txns) {
1968 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1969 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1980 /** Add a page to the txn's dirty list */
1982 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1985 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1987 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
1988 insert = mdb_mid2l_append;
1990 insert = mdb_mid2l_insert;
1992 mid.mid = mp->mp_pgno;
1994 rc = insert(txn->mt_u.dirty_list, &mid);
1995 mdb_tassert(txn, rc == 0);
1996 txn->mt_dirty_room--;
1999 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2000 * me_pghead and mt_next_pgno.
2002 * If there are free pages available from older transactions, they
2003 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2004 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2005 * and move me_pglast to say which records were consumed. Only this
2006 * function can create me_pghead and move me_pglast/mt_next_pgno.
2007 * @param[in] mc cursor A cursor handle identifying the transaction and
2008 * database for which we are allocating.
2009 * @param[in] num the number of pages to allocate.
2010 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2011 * will always be satisfied by a single contiguous chunk of memory.
2012 * @return 0 on success, non-zero on failure.
2015 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2017 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2018 /* Get at most <Max_retries> more freeDB records once me_pghead
2019 * has enough pages. If not enough, use new pages from the map.
2020 * If <Paranoid> and mc is updating the freeDB, only get new
2021 * records if me_pghead is empty. Then the freelist cannot play
2022 * catch-up with itself by growing while trying to save it.
2024 enum { Paranoid = 1, Max_retries = 500 };
2026 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2028 int rc, retry = num * 60;
2029 MDB_txn *txn = mc->mc_txn;
2030 MDB_env *env = txn->mt_env;
2031 pgno_t pgno, *mop = env->me_pghead;
2032 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2034 txnid_t oldest = 0, last;
2039 /* If there are any loose pages, just use them */
2040 if (num == 1 && txn->mt_loose_pgs) {
2041 np = txn->mt_loose_pgs;
2042 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2043 txn->mt_loose_count--;
2044 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2052 /* If our dirty list is already full, we can't do anything */
2053 if (txn->mt_dirty_room == 0) {
2058 for (op = MDB_FIRST;; op = MDB_NEXT) {
2063 /* Seek a big enough contiguous page range. Prefer
2064 * pages at the tail, just truncating the list.
2070 if (mop[i-n2] == pgno+n2)
2077 if (op == MDB_FIRST) { /* 1st iteration */
2078 /* Prepare to fetch more and coalesce */
2079 last = env->me_pglast;
2080 oldest = env->me_pgoldest;
2081 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2084 key.mv_data = &last; /* will look up last+1 */
2085 key.mv_size = sizeof(last);
2087 if (Paranoid && mc->mc_dbi == FREE_DBI)
2090 if (Paranoid && retry < 0 && mop_len)
2094 /* Do not fetch more if the record will be too recent */
2095 if (oldest <= last) {
2097 oldest = mdb_find_oldest(txn);
2098 env->me_pgoldest = oldest;
2104 rc = mdb_cursor_get(&m2, &key, NULL, op);
2106 if (rc == MDB_NOTFOUND)
2110 last = *(txnid_t*)key.mv_data;
2111 if (oldest <= last) {
2113 oldest = mdb_find_oldest(txn);
2114 env->me_pgoldest = oldest;
2120 np = m2.mc_pg[m2.mc_top];
2121 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2122 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2125 idl = (MDB_ID *) data.mv_data;
2128 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2133 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2135 mop = env->me_pghead;
2137 env->me_pglast = last;
2139 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2140 last, txn->mt_dbs[FREE_DBI].md_root, i));
2142 DPRINTF(("IDL %"Z"u", idl[j]));
2144 /* Merge in descending sorted order */
2145 mdb_midl_xmerge(mop, idl);
2149 /* Use new pages from the map when nothing suitable in the freeDB */
2151 pgno = txn->mt_next_pgno;
2152 if (pgno + num >= env->me_maxpg) {
2153 DPUTS("DB size maxed out");
2159 if (env->me_flags & MDB_WRITEMAP) {
2160 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2162 if (!(np = mdb_page_malloc(txn, num))) {
2168 mop[0] = mop_len -= num;
2169 /* Move any stragglers down */
2170 for (j = i-num; j < mop_len; )
2171 mop[++j] = mop[++i];
2173 txn->mt_next_pgno = pgno + num;
2176 mdb_page_dirty(txn, np);
2182 txn->mt_flags |= MDB_TXN_ERROR;
2186 /** Copy the used portions of a non-overflow page.
2187 * @param[in] dst page to copy into
2188 * @param[in] src page to copy from
2189 * @param[in] psize size of a page
2192 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2194 enum { Align = sizeof(pgno_t) };
2195 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2197 /* If page isn't full, just copy the used portion. Adjust
2198 * alignment so memcpy may copy words instead of bytes.
2200 if ((unused &= -Align) && !IS_LEAF2(src)) {
2201 upper = (upper + PAGEBASE) & -Align;
2202 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2203 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2206 memcpy(dst, src, psize - unused);
2210 /** Pull a page off the txn's spill list, if present.
2211 * If a page being referenced was spilled to disk in this txn, bring
2212 * it back and make it dirty/writable again.
2213 * @param[in] txn the transaction handle.
2214 * @param[in] mp the page being referenced. It must not be dirty.
2215 * @param[out] ret the writable page, if any. ret is unchanged if
2216 * mp wasn't spilled.
2219 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2221 MDB_env *env = txn->mt_env;
2224 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2226 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2227 if (!tx2->mt_spill_pgs)
2229 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2230 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2233 if (txn->mt_dirty_room == 0)
2234 return MDB_TXN_FULL;
2235 if (IS_OVERFLOW(mp))
2239 if (env->me_flags & MDB_WRITEMAP) {
2242 np = mdb_page_malloc(txn, num);
2246 memcpy(np, mp, num * env->me_psize);
2248 mdb_page_copy(np, mp, env->me_psize);
2251 /* If in current txn, this page is no longer spilled.
2252 * If it happens to be the last page, truncate the spill list.
2253 * Otherwise mark it as deleted by setting the LSB.
2255 if (x == txn->mt_spill_pgs[0])
2256 txn->mt_spill_pgs[0]--;
2258 txn->mt_spill_pgs[x] |= 1;
2259 } /* otherwise, if belonging to a parent txn, the
2260 * page remains spilled until child commits
2263 mdb_page_dirty(txn, np);
2264 np->mp_flags |= P_DIRTY;
2272 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2273 * @param[in] mc cursor pointing to the page to be touched
2274 * @return 0 on success, non-zero on failure.
2277 mdb_page_touch(MDB_cursor *mc)
2279 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2280 MDB_txn *txn = mc->mc_txn;
2281 MDB_cursor *m2, *m3;
2285 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2286 if (txn->mt_flags & MDB_TXN_SPILLS) {
2288 rc = mdb_page_unspill(txn, mp, &np);
2294 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2295 (rc = mdb_page_alloc(mc, 1, &np)))
2298 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2299 mp->mp_pgno, pgno));
2300 mdb_cassert(mc, mp->mp_pgno != pgno);
2301 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2302 /* Update the parent page, if any, to point to the new page */
2304 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2305 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2306 SETPGNO(node, pgno);
2308 mc->mc_db->md_root = pgno;
2310 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2311 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2313 /* If txn has a parent, make sure the page is in our
2317 unsigned x = mdb_mid2l_search(dl, pgno);
2318 if (x <= dl[0].mid && dl[x].mid == pgno) {
2319 if (mp != dl[x].mptr) { /* bad cursor? */
2320 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2321 txn->mt_flags |= MDB_TXN_ERROR;
2322 return MDB_CORRUPTED;
2327 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2329 np = mdb_page_malloc(txn, 1);
2334 rc = mdb_mid2l_insert(dl, &mid);
2335 mdb_cassert(mc, rc == 0);
2340 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2342 np->mp_flags |= P_DIRTY;
2345 /* Adjust cursors pointing to mp */
2346 mc->mc_pg[mc->mc_top] = np;
2347 m2 = txn->mt_cursors[mc->mc_dbi];
2348 if (mc->mc_flags & C_SUB) {
2349 for (; m2; m2=m2->mc_next) {
2350 m3 = &m2->mc_xcursor->mx_cursor;
2351 if (m3->mc_snum < mc->mc_snum) continue;
2352 if (m3->mc_pg[mc->mc_top] == mp)
2353 m3->mc_pg[mc->mc_top] = np;
2356 for (; m2; m2=m2->mc_next) {
2357 if (m2->mc_snum < mc->mc_snum) continue;
2358 if (m2->mc_pg[mc->mc_top] == mp) {
2359 m2->mc_pg[mc->mc_top] = np;
2360 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2362 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2364 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2365 if (!(leaf->mn_flags & F_SUBDATA))
2366 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2374 txn->mt_flags |= MDB_TXN_ERROR;
2379 mdb_env_sync(MDB_env *env, int force)
2382 if (env->me_flags & MDB_RDONLY)
2384 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2385 if (env->me_flags & MDB_WRITEMAP) {
2386 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2387 ? MS_ASYNC : MS_SYNC;
2388 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2391 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2395 #ifdef BROKEN_FDATASYNC
2396 if (env->me_flags & MDB_FSYNCONLY) {
2397 if (fsync(env->me_fd))
2401 if (MDB_FDATASYNC(env->me_fd))
2408 /** Back up parent txn's cursors, then grab the originals for tracking */
2410 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2412 MDB_cursor *mc, *bk;
2417 for (i = src->mt_numdbs; --i >= 0; ) {
2418 if ((mc = src->mt_cursors[i]) != NULL) {
2419 size = sizeof(MDB_cursor);
2421 size += sizeof(MDB_xcursor);
2422 for (; mc; mc = bk->mc_next) {
2428 mc->mc_db = &dst->mt_dbs[i];
2429 /* Kill pointers into src - and dst to reduce abuse: The
2430 * user may not use mc until dst ends. Otherwise we'd...
2432 mc->mc_txn = NULL; /* ...set this to dst */
2433 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2434 if ((mx = mc->mc_xcursor) != NULL) {
2435 *(MDB_xcursor *)(bk+1) = *mx;
2436 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2438 mc->mc_next = dst->mt_cursors[i];
2439 dst->mt_cursors[i] = mc;
2446 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2447 * @param[in] txn the transaction handle.
2448 * @param[in] merge true to keep changes to parent cursors, false to revert.
2449 * @return 0 on success, non-zero on failure.
2452 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2454 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2458 for (i = txn->mt_numdbs; --i >= 0; ) {
2459 for (mc = cursors[i]; mc; mc = next) {
2461 if ((bk = mc->mc_backup) != NULL) {
2463 /* Commit changes to parent txn */
2464 mc->mc_next = bk->mc_next;
2465 mc->mc_backup = bk->mc_backup;
2466 mc->mc_txn = bk->mc_txn;
2467 mc->mc_db = bk->mc_db;
2468 mc->mc_dbflag = bk->mc_dbflag;
2469 if ((mx = mc->mc_xcursor) != NULL)
2470 mx->mx_cursor.mc_txn = bk->mc_txn;
2472 /* Abort nested txn */
2474 if ((mx = mc->mc_xcursor) != NULL)
2475 *mx = *(MDB_xcursor *)(bk+1);
2479 /* Only malloced cursors are permanently tracked. */
2487 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2490 mdb_txn_reset0(MDB_txn *txn, const char *act);
2492 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2498 Pidset = F_SETLK, Pidcheck = F_GETLK
2502 /** Set or check a pid lock. Set returns 0 on success.
2503 * Check returns 0 if the process is certainly dead, nonzero if it may
2504 * be alive (the lock exists or an error happened so we do not know).
2506 * On Windows Pidset is a no-op, we merely check for the existence
2507 * of the process with the given pid. On POSIX we use a single byte
2508 * lock on the lockfile, set at an offset equal to the pid.
2511 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2513 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2516 if (op == Pidcheck) {
2517 h = OpenProcess(env->me_pidquery, FALSE, pid);
2518 /* No documented "no such process" code, but other program use this: */
2520 return ErrCode() != ERROR_INVALID_PARAMETER;
2521 /* A process exists until all handles to it close. Has it exited? */
2522 ret = WaitForSingleObject(h, 0) != 0;
2529 struct flock lock_info;
2530 memset(&lock_info, 0, sizeof(lock_info));
2531 lock_info.l_type = F_WRLCK;
2532 lock_info.l_whence = SEEK_SET;
2533 lock_info.l_start = pid;
2534 lock_info.l_len = 1;
2535 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2536 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2538 } else if ((rc = ErrCode()) == EINTR) {
2546 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2547 * @param[in] txn the transaction handle to initialize
2548 * @return 0 on success, non-zero on failure.
2551 mdb_txn_renew0(MDB_txn *txn)
2553 MDB_env *env = txn->mt_env;
2554 MDB_txninfo *ti = env->me_txns;
2556 unsigned int i, nr, flags = txn->mt_flags;
2558 int rc, new_notls = 0;
2560 if ((flags &= MDB_TXN_RDONLY) != 0) {
2562 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2563 txn->mt_txnid = meta->mm_txnid;
2564 txn->mt_u.reader = NULL;
2566 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2567 pthread_getspecific(env->me_txkey);
2569 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2570 return MDB_BAD_RSLOT;
2572 MDB_PID_T pid = env->me_pid;
2573 MDB_THR_T tid = pthread_self();
2574 mdb_mutexref_t rmutex = env->me_rmutex;
2576 if (!env->me_live_reader) {
2577 rc = mdb_reader_pid(env, Pidset, pid);
2580 env->me_live_reader = 1;
2583 if (LOCK_MUTEX(rc, env, rmutex))
2585 nr = ti->mti_numreaders;
2586 for (i=0; i<nr; i++)
2587 if (ti->mti_readers[i].mr_pid == 0)
2589 if (i == env->me_maxreaders) {
2590 UNLOCK_MUTEX(rmutex);
2591 return MDB_READERS_FULL;
2593 r = &ti->mti_readers[i];
2594 /* Claim the reader slot, carefully since other code
2595 * uses the reader table un-mutexed: First reset the
2596 * slot, next publish it in mti_numreaders. After
2597 * that, it is safe for mdb_env_close() to touch it.
2598 * When it will be closed, we can finally claim it.
2601 r->mr_txnid = (txnid_t)-1;
2604 ti->mti_numreaders = ++nr;
2605 env->me_close_readers = nr;
2607 UNLOCK_MUTEX(rmutex);
2609 new_notls = (env->me_flags & MDB_NOTLS);
2610 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2615 do /* LY: Retry on a race, ITS#7970. */
2616 r->mr_txnid = ti->mti_txnid;
2617 while(r->mr_txnid != ti->mti_txnid);
2618 txn->mt_txnid = r->mr_txnid;
2619 txn->mt_u.reader = r;
2620 meta = env->me_metas[txn->mt_txnid & 1];
2622 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2624 /* Not yet touching txn == env->me_txn0, it may be active */
2626 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2628 txn->mt_txnid = ti->mti_txnid;
2629 meta = env->me_metas[txn->mt_txnid & 1];
2631 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2632 txn->mt_txnid = meta->mm_txnid;
2636 if (txn->mt_txnid == mdb_debug_start)
2639 txn->mt_child = NULL;
2640 txn->mt_loose_pgs = NULL;
2641 txn->mt_loose_count = 0;
2642 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2643 txn->mt_u.dirty_list = env->me_dirty_list;
2644 txn->mt_u.dirty_list[0].mid = 0;
2645 txn->mt_free_pgs = env->me_free_pgs;
2646 txn->mt_free_pgs[0] = 0;
2647 txn->mt_spill_pgs = NULL;
2649 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2652 /* Copy the DB info and flags */
2653 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2655 /* Moved to here to avoid a data race in read TXNs */
2656 txn->mt_next_pgno = meta->mm_last_pg+1;
2658 txn->mt_flags = flags;
2661 txn->mt_numdbs = env->me_numdbs;
2662 for (i=2; i<txn->mt_numdbs; i++) {
2663 x = env->me_dbflags[i];
2664 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2665 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2667 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2669 if (env->me_maxpg < txn->mt_next_pgno) {
2670 mdb_txn_reset0(txn, "renew0-mapfail");
2672 txn->mt_u.reader->mr_pid = 0;
2673 txn->mt_u.reader = NULL;
2675 return MDB_MAP_RESIZED;
2682 mdb_txn_renew(MDB_txn *txn)
2686 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2689 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2690 DPUTS("environment had fatal error, must shutdown!");
2694 rc = mdb_txn_renew0(txn);
2695 if (rc == MDB_SUCCESS) {
2696 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2697 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2698 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2704 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2708 int rc, size, tsize;
2710 flags &= MDB_TXN_BEGIN_FLAGS;
2711 flags |= env->me_flags & MDB_WRITEMAP;
2713 if (env->me_flags & MDB_FATAL_ERROR) {
2714 DPUTS("environment had fatal error, must shutdown!");
2717 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2721 /* Nested transactions: Max 1 child, write txns only, no writemap */
2722 flags |= parent->mt_flags;
2723 if (parent->mt_child ||
2724 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2726 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2728 /* Child txns save MDB_pgstate and use own copy of cursors */
2729 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2730 size += tsize = sizeof(MDB_ntxn);
2731 } else if (flags & MDB_RDONLY) {
2732 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2733 size += tsize = sizeof(MDB_txn);
2735 /* Reuse preallocated write txn. However, do not touch it until
2736 * mdb_txn_renew0() succeeds, since it currently may be active.
2741 if ((txn = calloc(1, size)) == NULL) {
2742 DPRINTF(("calloc: %s", strerror(errno)));
2745 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2746 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2747 txn->mt_flags = flags;
2752 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2753 txn->mt_dbiseqs = parent->mt_dbiseqs;
2754 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2755 if (!txn->mt_u.dirty_list ||
2756 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2758 free(txn->mt_u.dirty_list);
2762 txn->mt_txnid = parent->mt_txnid;
2763 txn->mt_dirty_room = parent->mt_dirty_room;
2764 txn->mt_u.dirty_list[0].mid = 0;
2765 txn->mt_spill_pgs = NULL;
2766 txn->mt_next_pgno = parent->mt_next_pgno;
2767 parent->mt_child = txn;
2768 txn->mt_parent = parent;
2769 txn->mt_numdbs = parent->mt_numdbs;
2770 txn->mt_dbxs = parent->mt_dbxs;
2771 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2772 /* Copy parent's mt_dbflags, but clear DB_NEW */
2773 for (i=0; i<txn->mt_numdbs; i++)
2774 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2776 ntxn = (MDB_ntxn *)txn;
2777 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2778 if (env->me_pghead) {
2779 size = MDB_IDL_SIZEOF(env->me_pghead);
2780 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2782 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2787 rc = mdb_cursor_shadow(parent, txn);
2789 mdb_txn_reset0(txn, "beginchild-fail");
2790 } else { /* MDB_RDONLY */
2791 txn->mt_dbiseqs = env->me_dbiseqs;
2793 rc = mdb_txn_renew0(txn);
2796 if (txn != env->me_txn0)
2799 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2801 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2802 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2803 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2810 mdb_txn_env(MDB_txn *txn)
2812 if(!txn) return NULL;
2817 mdb_txn_id(MDB_txn *txn)
2820 return txn->mt_txnid;
2823 /** Export or close DBI handles opened in this txn. */
2825 mdb_dbis_update(MDB_txn *txn, int keep)
2828 MDB_dbi n = txn->mt_numdbs;
2829 MDB_env *env = txn->mt_env;
2830 unsigned char *tdbflags = txn->mt_dbflags;
2832 for (i = n; --i >= 2;) {
2833 if (tdbflags[i] & DB_NEW) {
2835 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2837 char *ptr = env->me_dbxs[i].md_name.mv_data;
2839 env->me_dbxs[i].md_name.mv_data = NULL;
2840 env->me_dbxs[i].md_name.mv_size = 0;
2841 env->me_dbflags[i] = 0;
2842 env->me_dbiseqs[i]++;
2848 if (keep && env->me_numdbs < n)
2852 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2853 * May be called twice for readonly txns: First reset it, then abort.
2854 * @param[in] txn the transaction handle to reset
2855 * @param[in] act why the transaction is being reset
2858 mdb_txn_reset0(MDB_txn *txn, const char *act)
2860 MDB_env *env = txn->mt_env;
2862 /* Close any DBI handles opened in this txn */
2863 mdb_dbis_update(txn, 0);
2865 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2866 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2867 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2869 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2870 if (txn->mt_u.reader) {
2871 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2872 if (!(env->me_flags & MDB_NOTLS))
2873 txn->mt_u.reader = NULL; /* txn does not own reader */
2875 txn->mt_numdbs = 0; /* close nothing if called again */
2876 txn->mt_dbxs = NULL; /* mark txn as reset */
2878 pgno_t *pghead = env->me_pghead;
2880 mdb_cursors_close(txn, 0);
2881 if (!(env->me_flags & MDB_WRITEMAP)) {
2882 mdb_dlist_free(txn);
2885 if (!txn->mt_parent) {
2886 mdb_midl_shrink(&txn->mt_free_pgs);
2887 env->me_free_pgs = txn->mt_free_pgs;
2889 env->me_pghead = NULL;
2893 /* The writer mutex was locked in mdb_txn_begin. */
2895 UNLOCK_MUTEX(env->me_wmutex);
2897 txn->mt_parent->mt_child = NULL;
2898 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2899 mdb_midl_free(txn->mt_free_pgs);
2900 mdb_midl_free(txn->mt_spill_pgs);
2901 free(txn->mt_u.dirty_list);
2904 mdb_midl_free(pghead);
2909 mdb_txn_reset(MDB_txn *txn)
2914 /* This call is only valid for read-only txns */
2915 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2918 mdb_txn_reset0(txn, "reset");
2922 mdb_txn_abort(MDB_txn *txn)
2928 mdb_txn_abort(txn->mt_child);
2930 mdb_txn_reset0(txn, "abort");
2931 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2932 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2933 txn->mt_u.reader->mr_pid = 0;
2935 if (txn != txn->mt_env->me_txn0)
2939 /** Save the freelist as of this transaction to the freeDB.
2940 * This changes the freelist. Keep trying until it stabilizes.
2943 mdb_freelist_save(MDB_txn *txn)
2945 /* env->me_pghead[] can grow and shrink during this call.
2946 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2947 * Page numbers cannot disappear from txn->mt_free_pgs[].
2950 MDB_env *env = txn->mt_env;
2951 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2952 txnid_t pglast = 0, head_id = 0;
2953 pgno_t freecnt = 0, *free_pgs, *mop;
2954 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2956 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2958 if (env->me_pghead) {
2959 /* Make sure first page of freeDB is touched and on freelist */
2960 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2961 if (rc && rc != MDB_NOTFOUND)
2965 if (!env->me_pghead && txn->mt_loose_pgs) {
2966 /* Put loose page numbers in mt_free_pgs, since
2967 * we may be unable to return them to me_pghead.
2969 MDB_page *mp = txn->mt_loose_pgs;
2970 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2972 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2973 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2974 txn->mt_loose_pgs = NULL;
2975 txn->mt_loose_count = 0;
2978 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2979 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2980 ? SSIZE_MAX : maxfree_1pg;
2983 /* Come back here after each Put() in case freelist changed */
2988 /* If using records from freeDB which we have not yet
2989 * deleted, delete them and any we reserved for me_pghead.
2991 while (pglast < env->me_pglast) {
2992 rc = mdb_cursor_first(&mc, &key, NULL);
2995 pglast = head_id = *(txnid_t *)key.mv_data;
2996 total_room = head_room = 0;
2997 mdb_tassert(txn, pglast <= env->me_pglast);
2998 rc = mdb_cursor_del(&mc, 0);
3003 /* Save the IDL of pages freed by this txn, to a single record */
3004 if (freecnt < txn->mt_free_pgs[0]) {
3006 /* Make sure last page of freeDB is touched and on freelist */
3007 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3008 if (rc && rc != MDB_NOTFOUND)
3011 free_pgs = txn->mt_free_pgs;
3012 /* Write to last page of freeDB */
3013 key.mv_size = sizeof(txn->mt_txnid);
3014 key.mv_data = &txn->mt_txnid;
3016 freecnt = free_pgs[0];
3017 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3018 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3021 /* Retry if mt_free_pgs[] grew during the Put() */
3022 free_pgs = txn->mt_free_pgs;
3023 } while (freecnt < free_pgs[0]);
3024 mdb_midl_sort(free_pgs);
3025 memcpy(data.mv_data, free_pgs, data.mv_size);
3028 unsigned int i = free_pgs[0];
3029 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3030 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3032 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3038 mop = env->me_pghead;
3039 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3041 /* Reserve records for me_pghead[]. Split it if multi-page,
3042 * to avoid searching freeDB for a page range. Use keys in
3043 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3045 if (total_room >= mop_len) {
3046 if (total_room == mop_len || --more < 0)
3048 } else if (head_room >= maxfree_1pg && head_id > 1) {
3049 /* Keep current record (overflow page), add a new one */
3053 /* (Re)write {key = head_id, IDL length = head_room} */
3054 total_room -= head_room;
3055 head_room = mop_len - total_room;
3056 if (head_room > maxfree_1pg && head_id > 1) {
3057 /* Overflow multi-page for part of me_pghead */
3058 head_room /= head_id; /* amortize page sizes */
3059 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3060 } else if (head_room < 0) {
3061 /* Rare case, not bothering to delete this record */
3064 key.mv_size = sizeof(head_id);
3065 key.mv_data = &head_id;
3066 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3067 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3070 /* IDL is initially empty, zero out at least the length */
3071 pgs = (pgno_t *)data.mv_data;
3072 j = head_room > clean_limit ? head_room : 0;
3076 total_room += head_room;
3079 /* Return loose page numbers to me_pghead, though usually none are
3080 * left at this point. The pages themselves remain in dirty_list.
3082 if (txn->mt_loose_pgs) {
3083 MDB_page *mp = txn->mt_loose_pgs;
3084 unsigned count = txn->mt_loose_count;
3086 /* Room for loose pages + temp IDL with same */
3087 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3089 mop = env->me_pghead;
3090 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3091 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3092 loose[ ++count ] = mp->mp_pgno;
3094 mdb_midl_sort(loose);
3095 mdb_midl_xmerge(mop, loose);
3096 txn->mt_loose_pgs = NULL;
3097 txn->mt_loose_count = 0;
3101 /* Fill in the reserved me_pghead records */
3107 rc = mdb_cursor_first(&mc, &key, &data);
3108 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3109 txnid_t id = *(txnid_t *)key.mv_data;
3110 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3113 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3115 if (len > mop_len) {
3117 data.mv_size = (len + 1) * sizeof(MDB_ID);
3119 data.mv_data = mop -= len;
3122 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3124 if (rc || !(mop_len -= len))
3131 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3132 * @param[in] txn the transaction that's being committed
3133 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3134 * @return 0 on success, non-zero on failure.
3137 mdb_page_flush(MDB_txn *txn, int keep)
3139 MDB_env *env = txn->mt_env;
3140 MDB_ID2L dl = txn->mt_u.dirty_list;
3141 unsigned psize = env->me_psize, j;
3142 int i, pagecount = dl[0].mid, rc;
3143 size_t size = 0, pos = 0;
3145 MDB_page *dp = NULL;
3149 struct iovec iov[MDB_COMMIT_PAGES];
3150 ssize_t wpos = 0, wsize = 0, wres;
3151 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3157 if (env->me_flags & MDB_WRITEMAP) {
3158 /* Clear dirty flags */
3159 while (++i <= pagecount) {
3161 /* Don't flush this page yet */
3162 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3163 dp->mp_flags &= ~P_KEEP;
3167 dp->mp_flags &= ~P_DIRTY;
3172 /* Write the pages */
3174 if (++i <= pagecount) {
3176 /* Don't flush this page yet */
3177 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3178 dp->mp_flags &= ~P_KEEP;
3183 /* clear dirty flag */
3184 dp->mp_flags &= ~P_DIRTY;
3187 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3192 /* Windows actually supports scatter/gather I/O, but only on
3193 * unbuffered file handles. Since we're relying on the OS page
3194 * cache for all our data, that's self-defeating. So we just
3195 * write pages one at a time. We use the ov structure to set
3196 * the write offset, to at least save the overhead of a Seek
3199 DPRINTF(("committing page %"Z"u", pgno));
3200 memset(&ov, 0, sizeof(ov));
3201 ov.Offset = pos & 0xffffffff;
3202 ov.OffsetHigh = pos >> 16 >> 16;
3203 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3205 DPRINTF(("WriteFile: %d", rc));
3209 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3210 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3213 /* Write previous page(s) */
3214 #ifdef MDB_USE_PWRITEV
3215 wres = pwritev(env->me_fd, iov, n, wpos);
3218 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3221 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3225 DPRINTF(("lseek: %s", strerror(rc)));
3228 wres = writev(env->me_fd, iov, n);
3231 if (wres != wsize) {
3236 DPRINTF(("Write error: %s", strerror(rc)));
3238 rc = EIO; /* TODO: Use which error code? */
3239 DPUTS("short write, filesystem full?");
3250 DPRINTF(("committing page %"Z"u", pgno));
3251 next_pos = pos + size;
3252 iov[n].iov_len = size;
3253 iov[n].iov_base = (char *)dp;
3259 /* MIPS has cache coherency issues, this is a no-op everywhere else
3260 * Note: for any size >= on-chip cache size, entire on-chip cache is
3263 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3265 for (i = keep; ++i <= pagecount; ) {
3267 /* This is a page we skipped above */
3270 dl[j].mid = dp->mp_pgno;
3273 mdb_dpage_free(env, dp);
3278 txn->mt_dirty_room += i - j;
3284 mdb_txn_commit(MDB_txn *txn)
3293 if (txn->mt_child) {
3294 rc = mdb_txn_commit(txn->mt_child);
3301 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3302 mdb_dbis_update(txn, 1);
3303 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3308 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3309 DPUTS("error flag is set, can't commit");
3311 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3316 if (txn->mt_parent) {
3317 MDB_txn *parent = txn->mt_parent;
3321 unsigned x, y, len, ps_len;
3323 /* Append our free list to parent's */
3324 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3327 mdb_midl_free(txn->mt_free_pgs);
3328 /* Failures after this must either undo the changes
3329 * to the parent or set MDB_TXN_ERROR in the parent.
3332 parent->mt_next_pgno = txn->mt_next_pgno;
3333 parent->mt_flags = txn->mt_flags;
3335 /* Merge our cursors into parent's and close them */
3336 mdb_cursors_close(txn, 1);
3338 /* Update parent's DB table. */
3339 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3340 parent->mt_numdbs = txn->mt_numdbs;
3341 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3342 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3343 for (i=2; i<txn->mt_numdbs; i++) {
3344 /* preserve parent's DB_NEW status */
3345 x = parent->mt_dbflags[i] & DB_NEW;
3346 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3349 dst = parent->mt_u.dirty_list;
3350 src = txn->mt_u.dirty_list;
3351 /* Remove anything in our dirty list from parent's spill list */
3352 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3354 pspill[0] = (pgno_t)-1;
3355 /* Mark our dirty pages as deleted in parent spill list */
3356 for (i=0, len=src[0].mid; ++i <= len; ) {
3357 MDB_ID pn = src[i].mid << 1;
3358 while (pn > pspill[x])
3360 if (pn == pspill[x]) {
3365 /* Squash deleted pagenums if we deleted any */
3366 for (x=y; ++x <= ps_len; )
3367 if (!(pspill[x] & 1))
3368 pspill[++y] = pspill[x];
3372 /* Find len = length of merging our dirty list with parent's */
3374 dst[0].mid = 0; /* simplify loops */
3375 if (parent->mt_parent) {
3376 len = x + src[0].mid;
3377 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3378 for (i = x; y && i; y--) {
3379 pgno_t yp = src[y].mid;
3380 while (yp < dst[i].mid)
3382 if (yp == dst[i].mid) {
3387 } else { /* Simplify the above for single-ancestor case */
3388 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3390 /* Merge our dirty list with parent's */
3392 for (i = len; y; dst[i--] = src[y--]) {
3393 pgno_t yp = src[y].mid;
3394 while (yp < dst[x].mid)
3395 dst[i--] = dst[x--];
3396 if (yp == dst[x].mid)
3397 free(dst[x--].mptr);
3399 mdb_tassert(txn, i == x);
3401 free(txn->mt_u.dirty_list);
3402 parent->mt_dirty_room = txn->mt_dirty_room;
3403 if (txn->mt_spill_pgs) {
3404 if (parent->mt_spill_pgs) {
3405 /* TODO: Prevent failure here, so parent does not fail */
3406 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3408 parent->mt_flags |= MDB_TXN_ERROR;
3409 mdb_midl_free(txn->mt_spill_pgs);
3410 mdb_midl_sort(parent->mt_spill_pgs);
3412 parent->mt_spill_pgs = txn->mt_spill_pgs;
3416 /* Append our loose page list to parent's */
3417 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3419 *lp = txn->mt_loose_pgs;
3420 parent->mt_loose_count += txn->mt_loose_count;
3422 parent->mt_child = NULL;
3423 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3428 if (txn != env->me_txn) {
3429 DPUTS("attempt to commit unknown transaction");
3434 mdb_cursors_close(txn, 0);
3436 if (!txn->mt_u.dirty_list[0].mid &&
3437 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3440 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3441 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3443 /* Update DB root pointers */
3444 if (txn->mt_numdbs > 2) {
3448 data.mv_size = sizeof(MDB_db);
3450 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3451 for (i = 2; i < txn->mt_numdbs; i++) {
3452 if (txn->mt_dbflags[i] & DB_DIRTY) {
3453 if (TXN_DBI_CHANGED(txn, i)) {
3457 data.mv_data = &txn->mt_dbs[i];
3458 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3466 rc = mdb_freelist_save(txn);
3470 mdb_midl_free(env->me_pghead);
3471 env->me_pghead = NULL;
3472 mdb_midl_shrink(&txn->mt_free_pgs);
3473 env->me_free_pgs = txn->mt_free_pgs;
3479 if ((rc = mdb_page_flush(txn, 0)) ||
3480 (rc = mdb_env_sync(env, 0)) ||
3481 (rc = mdb_env_write_meta(txn)))
3484 /* Free P_LOOSE pages left behind in dirty_list */
3485 if (!(env->me_flags & MDB_WRITEMAP))
3486 mdb_dlist_free(txn);
3491 mdb_dbis_update(txn, 1);
3494 UNLOCK_MUTEX(env->me_wmutex);
3495 if (txn != env->me_txn0)
3505 /** Read the environment parameters of a DB environment before
3506 * mapping it into memory.
3507 * @param[in] env the environment handle
3508 * @param[out] meta address of where to store the meta information
3509 * @return 0 on success, non-zero on failure.
3512 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3518 enum { Size = sizeof(pbuf) };
3520 /* We don't know the page size yet, so use a minimum value.
3521 * Read both meta pages so we can use the latest one.
3524 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3528 memset(&ov, 0, sizeof(ov));
3530 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3531 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3534 rc = pread(env->me_fd, &pbuf, Size, off);
3537 if (rc == 0 && off == 0)
3539 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3540 DPRINTF(("read: %s", mdb_strerror(rc)));
3544 p = (MDB_page *)&pbuf;
3546 if (!F_ISSET(p->mp_flags, P_META)) {
3547 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3552 if (m->mm_magic != MDB_MAGIC) {
3553 DPUTS("meta has invalid magic");
3557 if (m->mm_version != MDB_DATA_VERSION) {
3558 DPRINTF(("database is version %u, expected version %u",
3559 m->mm_version, MDB_DATA_VERSION));
3560 return MDB_VERSION_MISMATCH;
3563 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3569 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3571 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3573 meta->mm_magic = MDB_MAGIC;
3574 meta->mm_version = MDB_DATA_VERSION;
3575 meta->mm_mapsize = env->me_mapsize;
3576 meta->mm_psize = env->me_psize;
3577 meta->mm_last_pg = 1;
3578 meta->mm_flags = env->me_flags & 0xffff;
3579 meta->mm_flags |= MDB_INTEGERKEY;
3580 meta->mm_dbs[0].md_root = P_INVALID;
3581 meta->mm_dbs[1].md_root = P_INVALID;
3584 /** Write the environment parameters of a freshly created DB environment.
3585 * @param[in] env the environment handle
3586 * @param[in] meta the #MDB_meta to write
3587 * @return 0 on success, non-zero on failure.
3590 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3598 memset(&ov, 0, sizeof(ov));
3599 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3601 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3604 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3605 len = pwrite(fd, ptr, size, pos); \
3606 if (len == -1 && ErrCode() == EINTR) continue; \
3607 rc = (len >= 0); break; } while(1)
3610 DPUTS("writing new meta page");
3612 psize = env->me_psize;
3614 p = calloc(2, psize);
3619 p->mp_flags = P_META;
3620 *(MDB_meta *)METADATA(p) = *meta;
3622 q = (MDB_page *)((char *)p + psize);
3624 q->mp_flags = P_META;
3625 *(MDB_meta *)METADATA(q) = *meta;
3627 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3630 else if ((unsigned) len == psize * 2)
3638 /** Update the environment info to commit a transaction.
3639 * @param[in] txn the transaction that's being committed
3640 * @return 0 on success, non-zero on failure.
3643 mdb_env_write_meta(MDB_txn *txn)
3646 MDB_meta meta, metab, *mp;
3650 int rc, len, toggle;
3659 toggle = txn->mt_txnid & 1;
3660 DPRINTF(("writing meta page %d for root page %"Z"u",
3661 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3664 flags = env->me_flags;
3665 mp = env->me_metas[toggle];
3666 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3667 /* Persist any increases of mapsize config */
3668 if (mapsize < env->me_mapsize)
3669 mapsize = env->me_mapsize;
3671 if (flags & MDB_WRITEMAP) {
3672 mp->mm_mapsize = mapsize;
3673 mp->mm_dbs[0] = txn->mt_dbs[0];
3674 mp->mm_dbs[1] = txn->mt_dbs[1];
3675 mp->mm_last_pg = txn->mt_next_pgno - 1;
3676 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3677 !(defined(__i386__) || defined(__x86_64__))
3678 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3679 __sync_synchronize();
3681 mp->mm_txnid = txn->mt_txnid;
3682 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3683 unsigned meta_size = env->me_psize;
3684 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3687 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3688 if (meta_size < env->me_os_psize)
3689 meta_size += meta_size;
3694 if (MDB_MSYNC(ptr, meta_size, rc)) {
3701 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3702 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3704 meta.mm_mapsize = mapsize;
3705 meta.mm_dbs[0] = txn->mt_dbs[0];
3706 meta.mm_dbs[1] = txn->mt_dbs[1];
3707 meta.mm_last_pg = txn->mt_next_pgno - 1;
3708 meta.mm_txnid = txn->mt_txnid;
3710 off = offsetof(MDB_meta, mm_mapsize);
3711 ptr = (char *)&meta + off;
3712 len = sizeof(MDB_meta) - off;
3714 off += env->me_psize;
3717 /* Write to the SYNC fd */
3718 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3721 memset(&ov, 0, sizeof(ov));
3723 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3728 rc = pwrite(mfd, ptr, len, off);
3731 rc = rc < 0 ? ErrCode() : EIO;
3736 DPUTS("write failed, disk error?");
3737 /* On a failure, the pagecache still contains the new data.
3738 * Write some old data back, to prevent it from being used.
3739 * Use the non-SYNC fd; we know it will fail anyway.
3741 meta.mm_last_pg = metab.mm_last_pg;
3742 meta.mm_txnid = metab.mm_txnid;
3744 memset(&ov, 0, sizeof(ov));
3746 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3748 r2 = pwrite(env->me_fd, ptr, len, off);
3749 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3752 env->me_flags |= MDB_FATAL_ERROR;
3755 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3756 CACHEFLUSH(env->me_map + off, len, DCACHE);
3758 /* Memory ordering issues are irrelevant; since the entire writer
3759 * is wrapped by wmutex, all of these changes will become visible
3760 * after the wmutex is unlocked. Since the DB is multi-version,
3761 * readers will get consistent data regardless of how fresh or
3762 * how stale their view of these values is.
3765 env->me_txns->mti_txnid = txn->mt_txnid;
3770 /** Check both meta pages to see which one is newer.
3771 * @param[in] env the environment handle
3772 * @return meta toggle (0 or 1).
3775 mdb_env_pick_meta(const MDB_env *env)
3777 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3781 mdb_env_create(MDB_env **env)
3785 e = calloc(1, sizeof(MDB_env));
3789 e->me_maxreaders = DEFAULT_READERS;
3790 e->me_maxdbs = e->me_numdbs = 2;
3791 e->me_fd = INVALID_HANDLE_VALUE;
3792 e->me_lfd = INVALID_HANDLE_VALUE;
3793 e->me_mfd = INVALID_HANDLE_VALUE;
3794 #ifdef MDB_USE_POSIX_SEM
3795 e->me_rmutex = SEM_FAILED;
3796 e->me_wmutex = SEM_FAILED;
3798 e->me_pid = getpid();
3799 GET_PAGESIZE(e->me_os_psize);
3800 VGMEMP_CREATE(e,0,0);
3806 mdb_env_map(MDB_env *env, void *addr)
3809 unsigned int flags = env->me_flags;
3813 LONG sizelo, sizehi;
3816 if (flags & MDB_RDONLY) {
3817 /* Don't set explicit map size, use whatever exists */
3822 msize = env->me_mapsize;
3823 sizelo = msize & 0xffffffff;
3824 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3826 /* Windows won't create mappings for zero length files.
3827 * and won't map more than the file size.
3828 * Just set the maxsize right now.
3830 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3831 || !SetEndOfFile(env->me_fd)
3832 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3836 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3837 PAGE_READWRITE : PAGE_READONLY,
3838 sizehi, sizelo, NULL);
3841 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3842 FILE_MAP_WRITE : FILE_MAP_READ,
3844 rc = env->me_map ? 0 : ErrCode();
3849 int prot = PROT_READ;
3850 if (flags & MDB_WRITEMAP) {
3852 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3855 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3857 if (env->me_map == MAP_FAILED) {
3862 if (flags & MDB_NORDAHEAD) {
3863 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3865 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3867 #ifdef POSIX_MADV_RANDOM
3868 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3869 #endif /* POSIX_MADV_RANDOM */
3870 #endif /* MADV_RANDOM */
3874 /* Can happen because the address argument to mmap() is just a
3875 * hint. mmap() can pick another, e.g. if the range is in use.
3876 * The MAP_FIXED flag would prevent that, but then mmap could
3877 * instead unmap existing pages to make room for the new map.
3879 if (addr && env->me_map != addr)
3880 return EBUSY; /* TODO: Make a new MDB_* error code? */
3882 p = (MDB_page *)env->me_map;
3883 env->me_metas[0] = METADATA(p);
3884 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3890 mdb_env_set_mapsize(MDB_env *env, size_t size)
3892 /* If env is already open, caller is responsible for making
3893 * sure there are no active txns.
3901 meta = env->me_metas[mdb_env_pick_meta(env)];
3903 size = meta->mm_mapsize;
3905 /* Silently round up to minimum if the size is too small */
3906 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3910 munmap(env->me_map, env->me_mapsize);
3911 env->me_mapsize = size;
3912 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3913 rc = mdb_env_map(env, old);
3917 env->me_mapsize = size;
3919 env->me_maxpg = env->me_mapsize / env->me_psize;
3924 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3928 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3933 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3935 if (env->me_map || readers < 1)
3937 env->me_maxreaders = readers;
3942 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3944 if (!env || !readers)
3946 *readers = env->me_maxreaders;
3951 mdb_fsize(HANDLE fd, size_t *size)
3954 LARGE_INTEGER fsize;
3956 if (!GetFileSizeEx(fd, &fsize))
3959 *size = fsize.QuadPart;
3971 #ifdef BROKEN_FDATASYNC
3972 #include <sys/utsname.h>
3973 #include <sys/vfs.h>
3976 /** Further setup required for opening an LMDB environment
3979 mdb_env_open2(MDB_env *env)
3981 unsigned int flags = env->me_flags;
3982 int i, newenv = 0, rc;
3986 /* See if we should use QueryLimited */
3988 if ((rc & 0xff) > 5)
3989 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3991 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3994 #ifdef BROKEN_FDATASYNC
3995 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3996 * https://lkml.org/lkml/2012/9/3/83
3997 * Kernels after 3.6-rc6 are known good.
3998 * https://lkml.org/lkml/2012/9/10/556
3999 * See if the DB is on ext3/ext4, then check for new enough kernel
4000 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4005 fstatfs(env->me_fd, &st);
4006 while (st.f_type == 0xEF53) {
4010 if (uts.release[0] < '3') {
4011 if (!strncmp(uts.release, "2.6.32.", 7)) {
4012 i = atoi(uts.release+7);
4014 break; /* 2.6.32.60 and newer is OK */
4015 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4016 i = atoi(uts.release+7);
4018 break; /* 2.6.34.15 and newer is OK */
4020 } else if (uts.release[0] == '3') {
4021 i = atoi(uts.release+2);
4023 break; /* 3.6 and newer is OK */
4025 i = atoi(uts.release+4);
4027 break; /* 3.5.4 and newer is OK */
4028 } else if (i == 2) {
4029 i = atoi(uts.release+4);
4031 break; /* 3.2.30 and newer is OK */
4033 } else { /* 4.x and newer is OK */
4036 env->me_flags |= MDB_FSYNCONLY;
4042 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4045 DPUTS("new mdbenv");
4047 env->me_psize = env->me_os_psize;
4048 if (env->me_psize > MAX_PAGESIZE)
4049 env->me_psize = MAX_PAGESIZE;
4050 memset(&meta, 0, sizeof(meta));
4051 mdb_env_init_meta0(env, &meta);
4052 meta.mm_mapsize = DEFAULT_MAPSIZE;
4054 env->me_psize = meta.mm_psize;
4057 /* Was a mapsize configured? */
4058 if (!env->me_mapsize) {
4059 env->me_mapsize = meta.mm_mapsize;
4062 /* Make sure mapsize >= committed data size. Even when using
4063 * mm_mapsize, which could be broken in old files (ITS#7789).
4065 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4066 if (env->me_mapsize < minsize)
4067 env->me_mapsize = minsize;
4069 meta.mm_mapsize = env->me_mapsize;
4071 if (newenv && !(flags & MDB_FIXEDMAP)) {
4072 /* mdb_env_map() may grow the datafile. Write the metapages
4073 * first, so the file will be valid if initialization fails.
4074 * Except with FIXEDMAP, since we do not yet know mm_address.
4075 * We could fill in mm_address later, but then a different
4076 * program might end up doing that - one with a memory layout
4077 * and map address which does not suit the main program.
4079 rc = mdb_env_init_meta(env, &meta);
4085 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4090 if (flags & MDB_FIXEDMAP)
4091 meta.mm_address = env->me_map;
4092 i = mdb_env_init_meta(env, &meta);
4093 if (i != MDB_SUCCESS) {
4098 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4099 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4101 #if !(MDB_MAXKEYSIZE)
4102 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4104 env->me_maxpg = env->me_mapsize / env->me_psize;
4108 int toggle = mdb_env_pick_meta(env);
4109 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4111 DPRINTF(("opened database version %u, pagesize %u",
4112 env->me_metas[0]->mm_version, env->me_psize));
4113 DPRINTF(("using meta page %d", toggle));
4114 DPRINTF(("depth: %u", db->md_depth));
4115 DPRINTF(("entries: %"Z"u", db->md_entries));
4116 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4117 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4118 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4119 DPRINTF(("root: %"Z"u", db->md_root));
4127 /** Release a reader thread's slot in the reader lock table.
4128 * This function is called automatically when a thread exits.
4129 * @param[in] ptr This points to the slot in the reader lock table.
4132 mdb_env_reader_dest(void *ptr)
4134 MDB_reader *reader = ptr;
4140 /** Junk for arranging thread-specific callbacks on Windows. This is
4141 * necessarily platform and compiler-specific. Windows supports up
4142 * to 1088 keys. Let's assume nobody opens more than 64 environments
4143 * in a single process, for now. They can override this if needed.
4145 #ifndef MAX_TLS_KEYS
4146 #define MAX_TLS_KEYS 64
4148 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4149 static int mdb_tls_nkeys;
4151 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4155 case DLL_PROCESS_ATTACH: break;
4156 case DLL_THREAD_ATTACH: break;
4157 case DLL_THREAD_DETACH:
4158 for (i=0; i<mdb_tls_nkeys; i++) {
4159 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4161 mdb_env_reader_dest(r);
4165 case DLL_PROCESS_DETACH: break;
4170 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4172 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4176 /* Force some symbol references.
4177 * _tls_used forces the linker to create the TLS directory if not already done
4178 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4180 #pragma comment(linker, "/INCLUDE:_tls_used")
4181 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4182 #pragma const_seg(".CRT$XLB")
4183 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4184 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4187 #pragma comment(linker, "/INCLUDE:__tls_used")
4188 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4189 #pragma data_seg(".CRT$XLB")
4190 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4192 #endif /* WIN 32/64 */
4193 #endif /* !__GNUC__ */
4196 /** Downgrade the exclusive lock on the region back to shared */
4198 mdb_env_share_locks(MDB_env *env, int *excl)
4200 int rc = 0, toggle = mdb_env_pick_meta(env);
4202 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4207 /* First acquire a shared lock. The Unlock will
4208 * then release the existing exclusive lock.
4210 memset(&ov, 0, sizeof(ov));
4211 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4214 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4220 struct flock lock_info;
4221 /* The shared lock replaces the existing lock */
4222 memset((void *)&lock_info, 0, sizeof(lock_info));
4223 lock_info.l_type = F_RDLCK;
4224 lock_info.l_whence = SEEK_SET;
4225 lock_info.l_start = 0;
4226 lock_info.l_len = 1;
4227 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4228 (rc = ErrCode()) == EINTR) ;
4229 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4236 /** Try to get exclusive lock, otherwise shared.
4237 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4240 mdb_env_excl_lock(MDB_env *env, int *excl)
4244 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4248 memset(&ov, 0, sizeof(ov));
4249 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4256 struct flock lock_info;
4257 memset((void *)&lock_info, 0, sizeof(lock_info));
4258 lock_info.l_type = F_WRLCK;
4259 lock_info.l_whence = SEEK_SET;
4260 lock_info.l_start = 0;
4261 lock_info.l_len = 1;
4262 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4263 (rc = ErrCode()) == EINTR) ;
4267 # ifndef MDB_USE_POSIX_MUTEX
4268 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4271 lock_info.l_type = F_RDLCK;
4272 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4273 (rc = ErrCode()) == EINTR) ;
4283 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4285 * @(#) $Revision: 5.1 $
4286 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4287 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4289 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4293 * Please do not copyright this code. This code is in the public domain.
4295 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4296 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4297 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4298 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4299 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4300 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4301 * PERFORMANCE OF THIS SOFTWARE.
4304 * chongo <Landon Curt Noll> /\oo/\
4305 * http://www.isthe.com/chongo/
4307 * Share and Enjoy! :-)
4310 typedef unsigned long long mdb_hash_t;
4311 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4313 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4314 * @param[in] val value to hash
4315 * @param[in] hval initial value for hash
4316 * @return 64 bit hash
4318 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4319 * hval arg on the first call.
4322 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4324 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4325 unsigned char *end = s + val->mv_size;
4327 * FNV-1a hash each octet of the string
4330 /* xor the bottom with the current octet */
4331 hval ^= (mdb_hash_t)*s++;
4333 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4334 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4335 (hval << 7) + (hval << 8) + (hval << 40);
4337 /* return our new hash value */
4341 /** Hash the string and output the encoded hash.
4342 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4343 * very short name limits. We don't care about the encoding being reversible,
4344 * we just want to preserve as many bits of the input as possible in a
4345 * small printable string.
4346 * @param[in] str string to hash
4347 * @param[out] encbuf an array of 11 chars to hold the hash
4349 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4352 mdb_pack85(unsigned long l, char *out)
4356 for (i=0; i<5; i++) {
4357 *out++ = mdb_a85[l % 85];
4363 mdb_hash_enc(MDB_val *val, char *encbuf)
4365 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4367 mdb_pack85(h, encbuf);
4368 mdb_pack85(h>>32, encbuf+5);
4373 /** Open and/or initialize the lock region for the environment.
4374 * @param[in] env The LMDB environment.
4375 * @param[in] lpath The pathname of the file used for the lock region.
4376 * @param[in] mode The Unix permissions for the file, if we create it.
4377 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4378 * @return 0 on success, non-zero on failure.
4381 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4384 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4386 # define MDB_ERRCODE_ROFS EROFS
4387 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4388 # define MDB_CLOEXEC O_CLOEXEC
4391 # define MDB_CLOEXEC 0
4398 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4399 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4400 FILE_ATTRIBUTE_NORMAL, NULL);
4402 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4404 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4406 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4411 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4412 /* Lose record locks when exec*() */
4413 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4414 fcntl(env->me_lfd, F_SETFD, fdflags);
4417 if (!(env->me_flags & MDB_NOTLS)) {
4418 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4421 env->me_flags |= MDB_ENV_TXKEY;
4423 /* Windows TLS callbacks need help finding their TLS info. */
4424 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4428 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4432 /* Try to get exclusive lock. If we succeed, then
4433 * nobody is using the lock region and we should initialize it.
4435 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4438 size = GetFileSize(env->me_lfd, NULL);
4440 size = lseek(env->me_lfd, 0, SEEK_END);
4441 if (size == -1) goto fail_errno;
4443 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4444 if (size < rsize && *excl > 0) {
4446 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4447 || !SetEndOfFile(env->me_lfd))
4450 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4454 size = rsize - sizeof(MDB_txninfo);
4455 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4460 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4462 if (!mh) goto fail_errno;
4463 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4465 if (!env->me_txns) goto fail_errno;
4467 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4469 if (m == MAP_FAILED) goto fail_errno;
4475 BY_HANDLE_FILE_INFORMATION stbuf;
4484 if (!mdb_sec_inited) {
4485 InitializeSecurityDescriptor(&mdb_null_sd,
4486 SECURITY_DESCRIPTOR_REVISION);
4487 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4488 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4489 mdb_all_sa.bInheritHandle = FALSE;
4490 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4493 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4494 idbuf.volume = stbuf.dwVolumeSerialNumber;
4495 idbuf.nhigh = stbuf.nFileIndexHigh;
4496 idbuf.nlow = stbuf.nFileIndexLow;
4497 val.mv_data = &idbuf;
4498 val.mv_size = sizeof(idbuf);
4499 mdb_hash_enc(&val, encbuf);
4500 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4501 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4502 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4503 if (!env->me_rmutex) goto fail_errno;
4504 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4505 if (!env->me_wmutex) goto fail_errno;
4506 #elif defined(MDB_USE_POSIX_SEM)
4515 #if defined(__NetBSD__)
4516 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4518 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4519 idbuf.dev = stbuf.st_dev;
4520 idbuf.ino = stbuf.st_ino;
4521 val.mv_data = &idbuf;
4522 val.mv_size = sizeof(idbuf);
4523 mdb_hash_enc(&val, encbuf);
4524 #ifdef MDB_SHORT_SEMNAMES
4525 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4527 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4528 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4529 /* Clean up after a previous run, if needed: Try to
4530 * remove both semaphores before doing anything else.
4532 sem_unlink(env->me_txns->mti_rmname);
4533 sem_unlink(env->me_txns->mti_wmname);
4534 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4535 O_CREAT|O_EXCL, mode, 1);
4536 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4537 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4538 O_CREAT|O_EXCL, mode, 1);
4539 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4540 #else /* MDB_USE_POSIX_MUTEX: */
4541 pthread_mutexattr_t mattr;
4543 if ((rc = pthread_mutexattr_init(&mattr))
4544 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4545 #ifdef MDB_ROBUST_SUPPORTED
4546 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4548 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4549 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4551 pthread_mutexattr_destroy(&mattr);
4552 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4554 env->me_txns->mti_magic = MDB_MAGIC;
4555 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4556 env->me_txns->mti_txnid = 0;
4557 env->me_txns->mti_numreaders = 0;
4560 if (env->me_txns->mti_magic != MDB_MAGIC) {
4561 DPUTS("lock region has invalid magic");
4565 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4566 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4567 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4568 rc = MDB_VERSION_MISMATCH;
4572 if (rc && rc != EACCES && rc != EAGAIN) {
4576 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4577 if (!env->me_rmutex) goto fail_errno;
4578 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4579 if (!env->me_wmutex) goto fail_errno;
4580 #elif defined(MDB_USE_POSIX_SEM)
4581 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4582 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4583 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4584 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4595 /** The name of the lock file in the DB environment */
4596 #define LOCKNAME "/lock.mdb"
4597 /** The name of the data file in the DB environment */
4598 #define DATANAME "/data.mdb"
4599 /** The suffix of the lock file when no subdir is used */
4600 #define LOCKSUFF "-lock"
4601 /** Only a subset of the @ref mdb_env flags can be changed
4602 * at runtime. Changing other flags requires closing the
4603 * environment and re-opening it with the new flags.
4605 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4606 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4607 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4609 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4610 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4614 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4616 int oflags, rc, len, excl = -1;
4617 char *lpath, *dpath;
4619 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4623 if (flags & MDB_NOSUBDIR) {
4624 rc = len + sizeof(LOCKSUFF) + len + 1;
4626 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4631 if (flags & MDB_NOSUBDIR) {
4632 dpath = lpath + len + sizeof(LOCKSUFF);
4633 sprintf(lpath, "%s" LOCKSUFF, path);
4634 strcpy(dpath, path);
4636 dpath = lpath + len + sizeof(LOCKNAME);
4637 sprintf(lpath, "%s" LOCKNAME, path);
4638 sprintf(dpath, "%s" DATANAME, path);
4642 flags |= env->me_flags;
4643 if (flags & MDB_RDONLY) {
4644 /* silently ignore WRITEMAP when we're only getting read access */
4645 flags &= ~MDB_WRITEMAP;
4647 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4648 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4651 env->me_flags = flags |= MDB_ENV_ACTIVE;
4655 env->me_path = strdup(path);
4656 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4657 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4658 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4659 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4663 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4665 /* For RDONLY, get lockfile after we know datafile exists */
4666 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4667 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4673 if (F_ISSET(flags, MDB_RDONLY)) {
4674 oflags = GENERIC_READ;
4675 len = OPEN_EXISTING;
4677 oflags = GENERIC_READ|GENERIC_WRITE;
4680 mode = FILE_ATTRIBUTE_NORMAL;
4681 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4682 NULL, len, mode, NULL);
4684 if (F_ISSET(flags, MDB_RDONLY))
4687 oflags = O_RDWR | O_CREAT;
4689 env->me_fd = open(dpath, oflags, mode);
4691 if (env->me_fd == INVALID_HANDLE_VALUE) {
4696 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4697 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4702 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4703 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4704 env->me_mfd = env->me_fd;
4706 /* Synchronous fd for meta writes. Needed even with
4707 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4710 len = OPEN_EXISTING;
4711 env->me_mfd = CreateFile(dpath, oflags,
4712 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4713 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4716 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4718 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4723 DPRINTF(("opened dbenv %p", (void *) env));
4725 rc = mdb_env_share_locks(env, &excl);
4729 if (!(flags & MDB_RDONLY)) {
4731 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4732 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4733 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4734 (txn = calloc(1, size)))
4736 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4737 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4738 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4739 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4741 txn->mt_dbxs = env->me_dbxs;
4751 mdb_env_close0(env, excl);
4757 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4759 mdb_env_close0(MDB_env *env, int excl)
4763 if (!(env->me_flags & MDB_ENV_ACTIVE))
4766 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4768 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4769 free(env->me_dbxs[i].md_name.mv_data);
4774 free(env->me_dbiseqs);
4775 free(env->me_dbflags);
4777 free(env->me_dirty_list);
4779 mdb_midl_free(env->me_free_pgs);
4781 if (env->me_flags & MDB_ENV_TXKEY) {
4782 pthread_key_delete(env->me_txkey);
4784 /* Delete our key from the global list */
4785 for (i=0; i<mdb_tls_nkeys; i++)
4786 if (mdb_tls_keys[i] == env->me_txkey) {
4787 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4795 munmap(env->me_map, env->me_mapsize);
4797 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4798 (void) close(env->me_mfd);
4799 if (env->me_fd != INVALID_HANDLE_VALUE)
4800 (void) close(env->me_fd);
4802 MDB_PID_T pid = env->me_pid;
4803 /* Clearing readers is done in this function because
4804 * me_txkey with its destructor must be disabled first.
4806 * We skip the the reader mutex, so we touch only
4807 * data owned by this process (me_close_readers and
4808 * our readers), and clear each reader atomically.
4810 for (i = env->me_close_readers; --i >= 0; )
4811 if (env->me_txns->mti_readers[i].mr_pid == pid)
4812 env->me_txns->mti_readers[i].mr_pid = 0;
4814 if (env->me_rmutex) {
4815 CloseHandle(env->me_rmutex);
4816 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4818 /* Windows automatically destroys the mutexes when
4819 * the last handle closes.
4821 #elif defined(MDB_USE_POSIX_SEM)
4822 if (env->me_rmutex != SEM_FAILED) {
4823 sem_close(env->me_rmutex);
4824 if (env->me_wmutex != SEM_FAILED)
4825 sem_close(env->me_wmutex);
4826 /* If we have the filelock: If we are the
4827 * only remaining user, clean up semaphores.
4830 mdb_env_excl_lock(env, &excl);
4832 sem_unlink(env->me_txns->mti_rmname);
4833 sem_unlink(env->me_txns->mti_wmname);
4837 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4839 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4842 /* Unlock the lockfile. Windows would have unlocked it
4843 * after closing anyway, but not necessarily at once.
4845 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4848 (void) close(env->me_lfd);
4851 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4855 mdb_env_close(MDB_env *env)
4862 VGMEMP_DESTROY(env);
4863 while ((dp = env->me_dpages) != NULL) {
4864 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4865 env->me_dpages = dp->mp_next;
4869 mdb_env_close0(env, 0);
4873 /** Compare two items pointing at aligned size_t's */
4875 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4877 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4878 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4881 /** Compare two items pointing at aligned unsigned int's.
4883 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4884 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4887 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4889 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4890 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4893 /** Compare two items pointing at unsigned ints of unknown alignment.
4894 * Nodes and keys are guaranteed to be 2-byte aligned.
4897 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4899 #if BYTE_ORDER == LITTLE_ENDIAN
4900 unsigned short *u, *c;
4903 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4904 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4907 } while(!x && u > (unsigned short *)a->mv_data);
4910 unsigned short *u, *c, *end;
4913 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4914 u = (unsigned short *)a->mv_data;
4915 c = (unsigned short *)b->mv_data;
4918 } while(!x && u < end);
4923 /** Compare two items lexically */
4925 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4932 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4938 diff = memcmp(a->mv_data, b->mv_data, len);
4939 return diff ? diff : len_diff<0 ? -1 : len_diff;
4942 /** Compare two items in reverse byte order */
4944 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4946 const unsigned char *p1, *p2, *p1_lim;
4950 p1_lim = (const unsigned char *)a->mv_data;
4951 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4952 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4954 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4960 while (p1 > p1_lim) {
4961 diff = *--p1 - *--p2;
4965 return len_diff<0 ? -1 : len_diff;
4968 /** Search for key within a page, using binary search.
4969 * Returns the smallest entry larger or equal to the key.
4970 * If exactp is non-null, stores whether the found entry was an exact match
4971 * in *exactp (1 or 0).
4972 * Updates the cursor index with the index of the found entry.
4973 * If no entry larger or equal to the key is found, returns NULL.
4976 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4978 unsigned int i = 0, nkeys;
4981 MDB_page *mp = mc->mc_pg[mc->mc_top];
4982 MDB_node *node = NULL;
4987 nkeys = NUMKEYS(mp);
4989 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4990 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4993 low = IS_LEAF(mp) ? 0 : 1;
4995 cmp = mc->mc_dbx->md_cmp;
4997 /* Branch pages have no data, so if using integer keys,
4998 * alignment is guaranteed. Use faster mdb_cmp_int.
5000 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5001 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5008 nodekey.mv_size = mc->mc_db->md_pad;
5009 node = NODEPTR(mp, 0); /* fake */
5010 while (low <= high) {
5011 i = (low + high) >> 1;
5012 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5013 rc = cmp(key, &nodekey);
5014 DPRINTF(("found leaf index %u [%s], rc = %i",
5015 i, DKEY(&nodekey), rc));
5024 while (low <= high) {
5025 i = (low + high) >> 1;
5027 node = NODEPTR(mp, i);
5028 nodekey.mv_size = NODEKSZ(node);
5029 nodekey.mv_data = NODEKEY(node);
5031 rc = cmp(key, &nodekey);
5034 DPRINTF(("found leaf index %u [%s], rc = %i",
5035 i, DKEY(&nodekey), rc));
5037 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5038 i, DKEY(&nodekey), NODEPGNO(node), rc));
5049 if (rc > 0) { /* Found entry is less than the key. */
5050 i++; /* Skip to get the smallest entry larger than key. */
5052 node = NODEPTR(mp, i);
5055 *exactp = (rc == 0 && nkeys > 0);
5056 /* store the key index */
5057 mc->mc_ki[mc->mc_top] = i;
5059 /* There is no entry larger or equal to the key. */
5062 /* nodeptr is fake for LEAF2 */
5068 mdb_cursor_adjust(MDB_cursor *mc, func)
5072 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5073 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5080 /** Pop a page off the top of the cursor's stack. */
5082 mdb_cursor_pop(MDB_cursor *mc)
5086 MDB_page *top = mc->mc_pg[mc->mc_top];
5092 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5093 DDBI(mc), (void *) mc));
5097 /** Push a page onto the top of the cursor's stack. */
5099 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5101 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5102 DDBI(mc), (void *) mc));
5104 if (mc->mc_snum >= CURSOR_STACK) {
5105 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5106 return MDB_CURSOR_FULL;
5109 mc->mc_top = mc->mc_snum++;
5110 mc->mc_pg[mc->mc_top] = mp;
5111 mc->mc_ki[mc->mc_top] = 0;
5116 /** Find the address of the page corresponding to a given page number.
5117 * @param[in] txn the transaction for this access.
5118 * @param[in] pgno the page number for the page to retrieve.
5119 * @param[out] ret address of a pointer where the page's address will be stored.
5120 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5121 * @return 0 on success, non-zero on failure.
5124 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5126 MDB_env *env = txn->mt_env;
5130 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5134 MDB_ID2L dl = tx2->mt_u.dirty_list;
5136 /* Spilled pages were dirtied in this txn and flushed
5137 * because the dirty list got full. Bring this page
5138 * back in from the map (but don't unspill it here,
5139 * leave that unless page_touch happens again).
5141 if (tx2->mt_spill_pgs) {
5142 MDB_ID pn = pgno << 1;
5143 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5144 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5145 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5150 unsigned x = mdb_mid2l_search(dl, pgno);
5151 if (x <= dl[0].mid && dl[x].mid == pgno) {
5157 } while ((tx2 = tx2->mt_parent) != NULL);
5160 if (pgno < txn->mt_next_pgno) {
5162 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5164 DPRINTF(("page %"Z"u not found", pgno));
5165 txn->mt_flags |= MDB_TXN_ERROR;
5166 return MDB_PAGE_NOTFOUND;
5176 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5177 * The cursor is at the root page, set up the rest of it.
5180 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5182 MDB_page *mp = mc->mc_pg[mc->mc_top];
5186 while (IS_BRANCH(mp)) {
5190 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5191 mdb_cassert(mc, NUMKEYS(mp) > 1);
5192 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5194 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5196 if (flags & MDB_PS_LAST)
5197 i = NUMKEYS(mp) - 1;
5200 node = mdb_node_search(mc, key, &exact);
5202 i = NUMKEYS(mp) - 1;
5204 i = mc->mc_ki[mc->mc_top];
5206 mdb_cassert(mc, i > 0);
5210 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5213 mdb_cassert(mc, i < NUMKEYS(mp));
5214 node = NODEPTR(mp, i);
5216 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5219 mc->mc_ki[mc->mc_top] = i;
5220 if ((rc = mdb_cursor_push(mc, mp)))
5223 if (flags & MDB_PS_MODIFY) {
5224 if ((rc = mdb_page_touch(mc)) != 0)
5226 mp = mc->mc_pg[mc->mc_top];
5231 DPRINTF(("internal error, index points to a %02X page!?",
5233 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5234 return MDB_CORRUPTED;
5237 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5238 key ? DKEY(key) : "null"));
5239 mc->mc_flags |= C_INITIALIZED;
5240 mc->mc_flags &= ~C_EOF;
5245 /** Search for the lowest key under the current branch page.
5246 * This just bypasses a NUMKEYS check in the current page
5247 * before calling mdb_page_search_root(), because the callers
5248 * are all in situations where the current page is known to
5252 mdb_page_search_lowest(MDB_cursor *mc)
5254 MDB_page *mp = mc->mc_pg[mc->mc_top];
5255 MDB_node *node = NODEPTR(mp, 0);
5258 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5261 mc->mc_ki[mc->mc_top] = 0;
5262 if ((rc = mdb_cursor_push(mc, mp)))
5264 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5267 /** Search for the page a given key should be in.
5268 * Push it and its parent pages on the cursor stack.
5269 * @param[in,out] mc the cursor for this operation.
5270 * @param[in] key the key to search for, or NULL for first/last page.
5271 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5272 * are touched (updated with new page numbers).
5273 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5274 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5275 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5276 * @return 0 on success, non-zero on failure.
5279 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5284 /* Make sure the txn is still viable, then find the root from
5285 * the txn's db table and set it as the root of the cursor's stack.
5287 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5288 DPUTS("transaction has failed, must abort");
5291 /* Make sure we're using an up-to-date root */
5292 if (*mc->mc_dbflag & DB_STALE) {
5294 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5296 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5297 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5304 MDB_node *leaf = mdb_node_search(&mc2,
5305 &mc->mc_dbx->md_name, &exact);
5307 return MDB_NOTFOUND;
5308 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5309 return MDB_INCOMPATIBLE; /* not a named DB */
5310 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5313 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5315 /* The txn may not know this DBI, or another process may
5316 * have dropped and recreated the DB with other flags.
5318 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5319 return MDB_INCOMPATIBLE;
5320 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5322 *mc->mc_dbflag &= ~DB_STALE;
5324 root = mc->mc_db->md_root;
5326 if (root == P_INVALID) { /* Tree is empty. */
5327 DPUTS("tree is empty");
5328 return MDB_NOTFOUND;
5332 mdb_cassert(mc, root > 1);
5333 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5334 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5340 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5341 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5343 if (flags & MDB_PS_MODIFY) {
5344 if ((rc = mdb_page_touch(mc)))
5348 if (flags & MDB_PS_ROOTONLY)
5351 return mdb_page_search_root(mc, key, flags);
5355 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5357 MDB_txn *txn = mc->mc_txn;
5358 pgno_t pg = mp->mp_pgno;
5359 unsigned x = 0, ovpages = mp->mp_pages;
5360 MDB_env *env = txn->mt_env;
5361 MDB_IDL sl = txn->mt_spill_pgs;
5362 MDB_ID pn = pg << 1;
5365 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5366 /* If the page is dirty or on the spill list we just acquired it,
5367 * so we should give it back to our current free list, if any.
5368 * Otherwise put it onto the list of pages we freed in this txn.
5370 * Won't create me_pghead: me_pglast must be inited along with it.
5371 * Unsupported in nested txns: They would need to hide the page
5372 * range in ancestor txns' dirty and spilled lists.
5374 if (env->me_pghead &&
5376 ((mp->mp_flags & P_DIRTY) ||
5377 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5381 MDB_ID2 *dl, ix, iy;
5382 rc = mdb_midl_need(&env->me_pghead, ovpages);
5385 if (!(mp->mp_flags & P_DIRTY)) {
5386 /* This page is no longer spilled */
5393 /* Remove from dirty list */
5394 dl = txn->mt_u.dirty_list;
5396 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5402 mdb_cassert(mc, x > 1);
5404 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5405 txn->mt_flags |= MDB_TXN_ERROR;
5406 return MDB_CORRUPTED;
5409 if (!(env->me_flags & MDB_WRITEMAP))
5410 mdb_dpage_free(env, mp);
5412 /* Insert in me_pghead */
5413 mop = env->me_pghead;
5414 j = mop[0] + ovpages;
5415 for (i = mop[0]; i && mop[i] < pg; i--)
5421 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5425 mc->mc_db->md_overflow_pages -= ovpages;
5429 /** Return the data associated with a given node.
5430 * @param[in] txn The transaction for this operation.
5431 * @param[in] leaf The node being read.
5432 * @param[out] data Updated to point to the node's data.
5433 * @return 0 on success, non-zero on failure.
5436 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5438 MDB_page *omp; /* overflow page */
5442 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5443 data->mv_size = NODEDSZ(leaf);
5444 data->mv_data = NODEDATA(leaf);
5448 /* Read overflow data.
5450 data->mv_size = NODEDSZ(leaf);
5451 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5452 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5453 DPRINTF(("read overflow page %"Z"u failed", pgno));
5456 data->mv_data = METADATA(omp);
5462 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5463 MDB_val *key, MDB_val *data)
5470 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5472 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5475 if (txn->mt_flags & MDB_TXN_ERROR)
5478 mdb_cursor_init(&mc, txn, dbi, &mx);
5479 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5482 /** Find a sibling for a page.
5483 * Replaces the page at the top of the cursor's stack with the
5484 * specified sibling, if one exists.
5485 * @param[in] mc The cursor for this operation.
5486 * @param[in] move_right Non-zero if the right sibling is requested,
5487 * otherwise the left sibling.
5488 * @return 0 on success, non-zero on failure.
5491 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5497 if (mc->mc_snum < 2) {
5498 return MDB_NOTFOUND; /* root has no siblings */
5502 DPRINTF(("parent page is page %"Z"u, index %u",
5503 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5505 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5506 : (mc->mc_ki[mc->mc_top] == 0)) {
5507 DPRINTF(("no more keys left, moving to %s sibling",
5508 move_right ? "right" : "left"));
5509 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5510 /* undo cursor_pop before returning */
5517 mc->mc_ki[mc->mc_top]++;
5519 mc->mc_ki[mc->mc_top]--;
5520 DPRINTF(("just moving to %s index key %u",
5521 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5523 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5525 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5526 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5527 /* mc will be inconsistent if caller does mc_snum++ as above */
5528 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5532 mdb_cursor_push(mc, mp);
5534 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5539 /** Move the cursor to the next data item. */
5541 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5547 if (mc->mc_flags & C_EOF) {
5548 return MDB_NOTFOUND;
5551 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5553 mp = mc->mc_pg[mc->mc_top];
5555 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5556 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5557 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5558 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5559 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5560 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5561 if (rc == MDB_SUCCESS)
5562 MDB_GET_KEY(leaf, key);
5567 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5568 if (op == MDB_NEXT_DUP)
5569 return MDB_NOTFOUND;
5573 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5574 mdb_dbg_pgno(mp), (void *) mc));
5575 if (mc->mc_flags & C_DEL)
5578 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5579 DPUTS("=====> move to next sibling page");
5580 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5581 mc->mc_flags |= C_EOF;
5584 mp = mc->mc_pg[mc->mc_top];
5585 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5587 mc->mc_ki[mc->mc_top]++;
5590 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5591 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5594 key->mv_size = mc->mc_db->md_pad;
5595 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5599 mdb_cassert(mc, IS_LEAF(mp));
5600 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5602 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5603 mdb_xcursor_init1(mc, leaf);
5606 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5609 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5610 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5611 if (rc != MDB_SUCCESS)
5616 MDB_GET_KEY(leaf, key);
5620 /** Move the cursor to the previous data item. */
5622 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5628 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5630 mp = mc->mc_pg[mc->mc_top];
5632 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5633 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5634 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5635 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5636 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5637 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5638 if (rc == MDB_SUCCESS) {
5639 MDB_GET_KEY(leaf, key);
5640 mc->mc_flags &= ~C_EOF;
5646 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5647 if (op == MDB_PREV_DUP)
5648 return MDB_NOTFOUND;
5652 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5653 mdb_dbg_pgno(mp), (void *) mc));
5655 if (mc->mc_ki[mc->mc_top] == 0) {
5656 DPUTS("=====> move to prev sibling page");
5657 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5660 mp = mc->mc_pg[mc->mc_top];
5661 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5662 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5664 mc->mc_ki[mc->mc_top]--;
5666 mc->mc_flags &= ~C_EOF;
5668 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5669 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5672 key->mv_size = mc->mc_db->md_pad;
5673 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5677 mdb_cassert(mc, IS_LEAF(mp));
5678 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5680 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5681 mdb_xcursor_init1(mc, leaf);
5684 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5687 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5688 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5689 if (rc != MDB_SUCCESS)
5694 MDB_GET_KEY(leaf, key);
5698 /** Set the cursor on a specific data item. */
5700 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5701 MDB_cursor_op op, int *exactp)
5705 MDB_node *leaf = NULL;
5708 if (key->mv_size == 0)
5709 return MDB_BAD_VALSIZE;
5712 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5714 /* See if we're already on the right page */
5715 if (mc->mc_flags & C_INITIALIZED) {
5718 mp = mc->mc_pg[mc->mc_top];
5720 mc->mc_ki[mc->mc_top] = 0;
5721 return MDB_NOTFOUND;
5723 if (mp->mp_flags & P_LEAF2) {
5724 nodekey.mv_size = mc->mc_db->md_pad;
5725 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5727 leaf = NODEPTR(mp, 0);
5728 MDB_GET_KEY2(leaf, nodekey);
5730 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5732 /* Probably happens rarely, but first node on the page
5733 * was the one we wanted.
5735 mc->mc_ki[mc->mc_top] = 0;
5742 unsigned int nkeys = NUMKEYS(mp);
5744 if (mp->mp_flags & P_LEAF2) {
5745 nodekey.mv_data = LEAF2KEY(mp,
5746 nkeys-1, nodekey.mv_size);
5748 leaf = NODEPTR(mp, nkeys-1);
5749 MDB_GET_KEY2(leaf, nodekey);
5751 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5753 /* last node was the one we wanted */
5754 mc->mc_ki[mc->mc_top] = nkeys-1;
5760 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5761 /* This is definitely the right page, skip search_page */
5762 if (mp->mp_flags & P_LEAF2) {
5763 nodekey.mv_data = LEAF2KEY(mp,
5764 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5766 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5767 MDB_GET_KEY2(leaf, nodekey);
5769 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5771 /* current node was the one we wanted */
5781 /* If any parents have right-sibs, search.
5782 * Otherwise, there's nothing further.
5784 for (i=0; i<mc->mc_top; i++)
5786 NUMKEYS(mc->mc_pg[i])-1)
5788 if (i == mc->mc_top) {
5789 /* There are no other pages */
5790 mc->mc_ki[mc->mc_top] = nkeys;
5791 return MDB_NOTFOUND;
5795 /* There are no other pages */
5796 mc->mc_ki[mc->mc_top] = 0;
5797 if (op == MDB_SET_RANGE && !exactp) {
5801 return MDB_NOTFOUND;
5805 rc = mdb_page_search(mc, key, 0);
5806 if (rc != MDB_SUCCESS)
5809 mp = mc->mc_pg[mc->mc_top];
5810 mdb_cassert(mc, IS_LEAF(mp));
5813 leaf = mdb_node_search(mc, key, exactp);
5814 if (exactp != NULL && !*exactp) {
5815 /* MDB_SET specified and not an exact match. */
5816 return MDB_NOTFOUND;
5820 DPUTS("===> inexact leaf not found, goto sibling");
5821 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5822 mc->mc_flags |= C_EOF;
5823 return rc; /* no entries matched */
5825 mp = mc->mc_pg[mc->mc_top];
5826 mdb_cassert(mc, IS_LEAF(mp));
5827 leaf = NODEPTR(mp, 0);
5831 mc->mc_flags |= C_INITIALIZED;
5832 mc->mc_flags &= ~C_EOF;
5835 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5836 key->mv_size = mc->mc_db->md_pad;
5837 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5842 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5843 mdb_xcursor_init1(mc, leaf);
5846 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5847 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5848 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5851 if (op == MDB_GET_BOTH) {
5857 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5858 if (rc != MDB_SUCCESS)
5861 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5864 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5866 dcmp = mc->mc_dbx->md_dcmp;
5867 #if UINT_MAX < SIZE_MAX
5868 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5869 dcmp = mdb_cmp_clong;
5871 rc = dcmp(data, &olddata);
5873 if (op == MDB_GET_BOTH || rc > 0)
5874 return MDB_NOTFOUND;
5881 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5882 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5887 /* The key already matches in all other cases */
5888 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5889 MDB_GET_KEY(leaf, key);
5890 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5895 /** Move the cursor to the first item in the database. */
5897 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5903 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5905 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5906 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5907 if (rc != MDB_SUCCESS)
5910 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5912 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5913 mc->mc_flags |= C_INITIALIZED;
5914 mc->mc_flags &= ~C_EOF;
5916 mc->mc_ki[mc->mc_top] = 0;
5918 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5919 key->mv_size = mc->mc_db->md_pad;
5920 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5925 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5926 mdb_xcursor_init1(mc, leaf);
5927 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5931 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5935 MDB_GET_KEY(leaf, key);
5939 /** Move the cursor to the last item in the database. */
5941 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5947 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5949 if (!(mc->mc_flags & C_EOF)) {
5951 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5952 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5953 if (rc != MDB_SUCCESS)
5956 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5959 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5960 mc->mc_flags |= C_INITIALIZED|C_EOF;
5961 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5963 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5964 key->mv_size = mc->mc_db->md_pad;
5965 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5970 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5971 mdb_xcursor_init1(mc, leaf);
5972 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5976 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5981 MDB_GET_KEY(leaf, key);
5986 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5991 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5996 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6000 case MDB_GET_CURRENT:
6001 if (!(mc->mc_flags & C_INITIALIZED)) {
6004 MDB_page *mp = mc->mc_pg[mc->mc_top];
6005 int nkeys = NUMKEYS(mp);
6006 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6007 mc->mc_ki[mc->mc_top] = nkeys;
6013 key->mv_size = mc->mc_db->md_pad;
6014 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6016 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6017 MDB_GET_KEY(leaf, key);
6019 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6020 if (mc->mc_flags & C_DEL)
6021 mdb_xcursor_init1(mc, leaf);
6022 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6024 rc = mdb_node_read(mc->mc_txn, leaf, data);
6031 case MDB_GET_BOTH_RANGE:
6036 if (mc->mc_xcursor == NULL) {
6037 rc = MDB_INCOMPATIBLE;
6047 rc = mdb_cursor_set(mc, key, data, op,
6048 op == MDB_SET_RANGE ? NULL : &exact);
6051 case MDB_GET_MULTIPLE:
6052 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6056 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6057 rc = MDB_INCOMPATIBLE;
6061 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6062 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6065 case MDB_NEXT_MULTIPLE:
6070 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6071 rc = MDB_INCOMPATIBLE;
6074 if (!(mc->mc_flags & C_INITIALIZED))
6075 rc = mdb_cursor_first(mc, key, data);
6077 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6078 if (rc == MDB_SUCCESS) {
6079 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6082 mx = &mc->mc_xcursor->mx_cursor;
6083 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6085 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6086 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6094 case MDB_NEXT_NODUP:
6095 if (!(mc->mc_flags & C_INITIALIZED))
6096 rc = mdb_cursor_first(mc, key, data);
6098 rc = mdb_cursor_next(mc, key, data, op);
6102 case MDB_PREV_NODUP:
6103 if (!(mc->mc_flags & C_INITIALIZED)) {
6104 rc = mdb_cursor_last(mc, key, data);
6107 mc->mc_flags |= C_INITIALIZED;
6108 mc->mc_ki[mc->mc_top]++;
6110 rc = mdb_cursor_prev(mc, key, data, op);
6113 rc = mdb_cursor_first(mc, key, data);
6116 mfunc = mdb_cursor_first;
6118 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6122 if (mc->mc_xcursor == NULL) {
6123 rc = MDB_INCOMPATIBLE;
6127 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6128 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6129 MDB_GET_KEY(leaf, key);
6130 rc = mdb_node_read(mc->mc_txn, leaf, data);
6134 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6138 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6141 rc = mdb_cursor_last(mc, key, data);
6144 mfunc = mdb_cursor_last;
6147 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6152 if (mc->mc_flags & C_DEL)
6153 mc->mc_flags ^= C_DEL;
6158 /** Touch all the pages in the cursor stack. Set mc_top.
6159 * Makes sure all the pages are writable, before attempting a write operation.
6160 * @param[in] mc The cursor to operate on.
6163 mdb_cursor_touch(MDB_cursor *mc)
6165 int rc = MDB_SUCCESS;
6167 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6170 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6172 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6173 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6176 *mc->mc_dbflag |= DB_DIRTY;
6181 rc = mdb_page_touch(mc);
6182 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6183 mc->mc_top = mc->mc_snum-1;
6188 /** Do not spill pages to disk if txn is getting full, may fail instead */
6189 #define MDB_NOSPILL 0x8000
6192 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6196 MDB_node *leaf = NULL;
6197 MDB_page *fp, *mp, *sub_root = NULL;
6199 MDB_val xdata, *rdata, dkey, olddata;
6201 int do_sub = 0, insert_key, insert_data;
6202 unsigned int mcount = 0, dcount = 0, nospill;
6205 unsigned int nflags;
6208 if (mc == NULL || key == NULL)
6211 env = mc->mc_txn->mt_env;
6213 /* Check this first so counter will always be zero on any
6216 if (flags & MDB_MULTIPLE) {
6217 dcount = data[1].mv_size;
6218 data[1].mv_size = 0;
6219 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6220 return MDB_INCOMPATIBLE;
6223 nospill = flags & MDB_NOSPILL;
6224 flags &= ~MDB_NOSPILL;
6226 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6227 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6229 if (key->mv_size-1 >= ENV_MAXKEY(env))
6230 return MDB_BAD_VALSIZE;
6232 #if SIZE_MAX > MAXDATASIZE
6233 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6234 return MDB_BAD_VALSIZE;
6236 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6237 return MDB_BAD_VALSIZE;
6240 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6241 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6245 if (flags == MDB_CURRENT) {
6246 if (!(mc->mc_flags & C_INITIALIZED))
6249 } else if (mc->mc_db->md_root == P_INVALID) {
6250 /* new database, cursor has nothing to point to */
6253 mc->mc_flags &= ~C_INITIALIZED;
6258 if (flags & MDB_APPEND) {
6260 rc = mdb_cursor_last(mc, &k2, &d2);
6262 rc = mc->mc_dbx->md_cmp(key, &k2);
6265 mc->mc_ki[mc->mc_top]++;
6267 /* new key is <= last key */
6272 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6274 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6275 DPRINTF(("duplicate key [%s]", DKEY(key)));
6277 return MDB_KEYEXIST;
6279 if (rc && rc != MDB_NOTFOUND)
6283 if (mc->mc_flags & C_DEL)
6284 mc->mc_flags ^= C_DEL;
6286 /* Cursor is positioned, check for room in the dirty list */
6288 if (flags & MDB_MULTIPLE) {
6290 xdata.mv_size = data->mv_size * dcount;
6294 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6298 if (rc == MDB_NO_ROOT) {
6300 /* new database, write a root leaf page */
6301 DPUTS("allocating new root leaf page");
6302 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6305 mdb_cursor_push(mc, np);
6306 mc->mc_db->md_root = np->mp_pgno;
6307 mc->mc_db->md_depth++;
6308 *mc->mc_dbflag |= DB_DIRTY;
6309 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6311 np->mp_flags |= P_LEAF2;
6312 mc->mc_flags |= C_INITIALIZED;
6314 /* make sure all cursor pages are writable */
6315 rc2 = mdb_cursor_touch(mc);
6320 insert_key = insert_data = rc;
6322 /* The key does not exist */
6323 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6324 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6325 LEAFSIZE(key, data) > env->me_nodemax)
6327 /* Too big for a node, insert in sub-DB. Set up an empty
6328 * "old sub-page" for prep_subDB to expand to a full page.
6330 fp_flags = P_LEAF|P_DIRTY;
6332 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6333 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6334 olddata.mv_size = PAGEHDRSZ;
6338 /* there's only a key anyway, so this is a no-op */
6339 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6341 unsigned int ksize = mc->mc_db->md_pad;
6342 if (key->mv_size != ksize)
6343 return MDB_BAD_VALSIZE;
6344 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6345 memcpy(ptr, key->mv_data, ksize);
6347 /* if overwriting slot 0 of leaf, need to
6348 * update branch key if there is a parent page
6350 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6351 unsigned short top = mc->mc_top;
6353 /* slot 0 is always an empty key, find real slot */
6354 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6356 if (mc->mc_ki[mc->mc_top])
6357 rc2 = mdb_update_key(mc, key);
6368 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6369 olddata.mv_size = NODEDSZ(leaf);
6370 olddata.mv_data = NODEDATA(leaf);
6373 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6374 /* Prepare (sub-)page/sub-DB to accept the new item,
6375 * if needed. fp: old sub-page or a header faking
6376 * it. mp: new (sub-)page. offset: growth in page
6377 * size. xdata: node data with new page or DB.
6379 unsigned i, offset = 0;
6380 mp = fp = xdata.mv_data = env->me_pbuf;
6381 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6383 /* Was a single item before, must convert now */
6384 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6386 /* Just overwrite the current item */
6387 if (flags == MDB_CURRENT)
6389 dcmp = mc->mc_dbx->md_dcmp;
6390 #if UINT_MAX < SIZE_MAX
6391 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6392 dcmp = mdb_cmp_clong;
6394 /* does data match? */
6395 if (!dcmp(data, &olddata)) {
6396 if (flags & MDB_NODUPDATA)
6397 return MDB_KEYEXIST;
6402 /* Back up original data item */
6403 dkey.mv_size = olddata.mv_size;
6404 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6406 /* Make sub-page header for the dup items, with dummy body */
6407 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6408 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6409 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6410 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6411 fp->mp_flags |= P_LEAF2;
6412 fp->mp_pad = data->mv_size;
6413 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6415 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6416 (dkey.mv_size & 1) + (data->mv_size & 1);
6418 fp->mp_upper = xdata.mv_size - PAGEBASE;
6419 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6420 } else if (leaf->mn_flags & F_SUBDATA) {
6421 /* Data is on sub-DB, just store it */
6422 flags |= F_DUPDATA|F_SUBDATA;
6425 /* Data is on sub-page */
6426 fp = olddata.mv_data;
6429 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6430 offset = EVEN(NODESIZE + sizeof(indx_t) +
6434 offset = fp->mp_pad;
6435 if (SIZELEFT(fp) < offset) {
6436 offset *= 4; /* space for 4 more */
6439 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6441 fp->mp_flags |= P_DIRTY;
6442 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6443 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6447 xdata.mv_size = olddata.mv_size + offset;
6450 fp_flags = fp->mp_flags;
6451 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6452 /* Too big for a sub-page, convert to sub-DB */
6453 fp_flags &= ~P_SUBP;
6455 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6456 fp_flags |= P_LEAF2;
6457 dummy.md_pad = fp->mp_pad;
6458 dummy.md_flags = MDB_DUPFIXED;
6459 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6460 dummy.md_flags |= MDB_INTEGERKEY;
6466 dummy.md_branch_pages = 0;
6467 dummy.md_leaf_pages = 1;
6468 dummy.md_overflow_pages = 0;
6469 dummy.md_entries = NUMKEYS(fp);
6470 xdata.mv_size = sizeof(MDB_db);
6471 xdata.mv_data = &dummy;
6472 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6474 offset = env->me_psize - olddata.mv_size;
6475 flags |= F_DUPDATA|F_SUBDATA;
6476 dummy.md_root = mp->mp_pgno;
6480 mp->mp_flags = fp_flags | P_DIRTY;
6481 mp->mp_pad = fp->mp_pad;
6482 mp->mp_lower = fp->mp_lower;
6483 mp->mp_upper = fp->mp_upper + offset;
6484 if (fp_flags & P_LEAF2) {
6485 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6487 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6488 olddata.mv_size - fp->mp_upper - PAGEBASE);
6489 for (i=0; i<NUMKEYS(fp); i++)
6490 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6498 mdb_node_del(mc, 0);
6502 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6503 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6504 return MDB_INCOMPATIBLE;
6505 /* overflow page overwrites need special handling */
6506 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6509 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6511 memcpy(&pg, olddata.mv_data, sizeof(pg));
6512 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6514 ovpages = omp->mp_pages;
6516 /* Is the ov page large enough? */
6517 if (ovpages >= dpages) {
6518 if (!(omp->mp_flags & P_DIRTY) &&
6519 (level || (env->me_flags & MDB_WRITEMAP)))
6521 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6524 level = 0; /* dirty in this txn or clean */
6527 if (omp->mp_flags & P_DIRTY) {
6528 /* yes, overwrite it. Note in this case we don't
6529 * bother to try shrinking the page if the new data
6530 * is smaller than the overflow threshold.
6533 /* It is writable only in a parent txn */
6534 size_t sz = (size_t) env->me_psize * ovpages, off;
6535 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6541 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6542 mdb_cassert(mc, rc2 == 0);
6543 if (!(flags & MDB_RESERVE)) {
6544 /* Copy end of page, adjusting alignment so
6545 * compiler may copy words instead of bytes.
6547 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6548 memcpy((size_t *)((char *)np + off),
6549 (size_t *)((char *)omp + off), sz - off);
6552 memcpy(np, omp, sz); /* Copy beginning of page */
6555 SETDSZ(leaf, data->mv_size);
6556 if (F_ISSET(flags, MDB_RESERVE))
6557 data->mv_data = METADATA(omp);
6559 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6563 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6565 } else if (data->mv_size == olddata.mv_size) {
6566 /* same size, just replace it. Note that we could
6567 * also reuse this node if the new data is smaller,
6568 * but instead we opt to shrink the node in that case.
6570 if (F_ISSET(flags, MDB_RESERVE))
6571 data->mv_data = olddata.mv_data;
6572 else if (!(mc->mc_flags & C_SUB))
6573 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6575 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6580 mdb_node_del(mc, 0);
6586 nflags = flags & NODE_ADD_FLAGS;
6587 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6588 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6589 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6590 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6592 nflags |= MDB_SPLIT_REPLACE;
6593 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6595 /* There is room already in this leaf page. */
6596 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6597 if (rc == 0 && insert_key) {
6598 /* Adjust other cursors pointing to mp */
6599 MDB_cursor *m2, *m3;
6600 MDB_dbi dbi = mc->mc_dbi;
6601 unsigned i = mc->mc_top;
6602 MDB_page *mp = mc->mc_pg[i];
6604 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6605 if (mc->mc_flags & C_SUB)
6606 m3 = &m2->mc_xcursor->mx_cursor;
6609 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6610 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6617 if (rc == MDB_SUCCESS) {
6618 /* Now store the actual data in the child DB. Note that we're
6619 * storing the user data in the keys field, so there are strict
6620 * size limits on dupdata. The actual data fields of the child
6621 * DB are all zero size.
6624 int xflags, new_dupdata;
6629 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6630 if (flags & MDB_CURRENT) {
6631 xflags = MDB_CURRENT|MDB_NOSPILL;
6633 mdb_xcursor_init1(mc, leaf);
6634 xflags = (flags & MDB_NODUPDATA) ?
6635 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6638 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6639 new_dupdata = (int)dkey.mv_size;
6640 /* converted, write the original data first */
6642 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6645 /* we've done our job */
6648 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6649 /* Adjust other cursors pointing to mp */
6651 MDB_xcursor *mx = mc->mc_xcursor;
6652 unsigned i = mc->mc_top;
6653 MDB_page *mp = mc->mc_pg[i];
6655 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6656 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6657 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6658 if (m2->mc_pg[i] == mp) {
6659 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6660 mdb_xcursor_init2(m2, mx, new_dupdata);
6661 } else if (!insert_key) {
6662 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6663 if (!(n2->mn_flags & F_SUBDATA))
6664 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6669 ecount = mc->mc_xcursor->mx_db.md_entries;
6670 if (flags & MDB_APPENDDUP)
6671 xflags |= MDB_APPEND;
6672 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6673 if (flags & F_SUBDATA) {
6674 void *db = NODEDATA(leaf);
6675 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6677 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6679 /* Increment count unless we just replaced an existing item. */
6681 mc->mc_db->md_entries++;
6683 /* Invalidate txn if we created an empty sub-DB */
6686 /* If we succeeded and the key didn't exist before,
6687 * make sure the cursor is marked valid.
6689 mc->mc_flags |= C_INITIALIZED;
6691 if (flags & MDB_MULTIPLE) {
6694 /* let caller know how many succeeded, if any */
6695 data[1].mv_size = mcount;
6696 if (mcount < dcount) {
6697 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6698 insert_key = insert_data = 0;
6705 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6708 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6713 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6719 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6720 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6722 if (!(mc->mc_flags & C_INITIALIZED))
6725 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6726 return MDB_NOTFOUND;
6728 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6731 rc = mdb_cursor_touch(mc);
6735 mp = mc->mc_pg[mc->mc_top];
6738 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6740 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6741 if (flags & MDB_NODUPDATA) {
6742 /* mdb_cursor_del0() will subtract the final entry */
6743 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6745 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6746 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6748 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6751 /* If sub-DB still has entries, we're done */
6752 if (mc->mc_xcursor->mx_db.md_entries) {
6753 if (leaf->mn_flags & F_SUBDATA) {
6754 /* update subDB info */
6755 void *db = NODEDATA(leaf);
6756 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6759 /* shrink fake page */
6760 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6761 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6762 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6763 /* fix other sub-DB cursors pointed at fake pages on this page */
6764 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6765 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6766 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6767 if (m2->mc_pg[mc->mc_top] == mp) {
6768 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6769 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6771 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6772 if (!(n2->mn_flags & F_SUBDATA))
6773 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6778 mc->mc_db->md_entries--;
6779 mc->mc_flags |= C_DEL;
6782 /* otherwise fall thru and delete the sub-DB */
6785 if (leaf->mn_flags & F_SUBDATA) {
6786 /* add all the child DB's pages to the free list */
6787 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6792 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6793 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6794 rc = MDB_INCOMPATIBLE;
6798 /* add overflow pages to free list */
6799 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6803 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6804 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6805 (rc = mdb_ovpage_free(mc, omp)))
6810 return mdb_cursor_del0(mc);
6813 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6817 /** Allocate and initialize new pages for a database.
6818 * @param[in] mc a cursor on the database being added to.
6819 * @param[in] flags flags defining what type of page is being allocated.
6820 * @param[in] num the number of pages to allocate. This is usually 1,
6821 * unless allocating overflow pages for a large record.
6822 * @param[out] mp Address of a page, or NULL on failure.
6823 * @return 0 on success, non-zero on failure.
6826 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6831 if ((rc = mdb_page_alloc(mc, num, &np)))
6833 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6834 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6835 np->mp_flags = flags | P_DIRTY;
6836 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6837 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6840 mc->mc_db->md_branch_pages++;
6841 else if (IS_LEAF(np))
6842 mc->mc_db->md_leaf_pages++;
6843 else if (IS_OVERFLOW(np)) {
6844 mc->mc_db->md_overflow_pages += num;
6852 /** Calculate the size of a leaf node.
6853 * The size depends on the environment's page size; if a data item
6854 * is too large it will be put onto an overflow page and the node
6855 * size will only include the key and not the data. Sizes are always
6856 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6857 * of the #MDB_node headers.
6858 * @param[in] env The environment handle.
6859 * @param[in] key The key for the node.
6860 * @param[in] data The data for the node.
6861 * @return The number of bytes needed to store the node.
6864 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6868 sz = LEAFSIZE(key, data);
6869 if (sz > env->me_nodemax) {
6870 /* put on overflow page */
6871 sz -= data->mv_size - sizeof(pgno_t);
6874 return EVEN(sz + sizeof(indx_t));
6877 /** Calculate the size of a branch node.
6878 * The size should depend on the environment's page size but since
6879 * we currently don't support spilling large keys onto overflow
6880 * pages, it's simply the size of the #MDB_node header plus the
6881 * size of the key. Sizes are always rounded up to an even number
6882 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6883 * @param[in] env The environment handle.
6884 * @param[in] key The key for the node.
6885 * @return The number of bytes needed to store the node.
6888 mdb_branch_size(MDB_env *env, MDB_val *key)
6893 if (sz > env->me_nodemax) {
6894 /* put on overflow page */
6895 /* not implemented */
6896 /* sz -= key->size - sizeof(pgno_t); */
6899 return sz + sizeof(indx_t);
6902 /** Add a node to the page pointed to by the cursor.
6903 * @param[in] mc The cursor for this operation.
6904 * @param[in] indx The index on the page where the new node should be added.
6905 * @param[in] key The key for the new node.
6906 * @param[in] data The data for the new node, if any.
6907 * @param[in] pgno The page number, if adding a branch node.
6908 * @param[in] flags Flags for the node.
6909 * @return 0 on success, non-zero on failure. Possible errors are:
6911 * <li>ENOMEM - failed to allocate overflow pages for the node.
6912 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6913 * should never happen since all callers already calculate the
6914 * page's free space before calling this function.
6918 mdb_node_add(MDB_cursor *mc, indx_t indx,
6919 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6922 size_t node_size = NODESIZE;
6926 MDB_page *mp = mc->mc_pg[mc->mc_top];
6927 MDB_page *ofp = NULL; /* overflow page */
6930 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6932 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6933 IS_LEAF(mp) ? "leaf" : "branch",
6934 IS_SUBP(mp) ? "sub-" : "",
6935 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6936 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6939 /* Move higher keys up one slot. */
6940 int ksize = mc->mc_db->md_pad, dif;
6941 char *ptr = LEAF2KEY(mp, indx, ksize);
6942 dif = NUMKEYS(mp) - indx;
6944 memmove(ptr+ksize, ptr, dif*ksize);
6945 /* insert new key */
6946 memcpy(ptr, key->mv_data, ksize);
6948 /* Just using these for counting */
6949 mp->mp_lower += sizeof(indx_t);
6950 mp->mp_upper -= ksize - sizeof(indx_t);
6954 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6956 node_size += key->mv_size;
6958 mdb_cassert(mc, data);
6959 if (F_ISSET(flags, F_BIGDATA)) {
6960 /* Data already on overflow page. */
6961 node_size += sizeof(pgno_t);
6962 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6963 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6965 /* Put data on overflow page. */
6966 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6967 data->mv_size, node_size+data->mv_size));
6968 node_size = EVEN(node_size + sizeof(pgno_t));
6969 if ((ssize_t)node_size > room)
6971 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6973 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6977 node_size += data->mv_size;
6980 node_size = EVEN(node_size);
6981 if ((ssize_t)node_size > room)
6985 /* Move higher pointers up one slot. */
6986 for (i = NUMKEYS(mp); i > indx; i--)
6987 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6989 /* Adjust free space offsets. */
6990 ofs = mp->mp_upper - node_size;
6991 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6992 mp->mp_ptrs[indx] = ofs;
6994 mp->mp_lower += sizeof(indx_t);
6996 /* Write the node data. */
6997 node = NODEPTR(mp, indx);
6998 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6999 node->mn_flags = flags;
7001 SETDSZ(node,data->mv_size);
7006 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7009 mdb_cassert(mc, key);
7011 if (F_ISSET(flags, F_BIGDATA))
7012 memcpy(node->mn_data + key->mv_size, data->mv_data,
7014 else if (F_ISSET(flags, MDB_RESERVE))
7015 data->mv_data = node->mn_data + key->mv_size;
7017 memcpy(node->mn_data + key->mv_size, data->mv_data,
7020 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
7022 if (F_ISSET(flags, MDB_RESERVE))
7023 data->mv_data = METADATA(ofp);
7025 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
7032 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7033 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7034 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7035 DPRINTF(("node size = %"Z"u", node_size));
7036 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7037 return MDB_PAGE_FULL;
7040 /** Delete the specified node from a page.
7041 * @param[in] mc Cursor pointing to the node to delete.
7042 * @param[in] ksize The size of a node. Only used if the page is
7043 * part of a #MDB_DUPFIXED database.
7046 mdb_node_del(MDB_cursor *mc, int ksize)
7048 MDB_page *mp = mc->mc_pg[mc->mc_top];
7049 indx_t indx = mc->mc_ki[mc->mc_top];
7051 indx_t i, j, numkeys, ptr;
7055 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7056 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7057 numkeys = NUMKEYS(mp);
7058 mdb_cassert(mc, indx < numkeys);
7061 int x = numkeys - 1 - indx;
7062 base = LEAF2KEY(mp, indx, ksize);
7064 memmove(base, base + ksize, x * ksize);
7065 mp->mp_lower -= sizeof(indx_t);
7066 mp->mp_upper += ksize - sizeof(indx_t);
7070 node = NODEPTR(mp, indx);
7071 sz = NODESIZE + node->mn_ksize;
7073 if (F_ISSET(node->mn_flags, F_BIGDATA))
7074 sz += sizeof(pgno_t);
7076 sz += NODEDSZ(node);
7080 ptr = mp->mp_ptrs[indx];
7081 for (i = j = 0; i < numkeys; i++) {
7083 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7084 if (mp->mp_ptrs[i] < ptr)
7085 mp->mp_ptrs[j] += sz;
7090 base = (char *)mp + mp->mp_upper + PAGEBASE;
7091 memmove(base + sz, base, ptr - mp->mp_upper);
7093 mp->mp_lower -= sizeof(indx_t);
7097 /** Compact the main page after deleting a node on a subpage.
7098 * @param[in] mp The main page to operate on.
7099 * @param[in] indx The index of the subpage on the main page.
7102 mdb_node_shrink(MDB_page *mp, indx_t indx)
7108 indx_t i, numkeys, ptr;
7110 node = NODEPTR(mp, indx);
7111 sp = (MDB_page *)NODEDATA(node);
7112 delta = SIZELEFT(sp);
7113 xp = (MDB_page *)((char *)sp + delta);
7115 /* shift subpage upward */
7117 nsize = NUMKEYS(sp) * sp->mp_pad;
7119 return; /* do not make the node uneven-sized */
7120 memmove(METADATA(xp), METADATA(sp), nsize);
7123 numkeys = NUMKEYS(sp);
7124 for (i=numkeys-1; i>=0; i--)
7125 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7127 xp->mp_upper = sp->mp_lower;
7128 xp->mp_lower = sp->mp_lower;
7129 xp->mp_flags = sp->mp_flags;
7130 xp->mp_pad = sp->mp_pad;
7131 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7133 nsize = NODEDSZ(node) - delta;
7134 SETDSZ(node, nsize);
7136 /* shift lower nodes upward */
7137 ptr = mp->mp_ptrs[indx];
7138 numkeys = NUMKEYS(mp);
7139 for (i = 0; i < numkeys; i++) {
7140 if (mp->mp_ptrs[i] <= ptr)
7141 mp->mp_ptrs[i] += delta;
7144 base = (char *)mp + mp->mp_upper + PAGEBASE;
7145 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7146 mp->mp_upper += delta;
7149 /** Initial setup of a sorted-dups cursor.
7150 * Sorted duplicates are implemented as a sub-database for the given key.
7151 * The duplicate data items are actually keys of the sub-database.
7152 * Operations on the duplicate data items are performed using a sub-cursor
7153 * initialized when the sub-database is first accessed. This function does
7154 * the preliminary setup of the sub-cursor, filling in the fields that
7155 * depend only on the parent DB.
7156 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7159 mdb_xcursor_init0(MDB_cursor *mc)
7161 MDB_xcursor *mx = mc->mc_xcursor;
7163 mx->mx_cursor.mc_xcursor = NULL;
7164 mx->mx_cursor.mc_txn = mc->mc_txn;
7165 mx->mx_cursor.mc_db = &mx->mx_db;
7166 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7167 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7168 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7169 mx->mx_cursor.mc_snum = 0;
7170 mx->mx_cursor.mc_top = 0;
7171 mx->mx_cursor.mc_flags = C_SUB;
7172 mx->mx_dbx.md_name.mv_size = 0;
7173 mx->mx_dbx.md_name.mv_data = NULL;
7174 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7175 mx->mx_dbx.md_dcmp = NULL;
7176 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7179 /** Final setup of a sorted-dups cursor.
7180 * Sets up the fields that depend on the data from the main cursor.
7181 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7182 * @param[in] node The data containing the #MDB_db record for the
7183 * sorted-dup database.
7186 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7188 MDB_xcursor *mx = mc->mc_xcursor;
7190 if (node->mn_flags & F_SUBDATA) {
7191 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7192 mx->mx_cursor.mc_pg[0] = 0;
7193 mx->mx_cursor.mc_snum = 0;
7194 mx->mx_cursor.mc_top = 0;
7195 mx->mx_cursor.mc_flags = C_SUB;
7197 MDB_page *fp = NODEDATA(node);
7198 mx->mx_db.md_pad = 0;
7199 mx->mx_db.md_flags = 0;
7200 mx->mx_db.md_depth = 1;
7201 mx->mx_db.md_branch_pages = 0;
7202 mx->mx_db.md_leaf_pages = 1;
7203 mx->mx_db.md_overflow_pages = 0;
7204 mx->mx_db.md_entries = NUMKEYS(fp);
7205 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7206 mx->mx_cursor.mc_snum = 1;
7207 mx->mx_cursor.mc_top = 0;
7208 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7209 mx->mx_cursor.mc_pg[0] = fp;
7210 mx->mx_cursor.mc_ki[0] = 0;
7211 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7212 mx->mx_db.md_flags = MDB_DUPFIXED;
7213 mx->mx_db.md_pad = fp->mp_pad;
7214 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7215 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7218 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7219 mx->mx_db.md_root));
7220 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7221 #if UINT_MAX < SIZE_MAX
7222 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7223 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7228 /** Fixup a sorted-dups cursor due to underlying update.
7229 * Sets up some fields that depend on the data from the main cursor.
7230 * Almost the same as init1, but skips initialization steps if the
7231 * xcursor had already been used.
7232 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7233 * @param[in] src_mx The xcursor of an up-to-date cursor.
7234 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7237 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7239 MDB_xcursor *mx = mc->mc_xcursor;
7242 mx->mx_cursor.mc_snum = 1;
7243 mx->mx_cursor.mc_top = 0;
7244 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7245 mx->mx_cursor.mc_ki[0] = 0;
7246 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7247 #if UINT_MAX < SIZE_MAX
7248 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7250 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7253 mx->mx_db = src_mx->mx_db;
7254 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7255 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7256 mx->mx_db.md_root));
7259 /** Initialize a cursor for a given transaction and database. */
7261 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7264 mc->mc_backup = NULL;
7267 mc->mc_db = &txn->mt_dbs[dbi];
7268 mc->mc_dbx = &txn->mt_dbxs[dbi];
7269 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7275 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7276 mdb_tassert(txn, mx != NULL);
7277 mc->mc_xcursor = mx;
7278 mdb_xcursor_init0(mc);
7280 mc->mc_xcursor = NULL;
7282 if (*mc->mc_dbflag & DB_STALE) {
7283 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7288 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7291 size_t size = sizeof(MDB_cursor);
7293 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7296 if (txn->mt_flags & MDB_TXN_ERROR)
7299 /* Allow read access to the freelist */
7300 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7303 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7304 size += sizeof(MDB_xcursor);
7306 if ((mc = malloc(size)) != NULL) {
7307 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7308 if (txn->mt_cursors) {
7309 mc->mc_next = txn->mt_cursors[dbi];
7310 txn->mt_cursors[dbi] = mc;
7311 mc->mc_flags |= C_UNTRACK;
7323 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7325 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7328 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7331 if (txn->mt_flags & MDB_TXN_ERROR)
7334 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7338 /* Return the count of duplicate data items for the current key */
7340 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7344 if (mc == NULL || countp == NULL)
7347 if (mc->mc_xcursor == NULL)
7348 return MDB_INCOMPATIBLE;
7350 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7353 if (!(mc->mc_flags & C_INITIALIZED))
7356 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7357 return MDB_NOTFOUND;
7359 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7360 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7363 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7366 *countp = mc->mc_xcursor->mx_db.md_entries;
7372 mdb_cursor_close(MDB_cursor *mc)
7374 if (mc && !mc->mc_backup) {
7375 /* remove from txn, if tracked */
7376 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7377 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7378 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7380 *prev = mc->mc_next;
7387 mdb_cursor_txn(MDB_cursor *mc)
7389 if (!mc) return NULL;
7394 mdb_cursor_dbi(MDB_cursor *mc)
7399 /** Replace the key for a branch node with a new key.
7400 * @param[in] mc Cursor pointing to the node to operate on.
7401 * @param[in] key The new key to use.
7402 * @return 0 on success, non-zero on failure.
7405 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7411 int delta, ksize, oksize;
7412 indx_t ptr, i, numkeys, indx;
7415 indx = mc->mc_ki[mc->mc_top];
7416 mp = mc->mc_pg[mc->mc_top];
7417 node = NODEPTR(mp, indx);
7418 ptr = mp->mp_ptrs[indx];
7422 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7423 k2.mv_data = NODEKEY(node);
7424 k2.mv_size = node->mn_ksize;
7425 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7427 mdb_dkey(&k2, kbuf2),
7433 /* Sizes must be 2-byte aligned. */
7434 ksize = EVEN(key->mv_size);
7435 oksize = EVEN(node->mn_ksize);
7436 delta = ksize - oksize;
7438 /* Shift node contents if EVEN(key length) changed. */
7440 if (delta > 0 && SIZELEFT(mp) < delta) {
7442 /* not enough space left, do a delete and split */
7443 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7444 pgno = NODEPGNO(node);
7445 mdb_node_del(mc, 0);
7446 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7449 numkeys = NUMKEYS(mp);
7450 for (i = 0; i < numkeys; i++) {
7451 if (mp->mp_ptrs[i] <= ptr)
7452 mp->mp_ptrs[i] -= delta;
7455 base = (char *)mp + mp->mp_upper + PAGEBASE;
7456 len = ptr - mp->mp_upper + NODESIZE;
7457 memmove(base - delta, base, len);
7458 mp->mp_upper -= delta;
7460 node = NODEPTR(mp, indx);
7463 /* But even if no shift was needed, update ksize */
7464 if (node->mn_ksize != key->mv_size)
7465 node->mn_ksize = key->mv_size;
7468 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7474 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7476 /** Move a node from csrc to cdst.
7479 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7486 unsigned short flags;
7490 /* Mark src and dst as dirty. */
7491 if ((rc = mdb_page_touch(csrc)) ||
7492 (rc = mdb_page_touch(cdst)))
7495 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7496 key.mv_size = csrc->mc_db->md_pad;
7497 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7499 data.mv_data = NULL;
7503 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7504 mdb_cassert(csrc, !((size_t)srcnode & 1));
7505 srcpg = NODEPGNO(srcnode);
7506 flags = srcnode->mn_flags;
7507 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7508 unsigned int snum = csrc->mc_snum;
7510 /* must find the lowest key below src */
7511 rc = mdb_page_search_lowest(csrc);
7514 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7515 key.mv_size = csrc->mc_db->md_pad;
7516 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7518 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7519 key.mv_size = NODEKSZ(s2);
7520 key.mv_data = NODEKEY(s2);
7522 csrc->mc_snum = snum--;
7523 csrc->mc_top = snum;
7525 key.mv_size = NODEKSZ(srcnode);
7526 key.mv_data = NODEKEY(srcnode);
7528 data.mv_size = NODEDSZ(srcnode);
7529 data.mv_data = NODEDATA(srcnode);
7531 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7532 unsigned int snum = cdst->mc_snum;
7535 /* must find the lowest key below dst */
7536 mdb_cursor_copy(cdst, &mn);
7537 rc = mdb_page_search_lowest(&mn);
7540 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7541 bkey.mv_size = mn.mc_db->md_pad;
7542 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7544 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7545 bkey.mv_size = NODEKSZ(s2);
7546 bkey.mv_data = NODEKEY(s2);
7548 mn.mc_snum = snum--;
7551 rc = mdb_update_key(&mn, &bkey);
7556 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7557 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7558 csrc->mc_ki[csrc->mc_top],
7560 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7561 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7563 /* Add the node to the destination page.
7565 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7566 if (rc != MDB_SUCCESS)
7569 /* Delete the node from the source page.
7571 mdb_node_del(csrc, key.mv_size);
7574 /* Adjust other cursors pointing to mp */
7575 MDB_cursor *m2, *m3;
7576 MDB_dbi dbi = csrc->mc_dbi;
7579 mp = cdst->mc_pg[csrc->mc_top];
7580 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7581 if (csrc->mc_flags & C_SUB)
7582 m3 = &m2->mc_xcursor->mx_cursor;
7585 if (m3 == cdst) continue;
7586 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] >=
7587 cdst->mc_ki[csrc->mc_top]) {
7588 m3->mc_ki[csrc->mc_top]++;
7592 mp = csrc->mc_pg[csrc->mc_top];
7593 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7594 if (csrc->mc_flags & C_SUB)
7595 m3 = &m2->mc_xcursor->mx_cursor;
7598 if (m3 == csrc) continue;
7599 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7600 csrc->mc_ki[csrc->mc_top]) {
7601 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7602 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7607 /* Update the parent separators.
7609 if (csrc->mc_ki[csrc->mc_top] == 0) {
7610 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7611 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7612 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7614 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7615 key.mv_size = NODEKSZ(srcnode);
7616 key.mv_data = NODEKEY(srcnode);
7618 DPRINTF(("update separator for source page %"Z"u to [%s]",
7619 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7620 mdb_cursor_copy(csrc, &mn);
7623 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7626 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7628 indx_t ix = csrc->mc_ki[csrc->mc_top];
7629 nullkey.mv_size = 0;
7630 csrc->mc_ki[csrc->mc_top] = 0;
7631 rc = mdb_update_key(csrc, &nullkey);
7632 csrc->mc_ki[csrc->mc_top] = ix;
7633 mdb_cassert(csrc, rc == MDB_SUCCESS);
7637 if (cdst->mc_ki[cdst->mc_top] == 0) {
7638 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7639 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7640 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7642 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7643 key.mv_size = NODEKSZ(srcnode);
7644 key.mv_data = NODEKEY(srcnode);
7646 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7647 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7648 mdb_cursor_copy(cdst, &mn);
7651 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7654 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7656 indx_t ix = cdst->mc_ki[cdst->mc_top];
7657 nullkey.mv_size = 0;
7658 cdst->mc_ki[cdst->mc_top] = 0;
7659 rc = mdb_update_key(cdst, &nullkey);
7660 cdst->mc_ki[cdst->mc_top] = ix;
7661 mdb_cassert(cdst, rc == MDB_SUCCESS);
7668 /** Merge one page into another.
7669 * The nodes from the page pointed to by \b csrc will
7670 * be copied to the page pointed to by \b cdst and then
7671 * the \b csrc page will be freed.
7672 * @param[in] csrc Cursor pointing to the source page.
7673 * @param[in] cdst Cursor pointing to the destination page.
7674 * @return 0 on success, non-zero on failure.
7677 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7679 MDB_page *psrc, *pdst;
7686 psrc = csrc->mc_pg[csrc->mc_top];
7687 pdst = cdst->mc_pg[cdst->mc_top];
7689 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7691 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7692 mdb_cassert(csrc, cdst->mc_snum > 1);
7694 /* Mark dst as dirty. */
7695 if ((rc = mdb_page_touch(cdst)))
7698 /* Move all nodes from src to dst.
7700 j = nkeys = NUMKEYS(pdst);
7701 if (IS_LEAF2(psrc)) {
7702 key.mv_size = csrc->mc_db->md_pad;
7703 key.mv_data = METADATA(psrc);
7704 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7705 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7706 if (rc != MDB_SUCCESS)
7708 key.mv_data = (char *)key.mv_data + key.mv_size;
7711 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7712 srcnode = NODEPTR(psrc, i);
7713 if (i == 0 && IS_BRANCH(psrc)) {
7716 mdb_cursor_copy(csrc, &mn);
7717 /* must find the lowest key below src */
7718 rc = mdb_page_search_lowest(&mn);
7721 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7722 key.mv_size = mn.mc_db->md_pad;
7723 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7725 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7726 key.mv_size = NODEKSZ(s2);
7727 key.mv_data = NODEKEY(s2);
7730 key.mv_size = srcnode->mn_ksize;
7731 key.mv_data = NODEKEY(srcnode);
7734 data.mv_size = NODEDSZ(srcnode);
7735 data.mv_data = NODEDATA(srcnode);
7736 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7737 if (rc != MDB_SUCCESS)
7742 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7743 pdst->mp_pgno, NUMKEYS(pdst),
7744 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7746 /* Unlink the src page from parent and add to free list.
7749 mdb_node_del(csrc, 0);
7750 if (csrc->mc_ki[csrc->mc_top] == 0) {
7752 rc = mdb_update_key(csrc, &key);
7760 psrc = csrc->mc_pg[csrc->mc_top];
7761 /* If not operating on FreeDB, allow this page to be reused
7762 * in this txn. Otherwise just add to free list.
7764 rc = mdb_page_loose(csrc, psrc);
7768 csrc->mc_db->md_leaf_pages--;
7770 csrc->mc_db->md_branch_pages--;
7772 /* Adjust other cursors pointing to mp */
7773 MDB_cursor *m2, *m3;
7774 MDB_dbi dbi = csrc->mc_dbi;
7776 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7777 if (csrc->mc_flags & C_SUB)
7778 m3 = &m2->mc_xcursor->mx_cursor;
7781 if (m3 == csrc) continue;
7782 if (m3->mc_snum < csrc->mc_snum) continue;
7783 if (m3->mc_pg[csrc->mc_top] == psrc) {
7784 m3->mc_pg[csrc->mc_top] = pdst;
7785 m3->mc_ki[csrc->mc_top] += nkeys;
7790 unsigned int snum = cdst->mc_snum;
7791 uint16_t depth = cdst->mc_db->md_depth;
7792 mdb_cursor_pop(cdst);
7793 rc = mdb_rebalance(cdst);
7794 /* Did the tree height change? */
7795 if (depth != cdst->mc_db->md_depth)
7796 snum += cdst->mc_db->md_depth - depth;
7797 cdst->mc_snum = snum;
7798 cdst->mc_top = snum-1;
7803 /** Copy the contents of a cursor.
7804 * @param[in] csrc The cursor to copy from.
7805 * @param[out] cdst The cursor to copy to.
7808 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7812 cdst->mc_txn = csrc->mc_txn;
7813 cdst->mc_dbi = csrc->mc_dbi;
7814 cdst->mc_db = csrc->mc_db;
7815 cdst->mc_dbx = csrc->mc_dbx;
7816 cdst->mc_snum = csrc->mc_snum;
7817 cdst->mc_top = csrc->mc_top;
7818 cdst->mc_flags = csrc->mc_flags;
7820 for (i=0; i<csrc->mc_snum; i++) {
7821 cdst->mc_pg[i] = csrc->mc_pg[i];
7822 cdst->mc_ki[i] = csrc->mc_ki[i];
7826 /** Rebalance the tree after a delete operation.
7827 * @param[in] mc Cursor pointing to the page where rebalancing
7829 * @return 0 on success, non-zero on failure.
7832 mdb_rebalance(MDB_cursor *mc)
7836 unsigned int ptop, minkeys, thresh;
7840 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7845 thresh = FILL_THRESHOLD;
7847 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7848 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7849 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7850 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7852 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7853 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7854 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7855 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7859 if (mc->mc_snum < 2) {
7860 MDB_page *mp = mc->mc_pg[0];
7862 DPUTS("Can't rebalance a subpage, ignoring");
7865 if (NUMKEYS(mp) == 0) {
7866 DPUTS("tree is completely empty");
7867 mc->mc_db->md_root = P_INVALID;
7868 mc->mc_db->md_depth = 0;
7869 mc->mc_db->md_leaf_pages = 0;
7870 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7873 /* Adjust cursors pointing to mp */
7876 mc->mc_flags &= ~C_INITIALIZED;
7878 MDB_cursor *m2, *m3;
7879 MDB_dbi dbi = mc->mc_dbi;
7881 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7882 if (mc->mc_flags & C_SUB)
7883 m3 = &m2->mc_xcursor->mx_cursor;
7886 if (m3->mc_snum < mc->mc_snum) continue;
7887 if (m3->mc_pg[0] == mp) {
7890 m3->mc_flags &= ~C_INITIALIZED;
7894 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7896 DPUTS("collapsing root page!");
7897 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7900 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7901 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7904 mc->mc_db->md_depth--;
7905 mc->mc_db->md_branch_pages--;
7906 mc->mc_ki[0] = mc->mc_ki[1];
7907 for (i = 1; i<mc->mc_db->md_depth; i++) {
7908 mc->mc_pg[i] = mc->mc_pg[i+1];
7909 mc->mc_ki[i] = mc->mc_ki[i+1];
7912 /* Adjust other cursors pointing to mp */
7913 MDB_cursor *m2, *m3;
7914 MDB_dbi dbi = mc->mc_dbi;
7916 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7917 if (mc->mc_flags & C_SUB)
7918 m3 = &m2->mc_xcursor->mx_cursor;
7921 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7922 if (m3->mc_pg[0] == mp) {
7923 for (i=0; i<m3->mc_snum; i++) {
7924 m3->mc_pg[i] = m3->mc_pg[i+1];
7925 m3->mc_ki[i] = m3->mc_ki[i+1];
7933 DPUTS("root page doesn't need rebalancing");
7937 /* The parent (branch page) must have at least 2 pointers,
7938 * otherwise the tree is invalid.
7940 ptop = mc->mc_top-1;
7941 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7943 /* Leaf page fill factor is below the threshold.
7944 * Try to move keys from left or right neighbor, or
7945 * merge with a neighbor page.
7950 mdb_cursor_copy(mc, &mn);
7951 mn.mc_xcursor = NULL;
7953 oldki = mc->mc_ki[mc->mc_top];
7954 if (mc->mc_ki[ptop] == 0) {
7955 /* We're the leftmost leaf in our parent.
7957 DPUTS("reading right neighbor");
7959 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7960 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7963 mn.mc_ki[mn.mc_top] = 0;
7964 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7966 /* There is at least one neighbor to the left.
7968 DPUTS("reading left neighbor");
7970 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7971 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7974 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7975 mc->mc_ki[mc->mc_top] = 0;
7978 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7979 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7980 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7982 /* If the neighbor page is above threshold and has enough keys,
7983 * move one key from it. Otherwise we should try to merge them.
7984 * (A branch page must never have less than 2 keys.)
7986 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7987 rc = mdb_node_move(&mn, mc);
7988 if (mc->mc_ki[mc->mc_top-1]) {
7992 if (mc->mc_ki[ptop] == 0) {
7993 rc = mdb_page_merge(&mn, mc);
7996 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7997 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7998 /* We want mdb_rebalance to find mn when doing fixups */
7999 if (mc->mc_flags & C_SUB) {
8000 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8001 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8002 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8004 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8005 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8007 rc = mdb_page_merge(mc, &mn);
8008 if (mc->mc_flags & C_SUB)
8009 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8011 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8012 mdb_cursor_copy(&mn, mc);
8014 mc->mc_flags &= ~C_EOF;
8016 mc->mc_ki[mc->mc_top] = oldki;
8020 /** Complete a delete operation started by #mdb_cursor_del(). */
8022 mdb_cursor_del0(MDB_cursor *mc)
8028 MDB_cursor *m2, *m3;
8029 MDB_dbi dbi = mc->mc_dbi;
8031 ki = mc->mc_ki[mc->mc_top];
8032 mp = mc->mc_pg[mc->mc_top];
8033 mdb_node_del(mc, mc->mc_db->md_pad);
8034 mc->mc_db->md_entries--;
8036 /* Adjust other cursors pointing to mp */
8037 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8038 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8039 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8041 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8043 if (m3->mc_pg[mc->mc_top] == mp) {
8044 if (m3->mc_ki[mc->mc_top] >= ki) {
8045 m3->mc_flags |= C_DEL;
8046 if (m3->mc_ki[mc->mc_top] > ki)
8047 m3->mc_ki[mc->mc_top]--;
8048 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8049 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8054 rc = mdb_rebalance(mc);
8056 if (rc == MDB_SUCCESS) {
8057 /* DB is totally empty now, just bail out.
8058 * Other cursors adjustments were already done
8059 * by mdb_rebalance and aren't needed here.
8064 mp = mc->mc_pg[mc->mc_top];
8065 nkeys = NUMKEYS(mp);
8067 /* Adjust other cursors pointing to mp */
8068 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8069 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8070 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8072 if (m3->mc_snum < mc->mc_snum)
8074 if (m3->mc_pg[mc->mc_top] == mp) {
8075 /* if m3 points past last node in page, find next sibling */
8076 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8077 rc = mdb_cursor_sibling(m3, 1);
8078 if (rc == MDB_NOTFOUND) {
8079 m3->mc_flags |= C_EOF;
8085 mc->mc_flags |= C_DEL;
8089 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8094 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8095 MDB_val *key, MDB_val *data)
8097 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8100 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8101 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8103 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8104 /* must ignore any data */
8108 return mdb_del0(txn, dbi, key, data, 0);
8112 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8113 MDB_val *key, MDB_val *data, unsigned flags)
8118 MDB_val rdata, *xdata;
8122 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8124 mdb_cursor_init(&mc, txn, dbi, &mx);
8133 flags |= MDB_NODUPDATA;
8135 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8137 /* let mdb_page_split know about this cursor if needed:
8138 * delete will trigger a rebalance; if it needs to move
8139 * a node from one page to another, it will have to
8140 * update the parent's separator key(s). If the new sepkey
8141 * is larger than the current one, the parent page may
8142 * run out of space, triggering a split. We need this
8143 * cursor to be consistent until the end of the rebalance.
8145 mc.mc_flags |= C_UNTRACK;
8146 mc.mc_next = txn->mt_cursors[dbi];
8147 txn->mt_cursors[dbi] = &mc;
8148 rc = mdb_cursor_del(&mc, flags);
8149 txn->mt_cursors[dbi] = mc.mc_next;
8154 /** Split a page and insert a new node.
8155 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8156 * The cursor will be updated to point to the actual page and index where
8157 * the node got inserted after the split.
8158 * @param[in] newkey The key for the newly inserted node.
8159 * @param[in] newdata The data for the newly inserted node.
8160 * @param[in] newpgno The page number, if the new node is a branch node.
8161 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8162 * @return 0 on success, non-zero on failure.
8165 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8166 unsigned int nflags)
8169 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8172 int i, j, split_indx, nkeys, pmax;
8173 MDB_env *env = mc->mc_txn->mt_env;
8175 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8176 MDB_page *copy = NULL;
8177 MDB_page *mp, *rp, *pp;
8182 mp = mc->mc_pg[mc->mc_top];
8183 newindx = mc->mc_ki[mc->mc_top];
8184 nkeys = NUMKEYS(mp);
8186 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8187 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8188 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8190 /* Create a right sibling. */
8191 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8193 rp->mp_pad = mp->mp_pad;
8194 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8196 if (mc->mc_snum < 2) {
8197 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8199 /* shift current top to make room for new parent */
8200 mc->mc_pg[1] = mc->mc_pg[0];
8201 mc->mc_ki[1] = mc->mc_ki[0];
8204 mc->mc_db->md_root = pp->mp_pgno;
8205 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8206 new_root = mc->mc_db->md_depth++;
8208 /* Add left (implicit) pointer. */
8209 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8210 /* undo the pre-push */
8211 mc->mc_pg[0] = mc->mc_pg[1];
8212 mc->mc_ki[0] = mc->mc_ki[1];
8213 mc->mc_db->md_root = mp->mp_pgno;
8214 mc->mc_db->md_depth--;
8221 ptop = mc->mc_top-1;
8222 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8225 mc->mc_flags |= C_SPLITTING;
8226 mdb_cursor_copy(mc, &mn);
8227 mn.mc_pg[mn.mc_top] = rp;
8228 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8230 if (nflags & MDB_APPEND) {
8231 mn.mc_ki[mn.mc_top] = 0;
8233 split_indx = newindx;
8237 split_indx = (nkeys+1) / 2;
8242 unsigned int lsize, rsize, ksize;
8243 /* Move half of the keys to the right sibling */
8244 x = mc->mc_ki[mc->mc_top] - split_indx;
8245 ksize = mc->mc_db->md_pad;
8246 split = LEAF2KEY(mp, split_indx, ksize);
8247 rsize = (nkeys - split_indx) * ksize;
8248 lsize = (nkeys - split_indx) * sizeof(indx_t);
8249 mp->mp_lower -= lsize;
8250 rp->mp_lower += lsize;
8251 mp->mp_upper += rsize - lsize;
8252 rp->mp_upper -= rsize - lsize;
8253 sepkey.mv_size = ksize;
8254 if (newindx == split_indx) {
8255 sepkey.mv_data = newkey->mv_data;
8257 sepkey.mv_data = split;
8260 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8261 memcpy(rp->mp_ptrs, split, rsize);
8262 sepkey.mv_data = rp->mp_ptrs;
8263 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8264 memcpy(ins, newkey->mv_data, ksize);
8265 mp->mp_lower += sizeof(indx_t);
8266 mp->mp_upper -= ksize - sizeof(indx_t);
8269 memcpy(rp->mp_ptrs, split, x * ksize);
8270 ins = LEAF2KEY(rp, x, ksize);
8271 memcpy(ins, newkey->mv_data, ksize);
8272 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8273 rp->mp_lower += sizeof(indx_t);
8274 rp->mp_upper -= ksize - sizeof(indx_t);
8275 mc->mc_ki[mc->mc_top] = x;
8276 mc->mc_pg[mc->mc_top] = rp;
8279 int psize, nsize, k;
8280 /* Maximum free space in an empty page */
8281 pmax = env->me_psize - PAGEHDRSZ;
8283 nsize = mdb_leaf_size(env, newkey, newdata);
8285 nsize = mdb_branch_size(env, newkey);
8286 nsize = EVEN(nsize);
8288 /* grab a page to hold a temporary copy */
8289 copy = mdb_page_malloc(mc->mc_txn, 1);
8294 copy->mp_pgno = mp->mp_pgno;
8295 copy->mp_flags = mp->mp_flags;
8296 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8297 copy->mp_upper = env->me_psize - PAGEBASE;
8299 /* prepare to insert */
8300 for (i=0, j=0; i<nkeys; i++) {
8302 copy->mp_ptrs[j++] = 0;
8304 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8307 /* When items are relatively large the split point needs
8308 * to be checked, because being off-by-one will make the
8309 * difference between success or failure in mdb_node_add.
8311 * It's also relevant if a page happens to be laid out
8312 * such that one half of its nodes are all "small" and
8313 * the other half of its nodes are "large." If the new
8314 * item is also "large" and falls on the half with
8315 * "large" nodes, it also may not fit.
8317 * As a final tweak, if the new item goes on the last
8318 * spot on the page (and thus, onto the new page), bias
8319 * the split so the new page is emptier than the old page.
8320 * This yields better packing during sequential inserts.
8322 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8323 /* Find split point */
8325 if (newindx <= split_indx || newindx >= nkeys) {
8327 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8332 for (; i!=k; i+=j) {
8337 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8338 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8340 if (F_ISSET(node->mn_flags, F_BIGDATA))
8341 psize += sizeof(pgno_t);
8343 psize += NODEDSZ(node);
8345 psize = EVEN(psize);
8347 if (psize > pmax || i == k-j) {
8348 split_indx = i + (j<0);
8353 if (split_indx == newindx) {
8354 sepkey.mv_size = newkey->mv_size;
8355 sepkey.mv_data = newkey->mv_data;
8357 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8358 sepkey.mv_size = node->mn_ksize;
8359 sepkey.mv_data = NODEKEY(node);
8364 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8366 /* Copy separator key to the parent.
8368 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8372 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8377 if (mn.mc_snum == mc->mc_snum) {
8378 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8379 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8380 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8381 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8386 /* Right page might now have changed parent.
8387 * Check if left page also changed parent.
8389 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8390 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8391 for (i=0; i<ptop; i++) {
8392 mc->mc_pg[i] = mn.mc_pg[i];
8393 mc->mc_ki[i] = mn.mc_ki[i];
8395 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8396 if (mn.mc_ki[ptop]) {
8397 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8399 /* find right page's left sibling */
8400 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8401 mdb_cursor_sibling(mc, 0);
8406 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8409 mc->mc_flags ^= C_SPLITTING;
8410 if (rc != MDB_SUCCESS) {
8413 if (nflags & MDB_APPEND) {
8414 mc->mc_pg[mc->mc_top] = rp;
8415 mc->mc_ki[mc->mc_top] = 0;
8416 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8419 for (i=0; i<mc->mc_top; i++)
8420 mc->mc_ki[i] = mn.mc_ki[i];
8421 } else if (!IS_LEAF2(mp)) {
8423 mc->mc_pg[mc->mc_top] = rp;
8428 rkey.mv_data = newkey->mv_data;
8429 rkey.mv_size = newkey->mv_size;
8435 /* Update index for the new key. */
8436 mc->mc_ki[mc->mc_top] = j;
8438 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8439 rkey.mv_data = NODEKEY(node);
8440 rkey.mv_size = node->mn_ksize;
8442 xdata.mv_data = NODEDATA(node);
8443 xdata.mv_size = NODEDSZ(node);
8446 pgno = NODEPGNO(node);
8447 flags = node->mn_flags;
8450 if (!IS_LEAF(mp) && j == 0) {
8451 /* First branch index doesn't need key data. */
8455 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8461 mc->mc_pg[mc->mc_top] = copy;
8466 } while (i != split_indx);
8468 nkeys = NUMKEYS(copy);
8469 for (i=0; i<nkeys; i++)
8470 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8471 mp->mp_lower = copy->mp_lower;
8472 mp->mp_upper = copy->mp_upper;
8473 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8474 env->me_psize - copy->mp_upper - PAGEBASE);
8476 /* reset back to original page */
8477 if (newindx < split_indx) {
8478 mc->mc_pg[mc->mc_top] = mp;
8479 if (nflags & MDB_RESERVE) {
8480 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8481 if (!(node->mn_flags & F_BIGDATA))
8482 newdata->mv_data = NODEDATA(node);
8485 mc->mc_pg[mc->mc_top] = rp;
8487 /* Make sure mc_ki is still valid.
8489 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8490 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8491 for (i=0; i<=ptop; i++) {
8492 mc->mc_pg[i] = mn.mc_pg[i];
8493 mc->mc_ki[i] = mn.mc_ki[i];
8500 /* Adjust other cursors pointing to mp */
8501 MDB_cursor *m2, *m3;
8502 MDB_dbi dbi = mc->mc_dbi;
8503 int fixup = NUMKEYS(mp);
8505 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8506 if (mc->mc_flags & C_SUB)
8507 m3 = &m2->mc_xcursor->mx_cursor;
8512 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8514 if (m3->mc_flags & C_SPLITTING)
8519 for (k=new_root; k>=0; k--) {
8520 m3->mc_ki[k+1] = m3->mc_ki[k];
8521 m3->mc_pg[k+1] = m3->mc_pg[k];
8523 if (m3->mc_ki[0] >= split_indx) {
8528 m3->mc_pg[0] = mc->mc_pg[0];
8532 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8533 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8534 m3->mc_ki[mc->mc_top]++;
8535 if (m3->mc_ki[mc->mc_top] >= fixup) {
8536 m3->mc_pg[mc->mc_top] = rp;
8537 m3->mc_ki[mc->mc_top] -= fixup;
8538 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8540 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8541 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8546 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8549 if (copy) /* tmp page */
8550 mdb_page_free(env, copy);
8552 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8557 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8558 MDB_val *key, MDB_val *data, unsigned int flags)
8563 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8566 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8569 mdb_cursor_init(&mc, txn, dbi, &mx);
8570 return mdb_cursor_put(&mc, key, data, flags);
8574 #define MDB_WBUF (1024*1024)
8577 /** State needed for a compacting copy. */
8578 typedef struct mdb_copy {
8579 pthread_mutex_t mc_mutex;
8580 pthread_cond_t mc_cond;
8587 pgno_t mc_next_pgno;
8590 volatile int mc_new;
8595 /** Dedicated writer thread for compacting copy. */
8596 static THREAD_RET ESECT CALL_CONV
8597 mdb_env_copythr(void *arg)
8601 int toggle = 0, wsize, rc;
8604 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8607 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8610 pthread_mutex_lock(&my->mc_mutex);
8612 pthread_cond_signal(&my->mc_cond);
8615 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8616 if (my->mc_new < 0) {
8621 wsize = my->mc_wlen[toggle];
8622 ptr = my->mc_wbuf[toggle];
8625 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8629 } else if (len > 0) {
8643 /* If there's an overflow page tail, write it too */
8644 if (my->mc_olen[toggle]) {
8645 wsize = my->mc_olen[toggle];
8646 ptr = my->mc_over[toggle];
8647 my->mc_olen[toggle] = 0;
8650 my->mc_wlen[toggle] = 0;
8652 pthread_cond_signal(&my->mc_cond);
8654 pthread_cond_signal(&my->mc_cond);
8655 pthread_mutex_unlock(&my->mc_mutex);
8656 return (THREAD_RET)0;
8660 /** Tell the writer thread there's a buffer ready to write */
8662 mdb_env_cthr_toggle(mdb_copy *my, int st)
8664 int toggle = my->mc_toggle ^ 1;
8665 pthread_mutex_lock(&my->mc_mutex);
8666 if (my->mc_status) {
8667 pthread_mutex_unlock(&my->mc_mutex);
8668 return my->mc_status;
8670 while (my->mc_new == 1)
8671 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8673 my->mc_toggle = toggle;
8674 pthread_cond_signal(&my->mc_cond);
8675 pthread_mutex_unlock(&my->mc_mutex);
8679 /** Depth-first tree traversal for compacting copy. */
8681 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8684 MDB_txn *txn = my->mc_txn;
8686 MDB_page *mo, *mp, *leaf;
8691 /* Empty DB, nothing to do */
8692 if (*pg == P_INVALID)
8699 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8702 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8706 /* Make cursor pages writable */
8707 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8711 for (i=0; i<mc.mc_top; i++) {
8712 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8713 mc.mc_pg[i] = (MDB_page *)ptr;
8714 ptr += my->mc_env->me_psize;
8717 /* This is writable space for a leaf page. Usually not needed. */
8718 leaf = (MDB_page *)ptr;
8720 toggle = my->mc_toggle;
8721 while (mc.mc_snum > 0) {
8723 mp = mc.mc_pg[mc.mc_top];
8727 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8728 for (i=0; i<n; i++) {
8729 ni = NODEPTR(mp, i);
8730 if (ni->mn_flags & F_BIGDATA) {
8734 /* Need writable leaf */
8736 mc.mc_pg[mc.mc_top] = leaf;
8737 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8739 ni = NODEPTR(mp, i);
8742 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8743 rc = mdb_page_get(txn, pg, &omp, NULL);
8746 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8747 rc = mdb_env_cthr_toggle(my, 1);
8750 toggle = my->mc_toggle;
8752 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8753 memcpy(mo, omp, my->mc_env->me_psize);
8754 mo->mp_pgno = my->mc_next_pgno;
8755 my->mc_next_pgno += omp->mp_pages;
8756 my->mc_wlen[toggle] += my->mc_env->me_psize;
8757 if (omp->mp_pages > 1) {
8758 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8759 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8760 rc = mdb_env_cthr_toggle(my, 1);
8763 toggle = my->mc_toggle;
8765 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8766 } else if (ni->mn_flags & F_SUBDATA) {
8769 /* Need writable leaf */
8771 mc.mc_pg[mc.mc_top] = leaf;
8772 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8774 ni = NODEPTR(mp, i);
8777 memcpy(&db, NODEDATA(ni), sizeof(db));
8778 my->mc_toggle = toggle;
8779 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8782 toggle = my->mc_toggle;
8783 memcpy(NODEDATA(ni), &db, sizeof(db));
8788 mc.mc_ki[mc.mc_top]++;
8789 if (mc.mc_ki[mc.mc_top] < n) {
8792 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8794 rc = mdb_page_get(txn, pg, &mp, NULL);
8799 mc.mc_ki[mc.mc_top] = 0;
8800 if (IS_BRANCH(mp)) {
8801 /* Whenever we advance to a sibling branch page,
8802 * we must proceed all the way down to its first leaf.
8804 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8807 mc.mc_pg[mc.mc_top] = mp;
8811 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8812 rc = mdb_env_cthr_toggle(my, 1);
8815 toggle = my->mc_toggle;
8817 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8818 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8819 mo->mp_pgno = my->mc_next_pgno++;
8820 my->mc_wlen[toggle] += my->mc_env->me_psize;
8822 /* Update parent if there is one */
8823 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8824 SETPGNO(ni, mo->mp_pgno);
8825 mdb_cursor_pop(&mc);
8827 /* Otherwise we're done */
8837 /** Copy environment with compaction. */
8839 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8844 MDB_txn *txn = NULL;
8849 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8850 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8851 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8852 if (my.mc_wbuf[0] == NULL)
8855 pthread_mutex_init(&my.mc_mutex, NULL);
8856 pthread_cond_init(&my.mc_cond, NULL);
8857 #ifdef HAVE_MEMALIGN
8858 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8859 if (my.mc_wbuf[0] == NULL)
8862 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8867 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8868 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8873 my.mc_next_pgno = 2;
8879 THREAD_CREATE(thr, mdb_env_copythr, &my);
8881 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8885 mp = (MDB_page *)my.mc_wbuf[0];
8886 memset(mp, 0, 2*env->me_psize);
8888 mp->mp_flags = P_META;
8889 mm = (MDB_meta *)METADATA(mp);
8890 mdb_env_init_meta0(env, mm);
8891 mm->mm_address = env->me_metas[0]->mm_address;
8893 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8895 mp->mp_flags = P_META;
8896 *(MDB_meta *)METADATA(mp) = *mm;
8897 mm = (MDB_meta *)METADATA(mp);
8899 /* Count the number of free pages, subtract from lastpg to find
8900 * number of active pages
8903 MDB_ID freecount = 0;
8906 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8907 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8908 freecount += *(MDB_ID *)data.mv_data;
8909 freecount += txn->mt_dbs[0].md_branch_pages +
8910 txn->mt_dbs[0].md_leaf_pages +
8911 txn->mt_dbs[0].md_overflow_pages;
8913 /* Set metapage 1 */
8914 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8915 mm->mm_dbs[1] = txn->mt_dbs[1];
8916 if (mm->mm_last_pg > 1) {
8917 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8920 mm->mm_dbs[1].md_root = P_INVALID;
8923 my.mc_wlen[0] = env->me_psize * 2;
8925 pthread_mutex_lock(&my.mc_mutex);
8927 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8928 pthread_mutex_unlock(&my.mc_mutex);
8929 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8930 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8931 rc = mdb_env_cthr_toggle(&my, 1);
8932 mdb_env_cthr_toggle(&my, -1);
8933 pthread_mutex_lock(&my.mc_mutex);
8935 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8936 pthread_mutex_unlock(&my.mc_mutex);
8941 CloseHandle(my.mc_cond);
8942 CloseHandle(my.mc_mutex);
8943 _aligned_free(my.mc_wbuf[0]);
8945 pthread_cond_destroy(&my.mc_cond);
8946 pthread_mutex_destroy(&my.mc_mutex);
8947 free(my.mc_wbuf[0]);
8952 /** Copy environment as-is. */
8954 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8956 MDB_txn *txn = NULL;
8957 mdb_mutexref_t wmutex = NULL;
8963 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8967 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8970 /* Do the lock/unlock of the reader mutex before starting the
8971 * write txn. Otherwise other read txns could block writers.
8973 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8978 /* We must start the actual read txn after blocking writers */
8979 mdb_txn_reset0(txn, "reset-stage1");
8981 /* Temporarily block writers until we snapshot the meta pages */
8982 wmutex = env->me_wmutex;
8983 if (LOCK_MUTEX(rc, env, wmutex))
8986 rc = mdb_txn_renew0(txn);
8988 UNLOCK_MUTEX(wmutex);
8993 wsize = env->me_psize * 2;
8997 DO_WRITE(rc, fd, ptr, w2, len);
9001 } else if (len > 0) {
9007 /* Non-blocking or async handles are not supported */
9013 UNLOCK_MUTEX(wmutex);
9018 w2 = txn->mt_next_pgno * env->me_psize;
9021 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9028 if (wsize > MAX_WRITE)
9032 DO_WRITE(rc, fd, ptr, w2, len);
9036 } else if (len > 0) {
9053 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9055 if (flags & MDB_CP_COMPACT)
9056 return mdb_env_copyfd1(env, fd);
9058 return mdb_env_copyfd0(env, fd);
9062 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9064 return mdb_env_copyfd2(env, fd, 0);
9068 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9072 HANDLE newfd = INVALID_HANDLE_VALUE;
9074 if (env->me_flags & MDB_NOSUBDIR) {
9075 lpath = (char *)path;
9078 len += sizeof(DATANAME);
9079 lpath = malloc(len);
9082 sprintf(lpath, "%s" DATANAME, path);
9085 /* The destination path must exist, but the destination file must not.
9086 * We don't want the OS to cache the writes, since the source data is
9087 * already in the OS cache.
9090 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9091 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9093 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9095 if (newfd == INVALID_HANDLE_VALUE) {
9100 if (env->me_psize >= env->me_os_psize) {
9102 /* Set O_DIRECT if the file system supports it */
9103 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9104 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9106 #ifdef F_NOCACHE /* __APPLE__ */
9107 rc = fcntl(newfd, F_NOCACHE, 1);
9115 rc = mdb_env_copyfd2(env, newfd, flags);
9118 if (!(env->me_flags & MDB_NOSUBDIR))
9120 if (newfd != INVALID_HANDLE_VALUE)
9121 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9128 mdb_env_copy(MDB_env *env, const char *path)
9130 return mdb_env_copy2(env, path, 0);
9134 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9136 if ((flag & CHANGEABLE) != flag)
9139 env->me_flags |= flag;
9141 env->me_flags &= ~flag;
9146 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9151 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9156 mdb_env_set_userctx(MDB_env *env, void *ctx)
9160 env->me_userctx = ctx;
9165 mdb_env_get_userctx(MDB_env *env)
9167 return env ? env->me_userctx : NULL;
9171 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9176 env->me_assert_func = func;
9182 mdb_env_get_path(MDB_env *env, const char **arg)
9187 *arg = env->me_path;
9192 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9201 /** Common code for #mdb_stat() and #mdb_env_stat().
9202 * @param[in] env the environment to operate in.
9203 * @param[in] db the #MDB_db record containing the stats to return.
9204 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9205 * @return 0, this function always succeeds.
9208 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9210 arg->ms_psize = env->me_psize;
9211 arg->ms_depth = db->md_depth;
9212 arg->ms_branch_pages = db->md_branch_pages;
9213 arg->ms_leaf_pages = db->md_leaf_pages;
9214 arg->ms_overflow_pages = db->md_overflow_pages;
9215 arg->ms_entries = db->md_entries;
9221 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9225 if (env == NULL || arg == NULL)
9228 toggle = mdb_env_pick_meta(env);
9230 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9234 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9238 if (env == NULL || arg == NULL)
9241 toggle = mdb_env_pick_meta(env);
9242 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9243 arg->me_mapsize = env->me_mapsize;
9244 arg->me_maxreaders = env->me_maxreaders;
9245 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9247 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9248 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9252 /** Set the default comparison functions for a database.
9253 * Called immediately after a database is opened to set the defaults.
9254 * The user can then override them with #mdb_set_compare() or
9255 * #mdb_set_dupsort().
9256 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9257 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9260 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9262 uint16_t f = txn->mt_dbs[dbi].md_flags;
9264 txn->mt_dbxs[dbi].md_cmp =
9265 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9266 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9268 txn->mt_dbxs[dbi].md_dcmp =
9269 !(f & MDB_DUPSORT) ? 0 :
9270 ((f & MDB_INTEGERDUP)
9271 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9272 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9275 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9281 int rc, dbflag, exact;
9282 unsigned int unused = 0, seq;
9285 if ((flags & VALID_FLAGS) != flags)
9287 if (txn->mt_flags & MDB_TXN_ERROR)
9293 if (flags & PERSISTENT_FLAGS) {
9294 uint16_t f2 = flags & PERSISTENT_FLAGS;
9295 /* make sure flag changes get committed */
9296 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9297 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9298 txn->mt_flags |= MDB_TXN_DIRTY;
9301 mdb_default_cmp(txn, MAIN_DBI);
9305 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9306 mdb_default_cmp(txn, MAIN_DBI);
9309 /* Is the DB already open? */
9311 for (i=2; i<txn->mt_numdbs; i++) {
9312 if (!txn->mt_dbxs[i].md_name.mv_size) {
9313 /* Remember this free slot */
9314 if (!unused) unused = i;
9317 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9318 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9324 /* If no free slot and max hit, fail */
9325 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9326 return MDB_DBS_FULL;
9328 /* Cannot mix named databases with some mainDB flags */
9329 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9330 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9332 /* Find the DB info */
9333 dbflag = DB_NEW|DB_VALID;
9336 key.mv_data = (void *)name;
9337 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9338 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9339 if (rc == MDB_SUCCESS) {
9340 /* make sure this is actually a DB */
9341 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9342 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9343 return MDB_INCOMPATIBLE;
9344 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9345 /* Create if requested */
9346 data.mv_size = sizeof(MDB_db);
9347 data.mv_data = &dummy;
9348 memset(&dummy, 0, sizeof(dummy));
9349 dummy.md_root = P_INVALID;
9350 dummy.md_flags = flags & PERSISTENT_FLAGS;
9351 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9355 /* OK, got info, add to table */
9356 if (rc == MDB_SUCCESS) {
9357 unsigned int slot = unused ? unused : txn->mt_numdbs;
9358 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9359 txn->mt_dbxs[slot].md_name.mv_size = len;
9360 txn->mt_dbxs[slot].md_rel = NULL;
9361 txn->mt_dbflags[slot] = dbflag;
9362 /* txn-> and env-> are the same in read txns, use
9363 * tmp variable to avoid undefined assignment
9365 seq = ++txn->mt_env->me_dbiseqs[slot];
9366 txn->mt_dbiseqs[slot] = seq;
9368 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9370 mdb_default_cmp(txn, slot);
9379 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9381 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9384 if (txn->mt_flags & MDB_TXN_ERROR)
9387 if (txn->mt_dbflags[dbi] & DB_STALE) {
9390 /* Stale, must read the DB's root. cursor_init does it for us. */
9391 mdb_cursor_init(&mc, txn, dbi, &mx);
9393 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9396 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9399 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9401 ptr = env->me_dbxs[dbi].md_name.mv_data;
9402 /* If there was no name, this was already closed */
9404 env->me_dbxs[dbi].md_name.mv_data = NULL;
9405 env->me_dbxs[dbi].md_name.mv_size = 0;
9406 env->me_dbflags[dbi] = 0;
9407 env->me_dbiseqs[dbi]++;
9412 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9414 /* We could return the flags for the FREE_DBI too but what's the point? */
9415 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9417 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9421 /** Add all the DB's pages to the free list.
9422 * @param[in] mc Cursor on the DB to free.
9423 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9424 * @return 0 on success, non-zero on failure.
9427 mdb_drop0(MDB_cursor *mc, int subs)
9431 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9432 if (rc == MDB_SUCCESS) {
9433 MDB_txn *txn = mc->mc_txn;
9438 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9439 * This also avoids any P_LEAF2 pages, which have no nodes.
9441 if (mc->mc_flags & C_SUB)
9444 mdb_cursor_copy(mc, &mx);
9445 while (mc->mc_snum > 0) {
9446 MDB_page *mp = mc->mc_pg[mc->mc_top];
9447 unsigned n = NUMKEYS(mp);
9449 for (i=0; i<n; i++) {
9450 ni = NODEPTR(mp, i);
9451 if (ni->mn_flags & F_BIGDATA) {
9454 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9455 rc = mdb_page_get(txn, pg, &omp, NULL);
9458 mdb_cassert(mc, IS_OVERFLOW(omp));
9459 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9463 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9464 mdb_xcursor_init1(mc, ni);
9465 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9471 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9473 for (i=0; i<n; i++) {
9475 ni = NODEPTR(mp, i);
9478 mdb_midl_xappend(txn->mt_free_pgs, pg);
9483 mc->mc_ki[mc->mc_top] = i;
9484 rc = mdb_cursor_sibling(mc, 1);
9486 if (rc != MDB_NOTFOUND)
9488 /* no more siblings, go back to beginning
9489 * of previous level.
9493 for (i=1; i<mc->mc_snum; i++) {
9495 mc->mc_pg[i] = mx.mc_pg[i];
9500 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9503 txn->mt_flags |= MDB_TXN_ERROR;
9504 } else if (rc == MDB_NOTFOUND) {
9510 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9512 MDB_cursor *mc, *m2;
9515 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9518 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9521 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9524 rc = mdb_cursor_open(txn, dbi, &mc);
9528 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9529 /* Invalidate the dropped DB's cursors */
9530 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9531 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9535 /* Can't delete the main DB */
9536 if (del && dbi > MAIN_DBI) {
9537 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9539 txn->mt_dbflags[dbi] = DB_STALE;
9540 mdb_dbi_close(txn->mt_env, dbi);
9542 txn->mt_flags |= MDB_TXN_ERROR;
9545 /* reset the DB record, mark it dirty */
9546 txn->mt_dbflags[dbi] |= DB_DIRTY;
9547 txn->mt_dbs[dbi].md_depth = 0;
9548 txn->mt_dbs[dbi].md_branch_pages = 0;
9549 txn->mt_dbs[dbi].md_leaf_pages = 0;
9550 txn->mt_dbs[dbi].md_overflow_pages = 0;
9551 txn->mt_dbs[dbi].md_entries = 0;
9552 txn->mt_dbs[dbi].md_root = P_INVALID;
9554 txn->mt_flags |= MDB_TXN_DIRTY;
9557 mdb_cursor_close(mc);
9561 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9563 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9566 txn->mt_dbxs[dbi].md_cmp = cmp;
9570 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9572 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9575 txn->mt_dbxs[dbi].md_dcmp = cmp;
9579 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9581 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9584 txn->mt_dbxs[dbi].md_rel = rel;
9588 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9590 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9593 txn->mt_dbxs[dbi].md_relctx = ctx;
9598 mdb_env_get_maxkeysize(MDB_env *env)
9600 return ENV_MAXKEY(env);
9604 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9606 unsigned int i, rdrs;
9609 int rc = 0, first = 1;
9613 if (!env->me_txns) {
9614 return func("(no reader locks)\n", ctx);
9616 rdrs = env->me_txns->mti_numreaders;
9617 mr = env->me_txns->mti_readers;
9618 for (i=0; i<rdrs; i++) {
9620 txnid_t txnid = mr[i].mr_txnid;
9621 sprintf(buf, txnid == (txnid_t)-1 ?
9622 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9623 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9626 rc = func(" pid thread txnid\n", ctx);
9630 rc = func(buf, ctx);
9636 rc = func("(no active readers)\n", ctx);
9641 /** Insert pid into list if not already present.
9642 * return -1 if already present.
9645 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9647 /* binary search of pid in list */
9649 unsigned cursor = 1;
9651 unsigned n = ids[0];
9654 unsigned pivot = n >> 1;
9655 cursor = base + pivot + 1;
9656 val = pid - ids[cursor];
9661 } else if ( val > 0 ) {
9666 /* found, so it's a duplicate */
9675 for (n = ids[0]; n > cursor; n--)
9682 mdb_reader_check(MDB_env *env, int *dead)
9688 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9691 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9692 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9694 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9695 unsigned int i, j, rdrs;
9697 MDB_PID_T *pids, pid;
9698 int rc = MDB_SUCCESS, count = 0;
9700 rdrs = env->me_txns->mti_numreaders;
9701 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9705 mr = env->me_txns->mti_readers;
9706 for (i=0; i<rdrs; i++) {
9708 if (pid && pid != env->me_pid) {
9709 if (mdb_pid_insert(pids, pid) == 0) {
9710 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9711 /* Stale reader found */
9714 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9715 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9717 rdrs = 0; /* the above checked all readers */
9719 /* Recheck, a new process may have reused pid */
9720 if (mdb_reader_pid(env, Pidcheck, pid))
9725 if (mr[j].mr_pid == pid) {
9726 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9727 (unsigned) pid, mr[j].mr_txnid));
9732 UNLOCK_MUTEX(rmutex);
9743 #ifdef MDB_ROBUST_SUPPORTED
9744 /** Handle #LOCK_MUTEX0() failure.
9745 * Try to repair the lock file if the mutex owner died.
9746 * @param[in] env the environment handle
9747 * @param[in] mutex LOCK_MUTEX0() mutex
9748 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9749 * @return 0 on success with the mutex locked, or an error code on failure.
9751 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9753 int toggle, rlocked, rc2;
9755 if (rc == MDB_OWNERDEAD) {
9756 /* We own the mutex. Clean up after dead previous owner. */
9758 rlocked = (mutex == env->me_rmutex);
9760 /* Keep mti_txnid updated, otherwise next writer can
9761 * overwrite data which latest meta page refers to.
9763 toggle = mdb_env_pick_meta(env);
9764 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9765 /* env is hosed if the dead thread was ours */
9767 env->me_flags |= MDB_FATAL_ERROR;
9772 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9773 (rc ? "this process' env is hosed" : "recovering")));
9774 rc2 = mdb_reader_check0(env, rlocked, NULL);
9776 rc2 = mdb_mutex_consistent(mutex);
9777 if (rc || (rc = rc2)) {
9778 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9779 UNLOCK_MUTEX(mutex);
9785 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9790 #endif /* MDB_ROBUST_SUPPORTED */
projects / openldap / blob