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 */
2720 size = tsize = sizeof(MDB_txn);
2722 /* Nested transactions: Max 1 child, write txns only, no writemap */
2723 flags |= parent->mt_flags;
2724 if (parent->mt_child ||
2725 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2727 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2729 /* Child txns save MDB_pgstate and use own copy of cursors */
2730 size = tsize = sizeof(MDB_ntxn);
2731 size += env->me_maxdbs * sizeof(MDB_cursor *);
2732 } else if (!(flags & MDB_RDONLY)) {
2733 /* Reuse preallocated write txn. However, do not touch it until
2734 * mdb_txn_renew0() succeeds, since it currently may be active.
2739 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2741 if ((txn = calloc(1, size)) == NULL) {
2742 DPRINTF(("calloc: %s", strerror(errno)));
2745 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2746 if (flags & MDB_RDONLY) {
2747 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2748 txn->mt_dbiseqs = env->me_dbiseqs;
2750 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2752 txn->mt_dbiseqs = parent->mt_dbiseqs;
2753 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2755 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2756 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2759 txn->mt_flags = flags;
2764 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2765 if (!txn->mt_u.dirty_list ||
2766 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2768 free(txn->mt_u.dirty_list);
2772 txn->mt_txnid = parent->mt_txnid;
2773 txn->mt_dirty_room = parent->mt_dirty_room;
2774 txn->mt_u.dirty_list[0].mid = 0;
2775 txn->mt_spill_pgs = NULL;
2776 txn->mt_next_pgno = parent->mt_next_pgno;
2777 parent->mt_child = txn;
2778 txn->mt_parent = parent;
2779 txn->mt_numdbs = parent->mt_numdbs;
2780 txn->mt_dbxs = parent->mt_dbxs;
2781 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2782 /* Copy parent's mt_dbflags, but clear DB_NEW */
2783 for (i=0; i<txn->mt_numdbs; i++)
2784 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2786 ntxn = (MDB_ntxn *)txn;
2787 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2788 if (env->me_pghead) {
2789 size = MDB_IDL_SIZEOF(env->me_pghead);
2790 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2792 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2797 rc = mdb_cursor_shadow(parent, txn);
2799 mdb_txn_reset0(txn, "beginchild-fail");
2802 rc = mdb_txn_renew0(txn);
2805 if (txn != env->me_txn0)
2808 txn->mt_flags |= flags; /* for txn==me_txn0, no effect otherwise */
2810 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2811 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2812 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2819 mdb_txn_env(MDB_txn *txn)
2821 if(!txn) return NULL;
2826 mdb_txn_id(MDB_txn *txn)
2829 return txn->mt_txnid;
2832 /** Export or close DBI handles opened in this txn. */
2834 mdb_dbis_update(MDB_txn *txn, int keep)
2837 MDB_dbi n = txn->mt_numdbs;
2838 MDB_env *env = txn->mt_env;
2839 unsigned char *tdbflags = txn->mt_dbflags;
2841 for (i = n; --i >= 2;) {
2842 if (tdbflags[i] & DB_NEW) {
2844 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2846 char *ptr = env->me_dbxs[i].md_name.mv_data;
2848 env->me_dbxs[i].md_name.mv_data = NULL;
2849 env->me_dbxs[i].md_name.mv_size = 0;
2850 env->me_dbflags[i] = 0;
2851 env->me_dbiseqs[i]++;
2857 if (keep && env->me_numdbs < n)
2861 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2862 * May be called twice for readonly txns: First reset it, then abort.
2863 * @param[in] txn the transaction handle to reset
2864 * @param[in] act why the transaction is being reset
2867 mdb_txn_reset0(MDB_txn *txn, const char *act)
2869 MDB_env *env = txn->mt_env;
2871 /* Close any DBI handles opened in this txn */
2872 mdb_dbis_update(txn, 0);
2874 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2875 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2876 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2878 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2879 if (txn->mt_u.reader) {
2880 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2881 if (!(env->me_flags & MDB_NOTLS))
2882 txn->mt_u.reader = NULL; /* txn does not own reader */
2884 txn->mt_numdbs = 0; /* close nothing if called again */
2885 txn->mt_dbxs = NULL; /* mark txn as reset */
2887 pgno_t *pghead = env->me_pghead;
2889 mdb_cursors_close(txn, 0);
2890 if (!(env->me_flags & MDB_WRITEMAP)) {
2891 mdb_dlist_free(txn);
2894 if (!txn->mt_parent) {
2895 mdb_midl_shrink(&txn->mt_free_pgs);
2896 env->me_free_pgs = txn->mt_free_pgs;
2898 env->me_pghead = NULL;
2902 /* The writer mutex was locked in mdb_txn_begin. */
2904 UNLOCK_MUTEX(env->me_wmutex);
2906 txn->mt_parent->mt_child = NULL;
2907 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2908 mdb_midl_free(txn->mt_free_pgs);
2909 mdb_midl_free(txn->mt_spill_pgs);
2910 free(txn->mt_u.dirty_list);
2913 mdb_midl_free(pghead);
2918 mdb_txn_reset(MDB_txn *txn)
2923 /* This call is only valid for read-only txns */
2924 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2927 mdb_txn_reset0(txn, "reset");
2931 mdb_txn_abort(MDB_txn *txn)
2937 mdb_txn_abort(txn->mt_child);
2939 mdb_txn_reset0(txn, "abort");
2940 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2941 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2942 txn->mt_u.reader->mr_pid = 0;
2944 if (txn != txn->mt_env->me_txn0)
2948 /** Save the freelist as of this transaction to the freeDB.
2949 * This changes the freelist. Keep trying until it stabilizes.
2952 mdb_freelist_save(MDB_txn *txn)
2954 /* env->me_pghead[] can grow and shrink during this call.
2955 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2956 * Page numbers cannot disappear from txn->mt_free_pgs[].
2959 MDB_env *env = txn->mt_env;
2960 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2961 txnid_t pglast = 0, head_id = 0;
2962 pgno_t freecnt = 0, *free_pgs, *mop;
2963 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2965 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2967 if (env->me_pghead) {
2968 /* Make sure first page of freeDB is touched and on freelist */
2969 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2970 if (rc && rc != MDB_NOTFOUND)
2974 if (!env->me_pghead && txn->mt_loose_pgs) {
2975 /* Put loose page numbers in mt_free_pgs, since
2976 * we may be unable to return them to me_pghead.
2978 MDB_page *mp = txn->mt_loose_pgs;
2979 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2981 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2982 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2983 txn->mt_loose_pgs = NULL;
2984 txn->mt_loose_count = 0;
2987 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2988 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2989 ? SSIZE_MAX : maxfree_1pg;
2992 /* Come back here after each Put() in case freelist changed */
2997 /* If using records from freeDB which we have not yet
2998 * deleted, delete them and any we reserved for me_pghead.
3000 while (pglast < env->me_pglast) {
3001 rc = mdb_cursor_first(&mc, &key, NULL);
3004 pglast = head_id = *(txnid_t *)key.mv_data;
3005 total_room = head_room = 0;
3006 mdb_tassert(txn, pglast <= env->me_pglast);
3007 rc = mdb_cursor_del(&mc, 0);
3012 /* Save the IDL of pages freed by this txn, to a single record */
3013 if (freecnt < txn->mt_free_pgs[0]) {
3015 /* Make sure last page of freeDB is touched and on freelist */
3016 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3017 if (rc && rc != MDB_NOTFOUND)
3020 free_pgs = txn->mt_free_pgs;
3021 /* Write to last page of freeDB */
3022 key.mv_size = sizeof(txn->mt_txnid);
3023 key.mv_data = &txn->mt_txnid;
3025 freecnt = free_pgs[0];
3026 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3027 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3030 /* Retry if mt_free_pgs[] grew during the Put() */
3031 free_pgs = txn->mt_free_pgs;
3032 } while (freecnt < free_pgs[0]);
3033 mdb_midl_sort(free_pgs);
3034 memcpy(data.mv_data, free_pgs, data.mv_size);
3037 unsigned int i = free_pgs[0];
3038 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3039 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3041 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3047 mop = env->me_pghead;
3048 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3050 /* Reserve records for me_pghead[]. Split it if multi-page,
3051 * to avoid searching freeDB for a page range. Use keys in
3052 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3054 if (total_room >= mop_len) {
3055 if (total_room == mop_len || --more < 0)
3057 } else if (head_room >= maxfree_1pg && head_id > 1) {
3058 /* Keep current record (overflow page), add a new one */
3062 /* (Re)write {key = head_id, IDL length = head_room} */
3063 total_room -= head_room;
3064 head_room = mop_len - total_room;
3065 if (head_room > maxfree_1pg && head_id > 1) {
3066 /* Overflow multi-page for part of me_pghead */
3067 head_room /= head_id; /* amortize page sizes */
3068 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3069 } else if (head_room < 0) {
3070 /* Rare case, not bothering to delete this record */
3073 key.mv_size = sizeof(head_id);
3074 key.mv_data = &head_id;
3075 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3076 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3079 /* IDL is initially empty, zero out at least the length */
3080 pgs = (pgno_t *)data.mv_data;
3081 j = head_room > clean_limit ? head_room : 0;
3085 total_room += head_room;
3088 /* Return loose page numbers to me_pghead, though usually none are
3089 * left at this point. The pages themselves remain in dirty_list.
3091 if (txn->mt_loose_pgs) {
3092 MDB_page *mp = txn->mt_loose_pgs;
3093 unsigned count = txn->mt_loose_count;
3095 /* Room for loose pages + temp IDL with same */
3096 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3098 mop = env->me_pghead;
3099 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3100 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3101 loose[ ++count ] = mp->mp_pgno;
3103 mdb_midl_sort(loose);
3104 mdb_midl_xmerge(mop, loose);
3105 txn->mt_loose_pgs = NULL;
3106 txn->mt_loose_count = 0;
3110 /* Fill in the reserved me_pghead records */
3116 rc = mdb_cursor_first(&mc, &key, &data);
3117 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3118 txnid_t id = *(txnid_t *)key.mv_data;
3119 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3122 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3124 if (len > mop_len) {
3126 data.mv_size = (len + 1) * sizeof(MDB_ID);
3128 data.mv_data = mop -= len;
3131 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3133 if (rc || !(mop_len -= len))
3140 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3141 * @param[in] txn the transaction that's being committed
3142 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3143 * @return 0 on success, non-zero on failure.
3146 mdb_page_flush(MDB_txn *txn, int keep)
3148 MDB_env *env = txn->mt_env;
3149 MDB_ID2L dl = txn->mt_u.dirty_list;
3150 unsigned psize = env->me_psize, j;
3151 int i, pagecount = dl[0].mid, rc;
3152 size_t size = 0, pos = 0;
3154 MDB_page *dp = NULL;
3158 struct iovec iov[MDB_COMMIT_PAGES];
3159 ssize_t wpos = 0, wsize = 0, wres;
3160 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3166 if (env->me_flags & MDB_WRITEMAP) {
3167 /* Clear dirty flags */
3168 while (++i <= pagecount) {
3170 /* Don't flush this page yet */
3171 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3172 dp->mp_flags &= ~P_KEEP;
3176 dp->mp_flags &= ~P_DIRTY;
3181 /* Write the pages */
3183 if (++i <= pagecount) {
3185 /* Don't flush this page yet */
3186 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3187 dp->mp_flags &= ~P_KEEP;
3192 /* clear dirty flag */
3193 dp->mp_flags &= ~P_DIRTY;
3196 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3201 /* Windows actually supports scatter/gather I/O, but only on
3202 * unbuffered file handles. Since we're relying on the OS page
3203 * cache for all our data, that's self-defeating. So we just
3204 * write pages one at a time. We use the ov structure to set
3205 * the write offset, to at least save the overhead of a Seek
3208 DPRINTF(("committing page %"Z"u", pgno));
3209 memset(&ov, 0, sizeof(ov));
3210 ov.Offset = pos & 0xffffffff;
3211 ov.OffsetHigh = pos >> 16 >> 16;
3212 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3214 DPRINTF(("WriteFile: %d", rc));
3218 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3219 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3222 /* Write previous page(s) */
3223 #ifdef MDB_USE_PWRITEV
3224 wres = pwritev(env->me_fd, iov, n, wpos);
3227 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3230 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3234 DPRINTF(("lseek: %s", strerror(rc)));
3237 wres = writev(env->me_fd, iov, n);
3240 if (wres != wsize) {
3245 DPRINTF(("Write error: %s", strerror(rc)));
3247 rc = EIO; /* TODO: Use which error code? */
3248 DPUTS("short write, filesystem full?");
3259 DPRINTF(("committing page %"Z"u", pgno));
3260 next_pos = pos + size;
3261 iov[n].iov_len = size;
3262 iov[n].iov_base = (char *)dp;
3268 /* MIPS has cache coherency issues, this is a no-op everywhere else
3269 * Note: for any size >= on-chip cache size, entire on-chip cache is
3272 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3274 for (i = keep; ++i <= pagecount; ) {
3276 /* This is a page we skipped above */
3279 dl[j].mid = dp->mp_pgno;
3282 mdb_dpage_free(env, dp);
3287 txn->mt_dirty_room += i - j;
3293 mdb_txn_commit(MDB_txn *txn)
3299 if (txn == NULL || txn->mt_env == NULL)
3302 if (txn->mt_child) {
3303 rc = mdb_txn_commit(txn->mt_child);
3304 txn->mt_child = NULL;
3311 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3312 mdb_dbis_update(txn, 1);
3313 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3318 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3319 DPUTS("error flag is set, can't commit");
3321 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3326 if (txn->mt_parent) {
3327 MDB_txn *parent = txn->mt_parent;
3331 unsigned x, y, len, ps_len;
3333 /* Append our free list to parent's */
3334 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3337 mdb_midl_free(txn->mt_free_pgs);
3338 /* Failures after this must either undo the changes
3339 * to the parent or set MDB_TXN_ERROR in the parent.
3342 parent->mt_next_pgno = txn->mt_next_pgno;
3343 parent->mt_flags = txn->mt_flags;
3345 /* Merge our cursors into parent's and close them */
3346 mdb_cursors_close(txn, 1);
3348 /* Update parent's DB table. */
3349 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3350 parent->mt_numdbs = txn->mt_numdbs;
3351 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3352 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3353 for (i=2; i<txn->mt_numdbs; i++) {
3354 /* preserve parent's DB_NEW status */
3355 x = parent->mt_dbflags[i] & DB_NEW;
3356 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3359 dst = parent->mt_u.dirty_list;
3360 src = txn->mt_u.dirty_list;
3361 /* Remove anything in our dirty list from parent's spill list */
3362 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3364 pspill[0] = (pgno_t)-1;
3365 /* Mark our dirty pages as deleted in parent spill list */
3366 for (i=0, len=src[0].mid; ++i <= len; ) {
3367 MDB_ID pn = src[i].mid << 1;
3368 while (pn > pspill[x])
3370 if (pn == pspill[x]) {
3375 /* Squash deleted pagenums if we deleted any */
3376 for (x=y; ++x <= ps_len; )
3377 if (!(pspill[x] & 1))
3378 pspill[++y] = pspill[x];
3382 /* Find len = length of merging our dirty list with parent's */
3384 dst[0].mid = 0; /* simplify loops */
3385 if (parent->mt_parent) {
3386 len = x + src[0].mid;
3387 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3388 for (i = x; y && i; y--) {
3389 pgno_t yp = src[y].mid;
3390 while (yp < dst[i].mid)
3392 if (yp == dst[i].mid) {
3397 } else { /* Simplify the above for single-ancestor case */
3398 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3400 /* Merge our dirty list with parent's */
3402 for (i = len; y; dst[i--] = src[y--]) {
3403 pgno_t yp = src[y].mid;
3404 while (yp < dst[x].mid)
3405 dst[i--] = dst[x--];
3406 if (yp == dst[x].mid)
3407 free(dst[x--].mptr);
3409 mdb_tassert(txn, i == x);
3411 free(txn->mt_u.dirty_list);
3412 parent->mt_dirty_room = txn->mt_dirty_room;
3413 if (txn->mt_spill_pgs) {
3414 if (parent->mt_spill_pgs) {
3415 /* TODO: Prevent failure here, so parent does not fail */
3416 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3418 parent->mt_flags |= MDB_TXN_ERROR;
3419 mdb_midl_free(txn->mt_spill_pgs);
3420 mdb_midl_sort(parent->mt_spill_pgs);
3422 parent->mt_spill_pgs = txn->mt_spill_pgs;
3426 /* Append our loose page list to parent's */
3427 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3429 *lp = txn->mt_loose_pgs;
3430 parent->mt_loose_count += txn->mt_loose_count;
3432 parent->mt_child = NULL;
3433 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3438 if (txn != env->me_txn) {
3439 DPUTS("attempt to commit unknown transaction");
3444 mdb_cursors_close(txn, 0);
3446 if (!txn->mt_u.dirty_list[0].mid &&
3447 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3450 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3451 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3453 /* Update DB root pointers */
3454 if (txn->mt_numdbs > 2) {
3458 data.mv_size = sizeof(MDB_db);
3460 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3461 for (i = 2; i < txn->mt_numdbs; i++) {
3462 if (txn->mt_dbflags[i] & DB_DIRTY) {
3463 if (TXN_DBI_CHANGED(txn, i)) {
3467 data.mv_data = &txn->mt_dbs[i];
3468 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3476 rc = mdb_freelist_save(txn);
3480 mdb_midl_free(env->me_pghead);
3481 env->me_pghead = NULL;
3482 mdb_midl_shrink(&txn->mt_free_pgs);
3483 env->me_free_pgs = txn->mt_free_pgs;
3489 if ((rc = mdb_page_flush(txn, 0)) ||
3490 (rc = mdb_env_sync(env, 0)) ||
3491 (rc = mdb_env_write_meta(txn)))
3494 /* Free P_LOOSE pages left behind in dirty_list */
3495 if (!(env->me_flags & MDB_WRITEMAP))
3496 mdb_dlist_free(txn);
3501 mdb_dbis_update(txn, 1);
3504 UNLOCK_MUTEX(env->me_wmutex);
3505 if (txn != env->me_txn0)
3515 /** Read the environment parameters of a DB environment before
3516 * mapping it into memory.
3517 * @param[in] env the environment handle
3518 * @param[out] meta address of where to store the meta information
3519 * @return 0 on success, non-zero on failure.
3522 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3528 enum { Size = sizeof(pbuf) };
3530 /* We don't know the page size yet, so use a minimum value.
3531 * Read both meta pages so we can use the latest one.
3534 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3538 memset(&ov, 0, sizeof(ov));
3540 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3541 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3544 rc = pread(env->me_fd, &pbuf, Size, off);
3547 if (rc == 0 && off == 0)
3549 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3550 DPRINTF(("read: %s", mdb_strerror(rc)));
3554 p = (MDB_page *)&pbuf;
3556 if (!F_ISSET(p->mp_flags, P_META)) {
3557 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3562 if (m->mm_magic != MDB_MAGIC) {
3563 DPUTS("meta has invalid magic");
3567 if (m->mm_version != MDB_DATA_VERSION) {
3568 DPRINTF(("database is version %u, expected version %u",
3569 m->mm_version, MDB_DATA_VERSION));
3570 return MDB_VERSION_MISMATCH;
3573 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3579 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3581 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3583 meta->mm_magic = MDB_MAGIC;
3584 meta->mm_version = MDB_DATA_VERSION;
3585 meta->mm_mapsize = env->me_mapsize;
3586 meta->mm_psize = env->me_psize;
3587 meta->mm_last_pg = 1;
3588 meta->mm_flags = env->me_flags & 0xffff;
3589 meta->mm_flags |= MDB_INTEGERKEY;
3590 meta->mm_dbs[0].md_root = P_INVALID;
3591 meta->mm_dbs[1].md_root = P_INVALID;
3594 /** Write the environment parameters of a freshly created DB environment.
3595 * @param[in] env the environment handle
3596 * @param[in] meta the #MDB_meta to write
3597 * @return 0 on success, non-zero on failure.
3600 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3608 memset(&ov, 0, sizeof(ov));
3609 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3611 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3614 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3615 len = pwrite(fd, ptr, size, pos); \
3616 if (len == -1 && ErrCode() == EINTR) continue; \
3617 rc = (len >= 0); break; } while(1)
3620 DPUTS("writing new meta page");
3622 psize = env->me_psize;
3624 p = calloc(2, psize);
3629 p->mp_flags = P_META;
3630 *(MDB_meta *)METADATA(p) = *meta;
3632 q = (MDB_page *)((char *)p + psize);
3634 q->mp_flags = P_META;
3635 *(MDB_meta *)METADATA(q) = *meta;
3637 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3640 else if ((unsigned) len == psize * 2)
3648 /** Update the environment info to commit a transaction.
3649 * @param[in] txn the transaction that's being committed
3650 * @return 0 on success, non-zero on failure.
3653 mdb_env_write_meta(MDB_txn *txn)
3656 MDB_meta meta, metab, *mp;
3660 int rc, len, toggle;
3669 toggle = txn->mt_txnid & 1;
3670 DPRINTF(("writing meta page %d for root page %"Z"u",
3671 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3674 flags = env->me_flags;
3675 mp = env->me_metas[toggle];
3676 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3677 /* Persist any increases of mapsize config */
3678 if (mapsize < env->me_mapsize)
3679 mapsize = env->me_mapsize;
3681 if (flags & MDB_WRITEMAP) {
3682 mp->mm_mapsize = mapsize;
3683 mp->mm_dbs[0] = txn->mt_dbs[0];
3684 mp->mm_dbs[1] = txn->mt_dbs[1];
3685 mp->mm_last_pg = txn->mt_next_pgno - 1;
3686 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3687 !(defined(__i386__) || defined(__x86_64__))
3688 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3689 __sync_synchronize();
3691 mp->mm_txnid = txn->mt_txnid;
3692 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3693 unsigned meta_size = env->me_psize;
3694 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3697 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3698 if (meta_size < env->me_os_psize)
3699 meta_size += meta_size;
3704 if (MDB_MSYNC(ptr, meta_size, rc)) {
3711 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3712 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3714 meta.mm_mapsize = mapsize;
3715 meta.mm_dbs[0] = txn->mt_dbs[0];
3716 meta.mm_dbs[1] = txn->mt_dbs[1];
3717 meta.mm_last_pg = txn->mt_next_pgno - 1;
3718 meta.mm_txnid = txn->mt_txnid;
3720 off = offsetof(MDB_meta, mm_mapsize);
3721 ptr = (char *)&meta + off;
3722 len = sizeof(MDB_meta) - off;
3724 off += env->me_psize;
3727 /* Write to the SYNC fd */
3728 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3731 memset(&ov, 0, sizeof(ov));
3733 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3738 rc = pwrite(mfd, ptr, len, off);
3741 rc = rc < 0 ? ErrCode() : EIO;
3746 DPUTS("write failed, disk error?");
3747 /* On a failure, the pagecache still contains the new data.
3748 * Write some old data back, to prevent it from being used.
3749 * Use the non-SYNC fd; we know it will fail anyway.
3751 meta.mm_last_pg = metab.mm_last_pg;
3752 meta.mm_txnid = metab.mm_txnid;
3754 memset(&ov, 0, sizeof(ov));
3756 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3758 r2 = pwrite(env->me_fd, ptr, len, off);
3759 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3762 env->me_flags |= MDB_FATAL_ERROR;
3765 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3766 CACHEFLUSH(env->me_map + off, len, DCACHE);
3768 /* Memory ordering issues are irrelevant; since the entire writer
3769 * is wrapped by wmutex, all of these changes will become visible
3770 * after the wmutex is unlocked. Since the DB is multi-version,
3771 * readers will get consistent data regardless of how fresh or
3772 * how stale their view of these values is.
3775 env->me_txns->mti_txnid = txn->mt_txnid;
3780 /** Check both meta pages to see which one is newer.
3781 * @param[in] env the environment handle
3782 * @return meta toggle (0 or 1).
3785 mdb_env_pick_meta(const MDB_env *env)
3787 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3791 mdb_env_create(MDB_env **env)
3795 e = calloc(1, sizeof(MDB_env));
3799 e->me_maxreaders = DEFAULT_READERS;
3800 e->me_maxdbs = e->me_numdbs = 2;
3801 e->me_fd = INVALID_HANDLE_VALUE;
3802 e->me_lfd = INVALID_HANDLE_VALUE;
3803 e->me_mfd = INVALID_HANDLE_VALUE;
3804 #ifdef MDB_USE_POSIX_SEM
3805 e->me_rmutex = SEM_FAILED;
3806 e->me_wmutex = SEM_FAILED;
3808 e->me_pid = getpid();
3809 GET_PAGESIZE(e->me_os_psize);
3810 VGMEMP_CREATE(e,0,0);
3816 mdb_env_map(MDB_env *env, void *addr)
3819 unsigned int flags = env->me_flags;
3823 LONG sizelo, sizehi;
3826 if (flags & MDB_RDONLY) {
3827 /* Don't set explicit map size, use whatever exists */
3832 msize = env->me_mapsize;
3833 sizelo = msize & 0xffffffff;
3834 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3836 /* Windows won't create mappings for zero length files.
3837 * and won't map more than the file size.
3838 * Just set the maxsize right now.
3840 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3841 || !SetEndOfFile(env->me_fd)
3842 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3846 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3847 PAGE_READWRITE : PAGE_READONLY,
3848 sizehi, sizelo, NULL);
3851 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3852 FILE_MAP_WRITE : FILE_MAP_READ,
3854 rc = env->me_map ? 0 : ErrCode();
3859 int prot = PROT_READ;
3860 if (flags & MDB_WRITEMAP) {
3862 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3865 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3867 if (env->me_map == MAP_FAILED) {
3872 if (flags & MDB_NORDAHEAD) {
3873 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3875 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3877 #ifdef POSIX_MADV_RANDOM
3878 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3879 #endif /* POSIX_MADV_RANDOM */
3880 #endif /* MADV_RANDOM */
3884 /* Can happen because the address argument to mmap() is just a
3885 * hint. mmap() can pick another, e.g. if the range is in use.
3886 * The MAP_FIXED flag would prevent that, but then mmap could
3887 * instead unmap existing pages to make room for the new map.
3889 if (addr && env->me_map != addr)
3890 return EBUSY; /* TODO: Make a new MDB_* error code? */
3892 p = (MDB_page *)env->me_map;
3893 env->me_metas[0] = METADATA(p);
3894 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3900 mdb_env_set_mapsize(MDB_env *env, size_t size)
3902 /* If env is already open, caller is responsible for making
3903 * sure there are no active txns.
3911 meta = env->me_metas[mdb_env_pick_meta(env)];
3913 size = meta->mm_mapsize;
3915 /* Silently round up to minimum if the size is too small */
3916 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3920 munmap(env->me_map, env->me_mapsize);
3921 env->me_mapsize = size;
3922 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3923 rc = mdb_env_map(env, old);
3927 env->me_mapsize = size;
3929 env->me_maxpg = env->me_mapsize / env->me_psize;
3934 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3938 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3943 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3945 if (env->me_map || readers < 1)
3947 env->me_maxreaders = readers;
3952 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3954 if (!env || !readers)
3956 *readers = env->me_maxreaders;
3961 mdb_fsize(HANDLE fd, size_t *size)
3964 LARGE_INTEGER fsize;
3966 if (!GetFileSizeEx(fd, &fsize))
3969 *size = fsize.QuadPart;
3981 #ifdef BROKEN_FDATASYNC
3982 #include <sys/utsname.h>
3983 #include <sys/vfs.h>
3986 /** Further setup required for opening an LMDB environment
3989 mdb_env_open2(MDB_env *env)
3991 unsigned int flags = env->me_flags;
3992 int i, newenv = 0, rc;
3996 /* See if we should use QueryLimited */
3998 if ((rc & 0xff) > 5)
3999 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4001 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4004 #ifdef BROKEN_FDATASYNC
4005 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4006 * https://lkml.org/lkml/2012/9/3/83
4007 * Kernels after 3.6-rc6 are known good.
4008 * https://lkml.org/lkml/2012/9/10/556
4009 * See if the DB is on ext3/ext4, then check for new enough kernel
4010 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4015 fstatfs(env->me_fd, &st);
4016 while (st.f_type == 0xEF53) {
4020 if (uts.release[0] < '3') {
4021 if (!strncmp(uts.release, "2.6.32.", 7)) {
4022 i = atoi(uts.release+7);
4024 break; /* 2.6.32.60 and newer is OK */
4025 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4026 i = atoi(uts.release+7);
4028 break; /* 2.6.34.15 and newer is OK */
4030 } else if (uts.release[0] == '3') {
4031 i = atoi(uts.release+2);
4033 break; /* 3.6 and newer is OK */
4035 i = atoi(uts.release+4);
4037 break; /* 3.5.4 and newer is OK */
4038 } else if (i == 2) {
4039 i = atoi(uts.release+4);
4041 break; /* 3.2.30 and newer is OK */
4043 } else { /* 4.x and newer is OK */
4046 env->me_flags |= MDB_FSYNCONLY;
4052 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4055 DPUTS("new mdbenv");
4057 env->me_psize = env->me_os_psize;
4058 if (env->me_psize > MAX_PAGESIZE)
4059 env->me_psize = MAX_PAGESIZE;
4060 memset(&meta, 0, sizeof(meta));
4061 mdb_env_init_meta0(env, &meta);
4062 meta.mm_mapsize = DEFAULT_MAPSIZE;
4064 env->me_psize = meta.mm_psize;
4067 /* Was a mapsize configured? */
4068 if (!env->me_mapsize) {
4069 env->me_mapsize = meta.mm_mapsize;
4072 /* Make sure mapsize >= committed data size. Even when using
4073 * mm_mapsize, which could be broken in old files (ITS#7789).
4075 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4076 if (env->me_mapsize < minsize)
4077 env->me_mapsize = minsize;
4079 meta.mm_mapsize = env->me_mapsize;
4081 if (newenv && !(flags & MDB_FIXEDMAP)) {
4082 /* mdb_env_map() may grow the datafile. Write the metapages
4083 * first, so the file will be valid if initialization fails.
4084 * Except with FIXEDMAP, since we do not yet know mm_address.
4085 * We could fill in mm_address later, but then a different
4086 * program might end up doing that - one with a memory layout
4087 * and map address which does not suit the main program.
4089 rc = mdb_env_init_meta(env, &meta);
4095 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4100 if (flags & MDB_FIXEDMAP)
4101 meta.mm_address = env->me_map;
4102 i = mdb_env_init_meta(env, &meta);
4103 if (i != MDB_SUCCESS) {
4108 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4109 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4111 #if !(MDB_MAXKEYSIZE)
4112 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4114 env->me_maxpg = env->me_mapsize / env->me_psize;
4118 int toggle = mdb_env_pick_meta(env);
4119 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4121 DPRINTF(("opened database version %u, pagesize %u",
4122 env->me_metas[0]->mm_version, env->me_psize));
4123 DPRINTF(("using meta page %d", toggle));
4124 DPRINTF(("depth: %u", db->md_depth));
4125 DPRINTF(("entries: %"Z"u", db->md_entries));
4126 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4127 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4128 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4129 DPRINTF(("root: %"Z"u", db->md_root));
4137 /** Release a reader thread's slot in the reader lock table.
4138 * This function is called automatically when a thread exits.
4139 * @param[in] ptr This points to the slot in the reader lock table.
4142 mdb_env_reader_dest(void *ptr)
4144 MDB_reader *reader = ptr;
4150 /** Junk for arranging thread-specific callbacks on Windows. This is
4151 * necessarily platform and compiler-specific. Windows supports up
4152 * to 1088 keys. Let's assume nobody opens more than 64 environments
4153 * in a single process, for now. They can override this if needed.
4155 #ifndef MAX_TLS_KEYS
4156 #define MAX_TLS_KEYS 64
4158 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4159 static int mdb_tls_nkeys;
4161 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4165 case DLL_PROCESS_ATTACH: break;
4166 case DLL_THREAD_ATTACH: break;
4167 case DLL_THREAD_DETACH:
4168 for (i=0; i<mdb_tls_nkeys; i++) {
4169 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4171 mdb_env_reader_dest(r);
4175 case DLL_PROCESS_DETACH: break;
4180 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4182 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4186 /* Force some symbol references.
4187 * _tls_used forces the linker to create the TLS directory if not already done
4188 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4190 #pragma comment(linker, "/INCLUDE:_tls_used")
4191 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4192 #pragma const_seg(".CRT$XLB")
4193 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4194 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4197 #pragma comment(linker, "/INCLUDE:__tls_used")
4198 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4199 #pragma data_seg(".CRT$XLB")
4200 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4202 #endif /* WIN 32/64 */
4203 #endif /* !__GNUC__ */
4206 /** Downgrade the exclusive lock on the region back to shared */
4208 mdb_env_share_locks(MDB_env *env, int *excl)
4210 int rc = 0, toggle = mdb_env_pick_meta(env);
4212 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4217 /* First acquire a shared lock. The Unlock will
4218 * then release the existing exclusive lock.
4220 memset(&ov, 0, sizeof(ov));
4221 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4224 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4230 struct flock lock_info;
4231 /* The shared lock replaces the existing lock */
4232 memset((void *)&lock_info, 0, sizeof(lock_info));
4233 lock_info.l_type = F_RDLCK;
4234 lock_info.l_whence = SEEK_SET;
4235 lock_info.l_start = 0;
4236 lock_info.l_len = 1;
4237 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4238 (rc = ErrCode()) == EINTR) ;
4239 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4246 /** Try to get exclusive lock, otherwise shared.
4247 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4250 mdb_env_excl_lock(MDB_env *env, int *excl)
4254 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4258 memset(&ov, 0, sizeof(ov));
4259 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4266 struct flock lock_info;
4267 memset((void *)&lock_info, 0, sizeof(lock_info));
4268 lock_info.l_type = F_WRLCK;
4269 lock_info.l_whence = SEEK_SET;
4270 lock_info.l_start = 0;
4271 lock_info.l_len = 1;
4272 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4273 (rc = ErrCode()) == EINTR) ;
4277 # ifndef MDB_USE_POSIX_MUTEX
4278 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4281 lock_info.l_type = F_RDLCK;
4282 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4283 (rc = ErrCode()) == EINTR) ;
4293 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4295 * @(#) $Revision: 5.1 $
4296 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4297 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4299 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4303 * Please do not copyright this code. This code is in the public domain.
4305 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4306 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4307 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4308 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4309 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4310 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4311 * PERFORMANCE OF THIS SOFTWARE.
4314 * chongo <Landon Curt Noll> /\oo/\
4315 * http://www.isthe.com/chongo/
4317 * Share and Enjoy! :-)
4320 typedef unsigned long long mdb_hash_t;
4321 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4323 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4324 * @param[in] val value to hash
4325 * @param[in] hval initial value for hash
4326 * @return 64 bit hash
4328 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4329 * hval arg on the first call.
4332 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4334 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4335 unsigned char *end = s + val->mv_size;
4337 * FNV-1a hash each octet of the string
4340 /* xor the bottom with the current octet */
4341 hval ^= (mdb_hash_t)*s++;
4343 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4344 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4345 (hval << 7) + (hval << 8) + (hval << 40);
4347 /* return our new hash value */
4351 /** Hash the string and output the encoded hash.
4352 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4353 * very short name limits. We don't care about the encoding being reversible,
4354 * we just want to preserve as many bits of the input as possible in a
4355 * small printable string.
4356 * @param[in] str string to hash
4357 * @param[out] encbuf an array of 11 chars to hold the hash
4359 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4362 mdb_pack85(unsigned long l, char *out)
4366 for (i=0; i<5; i++) {
4367 *out++ = mdb_a85[l % 85];
4373 mdb_hash_enc(MDB_val *val, char *encbuf)
4375 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4377 mdb_pack85(h, encbuf);
4378 mdb_pack85(h>>32, encbuf+5);
4383 /** Open and/or initialize the lock region for the environment.
4384 * @param[in] env The LMDB environment.
4385 * @param[in] lpath The pathname of the file used for the lock region.
4386 * @param[in] mode The Unix permissions for the file, if we create it.
4387 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4388 * @return 0 on success, non-zero on failure.
4391 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4394 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4396 # define MDB_ERRCODE_ROFS EROFS
4397 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4398 # define MDB_CLOEXEC O_CLOEXEC
4401 # define MDB_CLOEXEC 0
4408 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4409 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4410 FILE_ATTRIBUTE_NORMAL, NULL);
4412 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4414 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4416 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4421 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4422 /* Lose record locks when exec*() */
4423 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4424 fcntl(env->me_lfd, F_SETFD, fdflags);
4427 if (!(env->me_flags & MDB_NOTLS)) {
4428 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4431 env->me_flags |= MDB_ENV_TXKEY;
4433 /* Windows TLS callbacks need help finding their TLS info. */
4434 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4438 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4442 /* Try to get exclusive lock. If we succeed, then
4443 * nobody is using the lock region and we should initialize it.
4445 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4448 size = GetFileSize(env->me_lfd, NULL);
4450 size = lseek(env->me_lfd, 0, SEEK_END);
4451 if (size == -1) goto fail_errno;
4453 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4454 if (size < rsize && *excl > 0) {
4456 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4457 || !SetEndOfFile(env->me_lfd))
4460 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4464 size = rsize - sizeof(MDB_txninfo);
4465 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4470 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4472 if (!mh) goto fail_errno;
4473 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4475 if (!env->me_txns) goto fail_errno;
4477 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4479 if (m == MAP_FAILED) goto fail_errno;
4485 BY_HANDLE_FILE_INFORMATION stbuf;
4494 if (!mdb_sec_inited) {
4495 InitializeSecurityDescriptor(&mdb_null_sd,
4496 SECURITY_DESCRIPTOR_REVISION);
4497 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4498 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4499 mdb_all_sa.bInheritHandle = FALSE;
4500 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4503 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4504 idbuf.volume = stbuf.dwVolumeSerialNumber;
4505 idbuf.nhigh = stbuf.nFileIndexHigh;
4506 idbuf.nlow = stbuf.nFileIndexLow;
4507 val.mv_data = &idbuf;
4508 val.mv_size = sizeof(idbuf);
4509 mdb_hash_enc(&val, encbuf);
4510 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4511 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4512 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4513 if (!env->me_rmutex) goto fail_errno;
4514 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4515 if (!env->me_wmutex) goto fail_errno;
4516 #elif defined(MDB_USE_POSIX_SEM)
4525 #if defined(__NetBSD__)
4526 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4528 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4529 idbuf.dev = stbuf.st_dev;
4530 idbuf.ino = stbuf.st_ino;
4531 val.mv_data = &idbuf;
4532 val.mv_size = sizeof(idbuf);
4533 mdb_hash_enc(&val, encbuf);
4534 #ifdef MDB_SHORT_SEMNAMES
4535 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4537 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4538 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4539 /* Clean up after a previous run, if needed: Try to
4540 * remove both semaphores before doing anything else.
4542 sem_unlink(env->me_txns->mti_rmname);
4543 sem_unlink(env->me_txns->mti_wmname);
4544 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4545 O_CREAT|O_EXCL, mode, 1);
4546 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4547 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4548 O_CREAT|O_EXCL, mode, 1);
4549 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4550 #else /* MDB_USE_POSIX_MUTEX: */
4551 pthread_mutexattr_t mattr;
4553 if ((rc = pthread_mutexattr_init(&mattr))
4554 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4555 #ifdef MDB_ROBUST_SUPPORTED
4556 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4558 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4559 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4561 pthread_mutexattr_destroy(&mattr);
4562 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4564 env->me_txns->mti_magic = MDB_MAGIC;
4565 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4566 env->me_txns->mti_txnid = 0;
4567 env->me_txns->mti_numreaders = 0;
4570 if (env->me_txns->mti_magic != MDB_MAGIC) {
4571 DPUTS("lock region has invalid magic");
4575 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4576 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4577 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4578 rc = MDB_VERSION_MISMATCH;
4582 if (rc && rc != EACCES && rc != EAGAIN) {
4586 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4587 if (!env->me_rmutex) goto fail_errno;
4588 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4589 if (!env->me_wmutex) goto fail_errno;
4590 #elif defined(MDB_USE_POSIX_SEM)
4591 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4592 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4593 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4594 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4605 /** The name of the lock file in the DB environment */
4606 #define LOCKNAME "/lock.mdb"
4607 /** The name of the data file in the DB environment */
4608 #define DATANAME "/data.mdb"
4609 /** The suffix of the lock file when no subdir is used */
4610 #define LOCKSUFF "-lock"
4611 /** Only a subset of the @ref mdb_env flags can be changed
4612 * at runtime. Changing other flags requires closing the
4613 * environment and re-opening it with the new flags.
4615 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4616 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4617 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4619 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4620 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4624 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4626 int oflags, rc, len, excl = -1;
4627 char *lpath, *dpath;
4629 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4633 if (flags & MDB_NOSUBDIR) {
4634 rc = len + sizeof(LOCKSUFF) + len + 1;
4636 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4641 if (flags & MDB_NOSUBDIR) {
4642 dpath = lpath + len + sizeof(LOCKSUFF);
4643 sprintf(lpath, "%s" LOCKSUFF, path);
4644 strcpy(dpath, path);
4646 dpath = lpath + len + sizeof(LOCKNAME);
4647 sprintf(lpath, "%s" LOCKNAME, path);
4648 sprintf(dpath, "%s" DATANAME, path);
4652 flags |= env->me_flags;
4653 if (flags & MDB_RDONLY) {
4654 /* silently ignore WRITEMAP when we're only getting read access */
4655 flags &= ~MDB_WRITEMAP;
4657 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4658 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4661 env->me_flags = flags |= MDB_ENV_ACTIVE;
4665 env->me_path = strdup(path);
4666 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4667 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4668 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4669 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4673 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4675 /* For RDONLY, get lockfile after we know datafile exists */
4676 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4677 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4683 if (F_ISSET(flags, MDB_RDONLY)) {
4684 oflags = GENERIC_READ;
4685 len = OPEN_EXISTING;
4687 oflags = GENERIC_READ|GENERIC_WRITE;
4690 mode = FILE_ATTRIBUTE_NORMAL;
4691 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4692 NULL, len, mode, NULL);
4694 if (F_ISSET(flags, MDB_RDONLY))
4697 oflags = O_RDWR | O_CREAT;
4699 env->me_fd = open(dpath, oflags, mode);
4701 if (env->me_fd == INVALID_HANDLE_VALUE) {
4706 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4707 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4712 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4713 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4714 env->me_mfd = env->me_fd;
4716 /* Synchronous fd for meta writes. Needed even with
4717 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4720 len = OPEN_EXISTING;
4721 env->me_mfd = CreateFile(dpath, oflags,
4722 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4723 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4726 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4728 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4733 DPRINTF(("opened dbenv %p", (void *) env));
4735 rc = mdb_env_share_locks(env, &excl);
4739 if (!(flags & MDB_RDONLY)) {
4741 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4742 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4743 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4744 (txn = calloc(1, size)))
4746 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4747 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4748 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4749 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4751 txn->mt_dbxs = env->me_dbxs;
4761 mdb_env_close0(env, excl);
4767 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4769 mdb_env_close0(MDB_env *env, int excl)
4773 if (!(env->me_flags & MDB_ENV_ACTIVE))
4776 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4778 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4779 free(env->me_dbxs[i].md_name.mv_data);
4784 free(env->me_dbiseqs);
4785 free(env->me_dbflags);
4787 free(env->me_dirty_list);
4789 mdb_midl_free(env->me_free_pgs);
4791 if (env->me_flags & MDB_ENV_TXKEY) {
4792 pthread_key_delete(env->me_txkey);
4794 /* Delete our key from the global list */
4795 for (i=0; i<mdb_tls_nkeys; i++)
4796 if (mdb_tls_keys[i] == env->me_txkey) {
4797 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4805 munmap(env->me_map, env->me_mapsize);
4807 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4808 (void) close(env->me_mfd);
4809 if (env->me_fd != INVALID_HANDLE_VALUE)
4810 (void) close(env->me_fd);
4812 MDB_PID_T pid = env->me_pid;
4813 /* Clearing readers is done in this function because
4814 * me_txkey with its destructor must be disabled first.
4816 * We skip the the reader mutex, so we touch only
4817 * data owned by this process (me_close_readers and
4818 * our readers), and clear each reader atomically.
4820 for (i = env->me_close_readers; --i >= 0; )
4821 if (env->me_txns->mti_readers[i].mr_pid == pid)
4822 env->me_txns->mti_readers[i].mr_pid = 0;
4824 if (env->me_rmutex) {
4825 CloseHandle(env->me_rmutex);
4826 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4828 /* Windows automatically destroys the mutexes when
4829 * the last handle closes.
4831 #elif defined(MDB_USE_POSIX_SEM)
4832 if (env->me_rmutex != SEM_FAILED) {
4833 sem_close(env->me_rmutex);
4834 if (env->me_wmutex != SEM_FAILED)
4835 sem_close(env->me_wmutex);
4836 /* If we have the filelock: If we are the
4837 * only remaining user, clean up semaphores.
4840 mdb_env_excl_lock(env, &excl);
4842 sem_unlink(env->me_txns->mti_rmname);
4843 sem_unlink(env->me_txns->mti_wmname);
4847 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4849 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4852 /* Unlock the lockfile. Windows would have unlocked it
4853 * after closing anyway, but not necessarily at once.
4855 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4858 (void) close(env->me_lfd);
4861 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4865 mdb_env_close(MDB_env *env)
4872 VGMEMP_DESTROY(env);
4873 while ((dp = env->me_dpages) != NULL) {
4874 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4875 env->me_dpages = dp->mp_next;
4879 mdb_env_close0(env, 0);
4883 /** Compare two items pointing at aligned size_t's */
4885 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4887 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4888 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4891 /** Compare two items pointing at aligned unsigned int's.
4893 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4894 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4897 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4899 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4900 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4903 /** Compare two items pointing at unsigned ints of unknown alignment.
4904 * Nodes and keys are guaranteed to be 2-byte aligned.
4907 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4909 #if BYTE_ORDER == LITTLE_ENDIAN
4910 unsigned short *u, *c;
4913 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4914 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4917 } while(!x && u > (unsigned short *)a->mv_data);
4920 unsigned short *u, *c, *end;
4923 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4924 u = (unsigned short *)a->mv_data;
4925 c = (unsigned short *)b->mv_data;
4928 } while(!x && u < end);
4933 /** Compare two items lexically */
4935 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4942 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4948 diff = memcmp(a->mv_data, b->mv_data, len);
4949 return diff ? diff : len_diff<0 ? -1 : len_diff;
4952 /** Compare two items in reverse byte order */
4954 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4956 const unsigned char *p1, *p2, *p1_lim;
4960 p1_lim = (const unsigned char *)a->mv_data;
4961 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4962 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4964 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4970 while (p1 > p1_lim) {
4971 diff = *--p1 - *--p2;
4975 return len_diff<0 ? -1 : len_diff;
4978 /** Search for key within a page, using binary search.
4979 * Returns the smallest entry larger or equal to the key.
4980 * If exactp is non-null, stores whether the found entry was an exact match
4981 * in *exactp (1 or 0).
4982 * Updates the cursor index with the index of the found entry.
4983 * If no entry larger or equal to the key is found, returns NULL.
4986 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4988 unsigned int i = 0, nkeys;
4991 MDB_page *mp = mc->mc_pg[mc->mc_top];
4992 MDB_node *node = NULL;
4997 nkeys = NUMKEYS(mp);
4999 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5000 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5003 low = IS_LEAF(mp) ? 0 : 1;
5005 cmp = mc->mc_dbx->md_cmp;
5007 /* Branch pages have no data, so if using integer keys,
5008 * alignment is guaranteed. Use faster mdb_cmp_int.
5010 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5011 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5018 nodekey.mv_size = mc->mc_db->md_pad;
5019 node = NODEPTR(mp, 0); /* fake */
5020 while (low <= high) {
5021 i = (low + high) >> 1;
5022 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5023 rc = cmp(key, &nodekey);
5024 DPRINTF(("found leaf index %u [%s], rc = %i",
5025 i, DKEY(&nodekey), rc));
5034 while (low <= high) {
5035 i = (low + high) >> 1;
5037 node = NODEPTR(mp, i);
5038 nodekey.mv_size = NODEKSZ(node);
5039 nodekey.mv_data = NODEKEY(node);
5041 rc = cmp(key, &nodekey);
5044 DPRINTF(("found leaf index %u [%s], rc = %i",
5045 i, DKEY(&nodekey), rc));
5047 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5048 i, DKEY(&nodekey), NODEPGNO(node), rc));
5059 if (rc > 0) { /* Found entry is less than the key. */
5060 i++; /* Skip to get the smallest entry larger than key. */
5062 node = NODEPTR(mp, i);
5065 *exactp = (rc == 0 && nkeys > 0);
5066 /* store the key index */
5067 mc->mc_ki[mc->mc_top] = i;
5069 /* There is no entry larger or equal to the key. */
5072 /* nodeptr is fake for LEAF2 */
5078 mdb_cursor_adjust(MDB_cursor *mc, func)
5082 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5083 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5090 /** Pop a page off the top of the cursor's stack. */
5092 mdb_cursor_pop(MDB_cursor *mc)
5096 MDB_page *top = mc->mc_pg[mc->mc_top];
5102 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5103 DDBI(mc), (void *) mc));
5107 /** Push a page onto the top of the cursor's stack. */
5109 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5111 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5112 DDBI(mc), (void *) mc));
5114 if (mc->mc_snum >= CURSOR_STACK) {
5115 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5116 return MDB_CURSOR_FULL;
5119 mc->mc_top = mc->mc_snum++;
5120 mc->mc_pg[mc->mc_top] = mp;
5121 mc->mc_ki[mc->mc_top] = 0;
5126 /** Find the address of the page corresponding to a given page number.
5127 * @param[in] txn the transaction for this access.
5128 * @param[in] pgno the page number for the page to retrieve.
5129 * @param[out] ret address of a pointer where the page's address will be stored.
5130 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5131 * @return 0 on success, non-zero on failure.
5134 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5136 MDB_env *env = txn->mt_env;
5140 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5144 MDB_ID2L dl = tx2->mt_u.dirty_list;
5146 /* Spilled pages were dirtied in this txn and flushed
5147 * because the dirty list got full. Bring this page
5148 * back in from the map (but don't unspill it here,
5149 * leave that unless page_touch happens again).
5151 if (tx2->mt_spill_pgs) {
5152 MDB_ID pn = pgno << 1;
5153 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5154 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5155 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5160 unsigned x = mdb_mid2l_search(dl, pgno);
5161 if (x <= dl[0].mid && dl[x].mid == pgno) {
5167 } while ((tx2 = tx2->mt_parent) != NULL);
5170 if (pgno < txn->mt_next_pgno) {
5172 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5174 DPRINTF(("page %"Z"u not found", pgno));
5175 txn->mt_flags |= MDB_TXN_ERROR;
5176 return MDB_PAGE_NOTFOUND;
5186 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5187 * The cursor is at the root page, set up the rest of it.
5190 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5192 MDB_page *mp = mc->mc_pg[mc->mc_top];
5196 while (IS_BRANCH(mp)) {
5200 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5201 mdb_cassert(mc, NUMKEYS(mp) > 1);
5202 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5204 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5206 if (flags & MDB_PS_LAST)
5207 i = NUMKEYS(mp) - 1;
5210 node = mdb_node_search(mc, key, &exact);
5212 i = NUMKEYS(mp) - 1;
5214 i = mc->mc_ki[mc->mc_top];
5216 mdb_cassert(mc, i > 0);
5220 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5223 mdb_cassert(mc, i < NUMKEYS(mp));
5224 node = NODEPTR(mp, i);
5226 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5229 mc->mc_ki[mc->mc_top] = i;
5230 if ((rc = mdb_cursor_push(mc, mp)))
5233 if (flags & MDB_PS_MODIFY) {
5234 if ((rc = mdb_page_touch(mc)) != 0)
5236 mp = mc->mc_pg[mc->mc_top];
5241 DPRINTF(("internal error, index points to a %02X page!?",
5243 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5244 return MDB_CORRUPTED;
5247 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5248 key ? DKEY(key) : "null"));
5249 mc->mc_flags |= C_INITIALIZED;
5250 mc->mc_flags &= ~C_EOF;
5255 /** Search for the lowest key under the current branch page.
5256 * This just bypasses a NUMKEYS check in the current page
5257 * before calling mdb_page_search_root(), because the callers
5258 * are all in situations where the current page is known to
5262 mdb_page_search_lowest(MDB_cursor *mc)
5264 MDB_page *mp = mc->mc_pg[mc->mc_top];
5265 MDB_node *node = NODEPTR(mp, 0);
5268 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5271 mc->mc_ki[mc->mc_top] = 0;
5272 if ((rc = mdb_cursor_push(mc, mp)))
5274 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5277 /** Search for the page a given key should be in.
5278 * Push it and its parent pages on the cursor stack.
5279 * @param[in,out] mc the cursor for this operation.
5280 * @param[in] key the key to search for, or NULL for first/last page.
5281 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5282 * are touched (updated with new page numbers).
5283 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5284 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5285 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5286 * @return 0 on success, non-zero on failure.
5289 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5294 /* Make sure the txn is still viable, then find the root from
5295 * the txn's db table and set it as the root of the cursor's stack.
5297 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5298 DPUTS("transaction has failed, must abort");
5301 /* Make sure we're using an up-to-date root */
5302 if (*mc->mc_dbflag & DB_STALE) {
5304 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5306 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5307 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5314 MDB_node *leaf = mdb_node_search(&mc2,
5315 &mc->mc_dbx->md_name, &exact);
5317 return MDB_NOTFOUND;
5318 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5319 return MDB_INCOMPATIBLE; /* not a named DB */
5320 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5323 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5325 /* The txn may not know this DBI, or another process may
5326 * have dropped and recreated the DB with other flags.
5328 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5329 return MDB_INCOMPATIBLE;
5330 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5332 *mc->mc_dbflag &= ~DB_STALE;
5334 root = mc->mc_db->md_root;
5336 if (root == P_INVALID) { /* Tree is empty. */
5337 DPUTS("tree is empty");
5338 return MDB_NOTFOUND;
5342 mdb_cassert(mc, root > 1);
5343 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5344 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5350 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5351 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5353 if (flags & MDB_PS_MODIFY) {
5354 if ((rc = mdb_page_touch(mc)))
5358 if (flags & MDB_PS_ROOTONLY)
5361 return mdb_page_search_root(mc, key, flags);
5365 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5367 MDB_txn *txn = mc->mc_txn;
5368 pgno_t pg = mp->mp_pgno;
5369 unsigned x = 0, ovpages = mp->mp_pages;
5370 MDB_env *env = txn->mt_env;
5371 MDB_IDL sl = txn->mt_spill_pgs;
5372 MDB_ID pn = pg << 1;
5375 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5376 /* If the page is dirty or on the spill list we just acquired it,
5377 * so we should give it back to our current free list, if any.
5378 * Otherwise put it onto the list of pages we freed in this txn.
5380 * Won't create me_pghead: me_pglast must be inited along with it.
5381 * Unsupported in nested txns: They would need to hide the page
5382 * range in ancestor txns' dirty and spilled lists.
5384 if (env->me_pghead &&
5386 ((mp->mp_flags & P_DIRTY) ||
5387 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5391 MDB_ID2 *dl, ix, iy;
5392 rc = mdb_midl_need(&env->me_pghead, ovpages);
5395 if (!(mp->mp_flags & P_DIRTY)) {
5396 /* This page is no longer spilled */
5403 /* Remove from dirty list */
5404 dl = txn->mt_u.dirty_list;
5406 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5412 mdb_cassert(mc, x > 1);
5414 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5415 txn->mt_flags |= MDB_TXN_ERROR;
5416 return MDB_CORRUPTED;
5419 if (!(env->me_flags & MDB_WRITEMAP))
5420 mdb_dpage_free(env, mp);
5422 /* Insert in me_pghead */
5423 mop = env->me_pghead;
5424 j = mop[0] + ovpages;
5425 for (i = mop[0]; i && mop[i] < pg; i--)
5431 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5435 mc->mc_db->md_overflow_pages -= ovpages;
5439 /** Return the data associated with a given node.
5440 * @param[in] txn The transaction for this operation.
5441 * @param[in] leaf The node being read.
5442 * @param[out] data Updated to point to the node's data.
5443 * @return 0 on success, non-zero on failure.
5446 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5448 MDB_page *omp; /* overflow page */
5452 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5453 data->mv_size = NODEDSZ(leaf);
5454 data->mv_data = NODEDATA(leaf);
5458 /* Read overflow data.
5460 data->mv_size = NODEDSZ(leaf);
5461 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5462 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5463 DPRINTF(("read overflow page %"Z"u failed", pgno));
5466 data->mv_data = METADATA(omp);
5472 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5473 MDB_val *key, MDB_val *data)
5480 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5482 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5485 if (txn->mt_flags & MDB_TXN_ERROR)
5488 mdb_cursor_init(&mc, txn, dbi, &mx);
5489 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5492 /** Find a sibling for a page.
5493 * Replaces the page at the top of the cursor's stack with the
5494 * specified sibling, if one exists.
5495 * @param[in] mc The cursor for this operation.
5496 * @param[in] move_right Non-zero if the right sibling is requested,
5497 * otherwise the left sibling.
5498 * @return 0 on success, non-zero on failure.
5501 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5507 if (mc->mc_snum < 2) {
5508 return MDB_NOTFOUND; /* root has no siblings */
5512 DPRINTF(("parent page is page %"Z"u, index %u",
5513 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5515 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5516 : (mc->mc_ki[mc->mc_top] == 0)) {
5517 DPRINTF(("no more keys left, moving to %s sibling",
5518 move_right ? "right" : "left"));
5519 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5520 /* undo cursor_pop before returning */
5527 mc->mc_ki[mc->mc_top]++;
5529 mc->mc_ki[mc->mc_top]--;
5530 DPRINTF(("just moving to %s index key %u",
5531 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5533 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5535 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5536 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5537 /* mc will be inconsistent if caller does mc_snum++ as above */
5538 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5542 mdb_cursor_push(mc, mp);
5544 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5549 /** Move the cursor to the next data item. */
5551 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5557 if (mc->mc_flags & C_EOF) {
5558 return MDB_NOTFOUND;
5561 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5563 mp = mc->mc_pg[mc->mc_top];
5565 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5566 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5567 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5568 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5569 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5570 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5571 if (rc == MDB_SUCCESS)
5572 MDB_GET_KEY(leaf, key);
5577 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5578 if (op == MDB_NEXT_DUP)
5579 return MDB_NOTFOUND;
5583 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5584 mdb_dbg_pgno(mp), (void *) mc));
5585 if (mc->mc_flags & C_DEL)
5588 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5589 DPUTS("=====> move to next sibling page");
5590 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5591 mc->mc_flags |= C_EOF;
5594 mp = mc->mc_pg[mc->mc_top];
5595 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5597 mc->mc_ki[mc->mc_top]++;
5600 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5601 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5604 key->mv_size = mc->mc_db->md_pad;
5605 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5609 mdb_cassert(mc, IS_LEAF(mp));
5610 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5612 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5613 mdb_xcursor_init1(mc, leaf);
5616 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5619 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5620 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5621 if (rc != MDB_SUCCESS)
5626 MDB_GET_KEY(leaf, key);
5630 /** Move the cursor to the previous data item. */
5632 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5638 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5640 mp = mc->mc_pg[mc->mc_top];
5642 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5643 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5644 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5645 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5646 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5647 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5648 if (rc == MDB_SUCCESS) {
5649 MDB_GET_KEY(leaf, key);
5650 mc->mc_flags &= ~C_EOF;
5656 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5657 if (op == MDB_PREV_DUP)
5658 return MDB_NOTFOUND;
5662 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5663 mdb_dbg_pgno(mp), (void *) mc));
5665 if (mc->mc_ki[mc->mc_top] == 0) {
5666 DPUTS("=====> move to prev sibling page");
5667 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5670 mp = mc->mc_pg[mc->mc_top];
5671 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5672 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5674 mc->mc_ki[mc->mc_top]--;
5676 mc->mc_flags &= ~C_EOF;
5678 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5679 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5682 key->mv_size = mc->mc_db->md_pad;
5683 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5687 mdb_cassert(mc, IS_LEAF(mp));
5688 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5690 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5691 mdb_xcursor_init1(mc, leaf);
5694 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5697 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5698 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5699 if (rc != MDB_SUCCESS)
5704 MDB_GET_KEY(leaf, key);
5708 /** Set the cursor on a specific data item. */
5710 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5711 MDB_cursor_op op, int *exactp)
5715 MDB_node *leaf = NULL;
5718 if (key->mv_size == 0)
5719 return MDB_BAD_VALSIZE;
5722 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5724 /* See if we're already on the right page */
5725 if (mc->mc_flags & C_INITIALIZED) {
5728 mp = mc->mc_pg[mc->mc_top];
5730 mc->mc_ki[mc->mc_top] = 0;
5731 return MDB_NOTFOUND;
5733 if (mp->mp_flags & P_LEAF2) {
5734 nodekey.mv_size = mc->mc_db->md_pad;
5735 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5737 leaf = NODEPTR(mp, 0);
5738 MDB_GET_KEY2(leaf, nodekey);
5740 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5742 /* Probably happens rarely, but first node on the page
5743 * was the one we wanted.
5745 mc->mc_ki[mc->mc_top] = 0;
5752 unsigned int nkeys = NUMKEYS(mp);
5754 if (mp->mp_flags & P_LEAF2) {
5755 nodekey.mv_data = LEAF2KEY(mp,
5756 nkeys-1, nodekey.mv_size);
5758 leaf = NODEPTR(mp, nkeys-1);
5759 MDB_GET_KEY2(leaf, nodekey);
5761 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5763 /* last node was the one we wanted */
5764 mc->mc_ki[mc->mc_top] = nkeys-1;
5770 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5771 /* This is definitely the right page, skip search_page */
5772 if (mp->mp_flags & P_LEAF2) {
5773 nodekey.mv_data = LEAF2KEY(mp,
5774 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5776 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5777 MDB_GET_KEY2(leaf, nodekey);
5779 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5781 /* current node was the one we wanted */
5791 /* If any parents have right-sibs, search.
5792 * Otherwise, there's nothing further.
5794 for (i=0; i<mc->mc_top; i++)
5796 NUMKEYS(mc->mc_pg[i])-1)
5798 if (i == mc->mc_top) {
5799 /* There are no other pages */
5800 mc->mc_ki[mc->mc_top] = nkeys;
5801 return MDB_NOTFOUND;
5805 /* There are no other pages */
5806 mc->mc_ki[mc->mc_top] = 0;
5807 if (op == MDB_SET_RANGE && !exactp) {
5811 return MDB_NOTFOUND;
5815 rc = mdb_page_search(mc, key, 0);
5816 if (rc != MDB_SUCCESS)
5819 mp = mc->mc_pg[mc->mc_top];
5820 mdb_cassert(mc, IS_LEAF(mp));
5823 leaf = mdb_node_search(mc, key, exactp);
5824 if (exactp != NULL && !*exactp) {
5825 /* MDB_SET specified and not an exact match. */
5826 return MDB_NOTFOUND;
5830 DPUTS("===> inexact leaf not found, goto sibling");
5831 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5832 mc->mc_flags |= C_EOF;
5833 return rc; /* no entries matched */
5835 mp = mc->mc_pg[mc->mc_top];
5836 mdb_cassert(mc, IS_LEAF(mp));
5837 leaf = NODEPTR(mp, 0);
5841 mc->mc_flags |= C_INITIALIZED;
5842 mc->mc_flags &= ~C_EOF;
5845 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5846 key->mv_size = mc->mc_db->md_pad;
5847 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5852 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5853 mdb_xcursor_init1(mc, leaf);
5856 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5857 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5858 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5861 if (op == MDB_GET_BOTH) {
5867 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5868 if (rc != MDB_SUCCESS)
5871 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5874 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5876 dcmp = mc->mc_dbx->md_dcmp;
5877 #if UINT_MAX < SIZE_MAX
5878 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5879 dcmp = mdb_cmp_clong;
5881 rc = dcmp(data, &olddata);
5883 if (op == MDB_GET_BOTH || rc > 0)
5884 return MDB_NOTFOUND;
5891 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5892 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5897 /* The key already matches in all other cases */
5898 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5899 MDB_GET_KEY(leaf, key);
5900 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5905 /** Move the cursor to the first item in the database. */
5907 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5913 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5915 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5916 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5917 if (rc != MDB_SUCCESS)
5920 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5922 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5923 mc->mc_flags |= C_INITIALIZED;
5924 mc->mc_flags &= ~C_EOF;
5926 mc->mc_ki[mc->mc_top] = 0;
5928 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5929 key->mv_size = mc->mc_db->md_pad;
5930 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5935 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5936 mdb_xcursor_init1(mc, leaf);
5937 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5941 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5945 MDB_GET_KEY(leaf, key);
5949 /** Move the cursor to the last item in the database. */
5951 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5957 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5959 if (!(mc->mc_flags & C_EOF)) {
5961 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5962 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5963 if (rc != MDB_SUCCESS)
5966 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5969 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5970 mc->mc_flags |= C_INITIALIZED|C_EOF;
5971 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5973 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5974 key->mv_size = mc->mc_db->md_pad;
5975 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5980 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5981 mdb_xcursor_init1(mc, leaf);
5982 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5986 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5991 MDB_GET_KEY(leaf, key);
5996 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6001 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6006 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6010 case MDB_GET_CURRENT:
6011 if (!(mc->mc_flags & C_INITIALIZED)) {
6014 MDB_page *mp = mc->mc_pg[mc->mc_top];
6015 int nkeys = NUMKEYS(mp);
6016 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6017 mc->mc_ki[mc->mc_top] = nkeys;
6023 key->mv_size = mc->mc_db->md_pad;
6024 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6026 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6027 MDB_GET_KEY(leaf, key);
6029 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6030 if (mc->mc_flags & C_DEL)
6031 mdb_xcursor_init1(mc, leaf);
6032 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6034 rc = mdb_node_read(mc->mc_txn, leaf, data);
6041 case MDB_GET_BOTH_RANGE:
6046 if (mc->mc_xcursor == NULL) {
6047 rc = MDB_INCOMPATIBLE;
6057 rc = mdb_cursor_set(mc, key, data, op,
6058 op == MDB_SET_RANGE ? NULL : &exact);
6061 case MDB_GET_MULTIPLE:
6062 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6066 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6067 rc = MDB_INCOMPATIBLE;
6071 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6072 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6075 case MDB_NEXT_MULTIPLE:
6080 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6081 rc = MDB_INCOMPATIBLE;
6084 if (!(mc->mc_flags & C_INITIALIZED))
6085 rc = mdb_cursor_first(mc, key, data);
6087 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6088 if (rc == MDB_SUCCESS) {
6089 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6092 mx = &mc->mc_xcursor->mx_cursor;
6093 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6095 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6096 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6104 case MDB_NEXT_NODUP:
6105 if (!(mc->mc_flags & C_INITIALIZED))
6106 rc = mdb_cursor_first(mc, key, data);
6108 rc = mdb_cursor_next(mc, key, data, op);
6112 case MDB_PREV_NODUP:
6113 if (!(mc->mc_flags & C_INITIALIZED)) {
6114 rc = mdb_cursor_last(mc, key, data);
6117 mc->mc_flags |= C_INITIALIZED;
6118 mc->mc_ki[mc->mc_top]++;
6120 rc = mdb_cursor_prev(mc, key, data, op);
6123 rc = mdb_cursor_first(mc, key, data);
6126 mfunc = mdb_cursor_first;
6128 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6132 if (mc->mc_xcursor == NULL) {
6133 rc = MDB_INCOMPATIBLE;
6137 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6138 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6139 MDB_GET_KEY(leaf, key);
6140 rc = mdb_node_read(mc->mc_txn, leaf, data);
6144 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6148 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6151 rc = mdb_cursor_last(mc, key, data);
6154 mfunc = mdb_cursor_last;
6157 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6162 if (mc->mc_flags & C_DEL)
6163 mc->mc_flags ^= C_DEL;
6168 /** Touch all the pages in the cursor stack. Set mc_top.
6169 * Makes sure all the pages are writable, before attempting a write operation.
6170 * @param[in] mc The cursor to operate on.
6173 mdb_cursor_touch(MDB_cursor *mc)
6175 int rc = MDB_SUCCESS;
6177 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6180 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6182 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6183 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6186 *mc->mc_dbflag |= DB_DIRTY;
6191 rc = mdb_page_touch(mc);
6192 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6193 mc->mc_top = mc->mc_snum-1;
6198 /** Do not spill pages to disk if txn is getting full, may fail instead */
6199 #define MDB_NOSPILL 0x8000
6202 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6206 MDB_node *leaf = NULL;
6207 MDB_page *fp, *mp, *sub_root = NULL;
6209 MDB_val xdata, *rdata, dkey, olddata;
6211 int do_sub = 0, insert_key, insert_data;
6212 unsigned int mcount = 0, dcount = 0, nospill;
6215 unsigned int nflags;
6218 if (mc == NULL || key == NULL)
6221 env = mc->mc_txn->mt_env;
6223 /* Check this first so counter will always be zero on any
6226 if (flags & MDB_MULTIPLE) {
6227 dcount = data[1].mv_size;
6228 data[1].mv_size = 0;
6229 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6230 return MDB_INCOMPATIBLE;
6233 nospill = flags & MDB_NOSPILL;
6234 flags &= ~MDB_NOSPILL;
6236 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6237 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6239 if (key->mv_size-1 >= ENV_MAXKEY(env))
6240 return MDB_BAD_VALSIZE;
6242 #if SIZE_MAX > MAXDATASIZE
6243 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6244 return MDB_BAD_VALSIZE;
6246 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6247 return MDB_BAD_VALSIZE;
6250 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6251 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6255 if (flags == MDB_CURRENT) {
6256 if (!(mc->mc_flags & C_INITIALIZED))
6259 } else if (mc->mc_db->md_root == P_INVALID) {
6260 /* new database, cursor has nothing to point to */
6263 mc->mc_flags &= ~C_INITIALIZED;
6268 if (flags & MDB_APPEND) {
6270 rc = mdb_cursor_last(mc, &k2, &d2);
6272 rc = mc->mc_dbx->md_cmp(key, &k2);
6275 mc->mc_ki[mc->mc_top]++;
6277 /* new key is <= last key */
6282 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6284 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6285 DPRINTF(("duplicate key [%s]", DKEY(key)));
6287 return MDB_KEYEXIST;
6289 if (rc && rc != MDB_NOTFOUND)
6293 if (mc->mc_flags & C_DEL)
6294 mc->mc_flags ^= C_DEL;
6296 /* Cursor is positioned, check for room in the dirty list */
6298 if (flags & MDB_MULTIPLE) {
6300 xdata.mv_size = data->mv_size * dcount;
6304 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6308 if (rc == MDB_NO_ROOT) {
6310 /* new database, write a root leaf page */
6311 DPUTS("allocating new root leaf page");
6312 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6315 mdb_cursor_push(mc, np);
6316 mc->mc_db->md_root = np->mp_pgno;
6317 mc->mc_db->md_depth++;
6318 *mc->mc_dbflag |= DB_DIRTY;
6319 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6321 np->mp_flags |= P_LEAF2;
6322 mc->mc_flags |= C_INITIALIZED;
6324 /* make sure all cursor pages are writable */
6325 rc2 = mdb_cursor_touch(mc);
6330 insert_key = insert_data = rc;
6332 /* The key does not exist */
6333 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6334 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6335 LEAFSIZE(key, data) > env->me_nodemax)
6337 /* Too big for a node, insert in sub-DB. Set up an empty
6338 * "old sub-page" for prep_subDB to expand to a full page.
6340 fp_flags = P_LEAF|P_DIRTY;
6342 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6343 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6344 olddata.mv_size = PAGEHDRSZ;
6348 /* there's only a key anyway, so this is a no-op */
6349 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6351 unsigned int ksize = mc->mc_db->md_pad;
6352 if (key->mv_size != ksize)
6353 return MDB_BAD_VALSIZE;
6354 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6355 memcpy(ptr, key->mv_data, ksize);
6357 /* if overwriting slot 0 of leaf, need to
6358 * update branch key if there is a parent page
6360 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6361 unsigned short top = mc->mc_top;
6363 /* slot 0 is always an empty key, find real slot */
6364 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6366 if (mc->mc_ki[mc->mc_top])
6367 rc2 = mdb_update_key(mc, key);
6378 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6379 olddata.mv_size = NODEDSZ(leaf);
6380 olddata.mv_data = NODEDATA(leaf);
6383 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6384 /* Prepare (sub-)page/sub-DB to accept the new item,
6385 * if needed. fp: old sub-page or a header faking
6386 * it. mp: new (sub-)page. offset: growth in page
6387 * size. xdata: node data with new page or DB.
6389 unsigned i, offset = 0;
6390 mp = fp = xdata.mv_data = env->me_pbuf;
6391 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6393 /* Was a single item before, must convert now */
6394 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6396 /* Just overwrite the current item */
6397 if (flags == MDB_CURRENT)
6399 dcmp = mc->mc_dbx->md_dcmp;
6400 #if UINT_MAX < SIZE_MAX
6401 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6402 dcmp = mdb_cmp_clong;
6404 /* does data match? */
6405 if (!dcmp(data, &olddata)) {
6406 if (flags & MDB_NODUPDATA)
6407 return MDB_KEYEXIST;
6412 /* Back up original data item */
6413 dkey.mv_size = olddata.mv_size;
6414 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6416 /* Make sub-page header for the dup items, with dummy body */
6417 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6418 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6419 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6420 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6421 fp->mp_flags |= P_LEAF2;
6422 fp->mp_pad = data->mv_size;
6423 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6425 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6426 (dkey.mv_size & 1) + (data->mv_size & 1);
6428 fp->mp_upper = xdata.mv_size - PAGEBASE;
6429 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6430 } else if (leaf->mn_flags & F_SUBDATA) {
6431 /* Data is on sub-DB, just store it */
6432 flags |= F_DUPDATA|F_SUBDATA;
6435 /* Data is on sub-page */
6436 fp = olddata.mv_data;
6439 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6440 offset = EVEN(NODESIZE + sizeof(indx_t) +
6444 offset = fp->mp_pad;
6445 if (SIZELEFT(fp) < offset) {
6446 offset *= 4; /* space for 4 more */
6449 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6451 fp->mp_flags |= P_DIRTY;
6452 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6453 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6457 xdata.mv_size = olddata.mv_size + offset;
6460 fp_flags = fp->mp_flags;
6461 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6462 /* Too big for a sub-page, convert to sub-DB */
6463 fp_flags &= ~P_SUBP;
6465 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6466 fp_flags |= P_LEAF2;
6467 dummy.md_pad = fp->mp_pad;
6468 dummy.md_flags = MDB_DUPFIXED;
6469 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6470 dummy.md_flags |= MDB_INTEGERKEY;
6476 dummy.md_branch_pages = 0;
6477 dummy.md_leaf_pages = 1;
6478 dummy.md_overflow_pages = 0;
6479 dummy.md_entries = NUMKEYS(fp);
6480 xdata.mv_size = sizeof(MDB_db);
6481 xdata.mv_data = &dummy;
6482 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6484 offset = env->me_psize - olddata.mv_size;
6485 flags |= F_DUPDATA|F_SUBDATA;
6486 dummy.md_root = mp->mp_pgno;
6490 mp->mp_flags = fp_flags | P_DIRTY;
6491 mp->mp_pad = fp->mp_pad;
6492 mp->mp_lower = fp->mp_lower;
6493 mp->mp_upper = fp->mp_upper + offset;
6494 if (fp_flags & P_LEAF2) {
6495 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6497 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6498 olddata.mv_size - fp->mp_upper - PAGEBASE);
6499 for (i=0; i<NUMKEYS(fp); i++)
6500 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6508 mdb_node_del(mc, 0);
6512 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6513 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6514 return MDB_INCOMPATIBLE;
6515 /* overflow page overwrites need special handling */
6516 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6519 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6521 memcpy(&pg, olddata.mv_data, sizeof(pg));
6522 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6524 ovpages = omp->mp_pages;
6526 /* Is the ov page large enough? */
6527 if (ovpages >= dpages) {
6528 if (!(omp->mp_flags & P_DIRTY) &&
6529 (level || (env->me_flags & MDB_WRITEMAP)))
6531 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6534 level = 0; /* dirty in this txn or clean */
6537 if (omp->mp_flags & P_DIRTY) {
6538 /* yes, overwrite it. Note in this case we don't
6539 * bother to try shrinking the page if the new data
6540 * is smaller than the overflow threshold.
6543 /* It is writable only in a parent txn */
6544 size_t sz = (size_t) env->me_psize * ovpages, off;
6545 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6551 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6552 mdb_cassert(mc, rc2 == 0);
6553 if (!(flags & MDB_RESERVE)) {
6554 /* Copy end of page, adjusting alignment so
6555 * compiler may copy words instead of bytes.
6557 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6558 memcpy((size_t *)((char *)np + off),
6559 (size_t *)((char *)omp + off), sz - off);
6562 memcpy(np, omp, sz); /* Copy beginning of page */
6565 SETDSZ(leaf, data->mv_size);
6566 if (F_ISSET(flags, MDB_RESERVE))
6567 data->mv_data = METADATA(omp);
6569 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6573 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6575 } else if (data->mv_size == olddata.mv_size) {
6576 /* same size, just replace it. Note that we could
6577 * also reuse this node if the new data is smaller,
6578 * but instead we opt to shrink the node in that case.
6580 if (F_ISSET(flags, MDB_RESERVE))
6581 data->mv_data = olddata.mv_data;
6582 else if (!(mc->mc_flags & C_SUB))
6583 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6585 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6590 mdb_node_del(mc, 0);
6596 nflags = flags & NODE_ADD_FLAGS;
6597 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6598 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6599 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6600 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6602 nflags |= MDB_SPLIT_REPLACE;
6603 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6605 /* There is room already in this leaf page. */
6606 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6607 if (rc == 0 && insert_key) {
6608 /* Adjust other cursors pointing to mp */
6609 MDB_cursor *m2, *m3;
6610 MDB_dbi dbi = mc->mc_dbi;
6611 unsigned i = mc->mc_top;
6612 MDB_page *mp = mc->mc_pg[i];
6614 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6615 if (mc->mc_flags & C_SUB)
6616 m3 = &m2->mc_xcursor->mx_cursor;
6619 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6620 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6627 if (rc == MDB_SUCCESS) {
6628 /* Now store the actual data in the child DB. Note that we're
6629 * storing the user data in the keys field, so there are strict
6630 * size limits on dupdata. The actual data fields of the child
6631 * DB are all zero size.
6634 int xflags, new_dupdata;
6639 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6640 if (flags & MDB_CURRENT) {
6641 xflags = MDB_CURRENT|MDB_NOSPILL;
6643 mdb_xcursor_init1(mc, leaf);
6644 xflags = (flags & MDB_NODUPDATA) ?
6645 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6648 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6649 new_dupdata = (int)dkey.mv_size;
6650 /* converted, write the original data first */
6652 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6655 /* we've done our job */
6658 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6659 /* Adjust other cursors pointing to mp */
6661 MDB_xcursor *mx = mc->mc_xcursor;
6662 unsigned i = mc->mc_top;
6663 MDB_page *mp = mc->mc_pg[i];
6665 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6666 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6667 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6668 if (m2->mc_pg[i] == mp) {
6669 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6670 mdb_xcursor_init2(m2, mx, new_dupdata);
6671 } else if (!insert_key) {
6672 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6673 if (!(n2->mn_flags & F_SUBDATA))
6674 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6679 ecount = mc->mc_xcursor->mx_db.md_entries;
6680 if (flags & MDB_APPENDDUP)
6681 xflags |= MDB_APPEND;
6682 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6683 if (flags & F_SUBDATA) {
6684 void *db = NODEDATA(leaf);
6685 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6687 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6689 /* Increment count unless we just replaced an existing item. */
6691 mc->mc_db->md_entries++;
6693 /* Invalidate txn if we created an empty sub-DB */
6696 /* If we succeeded and the key didn't exist before,
6697 * make sure the cursor is marked valid.
6699 mc->mc_flags |= C_INITIALIZED;
6701 if (flags & MDB_MULTIPLE) {
6704 /* let caller know how many succeeded, if any */
6705 data[1].mv_size = mcount;
6706 if (mcount < dcount) {
6707 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6708 insert_key = insert_data = 0;
6715 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6718 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6723 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6729 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6730 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6732 if (!(mc->mc_flags & C_INITIALIZED))
6735 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6736 return MDB_NOTFOUND;
6738 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6741 rc = mdb_cursor_touch(mc);
6745 mp = mc->mc_pg[mc->mc_top];
6748 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6750 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6751 if (flags & MDB_NODUPDATA) {
6752 /* mdb_cursor_del0() will subtract the final entry */
6753 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6755 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6756 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6758 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6761 /* If sub-DB still has entries, we're done */
6762 if (mc->mc_xcursor->mx_db.md_entries) {
6763 if (leaf->mn_flags & F_SUBDATA) {
6764 /* update subDB info */
6765 void *db = NODEDATA(leaf);
6766 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6769 /* shrink fake page */
6770 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6771 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6772 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6773 /* fix other sub-DB cursors pointed at fake pages on this page */
6774 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6775 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6776 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6777 if (m2->mc_pg[mc->mc_top] == mp) {
6778 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6779 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6781 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6782 if (!(n2->mn_flags & F_SUBDATA))
6783 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6788 mc->mc_db->md_entries--;
6789 mc->mc_flags |= C_DEL;
6792 /* otherwise fall thru and delete the sub-DB */
6795 if (leaf->mn_flags & F_SUBDATA) {
6796 /* add all the child DB's pages to the free list */
6797 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6802 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6803 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6804 rc = MDB_INCOMPATIBLE;
6808 /* add overflow pages to free list */
6809 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6813 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6814 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6815 (rc = mdb_ovpage_free(mc, omp)))
6820 return mdb_cursor_del0(mc);
6823 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6827 /** Allocate and initialize new pages for a database.
6828 * @param[in] mc a cursor on the database being added to.
6829 * @param[in] flags flags defining what type of page is being allocated.
6830 * @param[in] num the number of pages to allocate. This is usually 1,
6831 * unless allocating overflow pages for a large record.
6832 * @param[out] mp Address of a page, or NULL on failure.
6833 * @return 0 on success, non-zero on failure.
6836 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6841 if ((rc = mdb_page_alloc(mc, num, &np)))
6843 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6844 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6845 np->mp_flags = flags | P_DIRTY;
6846 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6847 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6850 mc->mc_db->md_branch_pages++;
6851 else if (IS_LEAF(np))
6852 mc->mc_db->md_leaf_pages++;
6853 else if (IS_OVERFLOW(np)) {
6854 mc->mc_db->md_overflow_pages += num;
6862 /** Calculate the size of a leaf node.
6863 * The size depends on the environment's page size; if a data item
6864 * is too large it will be put onto an overflow page and the node
6865 * size will only include the key and not the data. Sizes are always
6866 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6867 * of the #MDB_node headers.
6868 * @param[in] env The environment handle.
6869 * @param[in] key The key for the node.
6870 * @param[in] data The data for the node.
6871 * @return The number of bytes needed to store the node.
6874 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6878 sz = LEAFSIZE(key, data);
6879 if (sz > env->me_nodemax) {
6880 /* put on overflow page */
6881 sz -= data->mv_size - sizeof(pgno_t);
6884 return EVEN(sz + sizeof(indx_t));
6887 /** Calculate the size of a branch node.
6888 * The size should depend on the environment's page size but since
6889 * we currently don't support spilling large keys onto overflow
6890 * pages, it's simply the size of the #MDB_node header plus the
6891 * size of the key. Sizes are always rounded up to an even number
6892 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6893 * @param[in] env The environment handle.
6894 * @param[in] key The key for the node.
6895 * @return The number of bytes needed to store the node.
6898 mdb_branch_size(MDB_env *env, MDB_val *key)
6903 if (sz > env->me_nodemax) {
6904 /* put on overflow page */
6905 /* not implemented */
6906 /* sz -= key->size - sizeof(pgno_t); */
6909 return sz + sizeof(indx_t);
6912 /** Add a node to the page pointed to by the cursor.
6913 * @param[in] mc The cursor for this operation.
6914 * @param[in] indx The index on the page where the new node should be added.
6915 * @param[in] key The key for the new node.
6916 * @param[in] data The data for the new node, if any.
6917 * @param[in] pgno The page number, if adding a branch node.
6918 * @param[in] flags Flags for the node.
6919 * @return 0 on success, non-zero on failure. Possible errors are:
6921 * <li>ENOMEM - failed to allocate overflow pages for the node.
6922 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6923 * should never happen since all callers already calculate the
6924 * page's free space before calling this function.
6928 mdb_node_add(MDB_cursor *mc, indx_t indx,
6929 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6932 size_t node_size = NODESIZE;
6936 MDB_page *mp = mc->mc_pg[mc->mc_top];
6937 MDB_page *ofp = NULL; /* overflow page */
6940 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6942 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6943 IS_LEAF(mp) ? "leaf" : "branch",
6944 IS_SUBP(mp) ? "sub-" : "",
6945 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6946 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6949 /* Move higher keys up one slot. */
6950 int ksize = mc->mc_db->md_pad, dif;
6951 char *ptr = LEAF2KEY(mp, indx, ksize);
6952 dif = NUMKEYS(mp) - indx;
6954 memmove(ptr+ksize, ptr, dif*ksize);
6955 /* insert new key */
6956 memcpy(ptr, key->mv_data, ksize);
6958 /* Just using these for counting */
6959 mp->mp_lower += sizeof(indx_t);
6960 mp->mp_upper -= ksize - sizeof(indx_t);
6964 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6966 node_size += key->mv_size;
6968 mdb_cassert(mc, data);
6969 if (F_ISSET(flags, F_BIGDATA)) {
6970 /* Data already on overflow page. */
6971 node_size += sizeof(pgno_t);
6972 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6973 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6975 /* Put data on overflow page. */
6976 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6977 data->mv_size, node_size+data->mv_size));
6978 node_size = EVEN(node_size + sizeof(pgno_t));
6979 if ((ssize_t)node_size > room)
6981 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6983 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6987 node_size += data->mv_size;
6990 node_size = EVEN(node_size);
6991 if ((ssize_t)node_size > room)
6995 /* Move higher pointers up one slot. */
6996 for (i = NUMKEYS(mp); i > indx; i--)
6997 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6999 /* Adjust free space offsets. */
7000 ofs = mp->mp_upper - node_size;
7001 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7002 mp->mp_ptrs[indx] = ofs;
7004 mp->mp_lower += sizeof(indx_t);
7006 /* Write the node data. */
7007 node = NODEPTR(mp, indx);
7008 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7009 node->mn_flags = flags;
7011 SETDSZ(node,data->mv_size);
7016 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7019 mdb_cassert(mc, key);
7021 if (F_ISSET(flags, F_BIGDATA))
7022 memcpy(node->mn_data + key->mv_size, data->mv_data,
7024 else if (F_ISSET(flags, MDB_RESERVE))
7025 data->mv_data = node->mn_data + key->mv_size;
7027 memcpy(node->mn_data + key->mv_size, data->mv_data,
7030 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
7032 if (F_ISSET(flags, MDB_RESERVE))
7033 data->mv_data = METADATA(ofp);
7035 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
7042 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7043 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7044 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7045 DPRINTF(("node size = %"Z"u", node_size));
7046 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7047 return MDB_PAGE_FULL;
7050 /** Delete the specified node from a page.
7051 * @param[in] mc Cursor pointing to the node to delete.
7052 * @param[in] ksize The size of a node. Only used if the page is
7053 * part of a #MDB_DUPFIXED database.
7056 mdb_node_del(MDB_cursor *mc, int ksize)
7058 MDB_page *mp = mc->mc_pg[mc->mc_top];
7059 indx_t indx = mc->mc_ki[mc->mc_top];
7061 indx_t i, j, numkeys, ptr;
7065 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7066 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7067 numkeys = NUMKEYS(mp);
7068 mdb_cassert(mc, indx < numkeys);
7071 int x = numkeys - 1 - indx;
7072 base = LEAF2KEY(mp, indx, ksize);
7074 memmove(base, base + ksize, x * ksize);
7075 mp->mp_lower -= sizeof(indx_t);
7076 mp->mp_upper += ksize - sizeof(indx_t);
7080 node = NODEPTR(mp, indx);
7081 sz = NODESIZE + node->mn_ksize;
7083 if (F_ISSET(node->mn_flags, F_BIGDATA))
7084 sz += sizeof(pgno_t);
7086 sz += NODEDSZ(node);
7090 ptr = mp->mp_ptrs[indx];
7091 for (i = j = 0; i < numkeys; i++) {
7093 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7094 if (mp->mp_ptrs[i] < ptr)
7095 mp->mp_ptrs[j] += sz;
7100 base = (char *)mp + mp->mp_upper + PAGEBASE;
7101 memmove(base + sz, base, ptr - mp->mp_upper);
7103 mp->mp_lower -= sizeof(indx_t);
7107 /** Compact the main page after deleting a node on a subpage.
7108 * @param[in] mp The main page to operate on.
7109 * @param[in] indx The index of the subpage on the main page.
7112 mdb_node_shrink(MDB_page *mp, indx_t indx)
7118 indx_t i, numkeys, ptr;
7120 node = NODEPTR(mp, indx);
7121 sp = (MDB_page *)NODEDATA(node);
7122 delta = SIZELEFT(sp);
7123 xp = (MDB_page *)((char *)sp + delta);
7125 /* shift subpage upward */
7127 nsize = NUMKEYS(sp) * sp->mp_pad;
7129 return; /* do not make the node uneven-sized */
7130 memmove(METADATA(xp), METADATA(sp), nsize);
7133 numkeys = NUMKEYS(sp);
7134 for (i=numkeys-1; i>=0; i--)
7135 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7137 xp->mp_upper = sp->mp_lower;
7138 xp->mp_lower = sp->mp_lower;
7139 xp->mp_flags = sp->mp_flags;
7140 xp->mp_pad = sp->mp_pad;
7141 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7143 nsize = NODEDSZ(node) - delta;
7144 SETDSZ(node, nsize);
7146 /* shift lower nodes upward */
7147 ptr = mp->mp_ptrs[indx];
7148 numkeys = NUMKEYS(mp);
7149 for (i = 0; i < numkeys; i++) {
7150 if (mp->mp_ptrs[i] <= ptr)
7151 mp->mp_ptrs[i] += delta;
7154 base = (char *)mp + mp->mp_upper + PAGEBASE;
7155 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7156 mp->mp_upper += delta;
7159 /** Initial setup of a sorted-dups cursor.
7160 * Sorted duplicates are implemented as a sub-database for the given key.
7161 * The duplicate data items are actually keys of the sub-database.
7162 * Operations on the duplicate data items are performed using a sub-cursor
7163 * initialized when the sub-database is first accessed. This function does
7164 * the preliminary setup of the sub-cursor, filling in the fields that
7165 * depend only on the parent DB.
7166 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7169 mdb_xcursor_init0(MDB_cursor *mc)
7171 MDB_xcursor *mx = mc->mc_xcursor;
7173 mx->mx_cursor.mc_xcursor = NULL;
7174 mx->mx_cursor.mc_txn = mc->mc_txn;
7175 mx->mx_cursor.mc_db = &mx->mx_db;
7176 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7177 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7178 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7179 mx->mx_cursor.mc_snum = 0;
7180 mx->mx_cursor.mc_top = 0;
7181 mx->mx_cursor.mc_flags = C_SUB;
7182 mx->mx_dbx.md_name.mv_size = 0;
7183 mx->mx_dbx.md_name.mv_data = NULL;
7184 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7185 mx->mx_dbx.md_dcmp = NULL;
7186 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7189 /** Final setup of a sorted-dups cursor.
7190 * Sets up the fields that depend on the data from the main cursor.
7191 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7192 * @param[in] node The data containing the #MDB_db record for the
7193 * sorted-dup database.
7196 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7198 MDB_xcursor *mx = mc->mc_xcursor;
7200 if (node->mn_flags & F_SUBDATA) {
7201 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7202 mx->mx_cursor.mc_pg[0] = 0;
7203 mx->mx_cursor.mc_snum = 0;
7204 mx->mx_cursor.mc_top = 0;
7205 mx->mx_cursor.mc_flags = C_SUB;
7207 MDB_page *fp = NODEDATA(node);
7208 mx->mx_db.md_pad = 0;
7209 mx->mx_db.md_flags = 0;
7210 mx->mx_db.md_depth = 1;
7211 mx->mx_db.md_branch_pages = 0;
7212 mx->mx_db.md_leaf_pages = 1;
7213 mx->mx_db.md_overflow_pages = 0;
7214 mx->mx_db.md_entries = NUMKEYS(fp);
7215 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7216 mx->mx_cursor.mc_snum = 1;
7217 mx->mx_cursor.mc_top = 0;
7218 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7219 mx->mx_cursor.mc_pg[0] = fp;
7220 mx->mx_cursor.mc_ki[0] = 0;
7221 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7222 mx->mx_db.md_flags = MDB_DUPFIXED;
7223 mx->mx_db.md_pad = fp->mp_pad;
7224 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7225 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7228 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7229 mx->mx_db.md_root));
7230 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7231 #if UINT_MAX < SIZE_MAX
7232 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7233 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7238 /** Fixup a sorted-dups cursor due to underlying update.
7239 * Sets up some fields that depend on the data from the main cursor.
7240 * Almost the same as init1, but skips initialization steps if the
7241 * xcursor had already been used.
7242 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7243 * @param[in] src_mx The xcursor of an up-to-date cursor.
7244 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7247 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7249 MDB_xcursor *mx = mc->mc_xcursor;
7252 mx->mx_cursor.mc_snum = 1;
7253 mx->mx_cursor.mc_top = 0;
7254 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7255 mx->mx_cursor.mc_ki[0] = 0;
7256 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7257 #if UINT_MAX < SIZE_MAX
7258 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7260 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7263 mx->mx_db = src_mx->mx_db;
7264 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7265 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7266 mx->mx_db.md_root));
7269 /** Initialize a cursor for a given transaction and database. */
7271 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7274 mc->mc_backup = NULL;
7277 mc->mc_db = &txn->mt_dbs[dbi];
7278 mc->mc_dbx = &txn->mt_dbxs[dbi];
7279 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7285 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7286 mdb_tassert(txn, mx != NULL);
7287 mc->mc_xcursor = mx;
7288 mdb_xcursor_init0(mc);
7290 mc->mc_xcursor = NULL;
7292 if (*mc->mc_dbflag & DB_STALE) {
7293 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7298 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7301 size_t size = sizeof(MDB_cursor);
7303 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7306 if (txn->mt_flags & MDB_TXN_ERROR)
7309 /* Allow read access to the freelist */
7310 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7313 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7314 size += sizeof(MDB_xcursor);
7316 if ((mc = malloc(size)) != NULL) {
7317 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7318 if (txn->mt_cursors) {
7319 mc->mc_next = txn->mt_cursors[dbi];
7320 txn->mt_cursors[dbi] = mc;
7321 mc->mc_flags |= C_UNTRACK;
7333 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7335 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7338 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7341 if (txn->mt_flags & MDB_TXN_ERROR)
7344 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7348 /* Return the count of duplicate data items for the current key */
7350 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7354 if (mc == NULL || countp == NULL)
7357 if (mc->mc_xcursor == NULL)
7358 return MDB_INCOMPATIBLE;
7360 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7363 if (!(mc->mc_flags & C_INITIALIZED))
7366 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7367 return MDB_NOTFOUND;
7369 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7370 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7373 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7376 *countp = mc->mc_xcursor->mx_db.md_entries;
7382 mdb_cursor_close(MDB_cursor *mc)
7384 if (mc && !mc->mc_backup) {
7385 /* remove from txn, if tracked */
7386 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7387 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7388 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7390 *prev = mc->mc_next;
7397 mdb_cursor_txn(MDB_cursor *mc)
7399 if (!mc) return NULL;
7404 mdb_cursor_dbi(MDB_cursor *mc)
7409 /** Replace the key for a branch node with a new key.
7410 * @param[in] mc Cursor pointing to the node to operate on.
7411 * @param[in] key The new key to use.
7412 * @return 0 on success, non-zero on failure.
7415 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7421 int delta, ksize, oksize;
7422 indx_t ptr, i, numkeys, indx;
7425 indx = mc->mc_ki[mc->mc_top];
7426 mp = mc->mc_pg[mc->mc_top];
7427 node = NODEPTR(mp, indx);
7428 ptr = mp->mp_ptrs[indx];
7432 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7433 k2.mv_data = NODEKEY(node);
7434 k2.mv_size = node->mn_ksize;
7435 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7437 mdb_dkey(&k2, kbuf2),
7443 /* Sizes must be 2-byte aligned. */
7444 ksize = EVEN(key->mv_size);
7445 oksize = EVEN(node->mn_ksize);
7446 delta = ksize - oksize;
7448 /* Shift node contents if EVEN(key length) changed. */
7450 if (delta > 0 && SIZELEFT(mp) < delta) {
7452 /* not enough space left, do a delete and split */
7453 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7454 pgno = NODEPGNO(node);
7455 mdb_node_del(mc, 0);
7456 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7459 numkeys = NUMKEYS(mp);
7460 for (i = 0; i < numkeys; i++) {
7461 if (mp->mp_ptrs[i] <= ptr)
7462 mp->mp_ptrs[i] -= delta;
7465 base = (char *)mp + mp->mp_upper + PAGEBASE;
7466 len = ptr - mp->mp_upper + NODESIZE;
7467 memmove(base - delta, base, len);
7468 mp->mp_upper -= delta;
7470 node = NODEPTR(mp, indx);
7473 /* But even if no shift was needed, update ksize */
7474 if (node->mn_ksize != key->mv_size)
7475 node->mn_ksize = key->mv_size;
7478 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7484 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7486 /** Move a node from csrc to cdst.
7489 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7496 unsigned short flags;
7500 /* Mark src and dst as dirty. */
7501 if ((rc = mdb_page_touch(csrc)) ||
7502 (rc = mdb_page_touch(cdst)))
7505 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7506 key.mv_size = csrc->mc_db->md_pad;
7507 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7509 data.mv_data = NULL;
7513 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7514 mdb_cassert(csrc, !((size_t)srcnode & 1));
7515 srcpg = NODEPGNO(srcnode);
7516 flags = srcnode->mn_flags;
7517 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7518 unsigned int snum = csrc->mc_snum;
7520 /* must find the lowest key below src */
7521 rc = mdb_page_search_lowest(csrc);
7524 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7525 key.mv_size = csrc->mc_db->md_pad;
7526 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7528 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7529 key.mv_size = NODEKSZ(s2);
7530 key.mv_data = NODEKEY(s2);
7532 csrc->mc_snum = snum--;
7533 csrc->mc_top = snum;
7535 key.mv_size = NODEKSZ(srcnode);
7536 key.mv_data = NODEKEY(srcnode);
7538 data.mv_size = NODEDSZ(srcnode);
7539 data.mv_data = NODEDATA(srcnode);
7541 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7542 unsigned int snum = cdst->mc_snum;
7545 /* must find the lowest key below dst */
7546 mdb_cursor_copy(cdst, &mn);
7547 rc = mdb_page_search_lowest(&mn);
7550 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7551 bkey.mv_size = mn.mc_db->md_pad;
7552 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7554 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7555 bkey.mv_size = NODEKSZ(s2);
7556 bkey.mv_data = NODEKEY(s2);
7558 mn.mc_snum = snum--;
7561 rc = mdb_update_key(&mn, &bkey);
7566 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7567 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7568 csrc->mc_ki[csrc->mc_top],
7570 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7571 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7573 /* Add the node to the destination page.
7575 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7576 if (rc != MDB_SUCCESS)
7579 /* Delete the node from the source page.
7581 mdb_node_del(csrc, key.mv_size);
7584 /* Adjust other cursors pointing to mp */
7585 MDB_cursor *m2, *m3;
7586 MDB_dbi dbi = csrc->mc_dbi;
7589 mp = cdst->mc_pg[csrc->mc_top];
7590 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7591 if (csrc->mc_flags & C_SUB)
7592 m3 = &m2->mc_xcursor->mx_cursor;
7595 if (m3 == cdst) continue;
7596 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] >=
7597 cdst->mc_ki[csrc->mc_top]) {
7598 m3->mc_ki[csrc->mc_top]++;
7602 mp = csrc->mc_pg[csrc->mc_top];
7603 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7604 if (csrc->mc_flags & C_SUB)
7605 m3 = &m2->mc_xcursor->mx_cursor;
7608 if (m3 == csrc) continue;
7609 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7610 csrc->mc_ki[csrc->mc_top]) {
7611 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7612 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7617 /* Update the parent separators.
7619 if (csrc->mc_ki[csrc->mc_top] == 0) {
7620 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7621 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7622 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7624 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7625 key.mv_size = NODEKSZ(srcnode);
7626 key.mv_data = NODEKEY(srcnode);
7628 DPRINTF(("update separator for source page %"Z"u to [%s]",
7629 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7630 mdb_cursor_copy(csrc, &mn);
7633 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7636 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7638 indx_t ix = csrc->mc_ki[csrc->mc_top];
7639 nullkey.mv_size = 0;
7640 csrc->mc_ki[csrc->mc_top] = 0;
7641 rc = mdb_update_key(csrc, &nullkey);
7642 csrc->mc_ki[csrc->mc_top] = ix;
7643 mdb_cassert(csrc, rc == MDB_SUCCESS);
7647 if (cdst->mc_ki[cdst->mc_top] == 0) {
7648 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7649 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7650 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7652 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7653 key.mv_size = NODEKSZ(srcnode);
7654 key.mv_data = NODEKEY(srcnode);
7656 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7657 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7658 mdb_cursor_copy(cdst, &mn);
7661 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7664 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7666 indx_t ix = cdst->mc_ki[cdst->mc_top];
7667 nullkey.mv_size = 0;
7668 cdst->mc_ki[cdst->mc_top] = 0;
7669 rc = mdb_update_key(cdst, &nullkey);
7670 cdst->mc_ki[cdst->mc_top] = ix;
7671 mdb_cassert(cdst, rc == MDB_SUCCESS);
7678 /** Merge one page into another.
7679 * The nodes from the page pointed to by \b csrc will
7680 * be copied to the page pointed to by \b cdst and then
7681 * the \b csrc page will be freed.
7682 * @param[in] csrc Cursor pointing to the source page.
7683 * @param[in] cdst Cursor pointing to the destination page.
7684 * @return 0 on success, non-zero on failure.
7687 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7689 MDB_page *psrc, *pdst;
7696 psrc = csrc->mc_pg[csrc->mc_top];
7697 pdst = cdst->mc_pg[cdst->mc_top];
7699 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7701 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7702 mdb_cassert(csrc, cdst->mc_snum > 1);
7704 /* Mark dst as dirty. */
7705 if ((rc = mdb_page_touch(cdst)))
7708 /* Move all nodes from src to dst.
7710 j = nkeys = NUMKEYS(pdst);
7711 if (IS_LEAF2(psrc)) {
7712 key.mv_size = csrc->mc_db->md_pad;
7713 key.mv_data = METADATA(psrc);
7714 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7715 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7716 if (rc != MDB_SUCCESS)
7718 key.mv_data = (char *)key.mv_data + key.mv_size;
7721 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7722 srcnode = NODEPTR(psrc, i);
7723 if (i == 0 && IS_BRANCH(psrc)) {
7726 mdb_cursor_copy(csrc, &mn);
7727 /* must find the lowest key below src */
7728 rc = mdb_page_search_lowest(&mn);
7731 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7732 key.mv_size = mn.mc_db->md_pad;
7733 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7735 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7736 key.mv_size = NODEKSZ(s2);
7737 key.mv_data = NODEKEY(s2);
7740 key.mv_size = srcnode->mn_ksize;
7741 key.mv_data = NODEKEY(srcnode);
7744 data.mv_size = NODEDSZ(srcnode);
7745 data.mv_data = NODEDATA(srcnode);
7746 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7747 if (rc != MDB_SUCCESS)
7752 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7753 pdst->mp_pgno, NUMKEYS(pdst),
7754 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7756 /* Unlink the src page from parent and add to free list.
7759 mdb_node_del(csrc, 0);
7760 if (csrc->mc_ki[csrc->mc_top] == 0) {
7762 rc = mdb_update_key(csrc, &key);
7770 psrc = csrc->mc_pg[csrc->mc_top];
7771 /* If not operating on FreeDB, allow this page to be reused
7772 * in this txn. Otherwise just add to free list.
7774 rc = mdb_page_loose(csrc, psrc);
7778 csrc->mc_db->md_leaf_pages--;
7780 csrc->mc_db->md_branch_pages--;
7782 /* Adjust other cursors pointing to mp */
7783 MDB_cursor *m2, *m3;
7784 MDB_dbi dbi = csrc->mc_dbi;
7786 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7787 if (csrc->mc_flags & C_SUB)
7788 m3 = &m2->mc_xcursor->mx_cursor;
7791 if (m3 == csrc) continue;
7792 if (m3->mc_snum < csrc->mc_snum) continue;
7793 if (m3->mc_pg[csrc->mc_top] == psrc) {
7794 m3->mc_pg[csrc->mc_top] = pdst;
7795 m3->mc_ki[csrc->mc_top] += nkeys;
7800 unsigned int snum = cdst->mc_snum;
7801 uint16_t depth = cdst->mc_db->md_depth;
7802 mdb_cursor_pop(cdst);
7803 rc = mdb_rebalance(cdst);
7804 /* Did the tree height change? */
7805 if (depth != cdst->mc_db->md_depth)
7806 snum += cdst->mc_db->md_depth - depth;
7807 cdst->mc_snum = snum;
7808 cdst->mc_top = snum-1;
7813 /** Copy the contents of a cursor.
7814 * @param[in] csrc The cursor to copy from.
7815 * @param[out] cdst The cursor to copy to.
7818 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7822 cdst->mc_txn = csrc->mc_txn;
7823 cdst->mc_dbi = csrc->mc_dbi;
7824 cdst->mc_db = csrc->mc_db;
7825 cdst->mc_dbx = csrc->mc_dbx;
7826 cdst->mc_snum = csrc->mc_snum;
7827 cdst->mc_top = csrc->mc_top;
7828 cdst->mc_flags = csrc->mc_flags;
7830 for (i=0; i<csrc->mc_snum; i++) {
7831 cdst->mc_pg[i] = csrc->mc_pg[i];
7832 cdst->mc_ki[i] = csrc->mc_ki[i];
7836 /** Rebalance the tree after a delete operation.
7837 * @param[in] mc Cursor pointing to the page where rebalancing
7839 * @return 0 on success, non-zero on failure.
7842 mdb_rebalance(MDB_cursor *mc)
7846 unsigned int ptop, minkeys, thresh;
7850 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7855 thresh = FILL_THRESHOLD;
7857 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7858 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7859 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7860 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7862 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7863 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7864 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7865 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7869 if (mc->mc_snum < 2) {
7870 MDB_page *mp = mc->mc_pg[0];
7872 DPUTS("Can't rebalance a subpage, ignoring");
7875 if (NUMKEYS(mp) == 0) {
7876 DPUTS("tree is completely empty");
7877 mc->mc_db->md_root = P_INVALID;
7878 mc->mc_db->md_depth = 0;
7879 mc->mc_db->md_leaf_pages = 0;
7880 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7883 /* Adjust cursors pointing to mp */
7886 mc->mc_flags &= ~C_INITIALIZED;
7888 MDB_cursor *m2, *m3;
7889 MDB_dbi dbi = mc->mc_dbi;
7891 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7892 if (mc->mc_flags & C_SUB)
7893 m3 = &m2->mc_xcursor->mx_cursor;
7896 if (m3->mc_snum < mc->mc_snum) continue;
7897 if (m3->mc_pg[0] == mp) {
7900 m3->mc_flags &= ~C_INITIALIZED;
7904 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7906 DPUTS("collapsing root page!");
7907 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7910 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7911 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7914 mc->mc_db->md_depth--;
7915 mc->mc_db->md_branch_pages--;
7916 mc->mc_ki[0] = mc->mc_ki[1];
7917 for (i = 1; i<mc->mc_db->md_depth; i++) {
7918 mc->mc_pg[i] = mc->mc_pg[i+1];
7919 mc->mc_ki[i] = mc->mc_ki[i+1];
7922 /* Adjust other cursors pointing to mp */
7923 MDB_cursor *m2, *m3;
7924 MDB_dbi dbi = mc->mc_dbi;
7926 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7927 if (mc->mc_flags & C_SUB)
7928 m3 = &m2->mc_xcursor->mx_cursor;
7931 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7932 if (m3->mc_pg[0] == mp) {
7933 for (i=0; i<m3->mc_snum; i++) {
7934 m3->mc_pg[i] = m3->mc_pg[i+1];
7935 m3->mc_ki[i] = m3->mc_ki[i+1];
7943 DPUTS("root page doesn't need rebalancing");
7947 /* The parent (branch page) must have at least 2 pointers,
7948 * otherwise the tree is invalid.
7950 ptop = mc->mc_top-1;
7951 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7953 /* Leaf page fill factor is below the threshold.
7954 * Try to move keys from left or right neighbor, or
7955 * merge with a neighbor page.
7960 mdb_cursor_copy(mc, &mn);
7961 mn.mc_xcursor = NULL;
7963 oldki = mc->mc_ki[mc->mc_top];
7964 if (mc->mc_ki[ptop] == 0) {
7965 /* We're the leftmost leaf in our parent.
7967 DPUTS("reading right neighbor");
7969 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7970 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7973 mn.mc_ki[mn.mc_top] = 0;
7974 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7976 /* There is at least one neighbor to the left.
7978 DPUTS("reading left neighbor");
7980 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7981 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7984 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7985 mc->mc_ki[mc->mc_top] = 0;
7988 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7989 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7990 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7992 /* If the neighbor page is above threshold and has enough keys,
7993 * move one key from it. Otherwise we should try to merge them.
7994 * (A branch page must never have less than 2 keys.)
7996 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7997 rc = mdb_node_move(&mn, mc);
7998 if (mc->mc_ki[mc->mc_top-1]) {
8002 if (mc->mc_ki[ptop] == 0) {
8003 rc = mdb_page_merge(&mn, mc);
8006 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8007 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8008 /* We want mdb_rebalance to find mn when doing fixups */
8009 if (mc->mc_flags & C_SUB) {
8010 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8011 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8012 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8014 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8015 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8017 rc = mdb_page_merge(mc, &mn);
8018 if (mc->mc_flags & C_SUB)
8019 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8021 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8022 mdb_cursor_copy(&mn, mc);
8024 mc->mc_flags &= ~C_EOF;
8026 mc->mc_ki[mc->mc_top] = oldki;
8030 /** Complete a delete operation started by #mdb_cursor_del(). */
8032 mdb_cursor_del0(MDB_cursor *mc)
8038 MDB_cursor *m2, *m3;
8039 MDB_dbi dbi = mc->mc_dbi;
8041 ki = mc->mc_ki[mc->mc_top];
8042 mp = mc->mc_pg[mc->mc_top];
8043 mdb_node_del(mc, mc->mc_db->md_pad);
8044 mc->mc_db->md_entries--;
8046 /* Adjust other cursors pointing to mp */
8047 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8048 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8049 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8051 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8053 if (m3->mc_pg[mc->mc_top] == mp) {
8054 if (m3->mc_ki[mc->mc_top] >= ki) {
8055 m3->mc_flags |= C_DEL;
8056 if (m3->mc_ki[mc->mc_top] > ki)
8057 m3->mc_ki[mc->mc_top]--;
8058 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8059 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8064 rc = mdb_rebalance(mc);
8066 if (rc == MDB_SUCCESS) {
8067 /* DB is totally empty now, just bail out.
8068 * Other cursors adjustments were already done
8069 * by mdb_rebalance and aren't needed here.
8074 mp = mc->mc_pg[mc->mc_top];
8075 nkeys = NUMKEYS(mp);
8077 /* Adjust other cursors pointing to mp */
8078 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8079 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8080 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8082 if (m3->mc_snum < mc->mc_snum)
8084 if (m3->mc_pg[mc->mc_top] == mp) {
8085 /* if m3 points past last node in page, find next sibling */
8086 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8087 rc = mdb_cursor_sibling(m3, 1);
8088 if (rc == MDB_NOTFOUND) {
8089 m3->mc_flags |= C_EOF;
8095 mc->mc_flags |= C_DEL;
8099 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8104 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8105 MDB_val *key, MDB_val *data)
8107 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8110 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8111 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8113 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8114 /* must ignore any data */
8118 return mdb_del0(txn, dbi, key, data, 0);
8122 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8123 MDB_val *key, MDB_val *data, unsigned flags)
8128 MDB_val rdata, *xdata;
8132 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8134 mdb_cursor_init(&mc, txn, dbi, &mx);
8143 flags |= MDB_NODUPDATA;
8145 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8147 /* let mdb_page_split know about this cursor if needed:
8148 * delete will trigger a rebalance; if it needs to move
8149 * a node from one page to another, it will have to
8150 * update the parent's separator key(s). If the new sepkey
8151 * is larger than the current one, the parent page may
8152 * run out of space, triggering a split. We need this
8153 * cursor to be consistent until the end of the rebalance.
8155 mc.mc_flags |= C_UNTRACK;
8156 mc.mc_next = txn->mt_cursors[dbi];
8157 txn->mt_cursors[dbi] = &mc;
8158 rc = mdb_cursor_del(&mc, flags);
8159 txn->mt_cursors[dbi] = mc.mc_next;
8164 /** Split a page and insert a new node.
8165 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8166 * The cursor will be updated to point to the actual page and index where
8167 * the node got inserted after the split.
8168 * @param[in] newkey The key for the newly inserted node.
8169 * @param[in] newdata The data for the newly inserted node.
8170 * @param[in] newpgno The page number, if the new node is a branch node.
8171 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8172 * @return 0 on success, non-zero on failure.
8175 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8176 unsigned int nflags)
8179 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8182 int i, j, split_indx, nkeys, pmax;
8183 MDB_env *env = mc->mc_txn->mt_env;
8185 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8186 MDB_page *copy = NULL;
8187 MDB_page *mp, *rp, *pp;
8192 mp = mc->mc_pg[mc->mc_top];
8193 newindx = mc->mc_ki[mc->mc_top];
8194 nkeys = NUMKEYS(mp);
8196 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8197 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8198 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8200 /* Create a right sibling. */
8201 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8203 rp->mp_pad = mp->mp_pad;
8204 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8206 if (mc->mc_snum < 2) {
8207 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8209 /* shift current top to make room for new parent */
8210 mc->mc_pg[1] = mc->mc_pg[0];
8211 mc->mc_ki[1] = mc->mc_ki[0];
8214 mc->mc_db->md_root = pp->mp_pgno;
8215 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8216 new_root = mc->mc_db->md_depth++;
8218 /* Add left (implicit) pointer. */
8219 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8220 /* undo the pre-push */
8221 mc->mc_pg[0] = mc->mc_pg[1];
8222 mc->mc_ki[0] = mc->mc_ki[1];
8223 mc->mc_db->md_root = mp->mp_pgno;
8224 mc->mc_db->md_depth--;
8231 ptop = mc->mc_top-1;
8232 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8235 mc->mc_flags |= C_SPLITTING;
8236 mdb_cursor_copy(mc, &mn);
8237 mn.mc_pg[mn.mc_top] = rp;
8238 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8240 if (nflags & MDB_APPEND) {
8241 mn.mc_ki[mn.mc_top] = 0;
8243 split_indx = newindx;
8247 split_indx = (nkeys+1) / 2;
8252 unsigned int lsize, rsize, ksize;
8253 /* Move half of the keys to the right sibling */
8254 x = mc->mc_ki[mc->mc_top] - split_indx;
8255 ksize = mc->mc_db->md_pad;
8256 split = LEAF2KEY(mp, split_indx, ksize);
8257 rsize = (nkeys - split_indx) * ksize;
8258 lsize = (nkeys - split_indx) * sizeof(indx_t);
8259 mp->mp_lower -= lsize;
8260 rp->mp_lower += lsize;
8261 mp->mp_upper += rsize - lsize;
8262 rp->mp_upper -= rsize - lsize;
8263 sepkey.mv_size = ksize;
8264 if (newindx == split_indx) {
8265 sepkey.mv_data = newkey->mv_data;
8267 sepkey.mv_data = split;
8270 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8271 memcpy(rp->mp_ptrs, split, rsize);
8272 sepkey.mv_data = rp->mp_ptrs;
8273 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8274 memcpy(ins, newkey->mv_data, ksize);
8275 mp->mp_lower += sizeof(indx_t);
8276 mp->mp_upper -= ksize - sizeof(indx_t);
8279 memcpy(rp->mp_ptrs, split, x * ksize);
8280 ins = LEAF2KEY(rp, x, ksize);
8281 memcpy(ins, newkey->mv_data, ksize);
8282 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8283 rp->mp_lower += sizeof(indx_t);
8284 rp->mp_upper -= ksize - sizeof(indx_t);
8285 mc->mc_ki[mc->mc_top] = x;
8286 mc->mc_pg[mc->mc_top] = rp;
8289 int psize, nsize, k;
8290 /* Maximum free space in an empty page */
8291 pmax = env->me_psize - PAGEHDRSZ;
8293 nsize = mdb_leaf_size(env, newkey, newdata);
8295 nsize = mdb_branch_size(env, newkey);
8296 nsize = EVEN(nsize);
8298 /* grab a page to hold a temporary copy */
8299 copy = mdb_page_malloc(mc->mc_txn, 1);
8304 copy->mp_pgno = mp->mp_pgno;
8305 copy->mp_flags = mp->mp_flags;
8306 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8307 copy->mp_upper = env->me_psize - PAGEBASE;
8309 /* prepare to insert */
8310 for (i=0, j=0; i<nkeys; i++) {
8312 copy->mp_ptrs[j++] = 0;
8314 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8317 /* When items are relatively large the split point needs
8318 * to be checked, because being off-by-one will make the
8319 * difference between success or failure in mdb_node_add.
8321 * It's also relevant if a page happens to be laid out
8322 * such that one half of its nodes are all "small" and
8323 * the other half of its nodes are "large." If the new
8324 * item is also "large" and falls on the half with
8325 * "large" nodes, it also may not fit.
8327 * As a final tweak, if the new item goes on the last
8328 * spot on the page (and thus, onto the new page), bias
8329 * the split so the new page is emptier than the old page.
8330 * This yields better packing during sequential inserts.
8332 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8333 /* Find split point */
8335 if (newindx <= split_indx || newindx >= nkeys) {
8337 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8342 for (; i!=k; i+=j) {
8347 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8348 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8350 if (F_ISSET(node->mn_flags, F_BIGDATA))
8351 psize += sizeof(pgno_t);
8353 psize += NODEDSZ(node);
8355 psize = EVEN(psize);
8357 if (psize > pmax || i == k-j) {
8358 split_indx = i + (j<0);
8363 if (split_indx == newindx) {
8364 sepkey.mv_size = newkey->mv_size;
8365 sepkey.mv_data = newkey->mv_data;
8367 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8368 sepkey.mv_size = node->mn_ksize;
8369 sepkey.mv_data = NODEKEY(node);
8374 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8376 /* Copy separator key to the parent.
8378 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8382 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8387 if (mn.mc_snum == mc->mc_snum) {
8388 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8389 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8390 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8391 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8396 /* Right page might now have changed parent.
8397 * Check if left page also changed parent.
8399 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8400 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8401 for (i=0; i<ptop; i++) {
8402 mc->mc_pg[i] = mn.mc_pg[i];
8403 mc->mc_ki[i] = mn.mc_ki[i];
8405 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8406 if (mn.mc_ki[ptop]) {
8407 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8409 /* find right page's left sibling */
8410 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8411 mdb_cursor_sibling(mc, 0);
8416 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8419 mc->mc_flags ^= C_SPLITTING;
8420 if (rc != MDB_SUCCESS) {
8423 if (nflags & MDB_APPEND) {
8424 mc->mc_pg[mc->mc_top] = rp;
8425 mc->mc_ki[mc->mc_top] = 0;
8426 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8429 for (i=0; i<mc->mc_top; i++)
8430 mc->mc_ki[i] = mn.mc_ki[i];
8431 } else if (!IS_LEAF2(mp)) {
8433 mc->mc_pg[mc->mc_top] = rp;
8438 rkey.mv_data = newkey->mv_data;
8439 rkey.mv_size = newkey->mv_size;
8445 /* Update index for the new key. */
8446 mc->mc_ki[mc->mc_top] = j;
8448 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8449 rkey.mv_data = NODEKEY(node);
8450 rkey.mv_size = node->mn_ksize;
8452 xdata.mv_data = NODEDATA(node);
8453 xdata.mv_size = NODEDSZ(node);
8456 pgno = NODEPGNO(node);
8457 flags = node->mn_flags;
8460 if (!IS_LEAF(mp) && j == 0) {
8461 /* First branch index doesn't need key data. */
8465 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8471 mc->mc_pg[mc->mc_top] = copy;
8476 } while (i != split_indx);
8478 nkeys = NUMKEYS(copy);
8479 for (i=0; i<nkeys; i++)
8480 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8481 mp->mp_lower = copy->mp_lower;
8482 mp->mp_upper = copy->mp_upper;
8483 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8484 env->me_psize - copy->mp_upper - PAGEBASE);
8486 /* reset back to original page */
8487 if (newindx < split_indx) {
8488 mc->mc_pg[mc->mc_top] = mp;
8489 if (nflags & MDB_RESERVE) {
8490 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8491 if (!(node->mn_flags & F_BIGDATA))
8492 newdata->mv_data = NODEDATA(node);
8495 mc->mc_pg[mc->mc_top] = rp;
8497 /* Make sure mc_ki is still valid.
8499 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8500 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8501 for (i=0; i<=ptop; i++) {
8502 mc->mc_pg[i] = mn.mc_pg[i];
8503 mc->mc_ki[i] = mn.mc_ki[i];
8510 /* Adjust other cursors pointing to mp */
8511 MDB_cursor *m2, *m3;
8512 MDB_dbi dbi = mc->mc_dbi;
8513 int fixup = NUMKEYS(mp);
8515 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8516 if (mc->mc_flags & C_SUB)
8517 m3 = &m2->mc_xcursor->mx_cursor;
8522 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8524 if (m3->mc_flags & C_SPLITTING)
8529 for (k=new_root; k>=0; k--) {
8530 m3->mc_ki[k+1] = m3->mc_ki[k];
8531 m3->mc_pg[k+1] = m3->mc_pg[k];
8533 if (m3->mc_ki[0] >= split_indx) {
8538 m3->mc_pg[0] = mc->mc_pg[0];
8542 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8543 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8544 m3->mc_ki[mc->mc_top]++;
8545 if (m3->mc_ki[mc->mc_top] >= fixup) {
8546 m3->mc_pg[mc->mc_top] = rp;
8547 m3->mc_ki[mc->mc_top] -= fixup;
8548 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8550 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8551 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8556 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8559 if (copy) /* tmp page */
8560 mdb_page_free(env, copy);
8562 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8567 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8568 MDB_val *key, MDB_val *data, unsigned int flags)
8573 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8576 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8579 mdb_cursor_init(&mc, txn, dbi, &mx);
8580 return mdb_cursor_put(&mc, key, data, flags);
8584 #define MDB_WBUF (1024*1024)
8587 /** State needed for a compacting copy. */
8588 typedef struct mdb_copy {
8589 pthread_mutex_t mc_mutex;
8590 pthread_cond_t mc_cond;
8597 pgno_t mc_next_pgno;
8600 volatile int mc_new;
8605 /** Dedicated writer thread for compacting copy. */
8606 static THREAD_RET ESECT CALL_CONV
8607 mdb_env_copythr(void *arg)
8611 int toggle = 0, wsize, rc;
8614 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8617 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8620 pthread_mutex_lock(&my->mc_mutex);
8622 pthread_cond_signal(&my->mc_cond);
8625 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8626 if (my->mc_new < 0) {
8631 wsize = my->mc_wlen[toggle];
8632 ptr = my->mc_wbuf[toggle];
8635 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8639 } else if (len > 0) {
8653 /* If there's an overflow page tail, write it too */
8654 if (my->mc_olen[toggle]) {
8655 wsize = my->mc_olen[toggle];
8656 ptr = my->mc_over[toggle];
8657 my->mc_olen[toggle] = 0;
8660 my->mc_wlen[toggle] = 0;
8662 pthread_cond_signal(&my->mc_cond);
8664 pthread_cond_signal(&my->mc_cond);
8665 pthread_mutex_unlock(&my->mc_mutex);
8666 return (THREAD_RET)0;
8670 /** Tell the writer thread there's a buffer ready to write */
8672 mdb_env_cthr_toggle(mdb_copy *my, int st)
8674 int toggle = my->mc_toggle ^ 1;
8675 pthread_mutex_lock(&my->mc_mutex);
8676 if (my->mc_status) {
8677 pthread_mutex_unlock(&my->mc_mutex);
8678 return my->mc_status;
8680 while (my->mc_new == 1)
8681 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8683 my->mc_toggle = toggle;
8684 pthread_cond_signal(&my->mc_cond);
8685 pthread_mutex_unlock(&my->mc_mutex);
8689 /** Depth-first tree traversal for compacting copy. */
8691 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8694 MDB_txn *txn = my->mc_txn;
8696 MDB_page *mo, *mp, *leaf;
8701 /* Empty DB, nothing to do */
8702 if (*pg == P_INVALID)
8709 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8712 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8716 /* Make cursor pages writable */
8717 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8721 for (i=0; i<mc.mc_top; i++) {
8722 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8723 mc.mc_pg[i] = (MDB_page *)ptr;
8724 ptr += my->mc_env->me_psize;
8727 /* This is writable space for a leaf page. Usually not needed. */
8728 leaf = (MDB_page *)ptr;
8730 toggle = my->mc_toggle;
8731 while (mc.mc_snum > 0) {
8733 mp = mc.mc_pg[mc.mc_top];
8737 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8738 for (i=0; i<n; i++) {
8739 ni = NODEPTR(mp, i);
8740 if (ni->mn_flags & F_BIGDATA) {
8744 /* Need writable leaf */
8746 mc.mc_pg[mc.mc_top] = leaf;
8747 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8749 ni = NODEPTR(mp, i);
8752 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8753 rc = mdb_page_get(txn, pg, &omp, NULL);
8756 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8757 rc = mdb_env_cthr_toggle(my, 1);
8760 toggle = my->mc_toggle;
8762 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8763 memcpy(mo, omp, my->mc_env->me_psize);
8764 mo->mp_pgno = my->mc_next_pgno;
8765 my->mc_next_pgno += omp->mp_pages;
8766 my->mc_wlen[toggle] += my->mc_env->me_psize;
8767 if (omp->mp_pages > 1) {
8768 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8769 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8770 rc = mdb_env_cthr_toggle(my, 1);
8773 toggle = my->mc_toggle;
8775 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8776 } else if (ni->mn_flags & F_SUBDATA) {
8779 /* Need writable leaf */
8781 mc.mc_pg[mc.mc_top] = leaf;
8782 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8784 ni = NODEPTR(mp, i);
8787 memcpy(&db, NODEDATA(ni), sizeof(db));
8788 my->mc_toggle = toggle;
8789 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8792 toggle = my->mc_toggle;
8793 memcpy(NODEDATA(ni), &db, sizeof(db));
8798 mc.mc_ki[mc.mc_top]++;
8799 if (mc.mc_ki[mc.mc_top] < n) {
8802 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8804 rc = mdb_page_get(txn, pg, &mp, NULL);
8809 mc.mc_ki[mc.mc_top] = 0;
8810 if (IS_BRANCH(mp)) {
8811 /* Whenever we advance to a sibling branch page,
8812 * we must proceed all the way down to its first leaf.
8814 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8817 mc.mc_pg[mc.mc_top] = mp;
8821 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8822 rc = mdb_env_cthr_toggle(my, 1);
8825 toggle = my->mc_toggle;
8827 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8828 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8829 mo->mp_pgno = my->mc_next_pgno++;
8830 my->mc_wlen[toggle] += my->mc_env->me_psize;
8832 /* Update parent if there is one */
8833 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8834 SETPGNO(ni, mo->mp_pgno);
8835 mdb_cursor_pop(&mc);
8837 /* Otherwise we're done */
8847 /** Copy environment with compaction. */
8849 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8854 MDB_txn *txn = NULL;
8859 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8860 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8861 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8862 if (my.mc_wbuf[0] == NULL)
8865 pthread_mutex_init(&my.mc_mutex, NULL);
8866 pthread_cond_init(&my.mc_cond, NULL);
8867 #ifdef HAVE_MEMALIGN
8868 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8869 if (my.mc_wbuf[0] == NULL)
8872 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8877 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8878 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8883 my.mc_next_pgno = 2;
8889 THREAD_CREATE(thr, mdb_env_copythr, &my);
8891 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8895 mp = (MDB_page *)my.mc_wbuf[0];
8896 memset(mp, 0, 2*env->me_psize);
8898 mp->mp_flags = P_META;
8899 mm = (MDB_meta *)METADATA(mp);
8900 mdb_env_init_meta0(env, mm);
8901 mm->mm_address = env->me_metas[0]->mm_address;
8903 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8905 mp->mp_flags = P_META;
8906 *(MDB_meta *)METADATA(mp) = *mm;
8907 mm = (MDB_meta *)METADATA(mp);
8909 /* Count the number of free pages, subtract from lastpg to find
8910 * number of active pages
8913 MDB_ID freecount = 0;
8916 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8917 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8918 freecount += *(MDB_ID *)data.mv_data;
8919 freecount += txn->mt_dbs[0].md_branch_pages +
8920 txn->mt_dbs[0].md_leaf_pages +
8921 txn->mt_dbs[0].md_overflow_pages;
8923 /* Set metapage 1 */
8924 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8925 mm->mm_dbs[1] = txn->mt_dbs[1];
8926 if (mm->mm_last_pg > 1) {
8927 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8930 mm->mm_dbs[1].md_root = P_INVALID;
8933 my.mc_wlen[0] = env->me_psize * 2;
8935 pthread_mutex_lock(&my.mc_mutex);
8937 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8938 pthread_mutex_unlock(&my.mc_mutex);
8939 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8940 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8941 rc = mdb_env_cthr_toggle(&my, 1);
8942 mdb_env_cthr_toggle(&my, -1);
8943 pthread_mutex_lock(&my.mc_mutex);
8945 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8946 pthread_mutex_unlock(&my.mc_mutex);
8951 CloseHandle(my.mc_cond);
8952 CloseHandle(my.mc_mutex);
8953 _aligned_free(my.mc_wbuf[0]);
8955 pthread_cond_destroy(&my.mc_cond);
8956 pthread_mutex_destroy(&my.mc_mutex);
8957 free(my.mc_wbuf[0]);
8962 /** Copy environment as-is. */
8964 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8966 MDB_txn *txn = NULL;
8967 mdb_mutexref_t wmutex = NULL;
8973 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8977 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8980 /* Do the lock/unlock of the reader mutex before starting the
8981 * write txn. Otherwise other read txns could block writers.
8983 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8988 /* We must start the actual read txn after blocking writers */
8989 mdb_txn_reset0(txn, "reset-stage1");
8991 /* Temporarily block writers until we snapshot the meta pages */
8992 wmutex = env->me_wmutex;
8993 if (LOCK_MUTEX(rc, env, wmutex))
8996 rc = mdb_txn_renew0(txn);
8998 UNLOCK_MUTEX(wmutex);
9003 wsize = env->me_psize * 2;
9007 DO_WRITE(rc, fd, ptr, w2, len);
9011 } else if (len > 0) {
9017 /* Non-blocking or async handles are not supported */
9023 UNLOCK_MUTEX(wmutex);
9028 w2 = txn->mt_next_pgno * env->me_psize;
9031 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9038 if (wsize > MAX_WRITE)
9042 DO_WRITE(rc, fd, ptr, w2, len);
9046 } else if (len > 0) {
9063 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9065 if (flags & MDB_CP_COMPACT)
9066 return mdb_env_copyfd1(env, fd);
9068 return mdb_env_copyfd0(env, fd);
9072 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9074 return mdb_env_copyfd2(env, fd, 0);
9078 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9082 HANDLE newfd = INVALID_HANDLE_VALUE;
9084 if (env->me_flags & MDB_NOSUBDIR) {
9085 lpath = (char *)path;
9088 len += sizeof(DATANAME);
9089 lpath = malloc(len);
9092 sprintf(lpath, "%s" DATANAME, path);
9095 /* The destination path must exist, but the destination file must not.
9096 * We don't want the OS to cache the writes, since the source data is
9097 * already in the OS cache.
9100 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9101 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9103 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9105 if (newfd == INVALID_HANDLE_VALUE) {
9110 if (env->me_psize >= env->me_os_psize) {
9112 /* Set O_DIRECT if the file system supports it */
9113 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9114 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9116 #ifdef F_NOCACHE /* __APPLE__ */
9117 rc = fcntl(newfd, F_NOCACHE, 1);
9125 rc = mdb_env_copyfd2(env, newfd, flags);
9128 if (!(env->me_flags & MDB_NOSUBDIR))
9130 if (newfd != INVALID_HANDLE_VALUE)
9131 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9138 mdb_env_copy(MDB_env *env, const char *path)
9140 return mdb_env_copy2(env, path, 0);
9144 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9146 if ((flag & CHANGEABLE) != flag)
9149 env->me_flags |= flag;
9151 env->me_flags &= ~flag;
9156 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9161 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9166 mdb_env_set_userctx(MDB_env *env, void *ctx)
9170 env->me_userctx = ctx;
9175 mdb_env_get_userctx(MDB_env *env)
9177 return env ? env->me_userctx : NULL;
9181 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9186 env->me_assert_func = func;
9192 mdb_env_get_path(MDB_env *env, const char **arg)
9197 *arg = env->me_path;
9202 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9211 /** Common code for #mdb_stat() and #mdb_env_stat().
9212 * @param[in] env the environment to operate in.
9213 * @param[in] db the #MDB_db record containing the stats to return.
9214 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9215 * @return 0, this function always succeeds.
9218 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9220 arg->ms_psize = env->me_psize;
9221 arg->ms_depth = db->md_depth;
9222 arg->ms_branch_pages = db->md_branch_pages;
9223 arg->ms_leaf_pages = db->md_leaf_pages;
9224 arg->ms_overflow_pages = db->md_overflow_pages;
9225 arg->ms_entries = db->md_entries;
9231 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9235 if (env == NULL || arg == NULL)
9238 toggle = mdb_env_pick_meta(env);
9240 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9244 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9248 if (env == NULL || arg == NULL)
9251 toggle = mdb_env_pick_meta(env);
9252 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9253 arg->me_mapsize = env->me_mapsize;
9254 arg->me_maxreaders = env->me_maxreaders;
9255 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9257 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9258 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9262 /** Set the default comparison functions for a database.
9263 * Called immediately after a database is opened to set the defaults.
9264 * The user can then override them with #mdb_set_compare() or
9265 * #mdb_set_dupsort().
9266 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9267 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9270 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9272 uint16_t f = txn->mt_dbs[dbi].md_flags;
9274 txn->mt_dbxs[dbi].md_cmp =
9275 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9276 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9278 txn->mt_dbxs[dbi].md_dcmp =
9279 !(f & MDB_DUPSORT) ? 0 :
9280 ((f & MDB_INTEGERDUP)
9281 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9282 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9285 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9291 int rc, dbflag, exact;
9292 unsigned int unused = 0, seq;
9295 if ((flags & VALID_FLAGS) != flags)
9297 if (txn->mt_flags & MDB_TXN_ERROR)
9303 if (flags & PERSISTENT_FLAGS) {
9304 uint16_t f2 = flags & PERSISTENT_FLAGS;
9305 /* make sure flag changes get committed */
9306 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9307 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9308 txn->mt_flags |= MDB_TXN_DIRTY;
9311 mdb_default_cmp(txn, MAIN_DBI);
9315 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9316 mdb_default_cmp(txn, MAIN_DBI);
9319 /* Is the DB already open? */
9321 for (i=2; i<txn->mt_numdbs; i++) {
9322 if (!txn->mt_dbxs[i].md_name.mv_size) {
9323 /* Remember this free slot */
9324 if (!unused) unused = i;
9327 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9328 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9334 /* If no free slot and max hit, fail */
9335 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9336 return MDB_DBS_FULL;
9338 /* Cannot mix named databases with some mainDB flags */
9339 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9340 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9342 /* Find the DB info */
9343 dbflag = DB_NEW|DB_VALID;
9346 key.mv_data = (void *)name;
9347 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9348 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9349 if (rc == MDB_SUCCESS) {
9350 /* make sure this is actually a DB */
9351 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9352 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9353 return MDB_INCOMPATIBLE;
9354 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9355 /* Create if requested */
9356 data.mv_size = sizeof(MDB_db);
9357 data.mv_data = &dummy;
9358 memset(&dummy, 0, sizeof(dummy));
9359 dummy.md_root = P_INVALID;
9360 dummy.md_flags = flags & PERSISTENT_FLAGS;
9361 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9365 /* OK, got info, add to table */
9366 if (rc == MDB_SUCCESS) {
9367 unsigned int slot = unused ? unused : txn->mt_numdbs;
9368 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9369 txn->mt_dbxs[slot].md_name.mv_size = len;
9370 txn->mt_dbxs[slot].md_rel = NULL;
9371 txn->mt_dbflags[slot] = dbflag;
9372 /* txn-> and env-> are the same in read txns, use
9373 * tmp variable to avoid undefined assignment
9375 seq = ++txn->mt_env->me_dbiseqs[slot];
9376 txn->mt_dbiseqs[slot] = seq;
9378 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9380 mdb_default_cmp(txn, slot);
9389 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9391 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9394 if (txn->mt_flags & MDB_TXN_ERROR)
9397 if (txn->mt_dbflags[dbi] & DB_STALE) {
9400 /* Stale, must read the DB's root. cursor_init does it for us. */
9401 mdb_cursor_init(&mc, txn, dbi, &mx);
9403 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9406 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9409 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9411 ptr = env->me_dbxs[dbi].md_name.mv_data;
9412 /* If there was no name, this was already closed */
9414 env->me_dbxs[dbi].md_name.mv_data = NULL;
9415 env->me_dbxs[dbi].md_name.mv_size = 0;
9416 env->me_dbflags[dbi] = 0;
9417 env->me_dbiseqs[dbi]++;
9422 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9424 /* We could return the flags for the FREE_DBI too but what's the point? */
9425 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9427 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9431 /** Add all the DB's pages to the free list.
9432 * @param[in] mc Cursor on the DB to free.
9433 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9434 * @return 0 on success, non-zero on failure.
9437 mdb_drop0(MDB_cursor *mc, int subs)
9441 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9442 if (rc == MDB_SUCCESS) {
9443 MDB_txn *txn = mc->mc_txn;
9448 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9449 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9452 mdb_cursor_copy(mc, &mx);
9453 while (mc->mc_snum > 0) {
9454 MDB_page *mp = mc->mc_pg[mc->mc_top];
9455 unsigned n = NUMKEYS(mp);
9457 for (i=0; i<n; i++) {
9458 ni = NODEPTR(mp, i);
9459 if (ni->mn_flags & F_BIGDATA) {
9462 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9463 rc = mdb_page_get(txn, pg, &omp, NULL);
9466 mdb_cassert(mc, IS_OVERFLOW(omp));
9467 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9471 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9472 mdb_xcursor_init1(mc, ni);
9473 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9479 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9481 for (i=0; i<n; i++) {
9483 ni = NODEPTR(mp, i);
9486 mdb_midl_xappend(txn->mt_free_pgs, pg);
9491 mc->mc_ki[mc->mc_top] = i;
9492 rc = mdb_cursor_sibling(mc, 1);
9494 if (rc != MDB_NOTFOUND)
9496 /* no more siblings, go back to beginning
9497 * of previous level.
9501 for (i=1; i<mc->mc_snum; i++) {
9503 mc->mc_pg[i] = mx.mc_pg[i];
9508 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9511 txn->mt_flags |= MDB_TXN_ERROR;
9512 } else if (rc == MDB_NOTFOUND) {
9518 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9520 MDB_cursor *mc, *m2;
9523 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9526 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9529 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9532 rc = mdb_cursor_open(txn, dbi, &mc);
9536 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9537 /* Invalidate the dropped DB's cursors */
9538 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9539 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9543 /* Can't delete the main DB */
9544 if (del && dbi > MAIN_DBI) {
9545 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9547 txn->mt_dbflags[dbi] = DB_STALE;
9548 mdb_dbi_close(txn->mt_env, dbi);
9550 txn->mt_flags |= MDB_TXN_ERROR;
9553 /* reset the DB record, mark it dirty */
9554 txn->mt_dbflags[dbi] |= DB_DIRTY;
9555 txn->mt_dbs[dbi].md_depth = 0;
9556 txn->mt_dbs[dbi].md_branch_pages = 0;
9557 txn->mt_dbs[dbi].md_leaf_pages = 0;
9558 txn->mt_dbs[dbi].md_overflow_pages = 0;
9559 txn->mt_dbs[dbi].md_entries = 0;
9560 txn->mt_dbs[dbi].md_root = P_INVALID;
9562 txn->mt_flags |= MDB_TXN_DIRTY;
9565 mdb_cursor_close(mc);
9569 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9571 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9574 txn->mt_dbxs[dbi].md_cmp = cmp;
9578 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9580 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9583 txn->mt_dbxs[dbi].md_dcmp = cmp;
9587 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9589 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9592 txn->mt_dbxs[dbi].md_rel = rel;
9596 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9598 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9601 txn->mt_dbxs[dbi].md_relctx = ctx;
9606 mdb_env_get_maxkeysize(MDB_env *env)
9608 return ENV_MAXKEY(env);
9612 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9614 unsigned int i, rdrs;
9617 int rc = 0, first = 1;
9621 if (!env->me_txns) {
9622 return func("(no reader locks)\n", ctx);
9624 rdrs = env->me_txns->mti_numreaders;
9625 mr = env->me_txns->mti_readers;
9626 for (i=0; i<rdrs; i++) {
9628 txnid_t txnid = mr[i].mr_txnid;
9629 sprintf(buf, txnid == (txnid_t)-1 ?
9630 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9631 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9634 rc = func(" pid thread txnid\n", ctx);
9638 rc = func(buf, ctx);
9644 rc = func("(no active readers)\n", ctx);
9649 /** Insert pid into list if not already present.
9650 * return -1 if already present.
9653 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9655 /* binary search of pid in list */
9657 unsigned cursor = 1;
9659 unsigned n = ids[0];
9662 unsigned pivot = n >> 1;
9663 cursor = base + pivot + 1;
9664 val = pid - ids[cursor];
9669 } else if ( val > 0 ) {
9674 /* found, so it's a duplicate */
9683 for (n = ids[0]; n > cursor; n--)
9690 mdb_reader_check(MDB_env *env, int *dead)
9696 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9699 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9700 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9702 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9703 unsigned int i, j, rdrs;
9705 MDB_PID_T *pids, pid;
9706 int rc = MDB_SUCCESS, count = 0;
9708 rdrs = env->me_txns->mti_numreaders;
9709 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9713 mr = env->me_txns->mti_readers;
9714 for (i=0; i<rdrs; i++) {
9716 if (pid && pid != env->me_pid) {
9717 if (mdb_pid_insert(pids, pid) == 0) {
9718 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9719 /* Stale reader found */
9722 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9723 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9725 rdrs = 0; /* the above checked all readers */
9727 /* Recheck, a new process may have reused pid */
9728 if (mdb_reader_pid(env, Pidcheck, pid))
9733 if (mr[j].mr_pid == pid) {
9734 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9735 (unsigned) pid, mr[j].mr_txnid));
9740 UNLOCK_MUTEX(rmutex);
9751 #ifdef MDB_ROBUST_SUPPORTED
9752 /** Handle #LOCK_MUTEX0() failure.
9753 * Try to repair the lock file if the mutex owner died.
9754 * @param[in] env the environment handle
9755 * @param[in] mutex LOCK_MUTEX0() mutex
9756 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9757 * @return 0 on success with the mutex locked, or an error code on failure.
9759 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9761 int toggle, rlocked, rc2;
9763 if (rc == MDB_OWNERDEAD) {
9764 /* We own the mutex. Clean up after dead previous owner. */
9766 rlocked = (mutex == env->me_rmutex);
9768 /* Keep mti_txnid updated, otherwise next writer can
9769 * overwrite data which latest meta page refers to.
9771 toggle = mdb_env_pick_meta(env);
9772 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9773 /* env is hosed if the dead thread was ours */
9775 env->me_flags |= MDB_FATAL_ERROR;
9780 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9781 (rc ? "this process' env is hosed" : "recovering")));
9782 rc2 = mdb_reader_check0(env, rlocked, NULL);
9784 rc2 = mdb_mutex_consistent(mutex);
9785 if (rc || (rc = rc2)) {
9786 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9787 UNLOCK_MUTEX(mutex);
9793 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9798 #endif /* MDB_ROBUST_SUPPORTED */