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 */
1040 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1042 /** In write txns, array of cursors for each DB */
1043 MDB_cursor **mt_cursors;
1044 /** Array of flags for each DB */
1045 unsigned char *mt_dbflags;
1046 /** Number of DB records in use. This number only ever increments;
1047 * we don't decrement it when individual DB handles are closed.
1051 /** @defgroup mdb_txn Transaction Flags
1055 /** #mdb_txn_begin() flags */
1056 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1057 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1058 /* internal txn flags */
1059 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1060 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1061 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1062 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1064 unsigned int mt_flags; /**< @ref mdb_txn */
1065 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1066 * Includes ancestor txns' dirty pages not hidden by other txns'
1067 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1068 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1070 unsigned int mt_dirty_room;
1073 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1074 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1075 * raise this on a 64 bit machine.
1077 #define CURSOR_STACK 32
1081 /** Cursors are used for all DB operations.
1082 * A cursor holds a path of (page pointer, key index) from the DB
1083 * root to a position in the DB, plus other state. #MDB_DUPSORT
1084 * cursors include an xcursor to the current data item. Write txns
1085 * track their cursors and keep them up to date when data moves.
1086 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1087 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1090 /** Next cursor on this DB in this txn */
1091 MDB_cursor *mc_next;
1092 /** Backup of the original cursor if this cursor is a shadow */
1093 MDB_cursor *mc_backup;
1094 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1095 struct MDB_xcursor *mc_xcursor;
1096 /** The transaction that owns this cursor */
1098 /** The database handle this cursor operates on */
1100 /** The database record for this cursor */
1102 /** The database auxiliary record for this cursor */
1104 /** The @ref mt_dbflag for this database */
1105 unsigned char *mc_dbflag;
1106 unsigned short mc_snum; /**< number of pushed pages */
1107 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1108 /** @defgroup mdb_cursor Cursor Flags
1110 * Cursor state flags.
1113 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1114 #define C_EOF 0x02 /**< No more data */
1115 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1116 #define C_DEL 0x08 /**< last op was a cursor_del */
1117 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1118 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1120 unsigned int mc_flags; /**< @ref mdb_cursor */
1121 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1122 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1125 /** Context for sorted-dup records.
1126 * We could have gone to a fully recursive design, with arbitrarily
1127 * deep nesting of sub-databases. But for now we only handle these
1128 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1130 typedef struct MDB_xcursor {
1131 /** A sub-cursor for traversing the Dup DB */
1132 MDB_cursor mx_cursor;
1133 /** The database record for this Dup DB */
1135 /** The auxiliary DB record for this Dup DB */
1137 /** The @ref mt_dbflag for this Dup DB */
1138 unsigned char mx_dbflag;
1141 /** State of FreeDB old pages, stored in the MDB_env */
1142 typedef struct MDB_pgstate {
1143 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1144 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1147 /** The database environment. */
1149 HANDLE me_fd; /**< The main data file */
1150 HANDLE me_lfd; /**< The lock file */
1151 HANDLE me_mfd; /**< just for writing the meta pages */
1152 /** Failed to update the meta page. Probably an I/O error. */
1153 #define MDB_FATAL_ERROR 0x80000000U
1154 /** Some fields are initialized. */
1155 #define MDB_ENV_ACTIVE 0x20000000U
1156 /** me_txkey is set */
1157 #define MDB_ENV_TXKEY 0x10000000U
1158 /** fdatasync is unreliable */
1159 #define MDB_FSYNCONLY 0x08000000U
1160 uint32_t me_flags; /**< @ref mdb_env */
1161 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1162 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1163 unsigned int me_maxreaders; /**< size of the reader table */
1164 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1165 volatile int me_close_readers;
1166 MDB_dbi me_numdbs; /**< number of DBs opened */
1167 MDB_dbi me_maxdbs; /**< size of the DB table */
1168 MDB_PID_T me_pid; /**< process ID of this env */
1169 char *me_path; /**< path to the DB files */
1170 char *me_map; /**< the memory map of the data file */
1171 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1172 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1173 void *me_pbuf; /**< scratch area for DUPSORT put() */
1174 MDB_txn *me_txn; /**< current write transaction */
1175 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1176 size_t me_mapsize; /**< size of the data memory map */
1177 off_t me_size; /**< current file size */
1178 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1179 MDB_dbx *me_dbxs; /**< array of static DB info */
1180 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1181 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1182 pthread_key_t me_txkey; /**< thread-key for readers */
1183 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1184 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1185 # define me_pglast me_pgstate.mf_pglast
1186 # define me_pghead me_pgstate.mf_pghead
1187 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1188 /** IDL of pages that became unused in a write txn */
1189 MDB_IDL me_free_pgs;
1190 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1191 MDB_ID2L me_dirty_list;
1192 /** Max number of freelist items that can fit in a single overflow page */
1194 /** Max size of a node on a page */
1195 unsigned int me_nodemax;
1196 #if !(MDB_MAXKEYSIZE)
1197 unsigned int me_maxkey; /**< max size of a key */
1199 int me_live_reader; /**< have liveness lock in reader table */
1201 int me_pidquery; /**< Used in OpenProcess */
1203 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1204 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1205 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1207 mdb_mutex_t me_rmutex;
1208 mdb_mutex_t me_wmutex;
1210 void *me_userctx; /**< User-settable context */
1211 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1214 /** Nested transaction */
1215 typedef struct MDB_ntxn {
1216 MDB_txn mnt_txn; /**< the transaction */
1217 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1220 /** max number of pages to commit in one writev() call */
1221 #define MDB_COMMIT_PAGES 64
1222 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1223 #undef MDB_COMMIT_PAGES
1224 #define MDB_COMMIT_PAGES IOV_MAX
1227 /** max bytes to write in one call */
1228 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1230 /** Check \b txn and \b dbi arguments to a function */
1231 #define TXN_DBI_EXIST(txn, dbi, validity) \
1232 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1234 /** Check for misused \b dbi handles */
1235 #define TXN_DBI_CHANGED(txn, dbi) \
1236 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1238 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1239 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1240 static int mdb_page_touch(MDB_cursor *mc);
1242 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1243 static int mdb_page_search_root(MDB_cursor *mc,
1244 MDB_val *key, int modify);
1245 #define MDB_PS_MODIFY 1
1246 #define MDB_PS_ROOTONLY 2
1247 #define MDB_PS_FIRST 4
1248 #define MDB_PS_LAST 8
1249 static int mdb_page_search(MDB_cursor *mc,
1250 MDB_val *key, int flags);
1251 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1253 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1254 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1255 pgno_t newpgno, unsigned int nflags);
1257 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1258 static int mdb_env_pick_meta(const MDB_env *env);
1259 static int mdb_env_write_meta(MDB_txn *txn);
1260 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1261 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1263 static void mdb_env_close0(MDB_env *env, int excl);
1265 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1266 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1267 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1268 static void mdb_node_del(MDB_cursor *mc, int ksize);
1269 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1270 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1271 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1272 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1273 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1275 static int mdb_rebalance(MDB_cursor *mc);
1276 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1278 static void mdb_cursor_pop(MDB_cursor *mc);
1279 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1281 static int mdb_cursor_del0(MDB_cursor *mc);
1282 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1283 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1284 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1285 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1286 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1288 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1289 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1291 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1292 static void mdb_xcursor_init0(MDB_cursor *mc);
1293 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1294 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1296 static int mdb_drop0(MDB_cursor *mc, int subs);
1297 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1298 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1301 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1304 /** Compare two items pointing at size_t's of unknown alignment. */
1305 #ifdef MISALIGNED_OK
1306 # define mdb_cmp_clong mdb_cmp_long
1308 # define mdb_cmp_clong mdb_cmp_cint
1312 static SECURITY_DESCRIPTOR mdb_null_sd;
1313 static SECURITY_ATTRIBUTES mdb_all_sa;
1314 static int mdb_sec_inited;
1317 /** Return the library version info. */
1319 mdb_version(int *major, int *minor, int *patch)
1321 if (major) *major = MDB_VERSION_MAJOR;
1322 if (minor) *minor = MDB_VERSION_MINOR;
1323 if (patch) *patch = MDB_VERSION_PATCH;
1324 return MDB_VERSION_STRING;
1327 /** Table of descriptions for LMDB @ref errors */
1328 static char *const mdb_errstr[] = {
1329 "MDB_KEYEXIST: Key/data pair already exists",
1330 "MDB_NOTFOUND: No matching key/data pair found",
1331 "MDB_PAGE_NOTFOUND: Requested page not found",
1332 "MDB_CORRUPTED: Located page was wrong type",
1333 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1334 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1335 "MDB_INVALID: File is not an LMDB file",
1336 "MDB_MAP_FULL: Environment mapsize limit reached",
1337 "MDB_DBS_FULL: Environment maxdbs limit reached",
1338 "MDB_READERS_FULL: Environment maxreaders limit reached",
1339 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1340 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1341 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1342 "MDB_PAGE_FULL: Internal error - page has no more space",
1343 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1344 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1345 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1346 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1347 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1348 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1352 mdb_strerror(int err)
1355 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1356 * This works as long as no function between the call to mdb_strerror
1357 * and the actual use of the message uses more than 4K of stack.
1360 char buf[1024], *ptr = buf;
1364 return ("Successful return: 0");
1366 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1367 i = err - MDB_KEYEXIST;
1368 return mdb_errstr[i];
1372 /* These are the C-runtime error codes we use. The comment indicates
1373 * their numeric value, and the Win32 error they would correspond to
1374 * if the error actually came from a Win32 API. A major mess, we should
1375 * have used LMDB-specific error codes for everything.
1378 case ENOENT: /* 2, FILE_NOT_FOUND */
1379 case EIO: /* 5, ACCESS_DENIED */
1380 case ENOMEM: /* 12, INVALID_ACCESS */
1381 case EACCES: /* 13, INVALID_DATA */
1382 case EBUSY: /* 16, CURRENT_DIRECTORY */
1383 case EINVAL: /* 22, BAD_COMMAND */
1384 case ENOSPC: /* 28, OUT_OF_PAPER */
1385 return strerror(err);
1390 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1391 FORMAT_MESSAGE_IGNORE_INSERTS,
1392 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1395 return strerror(err);
1399 /** assert(3) variant in cursor context */
1400 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1401 /** assert(3) variant in transaction context */
1402 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1403 /** assert(3) variant in environment context */
1404 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1407 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1408 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1411 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1412 const char *func, const char *file, int line)
1415 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1416 file, line, expr_txt, func);
1417 if (env->me_assert_func)
1418 env->me_assert_func(env, buf);
1419 fprintf(stderr, "%s\n", buf);
1423 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1427 /** Return the page number of \b mp which may be sub-page, for debug output */
1429 mdb_dbg_pgno(MDB_page *mp)
1432 COPY_PGNO(ret, mp->mp_pgno);
1436 /** Display a key in hexadecimal and return the address of the result.
1437 * @param[in] key the key to display
1438 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1439 * @return The key in hexadecimal form.
1442 mdb_dkey(MDB_val *key, char *buf)
1445 unsigned char *c = key->mv_data;
1451 if (key->mv_size > DKBUF_MAXKEYSIZE)
1452 return "MDB_MAXKEYSIZE";
1453 /* may want to make this a dynamic check: if the key is mostly
1454 * printable characters, print it as-is instead of converting to hex.
1458 for (i=0; i<key->mv_size; i++)
1459 ptr += sprintf(ptr, "%02x", *c++);
1461 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1467 mdb_leafnode_type(MDB_node *n)
1469 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1470 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1471 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1474 /** Display all the keys in the page. */
1476 mdb_page_list(MDB_page *mp)
1478 pgno_t pgno = mdb_dbg_pgno(mp);
1479 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1481 unsigned int i, nkeys, nsize, total = 0;
1485 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1486 case P_BRANCH: type = "Branch page"; break;
1487 case P_LEAF: type = "Leaf page"; break;
1488 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1489 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1490 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1492 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1493 pgno, mp->mp_pages, state);
1496 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1497 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1500 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1504 nkeys = NUMKEYS(mp);
1505 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1507 for (i=0; i<nkeys; i++) {
1508 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1509 key.mv_size = nsize = mp->mp_pad;
1510 key.mv_data = LEAF2KEY(mp, i, nsize);
1512 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1515 node = NODEPTR(mp, i);
1516 key.mv_size = node->mn_ksize;
1517 key.mv_data = node->mn_data;
1518 nsize = NODESIZE + key.mv_size;
1519 if (IS_BRANCH(mp)) {
1520 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1524 if (F_ISSET(node->mn_flags, F_BIGDATA))
1525 nsize += sizeof(pgno_t);
1527 nsize += NODEDSZ(node);
1529 nsize += sizeof(indx_t);
1530 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1531 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1533 total = EVEN(total);
1535 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1536 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1540 mdb_cursor_chk(MDB_cursor *mc)
1546 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1547 for (i=0; i<mc->mc_top; i++) {
1549 node = NODEPTR(mp, mc->mc_ki[i]);
1550 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1553 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1559 /** Count all the pages in each DB and in the freelist
1560 * and make sure it matches the actual number of pages
1562 * All named DBs must be open for a correct count.
1564 static void mdb_audit(MDB_txn *txn)
1568 MDB_ID freecount, count;
1573 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1574 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1575 freecount += *(MDB_ID *)data.mv_data;
1576 mdb_tassert(txn, rc == MDB_NOTFOUND);
1579 for (i = 0; i<txn->mt_numdbs; i++) {
1581 if (!(txn->mt_dbflags[i] & DB_VALID))
1583 mdb_cursor_init(&mc, txn, i, &mx);
1584 if (txn->mt_dbs[i].md_root == P_INVALID)
1586 count += txn->mt_dbs[i].md_branch_pages +
1587 txn->mt_dbs[i].md_leaf_pages +
1588 txn->mt_dbs[i].md_overflow_pages;
1589 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1590 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1591 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1594 mp = mc.mc_pg[mc.mc_top];
1595 for (j=0; j<NUMKEYS(mp); j++) {
1596 MDB_node *leaf = NODEPTR(mp, j);
1597 if (leaf->mn_flags & F_SUBDATA) {
1599 memcpy(&db, NODEDATA(leaf), sizeof(db));
1600 count += db.md_branch_pages + db.md_leaf_pages +
1601 db.md_overflow_pages;
1605 mdb_tassert(txn, rc == MDB_NOTFOUND);
1608 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1609 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1610 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1616 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1618 return txn->mt_dbxs[dbi].md_cmp(a, b);
1622 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1624 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1625 #if UINT_MAX < SIZE_MAX
1626 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1627 dcmp = mdb_cmp_clong;
1632 /** Allocate memory for a page.
1633 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1636 mdb_page_malloc(MDB_txn *txn, unsigned num)
1638 MDB_env *env = txn->mt_env;
1639 MDB_page *ret = env->me_dpages;
1640 size_t psize = env->me_psize, sz = psize, off;
1641 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1642 * For a single page alloc, we init everything after the page header.
1643 * For multi-page, we init the final page; if the caller needed that
1644 * many pages they will be filling in at least up to the last page.
1648 VGMEMP_ALLOC(env, ret, sz);
1649 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1650 env->me_dpages = ret->mp_next;
1653 psize -= off = PAGEHDRSZ;
1658 if ((ret = malloc(sz)) != NULL) {
1659 VGMEMP_ALLOC(env, ret, sz);
1660 if (!(env->me_flags & MDB_NOMEMINIT)) {
1661 memset((char *)ret + off, 0, psize);
1665 txn->mt_flags |= MDB_TXN_ERROR;
1669 /** Free a single page.
1670 * Saves single pages to a list, for future reuse.
1671 * (This is not used for multi-page overflow pages.)
1674 mdb_page_free(MDB_env *env, MDB_page *mp)
1676 mp->mp_next = env->me_dpages;
1677 VGMEMP_FREE(env, mp);
1678 env->me_dpages = mp;
1681 /** Free a dirty page */
1683 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1685 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1686 mdb_page_free(env, dp);
1688 /* large pages just get freed directly */
1689 VGMEMP_FREE(env, dp);
1694 /** Return all dirty pages to dpage list */
1696 mdb_dlist_free(MDB_txn *txn)
1698 MDB_env *env = txn->mt_env;
1699 MDB_ID2L dl = txn->mt_u.dirty_list;
1700 unsigned i, n = dl[0].mid;
1702 for (i = 1; i <= n; i++) {
1703 mdb_dpage_free(env, dl[i].mptr);
1708 /** Loosen or free a single page.
1709 * Saves single pages to a list for future reuse
1710 * in this same txn. It has been pulled from the freeDB
1711 * and already resides on the dirty list, but has been
1712 * deleted. Use these pages first before pulling again
1715 * If the page wasn't dirtied in this txn, just add it
1716 * to this txn's free list.
1719 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1722 pgno_t pgno = mp->mp_pgno;
1723 MDB_txn *txn = mc->mc_txn;
1725 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1726 if (txn->mt_parent) {
1727 MDB_ID2 *dl = txn->mt_u.dirty_list;
1728 /* If txn has a parent, make sure the page is in our
1732 unsigned x = mdb_mid2l_search(dl, pgno);
1733 if (x <= dl[0].mid && dl[x].mid == pgno) {
1734 if (mp != dl[x].mptr) { /* bad cursor? */
1735 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1736 txn->mt_flags |= MDB_TXN_ERROR;
1737 return MDB_CORRUPTED;
1744 /* no parent txn, so it's just ours */
1749 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1751 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1752 txn->mt_loose_pgs = mp;
1753 txn->mt_loose_count++;
1754 mp->mp_flags |= P_LOOSE;
1756 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1764 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1765 * @param[in] mc A cursor handle for the current operation.
1766 * @param[in] pflags Flags of the pages to update:
1767 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1768 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1769 * @return 0 on success, non-zero on failure.
1772 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1774 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1775 MDB_txn *txn = mc->mc_txn;
1781 int rc = MDB_SUCCESS, level;
1783 /* Mark pages seen by cursors */
1784 if (mc->mc_flags & C_UNTRACK)
1785 mc = NULL; /* will find mc in mt_cursors */
1786 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1787 for (; mc; mc=mc->mc_next) {
1788 if (!(mc->mc_flags & C_INITIALIZED))
1790 for (m3 = mc;; m3 = &mx->mx_cursor) {
1792 for (j=0; j<m3->mc_snum; j++) {
1794 if ((mp->mp_flags & Mask) == pflags)
1795 mp->mp_flags ^= P_KEEP;
1797 mx = m3->mc_xcursor;
1798 /* Proceed to mx if it is at a sub-database */
1799 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1801 if (! (mp && (mp->mp_flags & P_LEAF)))
1803 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1804 if (!(leaf->mn_flags & F_SUBDATA))
1813 /* Mark dirty root pages */
1814 for (i=0; i<txn->mt_numdbs; i++) {
1815 if (txn->mt_dbflags[i] & DB_DIRTY) {
1816 pgno_t pgno = txn->mt_dbs[i].md_root;
1817 if (pgno == P_INVALID)
1819 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1821 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1822 dp->mp_flags ^= P_KEEP;
1830 static int mdb_page_flush(MDB_txn *txn, int keep);
1832 /** Spill pages from the dirty list back to disk.
1833 * This is intended to prevent running into #MDB_TXN_FULL situations,
1834 * but note that they may still occur in a few cases:
1835 * 1) our estimate of the txn size could be too small. Currently this
1836 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1837 * 2) child txns may run out of space if their parents dirtied a
1838 * lot of pages and never spilled them. TODO: we probably should do
1839 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1840 * the parent's dirty_room is below a given threshold.
1842 * Otherwise, if not using nested txns, it is expected that apps will
1843 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1844 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1845 * If the txn never references them again, they can be left alone.
1846 * If the txn only reads them, they can be used without any fuss.
1847 * If the txn writes them again, they can be dirtied immediately without
1848 * going thru all of the work of #mdb_page_touch(). Such references are
1849 * handled by #mdb_page_unspill().
1851 * Also note, we never spill DB root pages, nor pages of active cursors,
1852 * because we'll need these back again soon anyway. And in nested txns,
1853 * we can't spill a page in a child txn if it was already spilled in a
1854 * parent txn. That would alter the parent txns' data even though
1855 * the child hasn't committed yet, and we'd have no way to undo it if
1856 * the child aborted.
1858 * @param[in] m0 cursor A cursor handle identifying the transaction and
1859 * database for which we are checking space.
1860 * @param[in] key For a put operation, the key being stored.
1861 * @param[in] data For a put operation, the data being stored.
1862 * @return 0 on success, non-zero on failure.
1865 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1867 MDB_txn *txn = m0->mc_txn;
1869 MDB_ID2L dl = txn->mt_u.dirty_list;
1870 unsigned int i, j, need;
1873 if (m0->mc_flags & C_SUB)
1876 /* Estimate how much space this op will take */
1877 i = m0->mc_db->md_depth;
1878 /* Named DBs also dirty the main DB */
1879 if (m0->mc_dbi > MAIN_DBI)
1880 i += txn->mt_dbs[MAIN_DBI].md_depth;
1881 /* For puts, roughly factor in the key+data size */
1883 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1884 i += i; /* double it for good measure */
1887 if (txn->mt_dirty_room > i)
1890 if (!txn->mt_spill_pgs) {
1891 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1892 if (!txn->mt_spill_pgs)
1895 /* purge deleted slots */
1896 MDB_IDL sl = txn->mt_spill_pgs;
1897 unsigned int num = sl[0];
1899 for (i=1; i<=num; i++) {
1906 /* Preserve pages which may soon be dirtied again */
1907 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1910 /* Less aggressive spill - we originally spilled the entire dirty list,
1911 * with a few exceptions for cursor pages and DB root pages. But this
1912 * turns out to be a lot of wasted effort because in a large txn many
1913 * of those pages will need to be used again. So now we spill only 1/8th
1914 * of the dirty pages. Testing revealed this to be a good tradeoff,
1915 * better than 1/2, 1/4, or 1/10.
1917 if (need < MDB_IDL_UM_MAX / 8)
1918 need = MDB_IDL_UM_MAX / 8;
1920 /* Save the page IDs of all the pages we're flushing */
1921 /* flush from the tail forward, this saves a lot of shifting later on. */
1922 for (i=dl[0].mid; i && need; i--) {
1923 MDB_ID pn = dl[i].mid << 1;
1925 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1927 /* Can't spill twice, make sure it's not already in a parent's
1930 if (txn->mt_parent) {
1932 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1933 if (tx2->mt_spill_pgs) {
1934 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1935 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1936 dp->mp_flags |= P_KEEP;
1944 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1948 mdb_midl_sort(txn->mt_spill_pgs);
1950 /* Flush the spilled part of dirty list */
1951 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1954 /* Reset any dirty pages we kept that page_flush didn't see */
1955 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1958 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1962 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1964 mdb_find_oldest(MDB_txn *txn)
1967 txnid_t mr, oldest = txn->mt_txnid - 1;
1968 if (txn->mt_env->me_txns) {
1969 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1970 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1981 /** Add a page to the txn's dirty list */
1983 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1986 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1988 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
1989 insert = mdb_mid2l_append;
1991 insert = mdb_mid2l_insert;
1993 mid.mid = mp->mp_pgno;
1995 rc = insert(txn->mt_u.dirty_list, &mid);
1996 mdb_tassert(txn, rc == 0);
1997 txn->mt_dirty_room--;
2000 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2001 * me_pghead and mt_next_pgno.
2003 * If there are free pages available from older transactions, they
2004 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2005 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2006 * and move me_pglast to say which records were consumed. Only this
2007 * function can create me_pghead and move me_pglast/mt_next_pgno.
2008 * @param[in] mc cursor A cursor handle identifying the transaction and
2009 * database for which we are allocating.
2010 * @param[in] num the number of pages to allocate.
2011 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2012 * will always be satisfied by a single contiguous chunk of memory.
2013 * @return 0 on success, non-zero on failure.
2016 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2018 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2019 /* Get at most <Max_retries> more freeDB records once me_pghead
2020 * has enough pages. If not enough, use new pages from the map.
2021 * If <Paranoid> and mc is updating the freeDB, only get new
2022 * records if me_pghead is empty. Then the freelist cannot play
2023 * catch-up with itself by growing while trying to save it.
2025 enum { Paranoid = 1, Max_retries = 500 };
2027 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2029 int rc, retry = num * 60;
2030 MDB_txn *txn = mc->mc_txn;
2031 MDB_env *env = txn->mt_env;
2032 pgno_t pgno, *mop = env->me_pghead;
2033 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2035 txnid_t oldest = 0, last;
2040 /* If there are any loose pages, just use them */
2041 if (num == 1 && txn->mt_loose_pgs) {
2042 np = txn->mt_loose_pgs;
2043 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2044 txn->mt_loose_count--;
2045 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2053 /* If our dirty list is already full, we can't do anything */
2054 if (txn->mt_dirty_room == 0) {
2059 for (op = MDB_FIRST;; op = MDB_NEXT) {
2064 /* Seek a big enough contiguous page range. Prefer
2065 * pages at the tail, just truncating the list.
2071 if (mop[i-n2] == pgno+n2)
2078 if (op == MDB_FIRST) { /* 1st iteration */
2079 /* Prepare to fetch more and coalesce */
2080 last = env->me_pglast;
2081 oldest = env->me_pgoldest;
2082 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2085 key.mv_data = &last; /* will look up last+1 */
2086 key.mv_size = sizeof(last);
2088 if (Paranoid && mc->mc_dbi == FREE_DBI)
2091 if (Paranoid && retry < 0 && mop_len)
2095 /* Do not fetch more if the record will be too recent */
2096 if (oldest <= last) {
2098 oldest = mdb_find_oldest(txn);
2099 env->me_pgoldest = oldest;
2105 rc = mdb_cursor_get(&m2, &key, NULL, op);
2107 if (rc == MDB_NOTFOUND)
2111 last = *(txnid_t*)key.mv_data;
2112 if (oldest <= last) {
2114 oldest = mdb_find_oldest(txn);
2115 env->me_pgoldest = oldest;
2121 np = m2.mc_pg[m2.mc_top];
2122 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2123 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2126 idl = (MDB_ID *) data.mv_data;
2129 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2134 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2136 mop = env->me_pghead;
2138 env->me_pglast = last;
2140 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2141 last, txn->mt_dbs[FREE_DBI].md_root, i));
2143 DPRINTF(("IDL %"Z"u", idl[j]));
2145 /* Merge in descending sorted order */
2146 mdb_midl_xmerge(mop, idl);
2150 /* Use new pages from the map when nothing suitable in the freeDB */
2152 pgno = txn->mt_next_pgno;
2153 if (pgno + num >= env->me_maxpg) {
2154 DPUTS("DB size maxed out");
2160 if (env->me_flags & MDB_WRITEMAP) {
2161 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2163 if (!(np = mdb_page_malloc(txn, num))) {
2169 mop[0] = mop_len -= num;
2170 /* Move any stragglers down */
2171 for (j = i-num; j < mop_len; )
2172 mop[++j] = mop[++i];
2174 txn->mt_next_pgno = pgno + num;
2177 mdb_page_dirty(txn, np);
2183 txn->mt_flags |= MDB_TXN_ERROR;
2187 /** Copy the used portions of a non-overflow page.
2188 * @param[in] dst page to copy into
2189 * @param[in] src page to copy from
2190 * @param[in] psize size of a page
2193 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2195 enum { Align = sizeof(pgno_t) };
2196 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2198 /* If page isn't full, just copy the used portion. Adjust
2199 * alignment so memcpy may copy words instead of bytes.
2201 if ((unused &= -Align) && !IS_LEAF2(src)) {
2202 upper = (upper + PAGEBASE) & -Align;
2203 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2204 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2207 memcpy(dst, src, psize - unused);
2211 /** Pull a page off the txn's spill list, if present.
2212 * If a page being referenced was spilled to disk in this txn, bring
2213 * it back and make it dirty/writable again.
2214 * @param[in] txn the transaction handle.
2215 * @param[in] mp the page being referenced. It must not be dirty.
2216 * @param[out] ret the writable page, if any. ret is unchanged if
2217 * mp wasn't spilled.
2220 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2222 MDB_env *env = txn->mt_env;
2225 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2227 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2228 if (!tx2->mt_spill_pgs)
2230 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2231 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2234 if (txn->mt_dirty_room == 0)
2235 return MDB_TXN_FULL;
2236 if (IS_OVERFLOW(mp))
2240 if (env->me_flags & MDB_WRITEMAP) {
2243 np = mdb_page_malloc(txn, num);
2247 memcpy(np, mp, num * env->me_psize);
2249 mdb_page_copy(np, mp, env->me_psize);
2252 /* If in current txn, this page is no longer spilled.
2253 * If it happens to be the last page, truncate the spill list.
2254 * Otherwise mark it as deleted by setting the LSB.
2256 if (x == txn->mt_spill_pgs[0])
2257 txn->mt_spill_pgs[0]--;
2259 txn->mt_spill_pgs[x] |= 1;
2260 } /* otherwise, if belonging to a parent txn, the
2261 * page remains spilled until child commits
2264 mdb_page_dirty(txn, np);
2265 np->mp_flags |= P_DIRTY;
2273 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2274 * @param[in] mc cursor pointing to the page to be touched
2275 * @return 0 on success, non-zero on failure.
2278 mdb_page_touch(MDB_cursor *mc)
2280 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2281 MDB_txn *txn = mc->mc_txn;
2282 MDB_cursor *m2, *m3;
2286 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2287 if (txn->mt_flags & MDB_TXN_SPILLS) {
2289 rc = mdb_page_unspill(txn, mp, &np);
2295 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2296 (rc = mdb_page_alloc(mc, 1, &np)))
2299 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2300 mp->mp_pgno, pgno));
2301 mdb_cassert(mc, mp->mp_pgno != pgno);
2302 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2303 /* Update the parent page, if any, to point to the new page */
2305 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2306 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2307 SETPGNO(node, pgno);
2309 mc->mc_db->md_root = pgno;
2311 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2312 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2314 /* If txn has a parent, make sure the page is in our
2318 unsigned x = mdb_mid2l_search(dl, pgno);
2319 if (x <= dl[0].mid && dl[x].mid == pgno) {
2320 if (mp != dl[x].mptr) { /* bad cursor? */
2321 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2322 txn->mt_flags |= MDB_TXN_ERROR;
2323 return MDB_CORRUPTED;
2328 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2330 np = mdb_page_malloc(txn, 1);
2335 rc = mdb_mid2l_insert(dl, &mid);
2336 mdb_cassert(mc, rc == 0);
2341 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2343 np->mp_flags |= P_DIRTY;
2346 /* Adjust cursors pointing to mp */
2347 mc->mc_pg[mc->mc_top] = np;
2348 m2 = txn->mt_cursors[mc->mc_dbi];
2349 if (mc->mc_flags & C_SUB) {
2350 for (; m2; m2=m2->mc_next) {
2351 m3 = &m2->mc_xcursor->mx_cursor;
2352 if (m3->mc_snum < mc->mc_snum) continue;
2353 if (m3->mc_pg[mc->mc_top] == mp)
2354 m3->mc_pg[mc->mc_top] = np;
2357 for (; m2; m2=m2->mc_next) {
2358 if (m2->mc_snum < mc->mc_snum) continue;
2359 if (m2->mc_pg[mc->mc_top] == mp) {
2360 m2->mc_pg[mc->mc_top] = np;
2361 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2363 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2365 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2366 if (!(leaf->mn_flags & F_SUBDATA))
2367 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2375 txn->mt_flags |= MDB_TXN_ERROR;
2380 mdb_env_sync(MDB_env *env, int force)
2383 if (env->me_flags & MDB_RDONLY)
2385 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2386 if (env->me_flags & MDB_WRITEMAP) {
2387 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2388 ? MS_ASYNC : MS_SYNC;
2389 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2392 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2396 #ifdef BROKEN_FDATASYNC
2397 if (env->me_flags & MDB_FSYNCONLY) {
2398 if (fsync(env->me_fd))
2402 if (MDB_FDATASYNC(env->me_fd))
2409 /** Back up parent txn's cursors, then grab the originals for tracking */
2411 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2413 MDB_cursor *mc, *bk;
2418 for (i = src->mt_numdbs; --i >= 0; ) {
2419 if ((mc = src->mt_cursors[i]) != NULL) {
2420 size = sizeof(MDB_cursor);
2422 size += sizeof(MDB_xcursor);
2423 for (; mc; mc = bk->mc_next) {
2429 mc->mc_db = &dst->mt_dbs[i];
2430 /* Kill pointers into src - and dst to reduce abuse: The
2431 * user may not use mc until dst ends. Otherwise we'd...
2433 mc->mc_txn = NULL; /* ...set this to dst */
2434 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2435 if ((mx = mc->mc_xcursor) != NULL) {
2436 *(MDB_xcursor *)(bk+1) = *mx;
2437 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2439 mc->mc_next = dst->mt_cursors[i];
2440 dst->mt_cursors[i] = mc;
2447 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2448 * @param[in] txn the transaction handle.
2449 * @param[in] merge true to keep changes to parent cursors, false to revert.
2450 * @return 0 on success, non-zero on failure.
2453 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2455 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2459 for (i = txn->mt_numdbs; --i >= 0; ) {
2460 for (mc = cursors[i]; mc; mc = next) {
2462 if ((bk = mc->mc_backup) != NULL) {
2464 /* Commit changes to parent txn */
2465 mc->mc_next = bk->mc_next;
2466 mc->mc_backup = bk->mc_backup;
2467 mc->mc_txn = bk->mc_txn;
2468 mc->mc_db = bk->mc_db;
2469 mc->mc_dbflag = bk->mc_dbflag;
2470 if ((mx = mc->mc_xcursor) != NULL)
2471 mx->mx_cursor.mc_txn = bk->mc_txn;
2473 /* Abort nested txn */
2475 if ((mx = mc->mc_xcursor) != NULL)
2476 *mx = *(MDB_xcursor *)(bk+1);
2480 /* Only malloced cursors are permanently tracked. */
2488 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2491 mdb_txn_reset0(MDB_txn *txn, const char *act);
2493 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2499 Pidset = F_SETLK, Pidcheck = F_GETLK
2503 /** Set or check a pid lock. Set returns 0 on success.
2504 * Check returns 0 if the process is certainly dead, nonzero if it may
2505 * be alive (the lock exists or an error happened so we do not know).
2507 * On Windows Pidset is a no-op, we merely check for the existence
2508 * of the process with the given pid. On POSIX we use a single byte
2509 * lock on the lockfile, set at an offset equal to the pid.
2512 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2514 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2517 if (op == Pidcheck) {
2518 h = OpenProcess(env->me_pidquery, FALSE, pid);
2519 /* No documented "no such process" code, but other program use this: */
2521 return ErrCode() != ERROR_INVALID_PARAMETER;
2522 /* A process exists until all handles to it close. Has it exited? */
2523 ret = WaitForSingleObject(h, 0) != 0;
2530 struct flock lock_info;
2531 memset(&lock_info, 0, sizeof(lock_info));
2532 lock_info.l_type = F_WRLCK;
2533 lock_info.l_whence = SEEK_SET;
2534 lock_info.l_start = pid;
2535 lock_info.l_len = 1;
2536 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2537 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2539 } else if ((rc = ErrCode()) == EINTR) {
2547 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2548 * @param[in] txn the transaction handle to initialize
2549 * @return 0 on success, non-zero on failure.
2552 mdb_txn_renew0(MDB_txn *txn)
2554 MDB_env *env = txn->mt_env;
2555 MDB_txninfo *ti = env->me_txns;
2557 unsigned int i, nr, flags = txn->mt_flags;
2559 int rc, new_notls = 0;
2561 if ((flags &= MDB_TXN_RDONLY) != 0) {
2563 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2564 txn->mt_txnid = meta->mm_txnid;
2565 txn->mt_u.reader = NULL;
2567 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2568 pthread_getspecific(env->me_txkey);
2570 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2571 return MDB_BAD_RSLOT;
2573 MDB_PID_T pid = env->me_pid;
2574 MDB_THR_T tid = pthread_self();
2575 mdb_mutexref_t rmutex = env->me_rmutex;
2577 if (!env->me_live_reader) {
2578 rc = mdb_reader_pid(env, Pidset, pid);
2581 env->me_live_reader = 1;
2584 if (LOCK_MUTEX(rc, env, rmutex))
2586 nr = ti->mti_numreaders;
2587 for (i=0; i<nr; i++)
2588 if (ti->mti_readers[i].mr_pid == 0)
2590 if (i == env->me_maxreaders) {
2591 UNLOCK_MUTEX(rmutex);
2592 return MDB_READERS_FULL;
2594 r = &ti->mti_readers[i];
2595 /* Claim the reader slot, carefully since other code
2596 * uses the reader table un-mutexed: First reset the
2597 * slot, next publish it in mti_numreaders. After
2598 * that, it is safe for mdb_env_close() to touch it.
2599 * When it will be closed, we can finally claim it.
2602 r->mr_txnid = (txnid_t)-1;
2605 ti->mti_numreaders = ++nr;
2606 env->me_close_readers = nr;
2608 UNLOCK_MUTEX(rmutex);
2610 new_notls = (env->me_flags & MDB_NOTLS);
2611 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2616 do /* LY: Retry on a race, ITS#7970. */
2617 r->mr_txnid = ti->mti_txnid;
2618 while(r->mr_txnid != ti->mti_txnid);
2619 txn->mt_txnid = r->mr_txnid;
2620 txn->mt_u.reader = r;
2621 meta = env->me_metas[txn->mt_txnid & 1];
2623 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2625 /* Not yet touching txn == env->me_txn0, it may be active */
2627 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2629 txn->mt_txnid = ti->mti_txnid;
2630 meta = env->me_metas[txn->mt_txnid & 1];
2632 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2633 txn->mt_txnid = meta->mm_txnid;
2637 if (txn->mt_txnid == mdb_debug_start)
2640 txn->mt_child = NULL;
2641 txn->mt_loose_pgs = NULL;
2642 txn->mt_loose_count = 0;
2643 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2644 txn->mt_u.dirty_list = env->me_dirty_list;
2645 txn->mt_u.dirty_list[0].mid = 0;
2646 txn->mt_free_pgs = env->me_free_pgs;
2647 txn->mt_free_pgs[0] = 0;
2648 txn->mt_spill_pgs = NULL;
2650 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2653 /* Copy the DB info and flags */
2654 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2656 /* Moved to here to avoid a data race in read TXNs */
2657 txn->mt_next_pgno = meta->mm_last_pg+1;
2659 txn->mt_flags = flags;
2662 txn->mt_numdbs = env->me_numdbs;
2663 for (i=2; i<txn->mt_numdbs; i++) {
2664 x = env->me_dbflags[i];
2665 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2666 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2668 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2669 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2671 if (env->me_maxpg < txn->mt_next_pgno) {
2672 mdb_txn_reset0(txn, "renew0-mapfail");
2674 txn->mt_u.reader->mr_pid = 0;
2675 txn->mt_u.reader = NULL;
2677 return MDB_MAP_RESIZED;
2684 mdb_txn_renew(MDB_txn *txn)
2688 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2691 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2692 DPUTS("environment had fatal error, must shutdown!");
2696 rc = mdb_txn_renew0(txn);
2697 if (rc == MDB_SUCCESS) {
2698 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2699 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2700 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2706 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2710 int rc, size, tsize;
2712 flags &= MDB_TXN_BEGIN_FLAGS;
2713 flags |= env->me_flags & MDB_WRITEMAP;
2715 if (env->me_flags & MDB_FATAL_ERROR) {
2716 DPUTS("environment had fatal error, must shutdown!");
2719 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2723 /* Nested transactions: Max 1 child, write txns only, no writemap */
2724 flags |= parent->mt_flags;
2725 if (parent->mt_child ||
2726 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2728 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2730 /* Child txns save MDB_pgstate and use own copy of cursors */
2731 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2732 size += tsize = sizeof(MDB_ntxn);
2733 } else if (flags & MDB_RDONLY) {
2734 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2735 size += tsize = sizeof(MDB_txn);
2737 /* Reuse preallocated write txn. However, do not touch it until
2738 * mdb_txn_renew0() succeeds, since it currently may be active.
2743 if ((txn = calloc(1, size)) == NULL) {
2744 DPRINTF(("calloc: %s", strerror(errno)));
2747 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2748 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2749 txn->mt_flags = flags;
2754 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2755 txn->mt_dbiseqs = parent->mt_dbiseqs;
2756 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2757 if (!txn->mt_u.dirty_list ||
2758 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2760 free(txn->mt_u.dirty_list);
2764 txn->mt_txnid = parent->mt_txnid;
2765 txn->mt_dirty_room = parent->mt_dirty_room;
2766 txn->mt_u.dirty_list[0].mid = 0;
2767 txn->mt_spill_pgs = NULL;
2768 txn->mt_next_pgno = parent->mt_next_pgno;
2769 parent->mt_child = txn;
2770 txn->mt_parent = parent;
2771 txn->mt_numdbs = parent->mt_numdbs;
2772 txn->mt_dbxs = parent->mt_dbxs;
2773 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2774 /* Copy parent's mt_dbflags, but clear DB_NEW */
2775 for (i=0; i<txn->mt_numdbs; i++)
2776 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2778 ntxn = (MDB_ntxn *)txn;
2779 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2780 if (env->me_pghead) {
2781 size = MDB_IDL_SIZEOF(env->me_pghead);
2782 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2784 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2789 rc = mdb_cursor_shadow(parent, txn);
2791 mdb_txn_reset0(txn, "beginchild-fail");
2792 } else { /* MDB_RDONLY */
2793 txn->mt_dbiseqs = env->me_dbiseqs;
2795 rc = mdb_txn_renew0(txn);
2798 if (txn != env->me_txn0)
2801 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2803 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2804 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2805 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2812 mdb_txn_env(MDB_txn *txn)
2814 if(!txn) return NULL;
2819 mdb_txn_id(MDB_txn *txn)
2822 return txn->mt_txnid;
2825 /** Export or close DBI handles opened in this txn. */
2827 mdb_dbis_update(MDB_txn *txn, int keep)
2830 MDB_dbi n = txn->mt_numdbs;
2831 MDB_env *env = txn->mt_env;
2832 unsigned char *tdbflags = txn->mt_dbflags;
2834 for (i = n; --i >= 2;) {
2835 if (tdbflags[i] & DB_NEW) {
2837 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2839 char *ptr = env->me_dbxs[i].md_name.mv_data;
2841 env->me_dbxs[i].md_name.mv_data = NULL;
2842 env->me_dbxs[i].md_name.mv_size = 0;
2843 env->me_dbflags[i] = 0;
2844 env->me_dbiseqs[i]++;
2850 if (keep && env->me_numdbs < n)
2854 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2855 * May be called twice for readonly txns: First reset it, then abort.
2856 * @param[in] txn the transaction handle to reset
2857 * @param[in] act why the transaction is being reset
2860 mdb_txn_reset0(MDB_txn *txn, const char *act)
2862 MDB_env *env = txn->mt_env;
2864 /* Close any DBI handles opened in this txn */
2865 mdb_dbis_update(txn, 0);
2867 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2868 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2869 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2871 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2872 if (txn->mt_u.reader) {
2873 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2874 if (!(env->me_flags & MDB_NOTLS))
2875 txn->mt_u.reader = NULL; /* txn does not own reader */
2877 txn->mt_numdbs = 0; /* close nothing if called again */
2878 txn->mt_dbxs = NULL; /* mark txn as reset */
2880 pgno_t *pghead = env->me_pghead;
2882 mdb_cursors_close(txn, 0);
2883 if (!(env->me_flags & MDB_WRITEMAP)) {
2884 mdb_dlist_free(txn);
2887 if (!txn->mt_parent) {
2888 mdb_midl_shrink(&txn->mt_free_pgs);
2889 env->me_free_pgs = txn->mt_free_pgs;
2891 env->me_pghead = NULL;
2895 /* The writer mutex was locked in mdb_txn_begin. */
2897 UNLOCK_MUTEX(env->me_wmutex);
2899 txn->mt_parent->mt_child = NULL;
2900 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2901 mdb_midl_free(txn->mt_free_pgs);
2902 mdb_midl_free(txn->mt_spill_pgs);
2903 free(txn->mt_u.dirty_list);
2906 mdb_midl_free(pghead);
2911 mdb_txn_reset(MDB_txn *txn)
2916 /* This call is only valid for read-only txns */
2917 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2920 mdb_txn_reset0(txn, "reset");
2924 mdb_txn_abort(MDB_txn *txn)
2930 mdb_txn_abort(txn->mt_child);
2932 mdb_txn_reset0(txn, "abort");
2933 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2934 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2935 txn->mt_u.reader->mr_pid = 0;
2937 if (txn != txn->mt_env->me_txn0)
2941 /** Save the freelist as of this transaction to the freeDB.
2942 * This changes the freelist. Keep trying until it stabilizes.
2945 mdb_freelist_save(MDB_txn *txn)
2947 /* env->me_pghead[] can grow and shrink during this call.
2948 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2949 * Page numbers cannot disappear from txn->mt_free_pgs[].
2952 MDB_env *env = txn->mt_env;
2953 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2954 txnid_t pglast = 0, head_id = 0;
2955 pgno_t freecnt = 0, *free_pgs, *mop;
2956 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2958 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2960 if (env->me_pghead) {
2961 /* Make sure first page of freeDB is touched and on freelist */
2962 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2963 if (rc && rc != MDB_NOTFOUND)
2967 if (!env->me_pghead && txn->mt_loose_pgs) {
2968 /* Put loose page numbers in mt_free_pgs, since
2969 * we may be unable to return them to me_pghead.
2971 MDB_page *mp = txn->mt_loose_pgs;
2972 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2974 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2975 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2976 txn->mt_loose_pgs = NULL;
2977 txn->mt_loose_count = 0;
2980 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2981 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2982 ? SSIZE_MAX : maxfree_1pg;
2985 /* Come back here after each Put() in case freelist changed */
2990 /* If using records from freeDB which we have not yet
2991 * deleted, delete them and any we reserved for me_pghead.
2993 while (pglast < env->me_pglast) {
2994 rc = mdb_cursor_first(&mc, &key, NULL);
2997 pglast = head_id = *(txnid_t *)key.mv_data;
2998 total_room = head_room = 0;
2999 mdb_tassert(txn, pglast <= env->me_pglast);
3000 rc = mdb_cursor_del(&mc, 0);
3005 /* Save the IDL of pages freed by this txn, to a single record */
3006 if (freecnt < txn->mt_free_pgs[0]) {
3008 /* Make sure last page of freeDB is touched and on freelist */
3009 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3010 if (rc && rc != MDB_NOTFOUND)
3013 free_pgs = txn->mt_free_pgs;
3014 /* Write to last page of freeDB */
3015 key.mv_size = sizeof(txn->mt_txnid);
3016 key.mv_data = &txn->mt_txnid;
3018 freecnt = free_pgs[0];
3019 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3020 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3023 /* Retry if mt_free_pgs[] grew during the Put() */
3024 free_pgs = txn->mt_free_pgs;
3025 } while (freecnt < free_pgs[0]);
3026 mdb_midl_sort(free_pgs);
3027 memcpy(data.mv_data, free_pgs, data.mv_size);
3030 unsigned int i = free_pgs[0];
3031 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3032 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3034 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3040 mop = env->me_pghead;
3041 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3043 /* Reserve records for me_pghead[]. Split it if multi-page,
3044 * to avoid searching freeDB for a page range. Use keys in
3045 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3047 if (total_room >= mop_len) {
3048 if (total_room == mop_len || --more < 0)
3050 } else if (head_room >= maxfree_1pg && head_id > 1) {
3051 /* Keep current record (overflow page), add a new one */
3055 /* (Re)write {key = head_id, IDL length = head_room} */
3056 total_room -= head_room;
3057 head_room = mop_len - total_room;
3058 if (head_room > maxfree_1pg && head_id > 1) {
3059 /* Overflow multi-page for part of me_pghead */
3060 head_room /= head_id; /* amortize page sizes */
3061 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3062 } else if (head_room < 0) {
3063 /* Rare case, not bothering to delete this record */
3066 key.mv_size = sizeof(head_id);
3067 key.mv_data = &head_id;
3068 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3069 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3072 /* IDL is initially empty, zero out at least the length */
3073 pgs = (pgno_t *)data.mv_data;
3074 j = head_room > clean_limit ? head_room : 0;
3078 total_room += head_room;
3081 /* Return loose page numbers to me_pghead, though usually none are
3082 * left at this point. The pages themselves remain in dirty_list.
3084 if (txn->mt_loose_pgs) {
3085 MDB_page *mp = txn->mt_loose_pgs;
3086 unsigned count = txn->mt_loose_count;
3088 /* Room for loose pages + temp IDL with same */
3089 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3091 mop = env->me_pghead;
3092 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3093 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3094 loose[ ++count ] = mp->mp_pgno;
3096 mdb_midl_sort(loose);
3097 mdb_midl_xmerge(mop, loose);
3098 txn->mt_loose_pgs = NULL;
3099 txn->mt_loose_count = 0;
3103 /* Fill in the reserved me_pghead records */
3109 rc = mdb_cursor_first(&mc, &key, &data);
3110 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3111 txnid_t id = *(txnid_t *)key.mv_data;
3112 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3115 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3117 if (len > mop_len) {
3119 data.mv_size = (len + 1) * sizeof(MDB_ID);
3121 data.mv_data = mop -= len;
3124 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3126 if (rc || !(mop_len -= len))
3133 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3134 * @param[in] txn the transaction that's being committed
3135 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3136 * @return 0 on success, non-zero on failure.
3139 mdb_page_flush(MDB_txn *txn, int keep)
3141 MDB_env *env = txn->mt_env;
3142 MDB_ID2L dl = txn->mt_u.dirty_list;
3143 unsigned psize = env->me_psize, j;
3144 int i, pagecount = dl[0].mid, rc;
3145 size_t size = 0, pos = 0;
3147 MDB_page *dp = NULL;
3151 struct iovec iov[MDB_COMMIT_PAGES];
3152 ssize_t wpos = 0, wsize = 0, wres;
3153 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3159 if (env->me_flags & MDB_WRITEMAP) {
3160 /* Clear dirty flags */
3161 while (++i <= pagecount) {
3163 /* Don't flush this page yet */
3164 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3165 dp->mp_flags &= ~P_KEEP;
3169 dp->mp_flags &= ~P_DIRTY;
3174 /* Write the pages */
3176 if (++i <= pagecount) {
3178 /* Don't flush this page yet */
3179 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3180 dp->mp_flags &= ~P_KEEP;
3185 /* clear dirty flag */
3186 dp->mp_flags &= ~P_DIRTY;
3189 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3194 /* Windows actually supports scatter/gather I/O, but only on
3195 * unbuffered file handles. Since we're relying on the OS page
3196 * cache for all our data, that's self-defeating. So we just
3197 * write pages one at a time. We use the ov structure to set
3198 * the write offset, to at least save the overhead of a Seek
3201 DPRINTF(("committing page %"Z"u", pgno));
3202 memset(&ov, 0, sizeof(ov));
3203 ov.Offset = pos & 0xffffffff;
3204 ov.OffsetHigh = pos >> 16 >> 16;
3205 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3207 DPRINTF(("WriteFile: %d", rc));
3211 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3212 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3215 /* Write previous page(s) */
3216 #ifdef MDB_USE_PWRITEV
3217 wres = pwritev(env->me_fd, iov, n, wpos);
3220 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3223 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3227 DPRINTF(("lseek: %s", strerror(rc)));
3230 wres = writev(env->me_fd, iov, n);
3233 if (wres != wsize) {
3238 DPRINTF(("Write error: %s", strerror(rc)));
3240 rc = EIO; /* TODO: Use which error code? */
3241 DPUTS("short write, filesystem full?");
3252 DPRINTF(("committing page %"Z"u", pgno));
3253 next_pos = pos + size;
3254 iov[n].iov_len = size;
3255 iov[n].iov_base = (char *)dp;
3261 /* MIPS has cache coherency issues, this is a no-op everywhere else
3262 * Note: for any size >= on-chip cache size, entire on-chip cache is
3265 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3267 for (i = keep; ++i <= pagecount; ) {
3269 /* This is a page we skipped above */
3272 dl[j].mid = dp->mp_pgno;
3275 mdb_dpage_free(env, dp);
3280 txn->mt_dirty_room += i - j;
3286 mdb_txn_commit(MDB_txn *txn)
3295 if (txn->mt_child) {
3296 rc = mdb_txn_commit(txn->mt_child);
3303 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3304 mdb_dbis_update(txn, 1);
3305 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3310 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3311 DPUTS("error flag is set, can't commit");
3313 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3318 if (txn->mt_parent) {
3319 MDB_txn *parent = txn->mt_parent;
3323 unsigned x, y, len, ps_len;
3325 /* Append our free list to parent's */
3326 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3329 mdb_midl_free(txn->mt_free_pgs);
3330 /* Failures after this must either undo the changes
3331 * to the parent or set MDB_TXN_ERROR in the parent.
3334 parent->mt_next_pgno = txn->mt_next_pgno;
3335 parent->mt_flags = txn->mt_flags;
3337 /* Merge our cursors into parent's and close them */
3338 mdb_cursors_close(txn, 1);
3340 /* Update parent's DB table. */
3341 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3342 parent->mt_numdbs = txn->mt_numdbs;
3343 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3344 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3345 for (i=2; i<txn->mt_numdbs; i++) {
3346 /* preserve parent's DB_NEW status */
3347 x = parent->mt_dbflags[i] & DB_NEW;
3348 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3351 dst = parent->mt_u.dirty_list;
3352 src = txn->mt_u.dirty_list;
3353 /* Remove anything in our dirty list from parent's spill list */
3354 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3356 pspill[0] = (pgno_t)-1;
3357 /* Mark our dirty pages as deleted in parent spill list */
3358 for (i=0, len=src[0].mid; ++i <= len; ) {
3359 MDB_ID pn = src[i].mid << 1;
3360 while (pn > pspill[x])
3362 if (pn == pspill[x]) {
3367 /* Squash deleted pagenums if we deleted any */
3368 for (x=y; ++x <= ps_len; )
3369 if (!(pspill[x] & 1))
3370 pspill[++y] = pspill[x];
3374 /* Find len = length of merging our dirty list with parent's */
3376 dst[0].mid = 0; /* simplify loops */
3377 if (parent->mt_parent) {
3378 len = x + src[0].mid;
3379 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3380 for (i = x; y && i; y--) {
3381 pgno_t yp = src[y].mid;
3382 while (yp < dst[i].mid)
3384 if (yp == dst[i].mid) {
3389 } else { /* Simplify the above for single-ancestor case */
3390 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3392 /* Merge our dirty list with parent's */
3394 for (i = len; y; dst[i--] = src[y--]) {
3395 pgno_t yp = src[y].mid;
3396 while (yp < dst[x].mid)
3397 dst[i--] = dst[x--];
3398 if (yp == dst[x].mid)
3399 free(dst[x--].mptr);
3401 mdb_tassert(txn, i == x);
3403 free(txn->mt_u.dirty_list);
3404 parent->mt_dirty_room = txn->mt_dirty_room;
3405 if (txn->mt_spill_pgs) {
3406 if (parent->mt_spill_pgs) {
3407 /* TODO: Prevent failure here, so parent does not fail */
3408 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3410 parent->mt_flags |= MDB_TXN_ERROR;
3411 mdb_midl_free(txn->mt_spill_pgs);
3412 mdb_midl_sort(parent->mt_spill_pgs);
3414 parent->mt_spill_pgs = txn->mt_spill_pgs;
3418 /* Append our loose page list to parent's */
3419 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3421 *lp = txn->mt_loose_pgs;
3422 parent->mt_loose_count += txn->mt_loose_count;
3424 parent->mt_child = NULL;
3425 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3430 if (txn != env->me_txn) {
3431 DPUTS("attempt to commit unknown transaction");
3436 mdb_cursors_close(txn, 0);
3438 if (!txn->mt_u.dirty_list[0].mid &&
3439 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3442 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3443 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3445 /* Update DB root pointers */
3446 if (txn->mt_numdbs > 2) {
3450 data.mv_size = sizeof(MDB_db);
3452 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3453 for (i = 2; i < txn->mt_numdbs; i++) {
3454 if (txn->mt_dbflags[i] & DB_DIRTY) {
3455 if (TXN_DBI_CHANGED(txn, i)) {
3459 data.mv_data = &txn->mt_dbs[i];
3460 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3468 rc = mdb_freelist_save(txn);
3472 mdb_midl_free(env->me_pghead);
3473 env->me_pghead = NULL;
3474 mdb_midl_shrink(&txn->mt_free_pgs);
3475 env->me_free_pgs = txn->mt_free_pgs;
3481 if ((rc = mdb_page_flush(txn, 0)) ||
3482 (rc = mdb_env_sync(env, 0)) ||
3483 (rc = mdb_env_write_meta(txn)))
3486 /* Free P_LOOSE pages left behind in dirty_list */
3487 if (!(env->me_flags & MDB_WRITEMAP))
3488 mdb_dlist_free(txn);
3493 mdb_dbis_update(txn, 1);
3496 UNLOCK_MUTEX(env->me_wmutex);
3497 if (txn != env->me_txn0)
3507 /** Read the environment parameters of a DB environment before
3508 * mapping it into memory.
3509 * @param[in] env the environment handle
3510 * @param[out] meta address of where to store the meta information
3511 * @return 0 on success, non-zero on failure.
3514 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3520 enum { Size = sizeof(pbuf) };
3522 /* We don't know the page size yet, so use a minimum value.
3523 * Read both meta pages so we can use the latest one.
3526 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3530 memset(&ov, 0, sizeof(ov));
3532 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3533 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3536 rc = pread(env->me_fd, &pbuf, Size, off);
3539 if (rc == 0 && off == 0)
3541 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3542 DPRINTF(("read: %s", mdb_strerror(rc)));
3546 p = (MDB_page *)&pbuf;
3548 if (!F_ISSET(p->mp_flags, P_META)) {
3549 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3554 if (m->mm_magic != MDB_MAGIC) {
3555 DPUTS("meta has invalid magic");
3559 if (m->mm_version != MDB_DATA_VERSION) {
3560 DPRINTF(("database is version %u, expected version %u",
3561 m->mm_version, MDB_DATA_VERSION));
3562 return MDB_VERSION_MISMATCH;
3565 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3571 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3573 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3575 meta->mm_magic = MDB_MAGIC;
3576 meta->mm_version = MDB_DATA_VERSION;
3577 meta->mm_mapsize = env->me_mapsize;
3578 meta->mm_psize = env->me_psize;
3579 meta->mm_last_pg = 1;
3580 meta->mm_flags = env->me_flags & 0xffff;
3581 meta->mm_flags |= MDB_INTEGERKEY;
3582 meta->mm_dbs[0].md_root = P_INVALID;
3583 meta->mm_dbs[1].md_root = P_INVALID;
3586 /** Write the environment parameters of a freshly created DB environment.
3587 * @param[in] env the environment handle
3588 * @param[in] meta the #MDB_meta to write
3589 * @return 0 on success, non-zero on failure.
3592 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3600 memset(&ov, 0, sizeof(ov));
3601 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3603 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3606 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3607 len = pwrite(fd, ptr, size, pos); \
3608 if (len == -1 && ErrCode() == EINTR) continue; \
3609 rc = (len >= 0); break; } while(1)
3612 DPUTS("writing new meta page");
3614 psize = env->me_psize;
3616 p = calloc(2, psize);
3621 p->mp_flags = P_META;
3622 *(MDB_meta *)METADATA(p) = *meta;
3624 q = (MDB_page *)((char *)p + psize);
3626 q->mp_flags = P_META;
3627 *(MDB_meta *)METADATA(q) = *meta;
3629 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3632 else if ((unsigned) len == psize * 2)
3640 /** Update the environment info to commit a transaction.
3641 * @param[in] txn the transaction that's being committed
3642 * @return 0 on success, non-zero on failure.
3645 mdb_env_write_meta(MDB_txn *txn)
3648 MDB_meta meta, metab, *mp;
3652 int rc, len, toggle;
3661 toggle = txn->mt_txnid & 1;
3662 DPRINTF(("writing meta page %d for root page %"Z"u",
3663 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3666 flags = env->me_flags;
3667 mp = env->me_metas[toggle];
3668 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3669 /* Persist any increases of mapsize config */
3670 if (mapsize < env->me_mapsize)
3671 mapsize = env->me_mapsize;
3673 if (flags & MDB_WRITEMAP) {
3674 mp->mm_mapsize = mapsize;
3675 mp->mm_dbs[0] = txn->mt_dbs[0];
3676 mp->mm_dbs[1] = txn->mt_dbs[1];
3677 mp->mm_last_pg = txn->mt_next_pgno - 1;
3678 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3679 !(defined(__i386__) || defined(__x86_64__))
3680 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3681 __sync_synchronize();
3683 mp->mm_txnid = txn->mt_txnid;
3684 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3685 unsigned meta_size = env->me_psize;
3686 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3689 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3690 if (meta_size < env->me_os_psize)
3691 meta_size += meta_size;
3696 if (MDB_MSYNC(ptr, meta_size, rc)) {
3703 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3704 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3706 meta.mm_mapsize = mapsize;
3707 meta.mm_dbs[0] = txn->mt_dbs[0];
3708 meta.mm_dbs[1] = txn->mt_dbs[1];
3709 meta.mm_last_pg = txn->mt_next_pgno - 1;
3710 meta.mm_txnid = txn->mt_txnid;
3712 off = offsetof(MDB_meta, mm_mapsize);
3713 ptr = (char *)&meta + off;
3714 len = sizeof(MDB_meta) - off;
3716 off += env->me_psize;
3719 /* Write to the SYNC fd */
3720 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3723 memset(&ov, 0, sizeof(ov));
3725 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3730 rc = pwrite(mfd, ptr, len, off);
3733 rc = rc < 0 ? ErrCode() : EIO;
3738 DPUTS("write failed, disk error?");
3739 /* On a failure, the pagecache still contains the new data.
3740 * Write some old data back, to prevent it from being used.
3741 * Use the non-SYNC fd; we know it will fail anyway.
3743 meta.mm_last_pg = metab.mm_last_pg;
3744 meta.mm_txnid = metab.mm_txnid;
3746 memset(&ov, 0, sizeof(ov));
3748 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3750 r2 = pwrite(env->me_fd, ptr, len, off);
3751 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3754 env->me_flags |= MDB_FATAL_ERROR;
3757 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3758 CACHEFLUSH(env->me_map + off, len, DCACHE);
3760 /* Memory ordering issues are irrelevant; since the entire writer
3761 * is wrapped by wmutex, all of these changes will become visible
3762 * after the wmutex is unlocked. Since the DB is multi-version,
3763 * readers will get consistent data regardless of how fresh or
3764 * how stale their view of these values is.
3767 env->me_txns->mti_txnid = txn->mt_txnid;
3772 /** Check both meta pages to see which one is newer.
3773 * @param[in] env the environment handle
3774 * @return meta toggle (0 or 1).
3777 mdb_env_pick_meta(const MDB_env *env)
3779 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3783 mdb_env_create(MDB_env **env)
3787 e = calloc(1, sizeof(MDB_env));
3791 e->me_maxreaders = DEFAULT_READERS;
3792 e->me_maxdbs = e->me_numdbs = 2;
3793 e->me_fd = INVALID_HANDLE_VALUE;
3794 e->me_lfd = INVALID_HANDLE_VALUE;
3795 e->me_mfd = INVALID_HANDLE_VALUE;
3796 #ifdef MDB_USE_POSIX_SEM
3797 e->me_rmutex = SEM_FAILED;
3798 e->me_wmutex = SEM_FAILED;
3800 e->me_pid = getpid();
3801 GET_PAGESIZE(e->me_os_psize);
3802 VGMEMP_CREATE(e,0,0);
3808 mdb_env_map(MDB_env *env, void *addr)
3811 unsigned int flags = env->me_flags;
3815 LONG sizelo, sizehi;
3818 if (flags & MDB_RDONLY) {
3819 /* Don't set explicit map size, use whatever exists */
3824 msize = env->me_mapsize;
3825 sizelo = msize & 0xffffffff;
3826 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3828 /* Windows won't create mappings for zero length files.
3829 * and won't map more than the file size.
3830 * Just set the maxsize right now.
3832 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3833 || !SetEndOfFile(env->me_fd)
3834 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3838 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3839 PAGE_READWRITE : PAGE_READONLY,
3840 sizehi, sizelo, NULL);
3843 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3844 FILE_MAP_WRITE : FILE_MAP_READ,
3846 rc = env->me_map ? 0 : ErrCode();
3851 int prot = PROT_READ;
3852 if (flags & MDB_WRITEMAP) {
3854 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3857 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3859 if (env->me_map == MAP_FAILED) {
3864 if (flags & MDB_NORDAHEAD) {
3865 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3867 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3869 #ifdef POSIX_MADV_RANDOM
3870 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3871 #endif /* POSIX_MADV_RANDOM */
3872 #endif /* MADV_RANDOM */
3876 /* Can happen because the address argument to mmap() is just a
3877 * hint. mmap() can pick another, e.g. if the range is in use.
3878 * The MAP_FIXED flag would prevent that, but then mmap could
3879 * instead unmap existing pages to make room for the new map.
3881 if (addr && env->me_map != addr)
3882 return EBUSY; /* TODO: Make a new MDB_* error code? */
3884 p = (MDB_page *)env->me_map;
3885 env->me_metas[0] = METADATA(p);
3886 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3892 mdb_env_set_mapsize(MDB_env *env, size_t size)
3894 /* If env is already open, caller is responsible for making
3895 * sure there are no active txns.
3903 meta = env->me_metas[mdb_env_pick_meta(env)];
3905 size = meta->mm_mapsize;
3907 /* Silently round up to minimum if the size is too small */
3908 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3912 munmap(env->me_map, env->me_mapsize);
3913 env->me_mapsize = size;
3914 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3915 rc = mdb_env_map(env, old);
3919 env->me_mapsize = size;
3921 env->me_maxpg = env->me_mapsize / env->me_psize;
3926 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3930 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3935 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3937 if (env->me_map || readers < 1)
3939 env->me_maxreaders = readers;
3944 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3946 if (!env || !readers)
3948 *readers = env->me_maxreaders;
3953 mdb_fsize(HANDLE fd, size_t *size)
3956 LARGE_INTEGER fsize;
3958 if (!GetFileSizeEx(fd, &fsize))
3961 *size = fsize.QuadPart;
3973 #ifdef BROKEN_FDATASYNC
3974 #include <sys/utsname.h>
3975 #include <sys/vfs.h>
3978 /** Further setup required for opening an LMDB environment
3981 mdb_env_open2(MDB_env *env)
3983 unsigned int flags = env->me_flags;
3984 int i, newenv = 0, rc;
3988 /* See if we should use QueryLimited */
3990 if ((rc & 0xff) > 5)
3991 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3993 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3996 #ifdef BROKEN_FDATASYNC
3997 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3998 * https://lkml.org/lkml/2012/9/3/83
3999 * Kernels after 3.6-rc6 are known good.
4000 * https://lkml.org/lkml/2012/9/10/556
4001 * See if the DB is on ext3/ext4, then check for new enough kernel
4002 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4007 fstatfs(env->me_fd, &st);
4008 while (st.f_type == 0xEF53) {
4012 if (uts.release[0] < '3') {
4013 if (!strncmp(uts.release, "2.6.32.", 7)) {
4014 i = atoi(uts.release+7);
4016 break; /* 2.6.32.60 and newer is OK */
4017 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4018 i = atoi(uts.release+7);
4020 break; /* 2.6.34.15 and newer is OK */
4022 } else if (uts.release[0] == '3') {
4023 i = atoi(uts.release+2);
4025 break; /* 3.6 and newer is OK */
4027 i = atoi(uts.release+4);
4029 break; /* 3.5.4 and newer is OK */
4030 } else if (i == 2) {
4031 i = atoi(uts.release+4);
4033 break; /* 3.2.30 and newer is OK */
4035 } else { /* 4.x and newer is OK */
4038 env->me_flags |= MDB_FSYNCONLY;
4044 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4047 DPUTS("new mdbenv");
4049 env->me_psize = env->me_os_psize;
4050 if (env->me_psize > MAX_PAGESIZE)
4051 env->me_psize = MAX_PAGESIZE;
4052 memset(&meta, 0, sizeof(meta));
4053 mdb_env_init_meta0(env, &meta);
4054 meta.mm_mapsize = DEFAULT_MAPSIZE;
4056 env->me_psize = meta.mm_psize;
4059 /* Was a mapsize configured? */
4060 if (!env->me_mapsize) {
4061 env->me_mapsize = meta.mm_mapsize;
4064 /* Make sure mapsize >= committed data size. Even when using
4065 * mm_mapsize, which could be broken in old files (ITS#7789).
4067 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4068 if (env->me_mapsize < minsize)
4069 env->me_mapsize = minsize;
4071 meta.mm_mapsize = env->me_mapsize;
4073 if (newenv && !(flags & MDB_FIXEDMAP)) {
4074 /* mdb_env_map() may grow the datafile. Write the metapages
4075 * first, so the file will be valid if initialization fails.
4076 * Except with FIXEDMAP, since we do not yet know mm_address.
4077 * We could fill in mm_address later, but then a different
4078 * program might end up doing that - one with a memory layout
4079 * and map address which does not suit the main program.
4081 rc = mdb_env_init_meta(env, &meta);
4087 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4092 if (flags & MDB_FIXEDMAP)
4093 meta.mm_address = env->me_map;
4094 i = mdb_env_init_meta(env, &meta);
4095 if (i != MDB_SUCCESS) {
4100 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4101 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4103 #if !(MDB_MAXKEYSIZE)
4104 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4106 env->me_maxpg = env->me_mapsize / env->me_psize;
4110 int toggle = mdb_env_pick_meta(env);
4111 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4113 DPRINTF(("opened database version %u, pagesize %u",
4114 env->me_metas[0]->mm_version, env->me_psize));
4115 DPRINTF(("using meta page %d", toggle));
4116 DPRINTF(("depth: %u", db->md_depth));
4117 DPRINTF(("entries: %"Z"u", db->md_entries));
4118 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4119 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4120 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4121 DPRINTF(("root: %"Z"u", db->md_root));
4129 /** Release a reader thread's slot in the reader lock table.
4130 * This function is called automatically when a thread exits.
4131 * @param[in] ptr This points to the slot in the reader lock table.
4134 mdb_env_reader_dest(void *ptr)
4136 MDB_reader *reader = ptr;
4142 /** Junk for arranging thread-specific callbacks on Windows. This is
4143 * necessarily platform and compiler-specific. Windows supports up
4144 * to 1088 keys. Let's assume nobody opens more than 64 environments
4145 * in a single process, for now. They can override this if needed.
4147 #ifndef MAX_TLS_KEYS
4148 #define MAX_TLS_KEYS 64
4150 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4151 static int mdb_tls_nkeys;
4153 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4157 case DLL_PROCESS_ATTACH: break;
4158 case DLL_THREAD_ATTACH: break;
4159 case DLL_THREAD_DETACH:
4160 for (i=0; i<mdb_tls_nkeys; i++) {
4161 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4163 mdb_env_reader_dest(r);
4167 case DLL_PROCESS_DETACH: break;
4172 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4174 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4178 /* Force some symbol references.
4179 * _tls_used forces the linker to create the TLS directory if not already done
4180 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4182 #pragma comment(linker, "/INCLUDE:_tls_used")
4183 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4184 #pragma const_seg(".CRT$XLB")
4185 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4186 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4189 #pragma comment(linker, "/INCLUDE:__tls_used")
4190 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4191 #pragma data_seg(".CRT$XLB")
4192 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4194 #endif /* WIN 32/64 */
4195 #endif /* !__GNUC__ */
4198 /** Downgrade the exclusive lock on the region back to shared */
4200 mdb_env_share_locks(MDB_env *env, int *excl)
4202 int rc = 0, toggle = mdb_env_pick_meta(env);
4204 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4209 /* First acquire a shared lock. The Unlock will
4210 * then release the existing exclusive lock.
4212 memset(&ov, 0, sizeof(ov));
4213 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4216 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4222 struct flock lock_info;
4223 /* The shared lock replaces the existing lock */
4224 memset((void *)&lock_info, 0, sizeof(lock_info));
4225 lock_info.l_type = F_RDLCK;
4226 lock_info.l_whence = SEEK_SET;
4227 lock_info.l_start = 0;
4228 lock_info.l_len = 1;
4229 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4230 (rc = ErrCode()) == EINTR) ;
4231 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4238 /** Try to get exclusive lock, otherwise shared.
4239 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4242 mdb_env_excl_lock(MDB_env *env, int *excl)
4246 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4250 memset(&ov, 0, sizeof(ov));
4251 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4258 struct flock lock_info;
4259 memset((void *)&lock_info, 0, sizeof(lock_info));
4260 lock_info.l_type = F_WRLCK;
4261 lock_info.l_whence = SEEK_SET;
4262 lock_info.l_start = 0;
4263 lock_info.l_len = 1;
4264 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4265 (rc = ErrCode()) == EINTR) ;
4269 # ifndef MDB_USE_POSIX_MUTEX
4270 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4273 lock_info.l_type = F_RDLCK;
4274 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4275 (rc = ErrCode()) == EINTR) ;
4285 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4287 * @(#) $Revision: 5.1 $
4288 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4289 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4291 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4295 * Please do not copyright this code. This code is in the public domain.
4297 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4298 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4299 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4300 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4301 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4302 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4303 * PERFORMANCE OF THIS SOFTWARE.
4306 * chongo <Landon Curt Noll> /\oo/\
4307 * http://www.isthe.com/chongo/
4309 * Share and Enjoy! :-)
4312 typedef unsigned long long mdb_hash_t;
4313 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4315 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4316 * @param[in] val value to hash
4317 * @param[in] hval initial value for hash
4318 * @return 64 bit hash
4320 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4321 * hval arg on the first call.
4324 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4326 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4327 unsigned char *end = s + val->mv_size;
4329 * FNV-1a hash each octet of the string
4332 /* xor the bottom with the current octet */
4333 hval ^= (mdb_hash_t)*s++;
4335 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4336 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4337 (hval << 7) + (hval << 8) + (hval << 40);
4339 /* return our new hash value */
4343 /** Hash the string and output the encoded hash.
4344 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4345 * very short name limits. We don't care about the encoding being reversible,
4346 * we just want to preserve as many bits of the input as possible in a
4347 * small printable string.
4348 * @param[in] str string to hash
4349 * @param[out] encbuf an array of 11 chars to hold the hash
4351 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4354 mdb_pack85(unsigned long l, char *out)
4358 for (i=0; i<5; i++) {
4359 *out++ = mdb_a85[l % 85];
4365 mdb_hash_enc(MDB_val *val, char *encbuf)
4367 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4369 mdb_pack85(h, encbuf);
4370 mdb_pack85(h>>32, encbuf+5);
4375 /** Open and/or initialize the lock region for the environment.
4376 * @param[in] env The LMDB environment.
4377 * @param[in] lpath The pathname of the file used for the lock region.
4378 * @param[in] mode The Unix permissions for the file, if we create it.
4379 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4380 * @return 0 on success, non-zero on failure.
4383 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4386 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4388 # define MDB_ERRCODE_ROFS EROFS
4389 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4390 # define MDB_CLOEXEC O_CLOEXEC
4393 # define MDB_CLOEXEC 0
4400 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4401 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4402 FILE_ATTRIBUTE_NORMAL, NULL);
4404 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4406 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4408 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4413 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4414 /* Lose record locks when exec*() */
4415 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4416 fcntl(env->me_lfd, F_SETFD, fdflags);
4419 if (!(env->me_flags & MDB_NOTLS)) {
4420 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4423 env->me_flags |= MDB_ENV_TXKEY;
4425 /* Windows TLS callbacks need help finding their TLS info. */
4426 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4430 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4434 /* Try to get exclusive lock. If we succeed, then
4435 * nobody is using the lock region and we should initialize it.
4437 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4440 size = GetFileSize(env->me_lfd, NULL);
4442 size = lseek(env->me_lfd, 0, SEEK_END);
4443 if (size == -1) goto fail_errno;
4445 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4446 if (size < rsize && *excl > 0) {
4448 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4449 || !SetEndOfFile(env->me_lfd))
4452 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4456 size = rsize - sizeof(MDB_txninfo);
4457 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4462 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4464 if (!mh) goto fail_errno;
4465 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4467 if (!env->me_txns) goto fail_errno;
4469 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4471 if (m == MAP_FAILED) goto fail_errno;
4477 BY_HANDLE_FILE_INFORMATION stbuf;
4486 if (!mdb_sec_inited) {
4487 InitializeSecurityDescriptor(&mdb_null_sd,
4488 SECURITY_DESCRIPTOR_REVISION);
4489 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4490 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4491 mdb_all_sa.bInheritHandle = FALSE;
4492 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4495 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4496 idbuf.volume = stbuf.dwVolumeSerialNumber;
4497 idbuf.nhigh = stbuf.nFileIndexHigh;
4498 idbuf.nlow = stbuf.nFileIndexLow;
4499 val.mv_data = &idbuf;
4500 val.mv_size = sizeof(idbuf);
4501 mdb_hash_enc(&val, encbuf);
4502 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4503 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4504 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4505 if (!env->me_rmutex) goto fail_errno;
4506 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4507 if (!env->me_wmutex) goto fail_errno;
4508 #elif defined(MDB_USE_POSIX_SEM)
4517 #if defined(__NetBSD__)
4518 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4520 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4521 idbuf.dev = stbuf.st_dev;
4522 idbuf.ino = stbuf.st_ino;
4523 val.mv_data = &idbuf;
4524 val.mv_size = sizeof(idbuf);
4525 mdb_hash_enc(&val, encbuf);
4526 #ifdef MDB_SHORT_SEMNAMES
4527 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4529 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4530 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4531 /* Clean up after a previous run, if needed: Try to
4532 * remove both semaphores before doing anything else.
4534 sem_unlink(env->me_txns->mti_rmname);
4535 sem_unlink(env->me_txns->mti_wmname);
4536 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4537 O_CREAT|O_EXCL, mode, 1);
4538 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4539 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4540 O_CREAT|O_EXCL, mode, 1);
4541 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4542 #else /* MDB_USE_POSIX_MUTEX: */
4543 pthread_mutexattr_t mattr;
4545 if ((rc = pthread_mutexattr_init(&mattr))
4546 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4547 #ifdef MDB_ROBUST_SUPPORTED
4548 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4550 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4551 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4553 pthread_mutexattr_destroy(&mattr);
4554 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4556 env->me_txns->mti_magic = MDB_MAGIC;
4557 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4558 env->me_txns->mti_txnid = 0;
4559 env->me_txns->mti_numreaders = 0;
4562 if (env->me_txns->mti_magic != MDB_MAGIC) {
4563 DPUTS("lock region has invalid magic");
4567 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4568 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4569 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4570 rc = MDB_VERSION_MISMATCH;
4574 if (rc && rc != EACCES && rc != EAGAIN) {
4578 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4579 if (!env->me_rmutex) goto fail_errno;
4580 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4581 if (!env->me_wmutex) goto fail_errno;
4582 #elif defined(MDB_USE_POSIX_SEM)
4583 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4584 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4585 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4586 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4597 /** The name of the lock file in the DB environment */
4598 #define LOCKNAME "/lock.mdb"
4599 /** The name of the data file in the DB environment */
4600 #define DATANAME "/data.mdb"
4601 /** The suffix of the lock file when no subdir is used */
4602 #define LOCKSUFF "-lock"
4603 /** Only a subset of the @ref mdb_env flags can be changed
4604 * at runtime. Changing other flags requires closing the
4605 * environment and re-opening it with the new flags.
4607 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4608 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4609 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4611 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4612 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4616 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4618 int oflags, rc, len, excl = -1;
4619 char *lpath, *dpath;
4621 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4625 if (flags & MDB_NOSUBDIR) {
4626 rc = len + sizeof(LOCKSUFF) + len + 1;
4628 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4633 if (flags & MDB_NOSUBDIR) {
4634 dpath = lpath + len + sizeof(LOCKSUFF);
4635 sprintf(lpath, "%s" LOCKSUFF, path);
4636 strcpy(dpath, path);
4638 dpath = lpath + len + sizeof(LOCKNAME);
4639 sprintf(lpath, "%s" LOCKNAME, path);
4640 sprintf(dpath, "%s" DATANAME, path);
4644 flags |= env->me_flags;
4645 if (flags & MDB_RDONLY) {
4646 /* silently ignore WRITEMAP when we're only getting read access */
4647 flags &= ~MDB_WRITEMAP;
4649 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4650 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4653 env->me_flags = flags |= MDB_ENV_ACTIVE;
4657 env->me_path = strdup(path);
4658 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4659 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4660 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4661 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4665 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4667 /* For RDONLY, get lockfile after we know datafile exists */
4668 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4669 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4675 if (F_ISSET(flags, MDB_RDONLY)) {
4676 oflags = GENERIC_READ;
4677 len = OPEN_EXISTING;
4679 oflags = GENERIC_READ|GENERIC_WRITE;
4682 mode = FILE_ATTRIBUTE_NORMAL;
4683 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4684 NULL, len, mode, NULL);
4686 if (F_ISSET(flags, MDB_RDONLY))
4689 oflags = O_RDWR | O_CREAT;
4691 env->me_fd = open(dpath, oflags, mode);
4693 if (env->me_fd == INVALID_HANDLE_VALUE) {
4698 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4699 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4704 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4705 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4706 env->me_mfd = env->me_fd;
4708 /* Synchronous fd for meta writes. Needed even with
4709 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4712 len = OPEN_EXISTING;
4713 env->me_mfd = CreateFile(dpath, oflags,
4714 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4715 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4718 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4720 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4725 DPRINTF(("opened dbenv %p", (void *) env));
4727 rc = mdb_env_share_locks(env, &excl);
4731 if (!(flags & MDB_RDONLY)) {
4733 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4734 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4735 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4736 (txn = calloc(1, size)))
4738 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4739 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4740 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4741 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4743 txn->mt_dbxs = env->me_dbxs;
4753 mdb_env_close0(env, excl);
4759 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4761 mdb_env_close0(MDB_env *env, int excl)
4765 if (!(env->me_flags & MDB_ENV_ACTIVE))
4768 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4770 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4771 free(env->me_dbxs[i].md_name.mv_data);
4776 free(env->me_dbiseqs);
4777 free(env->me_dbflags);
4779 free(env->me_dirty_list);
4781 mdb_midl_free(env->me_free_pgs);
4783 if (env->me_flags & MDB_ENV_TXKEY) {
4784 pthread_key_delete(env->me_txkey);
4786 /* Delete our key from the global list */
4787 for (i=0; i<mdb_tls_nkeys; i++)
4788 if (mdb_tls_keys[i] == env->me_txkey) {
4789 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4797 munmap(env->me_map, env->me_mapsize);
4799 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4800 (void) close(env->me_mfd);
4801 if (env->me_fd != INVALID_HANDLE_VALUE)
4802 (void) close(env->me_fd);
4804 MDB_PID_T pid = env->me_pid;
4805 /* Clearing readers is done in this function because
4806 * me_txkey with its destructor must be disabled first.
4808 * We skip the the reader mutex, so we touch only
4809 * data owned by this process (me_close_readers and
4810 * our readers), and clear each reader atomically.
4812 for (i = env->me_close_readers; --i >= 0; )
4813 if (env->me_txns->mti_readers[i].mr_pid == pid)
4814 env->me_txns->mti_readers[i].mr_pid = 0;
4816 if (env->me_rmutex) {
4817 CloseHandle(env->me_rmutex);
4818 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4820 /* Windows automatically destroys the mutexes when
4821 * the last handle closes.
4823 #elif defined(MDB_USE_POSIX_SEM)
4824 if (env->me_rmutex != SEM_FAILED) {
4825 sem_close(env->me_rmutex);
4826 if (env->me_wmutex != SEM_FAILED)
4827 sem_close(env->me_wmutex);
4828 /* If we have the filelock: If we are the
4829 * only remaining user, clean up semaphores.
4832 mdb_env_excl_lock(env, &excl);
4834 sem_unlink(env->me_txns->mti_rmname);
4835 sem_unlink(env->me_txns->mti_wmname);
4839 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4841 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4844 /* Unlock the lockfile. Windows would have unlocked it
4845 * after closing anyway, but not necessarily at once.
4847 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4850 (void) close(env->me_lfd);
4853 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4857 mdb_env_close(MDB_env *env)
4864 VGMEMP_DESTROY(env);
4865 while ((dp = env->me_dpages) != NULL) {
4866 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4867 env->me_dpages = dp->mp_next;
4871 mdb_env_close0(env, 0);
4875 /** Compare two items pointing at aligned size_t's */
4877 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4879 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4880 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4883 /** Compare two items pointing at aligned unsigned int's.
4885 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4886 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4889 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4891 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4892 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4895 /** Compare two items pointing at unsigned ints of unknown alignment.
4896 * Nodes and keys are guaranteed to be 2-byte aligned.
4899 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4901 #if BYTE_ORDER == LITTLE_ENDIAN
4902 unsigned short *u, *c;
4905 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4906 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4909 } while(!x && u > (unsigned short *)a->mv_data);
4912 unsigned short *u, *c, *end;
4915 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4916 u = (unsigned short *)a->mv_data;
4917 c = (unsigned short *)b->mv_data;
4920 } while(!x && u < end);
4925 /** Compare two items lexically */
4927 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4934 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4940 diff = memcmp(a->mv_data, b->mv_data, len);
4941 return diff ? diff : len_diff<0 ? -1 : len_diff;
4944 /** Compare two items in reverse byte order */
4946 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4948 const unsigned char *p1, *p2, *p1_lim;
4952 p1_lim = (const unsigned char *)a->mv_data;
4953 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4954 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4956 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4962 while (p1 > p1_lim) {
4963 diff = *--p1 - *--p2;
4967 return len_diff<0 ? -1 : len_diff;
4970 /** Search for key within a page, using binary search.
4971 * Returns the smallest entry larger or equal to the key.
4972 * If exactp is non-null, stores whether the found entry was an exact match
4973 * in *exactp (1 or 0).
4974 * Updates the cursor index with the index of the found entry.
4975 * If no entry larger or equal to the key is found, returns NULL.
4978 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4980 unsigned int i = 0, nkeys;
4983 MDB_page *mp = mc->mc_pg[mc->mc_top];
4984 MDB_node *node = NULL;
4989 nkeys = NUMKEYS(mp);
4991 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4992 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4995 low = IS_LEAF(mp) ? 0 : 1;
4997 cmp = mc->mc_dbx->md_cmp;
4999 /* Branch pages have no data, so if using integer keys,
5000 * alignment is guaranteed. Use faster mdb_cmp_int.
5002 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5003 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5010 nodekey.mv_size = mc->mc_db->md_pad;
5011 node = NODEPTR(mp, 0); /* fake */
5012 while (low <= high) {
5013 i = (low + high) >> 1;
5014 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5015 rc = cmp(key, &nodekey);
5016 DPRINTF(("found leaf index %u [%s], rc = %i",
5017 i, DKEY(&nodekey), rc));
5026 while (low <= high) {
5027 i = (low + high) >> 1;
5029 node = NODEPTR(mp, i);
5030 nodekey.mv_size = NODEKSZ(node);
5031 nodekey.mv_data = NODEKEY(node);
5033 rc = cmp(key, &nodekey);
5036 DPRINTF(("found leaf index %u [%s], rc = %i",
5037 i, DKEY(&nodekey), rc));
5039 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5040 i, DKEY(&nodekey), NODEPGNO(node), rc));
5051 if (rc > 0) { /* Found entry is less than the key. */
5052 i++; /* Skip to get the smallest entry larger than key. */
5054 node = NODEPTR(mp, i);
5057 *exactp = (rc == 0 && nkeys > 0);
5058 /* store the key index */
5059 mc->mc_ki[mc->mc_top] = i;
5061 /* There is no entry larger or equal to the key. */
5064 /* nodeptr is fake for LEAF2 */
5070 mdb_cursor_adjust(MDB_cursor *mc, func)
5074 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5075 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5082 /** Pop a page off the top of the cursor's stack. */
5084 mdb_cursor_pop(MDB_cursor *mc)
5087 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5088 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5096 /** Push a page onto the top of the cursor's stack. */
5098 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5100 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5101 DDBI(mc), (void *) mc));
5103 if (mc->mc_snum >= CURSOR_STACK) {
5104 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5105 return MDB_CURSOR_FULL;
5108 mc->mc_top = mc->mc_snum++;
5109 mc->mc_pg[mc->mc_top] = mp;
5110 mc->mc_ki[mc->mc_top] = 0;
5115 /** Find the address of the page corresponding to a given page number.
5116 * @param[in] txn the transaction for this access.
5117 * @param[in] pgno the page number for the page to retrieve.
5118 * @param[out] ret address of a pointer where the page's address will be stored.
5119 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5120 * @return 0 on success, non-zero on failure.
5123 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5125 MDB_env *env = txn->mt_env;
5129 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5133 MDB_ID2L dl = tx2->mt_u.dirty_list;
5135 /* Spilled pages were dirtied in this txn and flushed
5136 * because the dirty list got full. Bring this page
5137 * back in from the map (but don't unspill it here,
5138 * leave that unless page_touch happens again).
5140 if (tx2->mt_spill_pgs) {
5141 MDB_ID pn = pgno << 1;
5142 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5143 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5144 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5149 unsigned x = mdb_mid2l_search(dl, pgno);
5150 if (x <= dl[0].mid && dl[x].mid == pgno) {
5156 } while ((tx2 = tx2->mt_parent) != NULL);
5159 if (pgno < txn->mt_next_pgno) {
5161 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5163 DPRINTF(("page %"Z"u not found", pgno));
5164 txn->mt_flags |= MDB_TXN_ERROR;
5165 return MDB_PAGE_NOTFOUND;
5175 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5176 * The cursor is at the root page, set up the rest of it.
5179 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5181 MDB_page *mp = mc->mc_pg[mc->mc_top];
5185 while (IS_BRANCH(mp)) {
5189 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5190 mdb_cassert(mc, NUMKEYS(mp) > 1);
5191 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5193 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5195 if (flags & MDB_PS_LAST)
5196 i = NUMKEYS(mp) - 1;
5199 node = mdb_node_search(mc, key, &exact);
5201 i = NUMKEYS(mp) - 1;
5203 i = mc->mc_ki[mc->mc_top];
5205 mdb_cassert(mc, i > 0);
5209 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5212 mdb_cassert(mc, i < NUMKEYS(mp));
5213 node = NODEPTR(mp, i);
5215 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5218 mc->mc_ki[mc->mc_top] = i;
5219 if ((rc = mdb_cursor_push(mc, mp)))
5222 if (flags & MDB_PS_MODIFY) {
5223 if ((rc = mdb_page_touch(mc)) != 0)
5225 mp = mc->mc_pg[mc->mc_top];
5230 DPRINTF(("internal error, index points to a %02X page!?",
5232 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5233 return MDB_CORRUPTED;
5236 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5237 key ? DKEY(key) : "null"));
5238 mc->mc_flags |= C_INITIALIZED;
5239 mc->mc_flags &= ~C_EOF;
5244 /** Search for the lowest key under the current branch page.
5245 * This just bypasses a NUMKEYS check in the current page
5246 * before calling mdb_page_search_root(), because the callers
5247 * are all in situations where the current page is known to
5251 mdb_page_search_lowest(MDB_cursor *mc)
5253 MDB_page *mp = mc->mc_pg[mc->mc_top];
5254 MDB_node *node = NODEPTR(mp, 0);
5257 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5260 mc->mc_ki[mc->mc_top] = 0;
5261 if ((rc = mdb_cursor_push(mc, mp)))
5263 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5266 /** Search for the page a given key should be in.
5267 * Push it and its parent pages on the cursor stack.
5268 * @param[in,out] mc the cursor for this operation.
5269 * @param[in] key the key to search for, or NULL for first/last page.
5270 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5271 * are touched (updated with new page numbers).
5272 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5273 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5274 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5275 * @return 0 on success, non-zero on failure.
5278 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5283 /* Make sure the txn is still viable, then find the root from
5284 * the txn's db table and set it as the root of the cursor's stack.
5286 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5287 DPUTS("transaction has failed, must abort");
5290 /* Make sure we're using an up-to-date root */
5291 if (*mc->mc_dbflag & DB_STALE) {
5293 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5295 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5296 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5303 MDB_node *leaf = mdb_node_search(&mc2,
5304 &mc->mc_dbx->md_name, &exact);
5306 return MDB_NOTFOUND;
5307 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5308 return MDB_INCOMPATIBLE; /* not a named DB */
5309 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5312 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5314 /* The txn may not know this DBI, or another process may
5315 * have dropped and recreated the DB with other flags.
5317 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5318 return MDB_INCOMPATIBLE;
5319 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5321 *mc->mc_dbflag &= ~DB_STALE;
5323 root = mc->mc_db->md_root;
5325 if (root == P_INVALID) { /* Tree is empty. */
5326 DPUTS("tree is empty");
5327 return MDB_NOTFOUND;
5331 mdb_cassert(mc, root > 1);
5332 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5333 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5339 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5340 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5342 if (flags & MDB_PS_MODIFY) {
5343 if ((rc = mdb_page_touch(mc)))
5347 if (flags & MDB_PS_ROOTONLY)
5350 return mdb_page_search_root(mc, key, flags);
5354 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5356 MDB_txn *txn = mc->mc_txn;
5357 pgno_t pg = mp->mp_pgno;
5358 unsigned x = 0, ovpages = mp->mp_pages;
5359 MDB_env *env = txn->mt_env;
5360 MDB_IDL sl = txn->mt_spill_pgs;
5361 MDB_ID pn = pg << 1;
5364 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5365 /* If the page is dirty or on the spill list we just acquired it,
5366 * so we should give it back to our current free list, if any.
5367 * Otherwise put it onto the list of pages we freed in this txn.
5369 * Won't create me_pghead: me_pglast must be inited along with it.
5370 * Unsupported in nested txns: They would need to hide the page
5371 * range in ancestor txns' dirty and spilled lists.
5373 if (env->me_pghead &&
5375 ((mp->mp_flags & P_DIRTY) ||
5376 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5380 MDB_ID2 *dl, ix, iy;
5381 rc = mdb_midl_need(&env->me_pghead, ovpages);
5384 if (!(mp->mp_flags & P_DIRTY)) {
5385 /* This page is no longer spilled */
5392 /* Remove from dirty list */
5393 dl = txn->mt_u.dirty_list;
5395 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5401 mdb_cassert(mc, x > 1);
5403 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5404 txn->mt_flags |= MDB_TXN_ERROR;
5405 return MDB_CORRUPTED;
5408 if (!(env->me_flags & MDB_WRITEMAP))
5409 mdb_dpage_free(env, mp);
5411 /* Insert in me_pghead */
5412 mop = env->me_pghead;
5413 j = mop[0] + ovpages;
5414 for (i = mop[0]; i && mop[i] < pg; i--)
5420 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5424 mc->mc_db->md_overflow_pages -= ovpages;
5428 /** Return the data associated with a given node.
5429 * @param[in] txn The transaction for this operation.
5430 * @param[in] leaf The node being read.
5431 * @param[out] data Updated to point to the node's data.
5432 * @return 0 on success, non-zero on failure.
5435 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5437 MDB_page *omp; /* overflow page */
5441 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5442 data->mv_size = NODEDSZ(leaf);
5443 data->mv_data = NODEDATA(leaf);
5447 /* Read overflow data.
5449 data->mv_size = NODEDSZ(leaf);
5450 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5451 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5452 DPRINTF(("read overflow page %"Z"u failed", pgno));
5455 data->mv_data = METADATA(omp);
5461 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5462 MDB_val *key, MDB_val *data)
5469 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5471 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5474 if (txn->mt_flags & MDB_TXN_ERROR)
5477 mdb_cursor_init(&mc, txn, dbi, &mx);
5478 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5481 /** Find a sibling for a page.
5482 * Replaces the page at the top of the cursor's stack with the
5483 * specified sibling, if one exists.
5484 * @param[in] mc The cursor for this operation.
5485 * @param[in] move_right Non-zero if the right sibling is requested,
5486 * otherwise the left sibling.
5487 * @return 0 on success, non-zero on failure.
5490 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5496 if (mc->mc_snum < 2) {
5497 return MDB_NOTFOUND; /* root has no siblings */
5501 DPRINTF(("parent page is page %"Z"u, index %u",
5502 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5504 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5505 : (mc->mc_ki[mc->mc_top] == 0)) {
5506 DPRINTF(("no more keys left, moving to %s sibling",
5507 move_right ? "right" : "left"));
5508 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5509 /* undo cursor_pop before returning */
5516 mc->mc_ki[mc->mc_top]++;
5518 mc->mc_ki[mc->mc_top]--;
5519 DPRINTF(("just moving to %s index key %u",
5520 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5522 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5524 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5525 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5526 /* mc will be inconsistent if caller does mc_snum++ as above */
5527 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5531 mdb_cursor_push(mc, mp);
5533 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5538 /** Move the cursor to the next data item. */
5540 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5546 if (mc->mc_flags & C_EOF) {
5547 return MDB_NOTFOUND;
5550 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5552 mp = mc->mc_pg[mc->mc_top];
5554 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5555 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5556 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5557 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5558 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5559 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5560 if (rc == MDB_SUCCESS)
5561 MDB_GET_KEY(leaf, key);
5566 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5567 if (op == MDB_NEXT_DUP)
5568 return MDB_NOTFOUND;
5572 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5573 mdb_dbg_pgno(mp), (void *) mc));
5574 if (mc->mc_flags & C_DEL)
5577 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5578 DPUTS("=====> move to next sibling page");
5579 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5580 mc->mc_flags |= C_EOF;
5583 mp = mc->mc_pg[mc->mc_top];
5584 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5586 mc->mc_ki[mc->mc_top]++;
5589 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5590 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5593 key->mv_size = mc->mc_db->md_pad;
5594 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5598 mdb_cassert(mc, IS_LEAF(mp));
5599 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5601 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5602 mdb_xcursor_init1(mc, leaf);
5605 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5608 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5609 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5610 if (rc != MDB_SUCCESS)
5615 MDB_GET_KEY(leaf, key);
5619 /** Move the cursor to the previous data item. */
5621 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5627 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5629 mp = mc->mc_pg[mc->mc_top];
5631 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5632 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5633 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5634 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5635 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5636 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5637 if (rc == MDB_SUCCESS) {
5638 MDB_GET_KEY(leaf, key);
5639 mc->mc_flags &= ~C_EOF;
5645 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5646 if (op == MDB_PREV_DUP)
5647 return MDB_NOTFOUND;
5651 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5652 mdb_dbg_pgno(mp), (void *) mc));
5654 if (mc->mc_ki[mc->mc_top] == 0) {
5655 DPUTS("=====> move to prev sibling page");
5656 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5659 mp = mc->mc_pg[mc->mc_top];
5660 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5661 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5663 mc->mc_ki[mc->mc_top]--;
5665 mc->mc_flags &= ~C_EOF;
5667 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5668 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5671 key->mv_size = mc->mc_db->md_pad;
5672 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5676 mdb_cassert(mc, IS_LEAF(mp));
5677 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5679 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5680 mdb_xcursor_init1(mc, leaf);
5683 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5686 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5687 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5688 if (rc != MDB_SUCCESS)
5693 MDB_GET_KEY(leaf, key);
5697 /** Set the cursor on a specific data item. */
5699 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5700 MDB_cursor_op op, int *exactp)
5704 MDB_node *leaf = NULL;
5707 if (key->mv_size == 0)
5708 return MDB_BAD_VALSIZE;
5711 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5713 /* See if we're already on the right page */
5714 if (mc->mc_flags & C_INITIALIZED) {
5717 mp = mc->mc_pg[mc->mc_top];
5719 mc->mc_ki[mc->mc_top] = 0;
5720 return MDB_NOTFOUND;
5722 if (mp->mp_flags & P_LEAF2) {
5723 nodekey.mv_size = mc->mc_db->md_pad;
5724 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5726 leaf = NODEPTR(mp, 0);
5727 MDB_GET_KEY2(leaf, nodekey);
5729 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5731 /* Probably happens rarely, but first node on the page
5732 * was the one we wanted.
5734 mc->mc_ki[mc->mc_top] = 0;
5741 unsigned int nkeys = NUMKEYS(mp);
5743 if (mp->mp_flags & P_LEAF2) {
5744 nodekey.mv_data = LEAF2KEY(mp,
5745 nkeys-1, nodekey.mv_size);
5747 leaf = NODEPTR(mp, nkeys-1);
5748 MDB_GET_KEY2(leaf, nodekey);
5750 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5752 /* last node was the one we wanted */
5753 mc->mc_ki[mc->mc_top] = nkeys-1;
5759 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5760 /* This is definitely the right page, skip search_page */
5761 if (mp->mp_flags & P_LEAF2) {
5762 nodekey.mv_data = LEAF2KEY(mp,
5763 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5765 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5766 MDB_GET_KEY2(leaf, nodekey);
5768 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5770 /* current node was the one we wanted */
5780 /* If any parents have right-sibs, search.
5781 * Otherwise, there's nothing further.
5783 for (i=0; i<mc->mc_top; i++)
5785 NUMKEYS(mc->mc_pg[i])-1)
5787 if (i == mc->mc_top) {
5788 /* There are no other pages */
5789 mc->mc_ki[mc->mc_top] = nkeys;
5790 return MDB_NOTFOUND;
5794 /* There are no other pages */
5795 mc->mc_ki[mc->mc_top] = 0;
5796 if (op == MDB_SET_RANGE && !exactp) {
5800 return MDB_NOTFOUND;
5804 rc = mdb_page_search(mc, key, 0);
5805 if (rc != MDB_SUCCESS)
5808 mp = mc->mc_pg[mc->mc_top];
5809 mdb_cassert(mc, IS_LEAF(mp));
5812 leaf = mdb_node_search(mc, key, exactp);
5813 if (exactp != NULL && !*exactp) {
5814 /* MDB_SET specified and not an exact match. */
5815 return MDB_NOTFOUND;
5819 DPUTS("===> inexact leaf not found, goto sibling");
5820 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5821 mc->mc_flags |= C_EOF;
5822 return rc; /* no entries matched */
5824 mp = mc->mc_pg[mc->mc_top];
5825 mdb_cassert(mc, IS_LEAF(mp));
5826 leaf = NODEPTR(mp, 0);
5830 mc->mc_flags |= C_INITIALIZED;
5831 mc->mc_flags &= ~C_EOF;
5834 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5835 key->mv_size = mc->mc_db->md_pad;
5836 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5841 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5842 mdb_xcursor_init1(mc, leaf);
5845 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5846 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5847 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5850 if (op == MDB_GET_BOTH) {
5856 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5857 if (rc != MDB_SUCCESS)
5860 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5863 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5865 dcmp = mc->mc_dbx->md_dcmp;
5866 #if UINT_MAX < SIZE_MAX
5867 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5868 dcmp = mdb_cmp_clong;
5870 rc = dcmp(data, &olddata);
5872 if (op == MDB_GET_BOTH || rc > 0)
5873 return MDB_NOTFOUND;
5880 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5881 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5886 /* The key already matches in all other cases */
5887 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5888 MDB_GET_KEY(leaf, key);
5889 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5894 /** Move the cursor to the first item in the database. */
5896 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5902 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5904 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5905 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5906 if (rc != MDB_SUCCESS)
5909 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5911 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5912 mc->mc_flags |= C_INITIALIZED;
5913 mc->mc_flags &= ~C_EOF;
5915 mc->mc_ki[mc->mc_top] = 0;
5917 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5918 key->mv_size = mc->mc_db->md_pad;
5919 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5924 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5925 mdb_xcursor_init1(mc, leaf);
5926 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5930 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5934 MDB_GET_KEY(leaf, key);
5938 /** Move the cursor to the last item in the database. */
5940 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5946 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5948 if (!(mc->mc_flags & C_EOF)) {
5950 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5951 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5952 if (rc != MDB_SUCCESS)
5955 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5958 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5959 mc->mc_flags |= C_INITIALIZED|C_EOF;
5960 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5962 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5963 key->mv_size = mc->mc_db->md_pad;
5964 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5969 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5970 mdb_xcursor_init1(mc, leaf);
5971 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5975 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5980 MDB_GET_KEY(leaf, key);
5985 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5990 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5995 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5999 case MDB_GET_CURRENT:
6000 if (!(mc->mc_flags & C_INITIALIZED)) {
6003 MDB_page *mp = mc->mc_pg[mc->mc_top];
6004 int nkeys = NUMKEYS(mp);
6005 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6006 mc->mc_ki[mc->mc_top] = nkeys;
6012 key->mv_size = mc->mc_db->md_pad;
6013 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6015 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6016 MDB_GET_KEY(leaf, key);
6018 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6019 if (mc->mc_flags & C_DEL)
6020 mdb_xcursor_init1(mc, leaf);
6021 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6023 rc = mdb_node_read(mc->mc_txn, leaf, data);
6030 case MDB_GET_BOTH_RANGE:
6035 if (mc->mc_xcursor == NULL) {
6036 rc = MDB_INCOMPATIBLE;
6046 rc = mdb_cursor_set(mc, key, data, op,
6047 op == MDB_SET_RANGE ? NULL : &exact);
6050 case MDB_GET_MULTIPLE:
6051 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6055 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6056 rc = MDB_INCOMPATIBLE;
6060 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6061 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6064 case MDB_NEXT_MULTIPLE:
6069 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6070 rc = MDB_INCOMPATIBLE;
6073 if (!(mc->mc_flags & C_INITIALIZED))
6074 rc = mdb_cursor_first(mc, key, data);
6076 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6077 if (rc == MDB_SUCCESS) {
6078 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6081 mx = &mc->mc_xcursor->mx_cursor;
6082 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6084 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6085 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6093 case MDB_NEXT_NODUP:
6094 if (!(mc->mc_flags & C_INITIALIZED))
6095 rc = mdb_cursor_first(mc, key, data);
6097 rc = mdb_cursor_next(mc, key, data, op);
6101 case MDB_PREV_NODUP:
6102 if (!(mc->mc_flags & C_INITIALIZED)) {
6103 rc = mdb_cursor_last(mc, key, data);
6106 mc->mc_flags |= C_INITIALIZED;
6107 mc->mc_ki[mc->mc_top]++;
6109 rc = mdb_cursor_prev(mc, key, data, op);
6112 rc = mdb_cursor_first(mc, key, data);
6115 mfunc = mdb_cursor_first;
6117 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6121 if (mc->mc_xcursor == NULL) {
6122 rc = MDB_INCOMPATIBLE;
6126 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6127 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6128 MDB_GET_KEY(leaf, key);
6129 rc = mdb_node_read(mc->mc_txn, leaf, data);
6133 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6137 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6140 rc = mdb_cursor_last(mc, key, data);
6143 mfunc = mdb_cursor_last;
6146 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6151 if (mc->mc_flags & C_DEL)
6152 mc->mc_flags ^= C_DEL;
6157 /** Touch all the pages in the cursor stack. Set mc_top.
6158 * Makes sure all the pages are writable, before attempting a write operation.
6159 * @param[in] mc The cursor to operate on.
6162 mdb_cursor_touch(MDB_cursor *mc)
6164 int rc = MDB_SUCCESS;
6166 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6169 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6171 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6172 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6175 *mc->mc_dbflag |= DB_DIRTY;
6180 rc = mdb_page_touch(mc);
6181 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6182 mc->mc_top = mc->mc_snum-1;
6187 /** Do not spill pages to disk if txn is getting full, may fail instead */
6188 #define MDB_NOSPILL 0x8000
6191 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6195 MDB_node *leaf = NULL;
6196 MDB_page *fp, *mp, *sub_root = NULL;
6198 MDB_val xdata, *rdata, dkey, olddata;
6200 int do_sub = 0, insert_key, insert_data;
6201 unsigned int mcount = 0, dcount = 0, nospill;
6204 unsigned int nflags;
6207 if (mc == NULL || key == NULL)
6210 env = mc->mc_txn->mt_env;
6212 /* Check this first so counter will always be zero on any
6215 if (flags & MDB_MULTIPLE) {
6216 dcount = data[1].mv_size;
6217 data[1].mv_size = 0;
6218 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6219 return MDB_INCOMPATIBLE;
6222 nospill = flags & MDB_NOSPILL;
6223 flags &= ~MDB_NOSPILL;
6225 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6226 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6228 if (key->mv_size-1 >= ENV_MAXKEY(env))
6229 return MDB_BAD_VALSIZE;
6231 #if SIZE_MAX > MAXDATASIZE
6232 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6233 return MDB_BAD_VALSIZE;
6235 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6236 return MDB_BAD_VALSIZE;
6239 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6240 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6244 if (flags == MDB_CURRENT) {
6245 if (!(mc->mc_flags & C_INITIALIZED))
6248 } else if (mc->mc_db->md_root == P_INVALID) {
6249 /* new database, cursor has nothing to point to */
6252 mc->mc_flags &= ~C_INITIALIZED;
6257 if (flags & MDB_APPEND) {
6259 rc = mdb_cursor_last(mc, &k2, &d2);
6261 rc = mc->mc_dbx->md_cmp(key, &k2);
6264 mc->mc_ki[mc->mc_top]++;
6266 /* new key is <= last key */
6271 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6273 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6274 DPRINTF(("duplicate key [%s]", DKEY(key)));
6276 return MDB_KEYEXIST;
6278 if (rc && rc != MDB_NOTFOUND)
6282 if (mc->mc_flags & C_DEL)
6283 mc->mc_flags ^= C_DEL;
6285 /* Cursor is positioned, check for room in the dirty list */
6287 if (flags & MDB_MULTIPLE) {
6289 xdata.mv_size = data->mv_size * dcount;
6293 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6297 if (rc == MDB_NO_ROOT) {
6299 /* new database, write a root leaf page */
6300 DPUTS("allocating new root leaf page");
6301 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6304 mdb_cursor_push(mc, np);
6305 mc->mc_db->md_root = np->mp_pgno;
6306 mc->mc_db->md_depth++;
6307 *mc->mc_dbflag |= DB_DIRTY;
6308 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6310 np->mp_flags |= P_LEAF2;
6311 mc->mc_flags |= C_INITIALIZED;
6313 /* make sure all cursor pages are writable */
6314 rc2 = mdb_cursor_touch(mc);
6319 insert_key = insert_data = rc;
6321 /* The key does not exist */
6322 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6323 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6324 LEAFSIZE(key, data) > env->me_nodemax)
6326 /* Too big for a node, insert in sub-DB. Set up an empty
6327 * "old sub-page" for prep_subDB to expand to a full page.
6329 fp_flags = P_LEAF|P_DIRTY;
6331 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6332 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6333 olddata.mv_size = PAGEHDRSZ;
6337 /* there's only a key anyway, so this is a no-op */
6338 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6340 unsigned int ksize = mc->mc_db->md_pad;
6341 if (key->mv_size != ksize)
6342 return MDB_BAD_VALSIZE;
6343 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6344 memcpy(ptr, key->mv_data, ksize);
6346 /* if overwriting slot 0 of leaf, need to
6347 * update branch key if there is a parent page
6349 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6350 unsigned short top = mc->mc_top;
6352 /* slot 0 is always an empty key, find real slot */
6353 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6355 if (mc->mc_ki[mc->mc_top])
6356 rc2 = mdb_update_key(mc, key);
6367 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6368 olddata.mv_size = NODEDSZ(leaf);
6369 olddata.mv_data = NODEDATA(leaf);
6372 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6373 /* Prepare (sub-)page/sub-DB to accept the new item,
6374 * if needed. fp: old sub-page or a header faking
6375 * it. mp: new (sub-)page. offset: growth in page
6376 * size. xdata: node data with new page or DB.
6378 unsigned i, offset = 0;
6379 mp = fp = xdata.mv_data = env->me_pbuf;
6380 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6382 /* Was a single item before, must convert now */
6383 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6385 /* Just overwrite the current item */
6386 if (flags == MDB_CURRENT)
6388 dcmp = mc->mc_dbx->md_dcmp;
6389 #if UINT_MAX < SIZE_MAX
6390 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6391 dcmp = mdb_cmp_clong;
6393 /* does data match? */
6394 if (!dcmp(data, &olddata)) {
6395 if (flags & MDB_NODUPDATA)
6396 return MDB_KEYEXIST;
6401 /* Back up original data item */
6402 dkey.mv_size = olddata.mv_size;
6403 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6405 /* Make sub-page header for the dup items, with dummy body */
6406 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6407 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6408 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6409 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6410 fp->mp_flags |= P_LEAF2;
6411 fp->mp_pad = data->mv_size;
6412 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6414 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6415 (dkey.mv_size & 1) + (data->mv_size & 1);
6417 fp->mp_upper = xdata.mv_size - PAGEBASE;
6418 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6419 } else if (leaf->mn_flags & F_SUBDATA) {
6420 /* Data is on sub-DB, just store it */
6421 flags |= F_DUPDATA|F_SUBDATA;
6424 /* Data is on sub-page */
6425 fp = olddata.mv_data;
6428 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6429 offset = EVEN(NODESIZE + sizeof(indx_t) +
6433 offset = fp->mp_pad;
6434 if (SIZELEFT(fp) < offset) {
6435 offset *= 4; /* space for 4 more */
6438 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6440 fp->mp_flags |= P_DIRTY;
6441 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6442 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6446 xdata.mv_size = olddata.mv_size + offset;
6449 fp_flags = fp->mp_flags;
6450 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6451 /* Too big for a sub-page, convert to sub-DB */
6452 fp_flags &= ~P_SUBP;
6454 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6455 fp_flags |= P_LEAF2;
6456 dummy.md_pad = fp->mp_pad;
6457 dummy.md_flags = MDB_DUPFIXED;
6458 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6459 dummy.md_flags |= MDB_INTEGERKEY;
6465 dummy.md_branch_pages = 0;
6466 dummy.md_leaf_pages = 1;
6467 dummy.md_overflow_pages = 0;
6468 dummy.md_entries = NUMKEYS(fp);
6469 xdata.mv_size = sizeof(MDB_db);
6470 xdata.mv_data = &dummy;
6471 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6473 offset = env->me_psize - olddata.mv_size;
6474 flags |= F_DUPDATA|F_SUBDATA;
6475 dummy.md_root = mp->mp_pgno;
6479 mp->mp_flags = fp_flags | P_DIRTY;
6480 mp->mp_pad = fp->mp_pad;
6481 mp->mp_lower = fp->mp_lower;
6482 mp->mp_upper = fp->mp_upper + offset;
6483 if (fp_flags & P_LEAF2) {
6484 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6486 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6487 olddata.mv_size - fp->mp_upper - PAGEBASE);
6488 for (i=0; i<NUMKEYS(fp); i++)
6489 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6497 mdb_node_del(mc, 0);
6501 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6502 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6503 return MDB_INCOMPATIBLE;
6504 /* overflow page overwrites need special handling */
6505 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6508 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6510 memcpy(&pg, olddata.mv_data, sizeof(pg));
6511 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6513 ovpages = omp->mp_pages;
6515 /* Is the ov page large enough? */
6516 if (ovpages >= dpages) {
6517 if (!(omp->mp_flags & P_DIRTY) &&
6518 (level || (env->me_flags & MDB_WRITEMAP)))
6520 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6523 level = 0; /* dirty in this txn or clean */
6526 if (omp->mp_flags & P_DIRTY) {
6527 /* yes, overwrite it. Note in this case we don't
6528 * bother to try shrinking the page if the new data
6529 * is smaller than the overflow threshold.
6532 /* It is writable only in a parent txn */
6533 size_t sz = (size_t) env->me_psize * ovpages, off;
6534 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6540 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6541 mdb_cassert(mc, rc2 == 0);
6542 if (!(flags & MDB_RESERVE)) {
6543 /* Copy end of page, adjusting alignment so
6544 * compiler may copy words instead of bytes.
6546 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6547 memcpy((size_t *)((char *)np + off),
6548 (size_t *)((char *)omp + off), sz - off);
6551 memcpy(np, omp, sz); /* Copy beginning of page */
6554 SETDSZ(leaf, data->mv_size);
6555 if (F_ISSET(flags, MDB_RESERVE))
6556 data->mv_data = METADATA(omp);
6558 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6562 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6564 } else if (data->mv_size == olddata.mv_size) {
6565 /* same size, just replace it. Note that we could
6566 * also reuse this node if the new data is smaller,
6567 * but instead we opt to shrink the node in that case.
6569 if (F_ISSET(flags, MDB_RESERVE))
6570 data->mv_data = olddata.mv_data;
6571 else if (!(mc->mc_flags & C_SUB))
6572 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6574 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6579 mdb_node_del(mc, 0);
6585 nflags = flags & NODE_ADD_FLAGS;
6586 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6587 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6588 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6589 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6591 nflags |= MDB_SPLIT_REPLACE;
6592 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6594 /* There is room already in this leaf page. */
6595 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6596 if (rc == 0 && insert_key) {
6597 /* Adjust other cursors pointing to mp */
6598 MDB_cursor *m2, *m3;
6599 MDB_dbi dbi = mc->mc_dbi;
6600 unsigned i = mc->mc_top;
6601 MDB_page *mp = mc->mc_pg[i];
6603 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6604 if (mc->mc_flags & C_SUB)
6605 m3 = &m2->mc_xcursor->mx_cursor;
6608 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6609 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6616 if (rc == MDB_SUCCESS) {
6617 /* Now store the actual data in the child DB. Note that we're
6618 * storing the user data in the keys field, so there are strict
6619 * size limits on dupdata. The actual data fields of the child
6620 * DB are all zero size.
6623 int xflags, new_dupdata;
6628 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6629 if (flags & MDB_CURRENT) {
6630 xflags = MDB_CURRENT|MDB_NOSPILL;
6632 mdb_xcursor_init1(mc, leaf);
6633 xflags = (flags & MDB_NODUPDATA) ?
6634 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6637 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6638 new_dupdata = (int)dkey.mv_size;
6639 /* converted, write the original data first */
6641 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6644 /* we've done our job */
6647 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6648 /* Adjust other cursors pointing to mp */
6650 MDB_xcursor *mx = mc->mc_xcursor;
6651 unsigned i = mc->mc_top;
6652 MDB_page *mp = mc->mc_pg[i];
6654 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6655 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6656 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6657 if (m2->mc_pg[i] == mp) {
6658 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6659 mdb_xcursor_init2(m2, mx, new_dupdata);
6660 } else if (!insert_key) {
6661 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6662 if (!(n2->mn_flags & F_SUBDATA))
6663 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6668 ecount = mc->mc_xcursor->mx_db.md_entries;
6669 if (flags & MDB_APPENDDUP)
6670 xflags |= MDB_APPEND;
6671 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6672 if (flags & F_SUBDATA) {
6673 void *db = NODEDATA(leaf);
6674 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6676 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6678 /* Increment count unless we just replaced an existing item. */
6680 mc->mc_db->md_entries++;
6682 /* Invalidate txn if we created an empty sub-DB */
6685 /* If we succeeded and the key didn't exist before,
6686 * make sure the cursor is marked valid.
6688 mc->mc_flags |= C_INITIALIZED;
6690 if (flags & MDB_MULTIPLE) {
6693 /* let caller know how many succeeded, if any */
6694 data[1].mv_size = mcount;
6695 if (mcount < dcount) {
6696 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6697 insert_key = insert_data = 0;
6704 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6707 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6712 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6718 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6719 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6721 if (!(mc->mc_flags & C_INITIALIZED))
6724 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6725 return MDB_NOTFOUND;
6727 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6730 rc = mdb_cursor_touch(mc);
6734 mp = mc->mc_pg[mc->mc_top];
6737 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6739 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6740 if (flags & MDB_NODUPDATA) {
6741 /* mdb_cursor_del0() will subtract the final entry */
6742 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6744 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6745 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6747 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6750 /* If sub-DB still has entries, we're done */
6751 if (mc->mc_xcursor->mx_db.md_entries) {
6752 if (leaf->mn_flags & F_SUBDATA) {
6753 /* update subDB info */
6754 void *db = NODEDATA(leaf);
6755 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6758 /* shrink fake page */
6759 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6760 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6761 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6762 /* fix other sub-DB cursors pointed at fake pages on this page */
6763 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6764 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6765 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6766 if (m2->mc_pg[mc->mc_top] == mp) {
6767 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6768 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6770 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6771 if (!(n2->mn_flags & F_SUBDATA))
6772 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6777 mc->mc_db->md_entries--;
6778 mc->mc_flags |= C_DEL;
6781 /* otherwise fall thru and delete the sub-DB */
6784 if (leaf->mn_flags & F_SUBDATA) {
6785 /* add all the child DB's pages to the free list */
6786 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6791 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6792 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6793 rc = MDB_INCOMPATIBLE;
6797 /* add overflow pages to free list */
6798 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6802 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6803 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6804 (rc = mdb_ovpage_free(mc, omp)))
6809 return mdb_cursor_del0(mc);
6812 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6816 /** Allocate and initialize new pages for a database.
6817 * @param[in] mc a cursor on the database being added to.
6818 * @param[in] flags flags defining what type of page is being allocated.
6819 * @param[in] num the number of pages to allocate. This is usually 1,
6820 * unless allocating overflow pages for a large record.
6821 * @param[out] mp Address of a page, or NULL on failure.
6822 * @return 0 on success, non-zero on failure.
6825 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6830 if ((rc = mdb_page_alloc(mc, num, &np)))
6832 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6833 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6834 np->mp_flags = flags | P_DIRTY;
6835 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6836 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6839 mc->mc_db->md_branch_pages++;
6840 else if (IS_LEAF(np))
6841 mc->mc_db->md_leaf_pages++;
6842 else if (IS_OVERFLOW(np)) {
6843 mc->mc_db->md_overflow_pages += num;
6851 /** Calculate the size of a leaf node.
6852 * The size depends on the environment's page size; if a data item
6853 * is too large it will be put onto an overflow page and the node
6854 * size will only include the key and not the data. Sizes are always
6855 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6856 * of the #MDB_node headers.
6857 * @param[in] env The environment handle.
6858 * @param[in] key The key for the node.
6859 * @param[in] data The data for the node.
6860 * @return The number of bytes needed to store the node.
6863 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6867 sz = LEAFSIZE(key, data);
6868 if (sz > env->me_nodemax) {
6869 /* put on overflow page */
6870 sz -= data->mv_size - sizeof(pgno_t);
6873 return EVEN(sz + sizeof(indx_t));
6876 /** Calculate the size of a branch node.
6877 * The size should depend on the environment's page size but since
6878 * we currently don't support spilling large keys onto overflow
6879 * pages, it's simply the size of the #MDB_node header plus the
6880 * size of the key. Sizes are always rounded up to an even number
6881 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6882 * @param[in] env The environment handle.
6883 * @param[in] key The key for the node.
6884 * @return The number of bytes needed to store the node.
6887 mdb_branch_size(MDB_env *env, MDB_val *key)
6892 if (sz > env->me_nodemax) {
6893 /* put on overflow page */
6894 /* not implemented */
6895 /* sz -= key->size - sizeof(pgno_t); */
6898 return sz + sizeof(indx_t);
6901 /** Add a node to the page pointed to by the cursor.
6902 * @param[in] mc The cursor for this operation.
6903 * @param[in] indx The index on the page where the new node should be added.
6904 * @param[in] key The key for the new node.
6905 * @param[in] data The data for the new node, if any.
6906 * @param[in] pgno The page number, if adding a branch node.
6907 * @param[in] flags Flags for the node.
6908 * @return 0 on success, non-zero on failure. Possible errors are:
6910 * <li>ENOMEM - failed to allocate overflow pages for the node.
6911 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6912 * should never happen since all callers already calculate the
6913 * page's free space before calling this function.
6917 mdb_node_add(MDB_cursor *mc, indx_t indx,
6918 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6921 size_t node_size = NODESIZE;
6925 MDB_page *mp = mc->mc_pg[mc->mc_top];
6926 MDB_page *ofp = NULL; /* overflow page */
6929 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6931 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6932 IS_LEAF(mp) ? "leaf" : "branch",
6933 IS_SUBP(mp) ? "sub-" : "",
6934 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6935 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6938 /* Move higher keys up one slot. */
6939 int ksize = mc->mc_db->md_pad, dif;
6940 char *ptr = LEAF2KEY(mp, indx, ksize);
6941 dif = NUMKEYS(mp) - indx;
6943 memmove(ptr+ksize, ptr, dif*ksize);
6944 /* insert new key */
6945 memcpy(ptr, key->mv_data, ksize);
6947 /* Just using these for counting */
6948 mp->mp_lower += sizeof(indx_t);
6949 mp->mp_upper -= ksize - sizeof(indx_t);
6953 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6955 node_size += key->mv_size;
6957 mdb_cassert(mc, data);
6958 if (F_ISSET(flags, F_BIGDATA)) {
6959 /* Data already on overflow page. */
6960 node_size += sizeof(pgno_t);
6961 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6962 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6964 /* Put data on overflow page. */
6965 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6966 data->mv_size, node_size+data->mv_size));
6967 node_size = EVEN(node_size + sizeof(pgno_t));
6968 if ((ssize_t)node_size > room)
6970 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6972 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6976 node_size += data->mv_size;
6979 node_size = EVEN(node_size);
6980 if ((ssize_t)node_size > room)
6984 /* Move higher pointers up one slot. */
6985 for (i = NUMKEYS(mp); i > indx; i--)
6986 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6988 /* Adjust free space offsets. */
6989 ofs = mp->mp_upper - node_size;
6990 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6991 mp->mp_ptrs[indx] = ofs;
6993 mp->mp_lower += sizeof(indx_t);
6995 /* Write the node data. */
6996 node = NODEPTR(mp, indx);
6997 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6998 node->mn_flags = flags;
7000 SETDSZ(node,data->mv_size);
7005 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7008 mdb_cassert(mc, key);
7010 if (F_ISSET(flags, F_BIGDATA))
7011 memcpy(node->mn_data + key->mv_size, data->mv_data,
7013 else if (F_ISSET(flags, MDB_RESERVE))
7014 data->mv_data = node->mn_data + key->mv_size;
7016 memcpy(node->mn_data + key->mv_size, data->mv_data,
7019 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
7021 if (F_ISSET(flags, MDB_RESERVE))
7022 data->mv_data = METADATA(ofp);
7024 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
7031 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7032 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7033 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7034 DPRINTF(("node size = %"Z"u", node_size));
7035 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7036 return MDB_PAGE_FULL;
7039 /** Delete the specified node from a page.
7040 * @param[in] mc Cursor pointing to the node to delete.
7041 * @param[in] ksize The size of a node. Only used if the page is
7042 * part of a #MDB_DUPFIXED database.
7045 mdb_node_del(MDB_cursor *mc, int ksize)
7047 MDB_page *mp = mc->mc_pg[mc->mc_top];
7048 indx_t indx = mc->mc_ki[mc->mc_top];
7050 indx_t i, j, numkeys, ptr;
7054 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7055 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7056 numkeys = NUMKEYS(mp);
7057 mdb_cassert(mc, indx < numkeys);
7060 int x = numkeys - 1 - indx;
7061 base = LEAF2KEY(mp, indx, ksize);
7063 memmove(base, base + ksize, x * ksize);
7064 mp->mp_lower -= sizeof(indx_t);
7065 mp->mp_upper += ksize - sizeof(indx_t);
7069 node = NODEPTR(mp, indx);
7070 sz = NODESIZE + node->mn_ksize;
7072 if (F_ISSET(node->mn_flags, F_BIGDATA))
7073 sz += sizeof(pgno_t);
7075 sz += NODEDSZ(node);
7079 ptr = mp->mp_ptrs[indx];
7080 for (i = j = 0; i < numkeys; i++) {
7082 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7083 if (mp->mp_ptrs[i] < ptr)
7084 mp->mp_ptrs[j] += sz;
7089 base = (char *)mp + mp->mp_upper + PAGEBASE;
7090 memmove(base + sz, base, ptr - mp->mp_upper);
7092 mp->mp_lower -= sizeof(indx_t);
7096 /** Compact the main page after deleting a node on a subpage.
7097 * @param[in] mp The main page to operate on.
7098 * @param[in] indx The index of the subpage on the main page.
7101 mdb_node_shrink(MDB_page *mp, indx_t indx)
7106 indx_t delta, nsize, len, ptr;
7109 node = NODEPTR(mp, indx);
7110 sp = (MDB_page *)NODEDATA(node);
7111 delta = SIZELEFT(sp);
7112 nsize = NODEDSZ(node) - delta;
7114 /* Prepare to shift upward, set len = length(subpage part to shift) */
7118 return; /* do not make the node uneven-sized */
7120 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7121 for (i = NUMKEYS(sp); --i >= 0; )
7122 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7125 sp->mp_upper = sp->mp_lower;
7126 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7127 SETDSZ(node, nsize);
7129 /* Shift <lower nodes...initial part of subpage> upward */
7130 base = (char *)mp + mp->mp_upper + PAGEBASE;
7131 memmove(base + delta, base, (char *)sp + len - base);
7133 ptr = mp->mp_ptrs[indx];
7134 for (i = NUMKEYS(mp); --i >= 0; ) {
7135 if (mp->mp_ptrs[i] <= ptr)
7136 mp->mp_ptrs[i] += delta;
7138 mp->mp_upper += delta;
7141 /** Initial setup of a sorted-dups cursor.
7142 * Sorted duplicates are implemented as a sub-database for the given key.
7143 * The duplicate data items are actually keys of the sub-database.
7144 * Operations on the duplicate data items are performed using a sub-cursor
7145 * initialized when the sub-database is first accessed. This function does
7146 * the preliminary setup of the sub-cursor, filling in the fields that
7147 * depend only on the parent DB.
7148 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7151 mdb_xcursor_init0(MDB_cursor *mc)
7153 MDB_xcursor *mx = mc->mc_xcursor;
7155 mx->mx_cursor.mc_xcursor = NULL;
7156 mx->mx_cursor.mc_txn = mc->mc_txn;
7157 mx->mx_cursor.mc_db = &mx->mx_db;
7158 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7159 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7160 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7161 mx->mx_cursor.mc_snum = 0;
7162 mx->mx_cursor.mc_top = 0;
7163 mx->mx_cursor.mc_flags = C_SUB;
7164 mx->mx_dbx.md_name.mv_size = 0;
7165 mx->mx_dbx.md_name.mv_data = NULL;
7166 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7167 mx->mx_dbx.md_dcmp = NULL;
7168 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7171 /** Final setup of a sorted-dups cursor.
7172 * Sets up the fields that depend on the data from the main cursor.
7173 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7174 * @param[in] node The data containing the #MDB_db record for the
7175 * sorted-dup database.
7178 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7180 MDB_xcursor *mx = mc->mc_xcursor;
7182 if (node->mn_flags & F_SUBDATA) {
7183 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7184 mx->mx_cursor.mc_pg[0] = 0;
7185 mx->mx_cursor.mc_snum = 0;
7186 mx->mx_cursor.mc_top = 0;
7187 mx->mx_cursor.mc_flags = C_SUB;
7189 MDB_page *fp = NODEDATA(node);
7190 mx->mx_db.md_pad = 0;
7191 mx->mx_db.md_flags = 0;
7192 mx->mx_db.md_depth = 1;
7193 mx->mx_db.md_branch_pages = 0;
7194 mx->mx_db.md_leaf_pages = 1;
7195 mx->mx_db.md_overflow_pages = 0;
7196 mx->mx_db.md_entries = NUMKEYS(fp);
7197 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7198 mx->mx_cursor.mc_snum = 1;
7199 mx->mx_cursor.mc_top = 0;
7200 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7201 mx->mx_cursor.mc_pg[0] = fp;
7202 mx->mx_cursor.mc_ki[0] = 0;
7203 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7204 mx->mx_db.md_flags = MDB_DUPFIXED;
7205 mx->mx_db.md_pad = fp->mp_pad;
7206 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7207 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7210 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7211 mx->mx_db.md_root));
7212 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7213 #if UINT_MAX < SIZE_MAX
7214 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7215 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7220 /** Fixup a sorted-dups cursor due to underlying update.
7221 * Sets up some fields that depend on the data from the main cursor.
7222 * Almost the same as init1, but skips initialization steps if the
7223 * xcursor had already been used.
7224 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7225 * @param[in] src_mx The xcursor of an up-to-date cursor.
7226 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7229 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7231 MDB_xcursor *mx = mc->mc_xcursor;
7234 mx->mx_cursor.mc_snum = 1;
7235 mx->mx_cursor.mc_top = 0;
7236 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7237 mx->mx_cursor.mc_ki[0] = 0;
7238 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7239 #if UINT_MAX < SIZE_MAX
7240 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7242 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7245 mx->mx_db = src_mx->mx_db;
7246 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7247 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7248 mx->mx_db.md_root));
7251 /** Initialize a cursor for a given transaction and database. */
7253 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7256 mc->mc_backup = NULL;
7259 mc->mc_db = &txn->mt_dbs[dbi];
7260 mc->mc_dbx = &txn->mt_dbxs[dbi];
7261 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7267 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7268 mdb_tassert(txn, mx != NULL);
7269 mc->mc_xcursor = mx;
7270 mdb_xcursor_init0(mc);
7272 mc->mc_xcursor = NULL;
7274 if (*mc->mc_dbflag & DB_STALE) {
7275 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7280 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7283 size_t size = sizeof(MDB_cursor);
7285 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7288 if (txn->mt_flags & MDB_TXN_ERROR)
7291 /* Allow read access to the freelist */
7292 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7295 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7296 size += sizeof(MDB_xcursor);
7298 if ((mc = malloc(size)) != NULL) {
7299 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7300 if (txn->mt_cursors) {
7301 mc->mc_next = txn->mt_cursors[dbi];
7302 txn->mt_cursors[dbi] = mc;
7303 mc->mc_flags |= C_UNTRACK;
7315 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7317 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7320 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7323 if (txn->mt_flags & MDB_TXN_ERROR)
7326 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7330 /* Return the count of duplicate data items for the current key */
7332 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7336 if (mc == NULL || countp == NULL)
7339 if (mc->mc_xcursor == NULL)
7340 return MDB_INCOMPATIBLE;
7342 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7345 if (!(mc->mc_flags & C_INITIALIZED))
7348 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7349 return MDB_NOTFOUND;
7351 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7352 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7355 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7358 *countp = mc->mc_xcursor->mx_db.md_entries;
7364 mdb_cursor_close(MDB_cursor *mc)
7366 if (mc && !mc->mc_backup) {
7367 /* remove from txn, if tracked */
7368 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7369 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7370 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7372 *prev = mc->mc_next;
7379 mdb_cursor_txn(MDB_cursor *mc)
7381 if (!mc) return NULL;
7386 mdb_cursor_dbi(MDB_cursor *mc)
7391 /** Replace the key for a branch node with a new key.
7392 * @param[in] mc Cursor pointing to the node to operate on.
7393 * @param[in] key The new key to use.
7394 * @return 0 on success, non-zero on failure.
7397 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7403 int delta, ksize, oksize;
7404 indx_t ptr, i, numkeys, indx;
7407 indx = mc->mc_ki[mc->mc_top];
7408 mp = mc->mc_pg[mc->mc_top];
7409 node = NODEPTR(mp, indx);
7410 ptr = mp->mp_ptrs[indx];
7414 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7415 k2.mv_data = NODEKEY(node);
7416 k2.mv_size = node->mn_ksize;
7417 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7419 mdb_dkey(&k2, kbuf2),
7425 /* Sizes must be 2-byte aligned. */
7426 ksize = EVEN(key->mv_size);
7427 oksize = EVEN(node->mn_ksize);
7428 delta = ksize - oksize;
7430 /* Shift node contents if EVEN(key length) changed. */
7432 if (delta > 0 && SIZELEFT(mp) < delta) {
7434 /* not enough space left, do a delete and split */
7435 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7436 pgno = NODEPGNO(node);
7437 mdb_node_del(mc, 0);
7438 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7441 numkeys = NUMKEYS(mp);
7442 for (i = 0; i < numkeys; i++) {
7443 if (mp->mp_ptrs[i] <= ptr)
7444 mp->mp_ptrs[i] -= delta;
7447 base = (char *)mp + mp->mp_upper + PAGEBASE;
7448 len = ptr - mp->mp_upper + NODESIZE;
7449 memmove(base - delta, base, len);
7450 mp->mp_upper -= delta;
7452 node = NODEPTR(mp, indx);
7455 /* But even if no shift was needed, update ksize */
7456 if (node->mn_ksize != key->mv_size)
7457 node->mn_ksize = key->mv_size;
7460 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7466 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7468 /** Move a node from csrc to cdst.
7471 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7478 unsigned short flags;
7482 /* Mark src and dst as dirty. */
7483 if ((rc = mdb_page_touch(csrc)) ||
7484 (rc = mdb_page_touch(cdst)))
7487 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7488 key.mv_size = csrc->mc_db->md_pad;
7489 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7491 data.mv_data = NULL;
7495 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7496 mdb_cassert(csrc, !((size_t)srcnode & 1));
7497 srcpg = NODEPGNO(srcnode);
7498 flags = srcnode->mn_flags;
7499 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7500 unsigned int snum = csrc->mc_snum;
7502 /* must find the lowest key below src */
7503 rc = mdb_page_search_lowest(csrc);
7506 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7507 key.mv_size = csrc->mc_db->md_pad;
7508 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7510 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7511 key.mv_size = NODEKSZ(s2);
7512 key.mv_data = NODEKEY(s2);
7514 csrc->mc_snum = snum--;
7515 csrc->mc_top = snum;
7517 key.mv_size = NODEKSZ(srcnode);
7518 key.mv_data = NODEKEY(srcnode);
7520 data.mv_size = NODEDSZ(srcnode);
7521 data.mv_data = NODEDATA(srcnode);
7523 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7524 unsigned int snum = cdst->mc_snum;
7527 /* must find the lowest key below dst */
7528 mdb_cursor_copy(cdst, &mn);
7529 rc = mdb_page_search_lowest(&mn);
7532 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7533 bkey.mv_size = mn.mc_db->md_pad;
7534 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7536 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7537 bkey.mv_size = NODEKSZ(s2);
7538 bkey.mv_data = NODEKEY(s2);
7540 mn.mc_snum = snum--;
7543 rc = mdb_update_key(&mn, &bkey);
7548 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7549 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7550 csrc->mc_ki[csrc->mc_top],
7552 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7553 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7555 /* Add the node to the destination page.
7557 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7558 if (rc != MDB_SUCCESS)
7561 /* Delete the node from the source page.
7563 mdb_node_del(csrc, key.mv_size);
7566 /* Adjust other cursors pointing to mp */
7567 MDB_cursor *m2, *m3;
7568 MDB_dbi dbi = csrc->mc_dbi;
7571 mp = cdst->mc_pg[csrc->mc_top];
7572 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7573 if (csrc->mc_flags & C_SUB)
7574 m3 = &m2->mc_xcursor->mx_cursor;
7577 if (m3 == cdst) continue;
7578 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] >=
7579 cdst->mc_ki[csrc->mc_top]) {
7580 m3->mc_ki[csrc->mc_top]++;
7584 mp = csrc->mc_pg[csrc->mc_top];
7585 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7586 if (csrc->mc_flags & C_SUB)
7587 m3 = &m2->mc_xcursor->mx_cursor;
7590 if (m3 == csrc) continue;
7591 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7592 csrc->mc_ki[csrc->mc_top]) {
7593 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7594 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7599 /* Update the parent separators.
7601 if (csrc->mc_ki[csrc->mc_top] == 0) {
7602 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7603 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7604 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7606 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7607 key.mv_size = NODEKSZ(srcnode);
7608 key.mv_data = NODEKEY(srcnode);
7610 DPRINTF(("update separator for source page %"Z"u to [%s]",
7611 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7612 mdb_cursor_copy(csrc, &mn);
7615 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7618 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7620 indx_t ix = csrc->mc_ki[csrc->mc_top];
7621 nullkey.mv_size = 0;
7622 csrc->mc_ki[csrc->mc_top] = 0;
7623 rc = mdb_update_key(csrc, &nullkey);
7624 csrc->mc_ki[csrc->mc_top] = ix;
7625 mdb_cassert(csrc, rc == MDB_SUCCESS);
7629 if (cdst->mc_ki[cdst->mc_top] == 0) {
7630 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7631 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7632 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7634 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7635 key.mv_size = NODEKSZ(srcnode);
7636 key.mv_data = NODEKEY(srcnode);
7638 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7639 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7640 mdb_cursor_copy(cdst, &mn);
7643 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7646 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7648 indx_t ix = cdst->mc_ki[cdst->mc_top];
7649 nullkey.mv_size = 0;
7650 cdst->mc_ki[cdst->mc_top] = 0;
7651 rc = mdb_update_key(cdst, &nullkey);
7652 cdst->mc_ki[cdst->mc_top] = ix;
7653 mdb_cassert(cdst, rc == MDB_SUCCESS);
7660 /** Merge one page into another.
7661 * The nodes from the page pointed to by \b csrc will
7662 * be copied to the page pointed to by \b cdst and then
7663 * the \b csrc page will be freed.
7664 * @param[in] csrc Cursor pointing to the source page.
7665 * @param[in] cdst Cursor pointing to the destination page.
7666 * @return 0 on success, non-zero on failure.
7669 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7671 MDB_page *psrc, *pdst;
7678 psrc = csrc->mc_pg[csrc->mc_top];
7679 pdst = cdst->mc_pg[cdst->mc_top];
7681 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7683 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7684 mdb_cassert(csrc, cdst->mc_snum > 1);
7686 /* Mark dst as dirty. */
7687 if ((rc = mdb_page_touch(cdst)))
7690 /* Move all nodes from src to dst.
7692 j = nkeys = NUMKEYS(pdst);
7693 if (IS_LEAF2(psrc)) {
7694 key.mv_size = csrc->mc_db->md_pad;
7695 key.mv_data = METADATA(psrc);
7696 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7697 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7698 if (rc != MDB_SUCCESS)
7700 key.mv_data = (char *)key.mv_data + key.mv_size;
7703 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7704 srcnode = NODEPTR(psrc, i);
7705 if (i == 0 && IS_BRANCH(psrc)) {
7708 mdb_cursor_copy(csrc, &mn);
7709 /* must find the lowest key below src */
7710 rc = mdb_page_search_lowest(&mn);
7713 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7714 key.mv_size = mn.mc_db->md_pad;
7715 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7717 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7718 key.mv_size = NODEKSZ(s2);
7719 key.mv_data = NODEKEY(s2);
7722 key.mv_size = srcnode->mn_ksize;
7723 key.mv_data = NODEKEY(srcnode);
7726 data.mv_size = NODEDSZ(srcnode);
7727 data.mv_data = NODEDATA(srcnode);
7728 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7729 if (rc != MDB_SUCCESS)
7734 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7735 pdst->mp_pgno, NUMKEYS(pdst),
7736 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7738 /* Unlink the src page from parent and add to free list.
7741 mdb_node_del(csrc, 0);
7742 if (csrc->mc_ki[csrc->mc_top] == 0) {
7744 rc = mdb_update_key(csrc, &key);
7752 psrc = csrc->mc_pg[csrc->mc_top];
7753 /* If not operating on FreeDB, allow this page to be reused
7754 * in this txn. Otherwise just add to free list.
7756 rc = mdb_page_loose(csrc, psrc);
7760 csrc->mc_db->md_leaf_pages--;
7762 csrc->mc_db->md_branch_pages--;
7764 /* Adjust other cursors pointing to mp */
7765 MDB_cursor *m2, *m3;
7766 MDB_dbi dbi = csrc->mc_dbi;
7768 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7769 if (csrc->mc_flags & C_SUB)
7770 m3 = &m2->mc_xcursor->mx_cursor;
7773 if (m3 == csrc) continue;
7774 if (m3->mc_snum < csrc->mc_snum) continue;
7775 if (m3->mc_pg[csrc->mc_top] == psrc) {
7776 m3->mc_pg[csrc->mc_top] = pdst;
7777 m3->mc_ki[csrc->mc_top] += nkeys;
7782 unsigned int snum = cdst->mc_snum;
7783 uint16_t depth = cdst->mc_db->md_depth;
7784 mdb_cursor_pop(cdst);
7785 rc = mdb_rebalance(cdst);
7786 /* Did the tree height change? */
7787 if (depth != cdst->mc_db->md_depth)
7788 snum += cdst->mc_db->md_depth - depth;
7789 cdst->mc_snum = snum;
7790 cdst->mc_top = snum-1;
7795 /** Copy the contents of a cursor.
7796 * @param[in] csrc The cursor to copy from.
7797 * @param[out] cdst The cursor to copy to.
7800 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7804 cdst->mc_txn = csrc->mc_txn;
7805 cdst->mc_dbi = csrc->mc_dbi;
7806 cdst->mc_db = csrc->mc_db;
7807 cdst->mc_dbx = csrc->mc_dbx;
7808 cdst->mc_snum = csrc->mc_snum;
7809 cdst->mc_top = csrc->mc_top;
7810 cdst->mc_flags = csrc->mc_flags;
7812 for (i=0; i<csrc->mc_snum; i++) {
7813 cdst->mc_pg[i] = csrc->mc_pg[i];
7814 cdst->mc_ki[i] = csrc->mc_ki[i];
7818 /** Rebalance the tree after a delete operation.
7819 * @param[in] mc Cursor pointing to the page where rebalancing
7821 * @return 0 on success, non-zero on failure.
7824 mdb_rebalance(MDB_cursor *mc)
7828 unsigned int ptop, minkeys, thresh;
7832 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7837 thresh = FILL_THRESHOLD;
7839 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7840 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7841 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7842 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7844 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7845 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7846 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7847 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7851 if (mc->mc_snum < 2) {
7852 MDB_page *mp = mc->mc_pg[0];
7854 DPUTS("Can't rebalance a subpage, ignoring");
7857 if (NUMKEYS(mp) == 0) {
7858 DPUTS("tree is completely empty");
7859 mc->mc_db->md_root = P_INVALID;
7860 mc->mc_db->md_depth = 0;
7861 mc->mc_db->md_leaf_pages = 0;
7862 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7865 /* Adjust cursors pointing to mp */
7868 mc->mc_flags &= ~C_INITIALIZED;
7870 MDB_cursor *m2, *m3;
7871 MDB_dbi dbi = mc->mc_dbi;
7873 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7874 if (mc->mc_flags & C_SUB)
7875 m3 = &m2->mc_xcursor->mx_cursor;
7878 if (m3->mc_snum < mc->mc_snum) continue;
7879 if (m3->mc_pg[0] == mp) {
7882 m3->mc_flags &= ~C_INITIALIZED;
7886 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7888 DPUTS("collapsing root page!");
7889 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7892 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7893 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7896 mc->mc_db->md_depth--;
7897 mc->mc_db->md_branch_pages--;
7898 mc->mc_ki[0] = mc->mc_ki[1];
7899 for (i = 1; i<mc->mc_db->md_depth; i++) {
7900 mc->mc_pg[i] = mc->mc_pg[i+1];
7901 mc->mc_ki[i] = mc->mc_ki[i+1];
7904 /* Adjust other cursors pointing to mp */
7905 MDB_cursor *m2, *m3;
7906 MDB_dbi dbi = mc->mc_dbi;
7908 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7909 if (mc->mc_flags & C_SUB)
7910 m3 = &m2->mc_xcursor->mx_cursor;
7913 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7914 if (m3->mc_pg[0] == mp) {
7915 for (i=0; i<m3->mc_snum; i++) {
7916 m3->mc_pg[i] = m3->mc_pg[i+1];
7917 m3->mc_ki[i] = m3->mc_ki[i+1];
7925 DPUTS("root page doesn't need rebalancing");
7929 /* The parent (branch page) must have at least 2 pointers,
7930 * otherwise the tree is invalid.
7932 ptop = mc->mc_top-1;
7933 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7935 /* Leaf page fill factor is below the threshold.
7936 * Try to move keys from left or right neighbor, or
7937 * merge with a neighbor page.
7942 mdb_cursor_copy(mc, &mn);
7943 mn.mc_xcursor = NULL;
7945 oldki = mc->mc_ki[mc->mc_top];
7946 if (mc->mc_ki[ptop] == 0) {
7947 /* We're the leftmost leaf in our parent.
7949 DPUTS("reading right neighbor");
7951 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7952 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7955 mn.mc_ki[mn.mc_top] = 0;
7956 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7958 /* There is at least one neighbor to the left.
7960 DPUTS("reading left neighbor");
7962 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7963 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7966 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7967 mc->mc_ki[mc->mc_top] = 0;
7970 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7971 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7972 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7974 /* If the neighbor page is above threshold and has enough keys,
7975 * move one key from it. Otherwise we should try to merge them.
7976 * (A branch page must never have less than 2 keys.)
7978 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7979 rc = mdb_node_move(&mn, mc);
7980 if (mc->mc_ki[mc->mc_top-1]) {
7984 if (mc->mc_ki[ptop] == 0) {
7985 rc = mdb_page_merge(&mn, mc);
7988 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7989 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7990 /* We want mdb_rebalance to find mn when doing fixups */
7991 if (mc->mc_flags & C_SUB) {
7992 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7993 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
7994 dummy.mc_xcursor = (MDB_xcursor *)&mn;
7996 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7997 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
7999 rc = mdb_page_merge(mc, &mn);
8000 if (mc->mc_flags & C_SUB)
8001 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8003 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8004 mdb_cursor_copy(&mn, mc);
8006 mc->mc_flags &= ~C_EOF;
8008 mc->mc_ki[mc->mc_top] = oldki;
8012 /** Complete a delete operation started by #mdb_cursor_del(). */
8014 mdb_cursor_del0(MDB_cursor *mc)
8020 MDB_cursor *m2, *m3;
8021 MDB_dbi dbi = mc->mc_dbi;
8023 ki = mc->mc_ki[mc->mc_top];
8024 mp = mc->mc_pg[mc->mc_top];
8025 mdb_node_del(mc, mc->mc_db->md_pad);
8026 mc->mc_db->md_entries--;
8028 /* Adjust other cursors pointing to mp */
8029 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8030 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8031 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8033 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8035 if (m3->mc_pg[mc->mc_top] == mp) {
8036 if (m3->mc_ki[mc->mc_top] >= ki) {
8037 m3->mc_flags |= C_DEL;
8038 if (m3->mc_ki[mc->mc_top] > ki)
8039 m3->mc_ki[mc->mc_top]--;
8040 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8041 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8046 rc = mdb_rebalance(mc);
8048 if (rc == MDB_SUCCESS) {
8049 /* DB is totally empty now, just bail out.
8050 * Other cursors adjustments were already done
8051 * by mdb_rebalance and aren't needed here.
8056 mp = mc->mc_pg[mc->mc_top];
8057 nkeys = NUMKEYS(mp);
8059 /* Adjust other cursors pointing to mp */
8060 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8061 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8062 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8064 if (m3->mc_snum < mc->mc_snum)
8066 if (m3->mc_pg[mc->mc_top] == mp) {
8067 /* if m3 points past last node in page, find next sibling */
8068 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8069 rc = mdb_cursor_sibling(m3, 1);
8070 if (rc == MDB_NOTFOUND) {
8071 m3->mc_flags |= C_EOF;
8077 mc->mc_flags |= C_DEL;
8081 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8086 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8087 MDB_val *key, MDB_val *data)
8089 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8092 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8093 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8095 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8096 /* must ignore any data */
8100 return mdb_del0(txn, dbi, key, data, 0);
8104 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8105 MDB_val *key, MDB_val *data, unsigned flags)
8110 MDB_val rdata, *xdata;
8114 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8116 mdb_cursor_init(&mc, txn, dbi, &mx);
8125 flags |= MDB_NODUPDATA;
8127 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8129 /* let mdb_page_split know about this cursor if needed:
8130 * delete will trigger a rebalance; if it needs to move
8131 * a node from one page to another, it will have to
8132 * update the parent's separator key(s). If the new sepkey
8133 * is larger than the current one, the parent page may
8134 * run out of space, triggering a split. We need this
8135 * cursor to be consistent until the end of the rebalance.
8137 mc.mc_flags |= C_UNTRACK;
8138 mc.mc_next = txn->mt_cursors[dbi];
8139 txn->mt_cursors[dbi] = &mc;
8140 rc = mdb_cursor_del(&mc, flags);
8141 txn->mt_cursors[dbi] = mc.mc_next;
8146 /** Split a page and insert a new node.
8147 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8148 * The cursor will be updated to point to the actual page and index where
8149 * the node got inserted after the split.
8150 * @param[in] newkey The key for the newly inserted node.
8151 * @param[in] newdata The data for the newly inserted node.
8152 * @param[in] newpgno The page number, if the new node is a branch node.
8153 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8154 * @return 0 on success, non-zero on failure.
8157 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8158 unsigned int nflags)
8161 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8164 int i, j, split_indx, nkeys, pmax;
8165 MDB_env *env = mc->mc_txn->mt_env;
8167 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8168 MDB_page *copy = NULL;
8169 MDB_page *mp, *rp, *pp;
8174 mp = mc->mc_pg[mc->mc_top];
8175 newindx = mc->mc_ki[mc->mc_top];
8176 nkeys = NUMKEYS(mp);
8178 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8179 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8180 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8182 /* Create a right sibling. */
8183 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8185 rp->mp_pad = mp->mp_pad;
8186 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8188 if (mc->mc_snum < 2) {
8189 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8191 /* shift current top to make room for new parent */
8192 mc->mc_pg[1] = mc->mc_pg[0];
8193 mc->mc_ki[1] = mc->mc_ki[0];
8196 mc->mc_db->md_root = pp->mp_pgno;
8197 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8198 new_root = mc->mc_db->md_depth++;
8200 /* Add left (implicit) pointer. */
8201 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8202 /* undo the pre-push */
8203 mc->mc_pg[0] = mc->mc_pg[1];
8204 mc->mc_ki[0] = mc->mc_ki[1];
8205 mc->mc_db->md_root = mp->mp_pgno;
8206 mc->mc_db->md_depth--;
8213 ptop = mc->mc_top-1;
8214 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8217 mc->mc_flags |= C_SPLITTING;
8218 mdb_cursor_copy(mc, &mn);
8219 mn.mc_pg[mn.mc_top] = rp;
8220 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8222 if (nflags & MDB_APPEND) {
8223 mn.mc_ki[mn.mc_top] = 0;
8225 split_indx = newindx;
8229 split_indx = (nkeys+1) / 2;
8234 unsigned int lsize, rsize, ksize;
8235 /* Move half of the keys to the right sibling */
8236 x = mc->mc_ki[mc->mc_top] - split_indx;
8237 ksize = mc->mc_db->md_pad;
8238 split = LEAF2KEY(mp, split_indx, ksize);
8239 rsize = (nkeys - split_indx) * ksize;
8240 lsize = (nkeys - split_indx) * sizeof(indx_t);
8241 mp->mp_lower -= lsize;
8242 rp->mp_lower += lsize;
8243 mp->mp_upper += rsize - lsize;
8244 rp->mp_upper -= rsize - lsize;
8245 sepkey.mv_size = ksize;
8246 if (newindx == split_indx) {
8247 sepkey.mv_data = newkey->mv_data;
8249 sepkey.mv_data = split;
8252 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8253 memcpy(rp->mp_ptrs, split, rsize);
8254 sepkey.mv_data = rp->mp_ptrs;
8255 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8256 memcpy(ins, newkey->mv_data, ksize);
8257 mp->mp_lower += sizeof(indx_t);
8258 mp->mp_upper -= ksize - sizeof(indx_t);
8261 memcpy(rp->mp_ptrs, split, x * ksize);
8262 ins = LEAF2KEY(rp, x, ksize);
8263 memcpy(ins, newkey->mv_data, ksize);
8264 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8265 rp->mp_lower += sizeof(indx_t);
8266 rp->mp_upper -= ksize - sizeof(indx_t);
8267 mc->mc_ki[mc->mc_top] = x;
8268 mc->mc_pg[mc->mc_top] = rp;
8271 int psize, nsize, k;
8272 /* Maximum free space in an empty page */
8273 pmax = env->me_psize - PAGEHDRSZ;
8275 nsize = mdb_leaf_size(env, newkey, newdata);
8277 nsize = mdb_branch_size(env, newkey);
8278 nsize = EVEN(nsize);
8280 /* grab a page to hold a temporary copy */
8281 copy = mdb_page_malloc(mc->mc_txn, 1);
8286 copy->mp_pgno = mp->mp_pgno;
8287 copy->mp_flags = mp->mp_flags;
8288 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8289 copy->mp_upper = env->me_psize - PAGEBASE;
8291 /* prepare to insert */
8292 for (i=0, j=0; i<nkeys; i++) {
8294 copy->mp_ptrs[j++] = 0;
8296 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8299 /* When items are relatively large the split point needs
8300 * to be checked, because being off-by-one will make the
8301 * difference between success or failure in mdb_node_add.
8303 * It's also relevant if a page happens to be laid out
8304 * such that one half of its nodes are all "small" and
8305 * the other half of its nodes are "large." If the new
8306 * item is also "large" and falls on the half with
8307 * "large" nodes, it also may not fit.
8309 * As a final tweak, if the new item goes on the last
8310 * spot on the page (and thus, onto the new page), bias
8311 * the split so the new page is emptier than the old page.
8312 * This yields better packing during sequential inserts.
8314 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8315 /* Find split point */
8317 if (newindx <= split_indx || newindx >= nkeys) {
8319 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8324 for (; i!=k; i+=j) {
8329 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8330 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8332 if (F_ISSET(node->mn_flags, F_BIGDATA))
8333 psize += sizeof(pgno_t);
8335 psize += NODEDSZ(node);
8337 psize = EVEN(psize);
8339 if (psize > pmax || i == k-j) {
8340 split_indx = i + (j<0);
8345 if (split_indx == newindx) {
8346 sepkey.mv_size = newkey->mv_size;
8347 sepkey.mv_data = newkey->mv_data;
8349 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8350 sepkey.mv_size = node->mn_ksize;
8351 sepkey.mv_data = NODEKEY(node);
8356 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8358 /* Copy separator key to the parent.
8360 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8364 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8369 if (mn.mc_snum == mc->mc_snum) {
8370 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8371 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8372 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8373 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8378 /* Right page might now have changed parent.
8379 * Check if left page also changed parent.
8381 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8382 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8383 for (i=0; i<ptop; i++) {
8384 mc->mc_pg[i] = mn.mc_pg[i];
8385 mc->mc_ki[i] = mn.mc_ki[i];
8387 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8388 if (mn.mc_ki[ptop]) {
8389 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8391 /* find right page's left sibling */
8392 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8393 mdb_cursor_sibling(mc, 0);
8398 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8401 mc->mc_flags ^= C_SPLITTING;
8402 if (rc != MDB_SUCCESS) {
8405 if (nflags & MDB_APPEND) {
8406 mc->mc_pg[mc->mc_top] = rp;
8407 mc->mc_ki[mc->mc_top] = 0;
8408 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8411 for (i=0; i<mc->mc_top; i++)
8412 mc->mc_ki[i] = mn.mc_ki[i];
8413 } else if (!IS_LEAF2(mp)) {
8415 mc->mc_pg[mc->mc_top] = rp;
8420 rkey.mv_data = newkey->mv_data;
8421 rkey.mv_size = newkey->mv_size;
8427 /* Update index for the new key. */
8428 mc->mc_ki[mc->mc_top] = j;
8430 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8431 rkey.mv_data = NODEKEY(node);
8432 rkey.mv_size = node->mn_ksize;
8434 xdata.mv_data = NODEDATA(node);
8435 xdata.mv_size = NODEDSZ(node);
8438 pgno = NODEPGNO(node);
8439 flags = node->mn_flags;
8442 if (!IS_LEAF(mp) && j == 0) {
8443 /* First branch index doesn't need key data. */
8447 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8453 mc->mc_pg[mc->mc_top] = copy;
8458 } while (i != split_indx);
8460 nkeys = NUMKEYS(copy);
8461 for (i=0; i<nkeys; i++)
8462 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8463 mp->mp_lower = copy->mp_lower;
8464 mp->mp_upper = copy->mp_upper;
8465 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8466 env->me_psize - copy->mp_upper - PAGEBASE);
8468 /* reset back to original page */
8469 if (newindx < split_indx) {
8470 mc->mc_pg[mc->mc_top] = mp;
8471 if (nflags & MDB_RESERVE) {
8472 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8473 if (!(node->mn_flags & F_BIGDATA))
8474 newdata->mv_data = NODEDATA(node);
8477 mc->mc_pg[mc->mc_top] = rp;
8479 /* Make sure mc_ki is still valid.
8481 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8482 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8483 for (i=0; i<=ptop; i++) {
8484 mc->mc_pg[i] = mn.mc_pg[i];
8485 mc->mc_ki[i] = mn.mc_ki[i];
8492 /* Adjust other cursors pointing to mp */
8493 MDB_cursor *m2, *m3;
8494 MDB_dbi dbi = mc->mc_dbi;
8495 int fixup = NUMKEYS(mp);
8497 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8498 if (mc->mc_flags & C_SUB)
8499 m3 = &m2->mc_xcursor->mx_cursor;
8504 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8506 if (m3->mc_flags & C_SPLITTING)
8511 for (k=new_root; k>=0; k--) {
8512 m3->mc_ki[k+1] = m3->mc_ki[k];
8513 m3->mc_pg[k+1] = m3->mc_pg[k];
8515 if (m3->mc_ki[0] >= split_indx) {
8520 m3->mc_pg[0] = mc->mc_pg[0];
8524 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8525 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8526 m3->mc_ki[mc->mc_top]++;
8527 if (m3->mc_ki[mc->mc_top] >= fixup) {
8528 m3->mc_pg[mc->mc_top] = rp;
8529 m3->mc_ki[mc->mc_top] -= fixup;
8530 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8532 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8533 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8538 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8541 if (copy) /* tmp page */
8542 mdb_page_free(env, copy);
8544 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8549 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8550 MDB_val *key, MDB_val *data, unsigned int flags)
8555 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8558 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8561 mdb_cursor_init(&mc, txn, dbi, &mx);
8562 return mdb_cursor_put(&mc, key, data, flags);
8566 #define MDB_WBUF (1024*1024)
8569 /** State needed for a compacting copy. */
8570 typedef struct mdb_copy {
8571 pthread_mutex_t mc_mutex;
8572 pthread_cond_t mc_cond;
8579 pgno_t mc_next_pgno;
8582 volatile int mc_new;
8587 /** Dedicated writer thread for compacting copy. */
8588 static THREAD_RET ESECT CALL_CONV
8589 mdb_env_copythr(void *arg)
8593 int toggle = 0, wsize, rc;
8596 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8599 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8602 pthread_mutex_lock(&my->mc_mutex);
8604 pthread_cond_signal(&my->mc_cond);
8607 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8608 if (my->mc_new < 0) {
8613 wsize = my->mc_wlen[toggle];
8614 ptr = my->mc_wbuf[toggle];
8617 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8621 } else if (len > 0) {
8635 /* If there's an overflow page tail, write it too */
8636 if (my->mc_olen[toggle]) {
8637 wsize = my->mc_olen[toggle];
8638 ptr = my->mc_over[toggle];
8639 my->mc_olen[toggle] = 0;
8642 my->mc_wlen[toggle] = 0;
8644 pthread_cond_signal(&my->mc_cond);
8646 pthread_cond_signal(&my->mc_cond);
8647 pthread_mutex_unlock(&my->mc_mutex);
8648 return (THREAD_RET)0;
8652 /** Tell the writer thread there's a buffer ready to write */
8654 mdb_env_cthr_toggle(mdb_copy *my, int st)
8656 int toggle = my->mc_toggle ^ 1;
8657 pthread_mutex_lock(&my->mc_mutex);
8658 if (my->mc_status) {
8659 pthread_mutex_unlock(&my->mc_mutex);
8660 return my->mc_status;
8662 while (my->mc_new == 1)
8663 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8665 my->mc_toggle = toggle;
8666 pthread_cond_signal(&my->mc_cond);
8667 pthread_mutex_unlock(&my->mc_mutex);
8671 /** Depth-first tree traversal for compacting copy. */
8673 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8676 MDB_txn *txn = my->mc_txn;
8678 MDB_page *mo, *mp, *leaf;
8683 /* Empty DB, nothing to do */
8684 if (*pg == P_INVALID)
8691 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8694 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8698 /* Make cursor pages writable */
8699 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8703 for (i=0; i<mc.mc_top; i++) {
8704 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8705 mc.mc_pg[i] = (MDB_page *)ptr;
8706 ptr += my->mc_env->me_psize;
8709 /* This is writable space for a leaf page. Usually not needed. */
8710 leaf = (MDB_page *)ptr;
8712 toggle = my->mc_toggle;
8713 while (mc.mc_snum > 0) {
8715 mp = mc.mc_pg[mc.mc_top];
8719 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8720 for (i=0; i<n; i++) {
8721 ni = NODEPTR(mp, i);
8722 if (ni->mn_flags & F_BIGDATA) {
8726 /* Need writable leaf */
8728 mc.mc_pg[mc.mc_top] = leaf;
8729 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8731 ni = NODEPTR(mp, i);
8734 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8735 rc = mdb_page_get(txn, pg, &omp, NULL);
8738 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8739 rc = mdb_env_cthr_toggle(my, 1);
8742 toggle = my->mc_toggle;
8744 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8745 memcpy(mo, omp, my->mc_env->me_psize);
8746 mo->mp_pgno = my->mc_next_pgno;
8747 my->mc_next_pgno += omp->mp_pages;
8748 my->mc_wlen[toggle] += my->mc_env->me_psize;
8749 if (omp->mp_pages > 1) {
8750 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8751 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8752 rc = mdb_env_cthr_toggle(my, 1);
8755 toggle = my->mc_toggle;
8757 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8758 } else if (ni->mn_flags & F_SUBDATA) {
8761 /* Need writable leaf */
8763 mc.mc_pg[mc.mc_top] = leaf;
8764 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8766 ni = NODEPTR(mp, i);
8769 memcpy(&db, NODEDATA(ni), sizeof(db));
8770 my->mc_toggle = toggle;
8771 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8774 toggle = my->mc_toggle;
8775 memcpy(NODEDATA(ni), &db, sizeof(db));
8780 mc.mc_ki[mc.mc_top]++;
8781 if (mc.mc_ki[mc.mc_top] < n) {
8784 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8786 rc = mdb_page_get(txn, pg, &mp, NULL);
8791 mc.mc_ki[mc.mc_top] = 0;
8792 if (IS_BRANCH(mp)) {
8793 /* Whenever we advance to a sibling branch page,
8794 * we must proceed all the way down to its first leaf.
8796 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8799 mc.mc_pg[mc.mc_top] = mp;
8803 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8804 rc = mdb_env_cthr_toggle(my, 1);
8807 toggle = my->mc_toggle;
8809 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8810 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8811 mo->mp_pgno = my->mc_next_pgno++;
8812 my->mc_wlen[toggle] += my->mc_env->me_psize;
8814 /* Update parent if there is one */
8815 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8816 SETPGNO(ni, mo->mp_pgno);
8817 mdb_cursor_pop(&mc);
8819 /* Otherwise we're done */
8829 /** Copy environment with compaction. */
8831 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8836 MDB_txn *txn = NULL;
8841 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8842 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8843 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8844 if (my.mc_wbuf[0] == NULL)
8847 pthread_mutex_init(&my.mc_mutex, NULL);
8848 pthread_cond_init(&my.mc_cond, NULL);
8849 #ifdef HAVE_MEMALIGN
8850 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8851 if (my.mc_wbuf[0] == NULL)
8854 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8859 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8860 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8865 my.mc_next_pgno = 2;
8871 THREAD_CREATE(thr, mdb_env_copythr, &my);
8873 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8877 mp = (MDB_page *)my.mc_wbuf[0];
8878 memset(mp, 0, 2*env->me_psize);
8880 mp->mp_flags = P_META;
8881 mm = (MDB_meta *)METADATA(mp);
8882 mdb_env_init_meta0(env, mm);
8883 mm->mm_address = env->me_metas[0]->mm_address;
8885 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8887 mp->mp_flags = P_META;
8888 *(MDB_meta *)METADATA(mp) = *mm;
8889 mm = (MDB_meta *)METADATA(mp);
8891 /* Count the number of free pages, subtract from lastpg to find
8892 * number of active pages
8895 MDB_ID freecount = 0;
8898 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8899 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8900 freecount += *(MDB_ID *)data.mv_data;
8901 freecount += txn->mt_dbs[0].md_branch_pages +
8902 txn->mt_dbs[0].md_leaf_pages +
8903 txn->mt_dbs[0].md_overflow_pages;
8905 /* Set metapage 1 */
8906 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8907 mm->mm_dbs[1] = txn->mt_dbs[1];
8908 if (mm->mm_last_pg > 1) {
8909 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8912 mm->mm_dbs[1].md_root = P_INVALID;
8915 my.mc_wlen[0] = env->me_psize * 2;
8917 pthread_mutex_lock(&my.mc_mutex);
8919 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8920 pthread_mutex_unlock(&my.mc_mutex);
8921 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8922 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8923 rc = mdb_env_cthr_toggle(&my, 1);
8924 mdb_env_cthr_toggle(&my, -1);
8925 pthread_mutex_lock(&my.mc_mutex);
8927 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8928 pthread_mutex_unlock(&my.mc_mutex);
8933 CloseHandle(my.mc_cond);
8934 CloseHandle(my.mc_mutex);
8935 _aligned_free(my.mc_wbuf[0]);
8937 pthread_cond_destroy(&my.mc_cond);
8938 pthread_mutex_destroy(&my.mc_mutex);
8939 free(my.mc_wbuf[0]);
8944 /** Copy environment as-is. */
8946 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8948 MDB_txn *txn = NULL;
8949 mdb_mutexref_t wmutex = NULL;
8955 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8959 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8962 /* Do the lock/unlock of the reader mutex before starting the
8963 * write txn. Otherwise other read txns could block writers.
8965 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8970 /* We must start the actual read txn after blocking writers */
8971 mdb_txn_reset0(txn, "reset-stage1");
8973 /* Temporarily block writers until we snapshot the meta pages */
8974 wmutex = env->me_wmutex;
8975 if (LOCK_MUTEX(rc, env, wmutex))
8978 rc = mdb_txn_renew0(txn);
8980 UNLOCK_MUTEX(wmutex);
8985 wsize = env->me_psize * 2;
8989 DO_WRITE(rc, fd, ptr, w2, len);
8993 } else if (len > 0) {
8999 /* Non-blocking or async handles are not supported */
9005 UNLOCK_MUTEX(wmutex);
9010 w2 = txn->mt_next_pgno * env->me_psize;
9013 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9020 if (wsize > MAX_WRITE)
9024 DO_WRITE(rc, fd, ptr, w2, len);
9028 } else if (len > 0) {
9045 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9047 if (flags & MDB_CP_COMPACT)
9048 return mdb_env_copyfd1(env, fd);
9050 return mdb_env_copyfd0(env, fd);
9054 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9056 return mdb_env_copyfd2(env, fd, 0);
9060 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9064 HANDLE newfd = INVALID_HANDLE_VALUE;
9066 if (env->me_flags & MDB_NOSUBDIR) {
9067 lpath = (char *)path;
9070 len += sizeof(DATANAME);
9071 lpath = malloc(len);
9074 sprintf(lpath, "%s" DATANAME, path);
9077 /* The destination path must exist, but the destination file must not.
9078 * We don't want the OS to cache the writes, since the source data is
9079 * already in the OS cache.
9082 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9083 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9085 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9087 if (newfd == INVALID_HANDLE_VALUE) {
9092 if (env->me_psize >= env->me_os_psize) {
9094 /* Set O_DIRECT if the file system supports it */
9095 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9096 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9098 #ifdef F_NOCACHE /* __APPLE__ */
9099 rc = fcntl(newfd, F_NOCACHE, 1);
9107 rc = mdb_env_copyfd2(env, newfd, flags);
9110 if (!(env->me_flags & MDB_NOSUBDIR))
9112 if (newfd != INVALID_HANDLE_VALUE)
9113 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9120 mdb_env_copy(MDB_env *env, const char *path)
9122 return mdb_env_copy2(env, path, 0);
9126 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9128 if (flag & ~CHANGEABLE)
9131 env->me_flags |= flag;
9133 env->me_flags &= ~flag;
9138 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9143 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9148 mdb_env_set_userctx(MDB_env *env, void *ctx)
9152 env->me_userctx = ctx;
9157 mdb_env_get_userctx(MDB_env *env)
9159 return env ? env->me_userctx : NULL;
9163 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9168 env->me_assert_func = func;
9174 mdb_env_get_path(MDB_env *env, const char **arg)
9179 *arg = env->me_path;
9184 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9193 /** Common code for #mdb_stat() and #mdb_env_stat().
9194 * @param[in] env the environment to operate in.
9195 * @param[in] db the #MDB_db record containing the stats to return.
9196 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9197 * @return 0, this function always succeeds.
9200 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9202 arg->ms_psize = env->me_psize;
9203 arg->ms_depth = db->md_depth;
9204 arg->ms_branch_pages = db->md_branch_pages;
9205 arg->ms_leaf_pages = db->md_leaf_pages;
9206 arg->ms_overflow_pages = db->md_overflow_pages;
9207 arg->ms_entries = db->md_entries;
9213 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9217 if (env == NULL || arg == NULL)
9220 toggle = mdb_env_pick_meta(env);
9222 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9226 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9230 if (env == NULL || arg == NULL)
9233 toggle = mdb_env_pick_meta(env);
9234 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9235 arg->me_mapsize = env->me_mapsize;
9236 arg->me_maxreaders = env->me_maxreaders;
9237 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9239 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9240 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9244 /** Set the default comparison functions for a database.
9245 * Called immediately after a database is opened to set the defaults.
9246 * The user can then override them with #mdb_set_compare() or
9247 * #mdb_set_dupsort().
9248 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9249 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9252 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9254 uint16_t f = txn->mt_dbs[dbi].md_flags;
9256 txn->mt_dbxs[dbi].md_cmp =
9257 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9258 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9260 txn->mt_dbxs[dbi].md_dcmp =
9261 !(f & MDB_DUPSORT) ? 0 :
9262 ((f & MDB_INTEGERDUP)
9263 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9264 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9267 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9273 int rc, dbflag, exact;
9274 unsigned int unused = 0, seq;
9277 if (flags & ~VALID_FLAGS)
9279 if (txn->mt_flags & MDB_TXN_ERROR)
9285 if (flags & PERSISTENT_FLAGS) {
9286 uint16_t f2 = flags & PERSISTENT_FLAGS;
9287 /* make sure flag changes get committed */
9288 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9289 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9290 txn->mt_flags |= MDB_TXN_DIRTY;
9293 mdb_default_cmp(txn, MAIN_DBI);
9297 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9298 mdb_default_cmp(txn, MAIN_DBI);
9301 /* Is the DB already open? */
9303 for (i=2; i<txn->mt_numdbs; i++) {
9304 if (!txn->mt_dbxs[i].md_name.mv_size) {
9305 /* Remember this free slot */
9306 if (!unused) unused = i;
9309 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9310 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9316 /* If no free slot and max hit, fail */
9317 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9318 return MDB_DBS_FULL;
9320 /* Cannot mix named databases with some mainDB flags */
9321 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9322 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9324 /* Find the DB info */
9325 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9328 key.mv_data = (void *)name;
9329 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9330 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9331 if (rc == MDB_SUCCESS) {
9332 /* make sure this is actually a DB */
9333 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9334 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9335 return MDB_INCOMPATIBLE;
9336 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9337 /* Create if requested */
9338 data.mv_size = sizeof(MDB_db);
9339 data.mv_data = &dummy;
9340 memset(&dummy, 0, sizeof(dummy));
9341 dummy.md_root = P_INVALID;
9342 dummy.md_flags = flags & PERSISTENT_FLAGS;
9343 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9347 /* OK, got info, add to table */
9348 if (rc == MDB_SUCCESS) {
9349 unsigned int slot = unused ? unused : txn->mt_numdbs;
9350 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9351 txn->mt_dbxs[slot].md_name.mv_size = len;
9352 txn->mt_dbxs[slot].md_rel = NULL;
9353 txn->mt_dbflags[slot] = dbflag;
9354 /* txn-> and env-> are the same in read txns, use
9355 * tmp variable to avoid undefined assignment
9357 seq = ++txn->mt_env->me_dbiseqs[slot];
9358 txn->mt_dbiseqs[slot] = seq;
9360 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9362 mdb_default_cmp(txn, slot);
9371 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9373 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9376 if (txn->mt_flags & MDB_TXN_ERROR)
9379 if (txn->mt_dbflags[dbi] & DB_STALE) {
9382 /* Stale, must read the DB's root. cursor_init does it for us. */
9383 mdb_cursor_init(&mc, txn, dbi, &mx);
9385 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9388 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9391 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9393 ptr = env->me_dbxs[dbi].md_name.mv_data;
9394 /* If there was no name, this was already closed */
9396 env->me_dbxs[dbi].md_name.mv_data = NULL;
9397 env->me_dbxs[dbi].md_name.mv_size = 0;
9398 env->me_dbflags[dbi] = 0;
9399 env->me_dbiseqs[dbi]++;
9404 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9406 /* We could return the flags for the FREE_DBI too but what's the point? */
9407 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9409 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9413 /** Add all the DB's pages to the free list.
9414 * @param[in] mc Cursor on the DB to free.
9415 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9416 * @return 0 on success, non-zero on failure.
9419 mdb_drop0(MDB_cursor *mc, int subs)
9423 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9424 if (rc == MDB_SUCCESS) {
9425 MDB_txn *txn = mc->mc_txn;
9430 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9431 * This also avoids any P_LEAF2 pages, which have no nodes.
9433 if (mc->mc_flags & C_SUB)
9436 mdb_cursor_copy(mc, &mx);
9437 while (mc->mc_snum > 0) {
9438 MDB_page *mp = mc->mc_pg[mc->mc_top];
9439 unsigned n = NUMKEYS(mp);
9441 for (i=0; i<n; i++) {
9442 ni = NODEPTR(mp, i);
9443 if (ni->mn_flags & F_BIGDATA) {
9446 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9447 rc = mdb_page_get(txn, pg, &omp, NULL);
9450 mdb_cassert(mc, IS_OVERFLOW(omp));
9451 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9455 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9456 mdb_xcursor_init1(mc, ni);
9457 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9463 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9465 for (i=0; i<n; i++) {
9467 ni = NODEPTR(mp, i);
9470 mdb_midl_xappend(txn->mt_free_pgs, pg);
9475 mc->mc_ki[mc->mc_top] = i;
9476 rc = mdb_cursor_sibling(mc, 1);
9478 if (rc != MDB_NOTFOUND)
9480 /* no more siblings, go back to beginning
9481 * of previous level.
9485 for (i=1; i<mc->mc_snum; i++) {
9487 mc->mc_pg[i] = mx.mc_pg[i];
9492 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9495 txn->mt_flags |= MDB_TXN_ERROR;
9496 } else if (rc == MDB_NOTFOUND) {
9502 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9504 MDB_cursor *mc, *m2;
9507 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9510 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9513 if (TXN_DBI_CHANGED(txn, dbi))
9516 rc = mdb_cursor_open(txn, dbi, &mc);
9520 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9521 /* Invalidate the dropped DB's cursors */
9522 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9523 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9527 /* Can't delete the main DB */
9528 if (del && dbi > MAIN_DBI) {
9529 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9531 txn->mt_dbflags[dbi] = DB_STALE;
9532 mdb_dbi_close(txn->mt_env, dbi);
9534 txn->mt_flags |= MDB_TXN_ERROR;
9537 /* reset the DB record, mark it dirty */
9538 txn->mt_dbflags[dbi] |= DB_DIRTY;
9539 txn->mt_dbs[dbi].md_depth = 0;
9540 txn->mt_dbs[dbi].md_branch_pages = 0;
9541 txn->mt_dbs[dbi].md_leaf_pages = 0;
9542 txn->mt_dbs[dbi].md_overflow_pages = 0;
9543 txn->mt_dbs[dbi].md_entries = 0;
9544 txn->mt_dbs[dbi].md_root = P_INVALID;
9546 txn->mt_flags |= MDB_TXN_DIRTY;
9549 mdb_cursor_close(mc);
9553 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9555 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9558 txn->mt_dbxs[dbi].md_cmp = cmp;
9562 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9564 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9567 txn->mt_dbxs[dbi].md_dcmp = cmp;
9571 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9573 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9576 txn->mt_dbxs[dbi].md_rel = rel;
9580 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9582 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9585 txn->mt_dbxs[dbi].md_relctx = ctx;
9590 mdb_env_get_maxkeysize(MDB_env *env)
9592 return ENV_MAXKEY(env);
9596 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9598 unsigned int i, rdrs;
9601 int rc = 0, first = 1;
9605 if (!env->me_txns) {
9606 return func("(no reader locks)\n", ctx);
9608 rdrs = env->me_txns->mti_numreaders;
9609 mr = env->me_txns->mti_readers;
9610 for (i=0; i<rdrs; i++) {
9612 txnid_t txnid = mr[i].mr_txnid;
9613 sprintf(buf, txnid == (txnid_t)-1 ?
9614 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9615 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9618 rc = func(" pid thread txnid\n", ctx);
9622 rc = func(buf, ctx);
9628 rc = func("(no active readers)\n", ctx);
9633 /** Insert pid into list if not already present.
9634 * return -1 if already present.
9637 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9639 /* binary search of pid in list */
9641 unsigned cursor = 1;
9643 unsigned n = ids[0];
9646 unsigned pivot = n >> 1;
9647 cursor = base + pivot + 1;
9648 val = pid - ids[cursor];
9653 } else if ( val > 0 ) {
9658 /* found, so it's a duplicate */
9667 for (n = ids[0]; n > cursor; n--)
9674 mdb_reader_check(MDB_env *env, int *dead)
9680 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9683 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9684 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9686 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9687 unsigned int i, j, rdrs;
9689 MDB_PID_T *pids, pid;
9690 int rc = MDB_SUCCESS, count = 0;
9692 rdrs = env->me_txns->mti_numreaders;
9693 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9697 mr = env->me_txns->mti_readers;
9698 for (i=0; i<rdrs; i++) {
9700 if (pid && pid != env->me_pid) {
9701 if (mdb_pid_insert(pids, pid) == 0) {
9702 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9703 /* Stale reader found */
9706 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9707 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9709 rdrs = 0; /* the above checked all readers */
9711 /* Recheck, a new process may have reused pid */
9712 if (mdb_reader_pid(env, Pidcheck, pid))
9717 if (mr[j].mr_pid == pid) {
9718 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9719 (unsigned) pid, mr[j].mr_txnid));
9724 UNLOCK_MUTEX(rmutex);
9735 #ifdef MDB_ROBUST_SUPPORTED
9736 /** Handle #LOCK_MUTEX0() failure.
9737 * Try to repair the lock file if the mutex owner died.
9738 * @param[in] env the environment handle
9739 * @param[in] mutex LOCK_MUTEX0() mutex
9740 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9741 * @return 0 on success with the mutex locked, or an error code on failure.
9743 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9745 int toggle, rlocked, rc2;
9747 if (rc == MDB_OWNERDEAD) {
9748 /* We own the mutex. Clean up after dead previous owner. */
9750 rlocked = (mutex == env->me_rmutex);
9752 /* Keep mti_txnid updated, otherwise next writer can
9753 * overwrite data which latest meta page refers to.
9755 toggle = mdb_env_pick_meta(env);
9756 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9757 /* env is hosed if the dead thread was ours */
9759 env->me_flags |= MDB_FATAL_ERROR;
9764 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9765 (rc ? "this process' env is hosed" : "recovering")));
9766 rc2 = mdb_reader_check0(env, rlocked, NULL);
9768 rc2 = mdb_mutex_consistent(mutex);
9769 if (rc || (rc = rc2)) {
9770 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9771 UNLOCK_MUTEX(mutex);
9777 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9782 #endif /* MDB_ROBUST_SUPPORTED */