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 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1000 pgno_t mt_next_pgno; /**< next unallocated page */
1001 /** The ID of this transaction. IDs are integers incrementing from 1.
1002 * Only committed write transactions increment the ID. If a transaction
1003 * aborts, the ID may be re-used by the next writer.
1006 MDB_env *mt_env; /**< the DB environment */
1007 /** The list of pages that became unused during this transaction.
1009 MDB_IDL mt_free_pgs;
1010 /** The list of loose pages that became unused and may be reused
1011 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1013 MDB_page *mt_loose_pgs;
1014 /* #Number of loose pages (#mt_loose_pgs) */
1016 /** The sorted list of dirty pages we temporarily wrote to disk
1017 * because the dirty list was full. page numbers in here are
1018 * shifted left by 1, deleted slots have the LSB set.
1020 MDB_IDL mt_spill_pgs;
1022 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1023 MDB_ID2L dirty_list;
1024 /** For read txns: This thread/txn's reader table slot, or NULL. */
1027 /** Array of records for each DB known in the environment. */
1029 /** Array of MDB_db records for each known DB */
1031 /** Array of sequence numbers for each DB handle */
1032 unsigned int *mt_dbiseqs;
1033 /** @defgroup mt_dbflag Transaction DB Flags
1037 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1038 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1039 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1040 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1041 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1043 /** In write txns, array of cursors for each DB */
1044 MDB_cursor **mt_cursors;
1045 /** Array of flags for each DB */
1046 unsigned char *mt_dbflags;
1047 /** Number of DB records in use, or 0 when the txn is finished.
1048 * This number only ever increments until the txn finishes; we
1049 * don't decrement it when individual DB handles are closed.
1053 /** @defgroup mdb_txn Transaction Flags
1057 /** #mdb_txn_begin() flags */
1058 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1059 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1060 /* internal txn flags */
1061 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1062 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1063 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1064 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1065 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1066 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1067 /** most operations on the txn are currently illegal */
1068 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1070 unsigned int mt_flags; /**< @ref mdb_txn */
1071 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1072 * Includes ancestor txns' dirty pages not hidden by other txns'
1073 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1074 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1076 unsigned int mt_dirty_room;
1079 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1080 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1081 * raise this on a 64 bit machine.
1083 #define CURSOR_STACK 32
1087 /** Cursors are used for all DB operations.
1088 * A cursor holds a path of (page pointer, key index) from the DB
1089 * root to a position in the DB, plus other state. #MDB_DUPSORT
1090 * cursors include an xcursor to the current data item. Write txns
1091 * track their cursors and keep them up to date when data moves.
1092 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1093 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1096 /** Next cursor on this DB in this txn */
1097 MDB_cursor *mc_next;
1098 /** Backup of the original cursor if this cursor is a shadow */
1099 MDB_cursor *mc_backup;
1100 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1101 struct MDB_xcursor *mc_xcursor;
1102 /** The transaction that owns this cursor */
1104 /** The database handle this cursor operates on */
1106 /** The database record for this cursor */
1108 /** The database auxiliary record for this cursor */
1110 /** The @ref mt_dbflag for this database */
1111 unsigned char *mc_dbflag;
1112 unsigned short mc_snum; /**< number of pushed pages */
1113 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1114 /** @defgroup mdb_cursor Cursor Flags
1116 * Cursor state flags.
1119 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1120 #define C_EOF 0x02 /**< No more data */
1121 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1122 #define C_DEL 0x08 /**< last op was a cursor_del */
1123 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1124 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1126 unsigned int mc_flags; /**< @ref mdb_cursor */
1127 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1128 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1131 /** Context for sorted-dup records.
1132 * We could have gone to a fully recursive design, with arbitrarily
1133 * deep nesting of sub-databases. But for now we only handle these
1134 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1136 typedef struct MDB_xcursor {
1137 /** A sub-cursor for traversing the Dup DB */
1138 MDB_cursor mx_cursor;
1139 /** The database record for this Dup DB */
1141 /** The auxiliary DB record for this Dup DB */
1143 /** The @ref mt_dbflag for this Dup DB */
1144 unsigned char mx_dbflag;
1147 /** State of FreeDB old pages, stored in the MDB_env */
1148 typedef struct MDB_pgstate {
1149 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1150 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1153 /** The database environment. */
1155 HANDLE me_fd; /**< The main data file */
1156 HANDLE me_lfd; /**< The lock file */
1157 HANDLE me_mfd; /**< just for writing the meta pages */
1158 /** Failed to update the meta page. Probably an I/O error. */
1159 #define MDB_FATAL_ERROR 0x80000000U
1160 /** Some fields are initialized. */
1161 #define MDB_ENV_ACTIVE 0x20000000U
1162 /** me_txkey is set */
1163 #define MDB_ENV_TXKEY 0x10000000U
1164 /** fdatasync is unreliable */
1165 #define MDB_FSYNCONLY 0x08000000U
1166 uint32_t me_flags; /**< @ref mdb_env */
1167 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1168 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1169 unsigned int me_maxreaders; /**< size of the reader table */
1170 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1171 volatile int me_close_readers;
1172 MDB_dbi me_numdbs; /**< number of DBs opened */
1173 MDB_dbi me_maxdbs; /**< size of the DB table */
1174 MDB_PID_T me_pid; /**< process ID of this env */
1175 char *me_path; /**< path to the DB files */
1176 char *me_map; /**< the memory map of the data file */
1177 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1178 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1179 void *me_pbuf; /**< scratch area for DUPSORT put() */
1180 MDB_txn *me_txn; /**< current write transaction */
1181 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1182 size_t me_mapsize; /**< size of the data memory map */
1183 off_t me_size; /**< current file size */
1184 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1185 MDB_dbx *me_dbxs; /**< array of static DB info */
1186 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1187 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1188 pthread_key_t me_txkey; /**< thread-key for readers */
1189 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1190 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1191 # define me_pglast me_pgstate.mf_pglast
1192 # define me_pghead me_pgstate.mf_pghead
1193 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1194 /** IDL of pages that became unused in a write txn */
1195 MDB_IDL me_free_pgs;
1196 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1197 MDB_ID2L me_dirty_list;
1198 /** Max number of freelist items that can fit in a single overflow page */
1200 /** Max size of a node on a page */
1201 unsigned int me_nodemax;
1202 #if !(MDB_MAXKEYSIZE)
1203 unsigned int me_maxkey; /**< max size of a key */
1205 int me_live_reader; /**< have liveness lock in reader table */
1207 int me_pidquery; /**< Used in OpenProcess */
1209 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1210 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1211 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1213 mdb_mutex_t me_rmutex;
1214 mdb_mutex_t me_wmutex;
1216 void *me_userctx; /**< User-settable context */
1217 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1220 /** Nested transaction */
1221 typedef struct MDB_ntxn {
1222 MDB_txn mnt_txn; /**< the transaction */
1223 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1226 /** max number of pages to commit in one writev() call */
1227 #define MDB_COMMIT_PAGES 64
1228 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1229 #undef MDB_COMMIT_PAGES
1230 #define MDB_COMMIT_PAGES IOV_MAX
1233 /** max bytes to write in one call */
1234 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1236 /** Check \b txn and \b dbi arguments to a function */
1237 #define TXN_DBI_EXIST(txn, dbi, validity) \
1238 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1240 /** Check for misused \b dbi handles */
1241 #define TXN_DBI_CHANGED(txn, dbi) \
1242 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1244 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1245 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1246 static int mdb_page_touch(MDB_cursor *mc);
1248 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1249 "reset-tmp", "fail-begin", "fail-beginchild"}
1251 /* mdb_txn_end operation number, for logging */
1252 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1253 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1255 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1256 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1257 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1258 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1259 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1261 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1262 static int mdb_page_search_root(MDB_cursor *mc,
1263 MDB_val *key, int modify);
1264 #define MDB_PS_MODIFY 1
1265 #define MDB_PS_ROOTONLY 2
1266 #define MDB_PS_FIRST 4
1267 #define MDB_PS_LAST 8
1268 static int mdb_page_search(MDB_cursor *mc,
1269 MDB_val *key, int flags);
1270 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1272 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1273 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1274 pgno_t newpgno, unsigned int nflags);
1276 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1277 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1278 static int mdb_env_write_meta(MDB_txn *txn);
1279 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1280 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1282 static void mdb_env_close0(MDB_env *env, int excl);
1284 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1285 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1286 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1287 static void mdb_node_del(MDB_cursor *mc, int ksize);
1288 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1289 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1290 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1291 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1292 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1294 static int mdb_rebalance(MDB_cursor *mc);
1295 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1297 static void mdb_cursor_pop(MDB_cursor *mc);
1298 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1300 static int mdb_cursor_del0(MDB_cursor *mc);
1301 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1302 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1303 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1304 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1305 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1307 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1308 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1310 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1311 static void mdb_xcursor_init0(MDB_cursor *mc);
1312 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1313 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1315 static int mdb_drop0(MDB_cursor *mc, int subs);
1316 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1317 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1320 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1323 /** Compare two items pointing at size_t's of unknown alignment. */
1324 #ifdef MISALIGNED_OK
1325 # define mdb_cmp_clong mdb_cmp_long
1327 # define mdb_cmp_clong mdb_cmp_cint
1331 static SECURITY_DESCRIPTOR mdb_null_sd;
1332 static SECURITY_ATTRIBUTES mdb_all_sa;
1333 static int mdb_sec_inited;
1336 /** Return the library version info. */
1338 mdb_version(int *major, int *minor, int *patch)
1340 if (major) *major = MDB_VERSION_MAJOR;
1341 if (minor) *minor = MDB_VERSION_MINOR;
1342 if (patch) *patch = MDB_VERSION_PATCH;
1343 return MDB_VERSION_STRING;
1346 /** Table of descriptions for LMDB @ref errors */
1347 static char *const mdb_errstr[] = {
1348 "MDB_KEYEXIST: Key/data pair already exists",
1349 "MDB_NOTFOUND: No matching key/data pair found",
1350 "MDB_PAGE_NOTFOUND: Requested page not found",
1351 "MDB_CORRUPTED: Located page was wrong type",
1352 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1353 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1354 "MDB_INVALID: File is not an LMDB file",
1355 "MDB_MAP_FULL: Environment mapsize limit reached",
1356 "MDB_DBS_FULL: Environment maxdbs limit reached",
1357 "MDB_READERS_FULL: Environment maxreaders limit reached",
1358 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1359 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1360 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1361 "MDB_PAGE_FULL: Internal error - page has no more space",
1362 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1363 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1364 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1365 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1366 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1367 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1371 mdb_strerror(int err)
1374 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1375 * This works as long as no function between the call to mdb_strerror
1376 * and the actual use of the message uses more than 4K of stack.
1379 char buf[1024], *ptr = buf;
1383 return ("Successful return: 0");
1385 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1386 i = err - MDB_KEYEXIST;
1387 return mdb_errstr[i];
1391 /* These are the C-runtime error codes we use. The comment indicates
1392 * their numeric value, and the Win32 error they would correspond to
1393 * if the error actually came from a Win32 API. A major mess, we should
1394 * have used LMDB-specific error codes for everything.
1397 case ENOENT: /* 2, FILE_NOT_FOUND */
1398 case EIO: /* 5, ACCESS_DENIED */
1399 case ENOMEM: /* 12, INVALID_ACCESS */
1400 case EACCES: /* 13, INVALID_DATA */
1401 case EBUSY: /* 16, CURRENT_DIRECTORY */
1402 case EINVAL: /* 22, BAD_COMMAND */
1403 case ENOSPC: /* 28, OUT_OF_PAPER */
1404 return strerror(err);
1409 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1410 FORMAT_MESSAGE_IGNORE_INSERTS,
1411 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1414 return strerror(err);
1418 /** assert(3) variant in cursor context */
1419 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1420 /** assert(3) variant in transaction context */
1421 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1422 /** assert(3) variant in environment context */
1423 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1426 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1427 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1430 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1431 const char *func, const char *file, int line)
1434 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1435 file, line, expr_txt, func);
1436 if (env->me_assert_func)
1437 env->me_assert_func(env, buf);
1438 fprintf(stderr, "%s\n", buf);
1442 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1446 /** Return the page number of \b mp which may be sub-page, for debug output */
1448 mdb_dbg_pgno(MDB_page *mp)
1451 COPY_PGNO(ret, mp->mp_pgno);
1455 /** Display a key in hexadecimal and return the address of the result.
1456 * @param[in] key the key to display
1457 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1458 * @return The key in hexadecimal form.
1461 mdb_dkey(MDB_val *key, char *buf)
1464 unsigned char *c = key->mv_data;
1470 if (key->mv_size > DKBUF_MAXKEYSIZE)
1471 return "MDB_MAXKEYSIZE";
1472 /* may want to make this a dynamic check: if the key is mostly
1473 * printable characters, print it as-is instead of converting to hex.
1477 for (i=0; i<key->mv_size; i++)
1478 ptr += sprintf(ptr, "%02x", *c++);
1480 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1486 mdb_leafnode_type(MDB_node *n)
1488 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1489 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1490 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1493 /** Display all the keys in the page. */
1495 mdb_page_list(MDB_page *mp)
1497 pgno_t pgno = mdb_dbg_pgno(mp);
1498 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1500 unsigned int i, nkeys, nsize, total = 0;
1504 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1505 case P_BRANCH: type = "Branch page"; break;
1506 case P_LEAF: type = "Leaf page"; break;
1507 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1508 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1509 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1511 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1512 pgno, mp->mp_pages, state);
1515 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1516 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1519 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1523 nkeys = NUMKEYS(mp);
1524 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1526 for (i=0; i<nkeys; i++) {
1527 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1528 key.mv_size = nsize = mp->mp_pad;
1529 key.mv_data = LEAF2KEY(mp, i, nsize);
1531 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1534 node = NODEPTR(mp, i);
1535 key.mv_size = node->mn_ksize;
1536 key.mv_data = node->mn_data;
1537 nsize = NODESIZE + key.mv_size;
1538 if (IS_BRANCH(mp)) {
1539 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1543 if (F_ISSET(node->mn_flags, F_BIGDATA))
1544 nsize += sizeof(pgno_t);
1546 nsize += NODEDSZ(node);
1548 nsize += sizeof(indx_t);
1549 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1550 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1552 total = EVEN(total);
1554 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1555 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1559 mdb_cursor_chk(MDB_cursor *mc)
1565 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1566 for (i=0; i<mc->mc_top; i++) {
1568 node = NODEPTR(mp, mc->mc_ki[i]);
1569 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1572 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1578 /** Count all the pages in each DB and in the freelist
1579 * and make sure it matches the actual number of pages
1581 * All named DBs must be open for a correct count.
1583 static void mdb_audit(MDB_txn *txn)
1587 MDB_ID freecount, count;
1592 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1593 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1594 freecount += *(MDB_ID *)data.mv_data;
1595 mdb_tassert(txn, rc == MDB_NOTFOUND);
1598 for (i = 0; i<txn->mt_numdbs; i++) {
1600 if (!(txn->mt_dbflags[i] & DB_VALID))
1602 mdb_cursor_init(&mc, txn, i, &mx);
1603 if (txn->mt_dbs[i].md_root == P_INVALID)
1605 count += txn->mt_dbs[i].md_branch_pages +
1606 txn->mt_dbs[i].md_leaf_pages +
1607 txn->mt_dbs[i].md_overflow_pages;
1608 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1609 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1610 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1613 mp = mc.mc_pg[mc.mc_top];
1614 for (j=0; j<NUMKEYS(mp); j++) {
1615 MDB_node *leaf = NODEPTR(mp, j);
1616 if (leaf->mn_flags & F_SUBDATA) {
1618 memcpy(&db, NODEDATA(leaf), sizeof(db));
1619 count += db.md_branch_pages + db.md_leaf_pages +
1620 db.md_overflow_pages;
1624 mdb_tassert(txn, rc == MDB_NOTFOUND);
1627 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1628 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1629 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1635 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1637 return txn->mt_dbxs[dbi].md_cmp(a, b);
1641 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1643 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1644 #if UINT_MAX < SIZE_MAX
1645 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1646 dcmp = mdb_cmp_clong;
1651 /** Allocate memory for a page.
1652 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1655 mdb_page_malloc(MDB_txn *txn, unsigned num)
1657 MDB_env *env = txn->mt_env;
1658 MDB_page *ret = env->me_dpages;
1659 size_t psize = env->me_psize, sz = psize, off;
1660 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1661 * For a single page alloc, we init everything after the page header.
1662 * For multi-page, we init the final page; if the caller needed that
1663 * many pages they will be filling in at least up to the last page.
1667 VGMEMP_ALLOC(env, ret, sz);
1668 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1669 env->me_dpages = ret->mp_next;
1672 psize -= off = PAGEHDRSZ;
1677 if ((ret = malloc(sz)) != NULL) {
1678 VGMEMP_ALLOC(env, ret, sz);
1679 if (!(env->me_flags & MDB_NOMEMINIT)) {
1680 memset((char *)ret + off, 0, psize);
1684 txn->mt_flags |= MDB_TXN_ERROR;
1688 /** Free a single page.
1689 * Saves single pages to a list, for future reuse.
1690 * (This is not used for multi-page overflow pages.)
1693 mdb_page_free(MDB_env *env, MDB_page *mp)
1695 mp->mp_next = env->me_dpages;
1696 VGMEMP_FREE(env, mp);
1697 env->me_dpages = mp;
1700 /** Free a dirty page */
1702 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1704 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1705 mdb_page_free(env, dp);
1707 /* large pages just get freed directly */
1708 VGMEMP_FREE(env, dp);
1713 /** Return all dirty pages to dpage list */
1715 mdb_dlist_free(MDB_txn *txn)
1717 MDB_env *env = txn->mt_env;
1718 MDB_ID2L dl = txn->mt_u.dirty_list;
1719 unsigned i, n = dl[0].mid;
1721 for (i = 1; i <= n; i++) {
1722 mdb_dpage_free(env, dl[i].mptr);
1727 /** Loosen or free a single page.
1728 * Saves single pages to a list for future reuse
1729 * in this same txn. It has been pulled from the freeDB
1730 * and already resides on the dirty list, but has been
1731 * deleted. Use these pages first before pulling again
1734 * If the page wasn't dirtied in this txn, just add it
1735 * to this txn's free list.
1738 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1741 pgno_t pgno = mp->mp_pgno;
1742 MDB_txn *txn = mc->mc_txn;
1744 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1745 if (txn->mt_parent) {
1746 MDB_ID2 *dl = txn->mt_u.dirty_list;
1747 /* If txn has a parent, make sure the page is in our
1751 unsigned x = mdb_mid2l_search(dl, pgno);
1752 if (x <= dl[0].mid && dl[x].mid == pgno) {
1753 if (mp != dl[x].mptr) { /* bad cursor? */
1754 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1755 txn->mt_flags |= MDB_TXN_ERROR;
1756 return MDB_CORRUPTED;
1763 /* no parent txn, so it's just ours */
1768 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1770 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1771 txn->mt_loose_pgs = mp;
1772 txn->mt_loose_count++;
1773 mp->mp_flags |= P_LOOSE;
1775 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1783 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1784 * @param[in] mc A cursor handle for the current operation.
1785 * @param[in] pflags Flags of the pages to update:
1786 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1787 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1788 * @return 0 on success, non-zero on failure.
1791 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1793 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1794 MDB_txn *txn = mc->mc_txn;
1800 int rc = MDB_SUCCESS, level;
1802 /* Mark pages seen by cursors */
1803 if (mc->mc_flags & C_UNTRACK)
1804 mc = NULL; /* will find mc in mt_cursors */
1805 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1806 for (; mc; mc=mc->mc_next) {
1807 if (!(mc->mc_flags & C_INITIALIZED))
1809 for (m3 = mc;; m3 = &mx->mx_cursor) {
1811 for (j=0; j<m3->mc_snum; j++) {
1813 if ((mp->mp_flags & Mask) == pflags)
1814 mp->mp_flags ^= P_KEEP;
1816 mx = m3->mc_xcursor;
1817 /* Proceed to mx if it is at a sub-database */
1818 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1820 if (! (mp && (mp->mp_flags & P_LEAF)))
1822 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1823 if (!(leaf->mn_flags & F_SUBDATA))
1832 /* Mark dirty root pages */
1833 for (i=0; i<txn->mt_numdbs; i++) {
1834 if (txn->mt_dbflags[i] & DB_DIRTY) {
1835 pgno_t pgno = txn->mt_dbs[i].md_root;
1836 if (pgno == P_INVALID)
1838 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1840 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1841 dp->mp_flags ^= P_KEEP;
1849 static int mdb_page_flush(MDB_txn *txn, int keep);
1851 /** Spill pages from the dirty list back to disk.
1852 * This is intended to prevent running into #MDB_TXN_FULL situations,
1853 * but note that they may still occur in a few cases:
1854 * 1) our estimate of the txn size could be too small. Currently this
1855 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1856 * 2) child txns may run out of space if their parents dirtied a
1857 * lot of pages and never spilled them. TODO: we probably should do
1858 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1859 * the parent's dirty_room is below a given threshold.
1861 * Otherwise, if not using nested txns, it is expected that apps will
1862 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1863 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1864 * If the txn never references them again, they can be left alone.
1865 * If the txn only reads them, they can be used without any fuss.
1866 * If the txn writes them again, they can be dirtied immediately without
1867 * going thru all of the work of #mdb_page_touch(). Such references are
1868 * handled by #mdb_page_unspill().
1870 * Also note, we never spill DB root pages, nor pages of active cursors,
1871 * because we'll need these back again soon anyway. And in nested txns,
1872 * we can't spill a page in a child txn if it was already spilled in a
1873 * parent txn. That would alter the parent txns' data even though
1874 * the child hasn't committed yet, and we'd have no way to undo it if
1875 * the child aborted.
1877 * @param[in] m0 cursor A cursor handle identifying the transaction and
1878 * database for which we are checking space.
1879 * @param[in] key For a put operation, the key being stored.
1880 * @param[in] data For a put operation, the data being stored.
1881 * @return 0 on success, non-zero on failure.
1884 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1886 MDB_txn *txn = m0->mc_txn;
1888 MDB_ID2L dl = txn->mt_u.dirty_list;
1889 unsigned int i, j, need;
1892 if (m0->mc_flags & C_SUB)
1895 /* Estimate how much space this op will take */
1896 i = m0->mc_db->md_depth;
1897 /* Named DBs also dirty the main DB */
1898 if (m0->mc_dbi > MAIN_DBI)
1899 i += txn->mt_dbs[MAIN_DBI].md_depth;
1900 /* For puts, roughly factor in the key+data size */
1902 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1903 i += i; /* double it for good measure */
1906 if (txn->mt_dirty_room > i)
1909 if (!txn->mt_spill_pgs) {
1910 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1911 if (!txn->mt_spill_pgs)
1914 /* purge deleted slots */
1915 MDB_IDL sl = txn->mt_spill_pgs;
1916 unsigned int num = sl[0];
1918 for (i=1; i<=num; i++) {
1925 /* Preserve pages which may soon be dirtied again */
1926 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1929 /* Less aggressive spill - we originally spilled the entire dirty list,
1930 * with a few exceptions for cursor pages and DB root pages. But this
1931 * turns out to be a lot of wasted effort because in a large txn many
1932 * of those pages will need to be used again. So now we spill only 1/8th
1933 * of the dirty pages. Testing revealed this to be a good tradeoff,
1934 * better than 1/2, 1/4, or 1/10.
1936 if (need < MDB_IDL_UM_MAX / 8)
1937 need = MDB_IDL_UM_MAX / 8;
1939 /* Save the page IDs of all the pages we're flushing */
1940 /* flush from the tail forward, this saves a lot of shifting later on. */
1941 for (i=dl[0].mid; i && need; i--) {
1942 MDB_ID pn = dl[i].mid << 1;
1944 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1946 /* Can't spill twice, make sure it's not already in a parent's
1949 if (txn->mt_parent) {
1951 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1952 if (tx2->mt_spill_pgs) {
1953 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1954 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1955 dp->mp_flags |= P_KEEP;
1963 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1967 mdb_midl_sort(txn->mt_spill_pgs);
1969 /* Flush the spilled part of dirty list */
1970 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1973 /* Reset any dirty pages we kept that page_flush didn't see */
1974 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1977 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1981 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1983 mdb_find_oldest(MDB_txn *txn)
1986 txnid_t mr, oldest = txn->mt_txnid - 1;
1987 if (txn->mt_env->me_txns) {
1988 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1989 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2000 /** Add a page to the txn's dirty list */
2002 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2005 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2007 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2008 insert = mdb_mid2l_append;
2010 insert = mdb_mid2l_insert;
2012 mid.mid = mp->mp_pgno;
2014 rc = insert(txn->mt_u.dirty_list, &mid);
2015 mdb_tassert(txn, rc == 0);
2016 txn->mt_dirty_room--;
2019 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2020 * me_pghead and mt_next_pgno.
2022 * If there are free pages available from older transactions, they
2023 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2024 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2025 * and move me_pglast to say which records were consumed. Only this
2026 * function can create me_pghead and move me_pglast/mt_next_pgno.
2027 * @param[in] mc cursor A cursor handle identifying the transaction and
2028 * database for which we are allocating.
2029 * @param[in] num the number of pages to allocate.
2030 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2031 * will always be satisfied by a single contiguous chunk of memory.
2032 * @return 0 on success, non-zero on failure.
2035 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2037 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2038 /* Get at most <Max_retries> more freeDB records once me_pghead
2039 * has enough pages. If not enough, use new pages from the map.
2040 * If <Paranoid> and mc is updating the freeDB, only get new
2041 * records if me_pghead is empty. Then the freelist cannot play
2042 * catch-up with itself by growing while trying to save it.
2044 enum { Paranoid = 1, Max_retries = 500 };
2046 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2048 int rc, retry = num * 60;
2049 MDB_txn *txn = mc->mc_txn;
2050 MDB_env *env = txn->mt_env;
2051 pgno_t pgno, *mop = env->me_pghead;
2052 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2054 txnid_t oldest = 0, last;
2059 /* If there are any loose pages, just use them */
2060 if (num == 1 && txn->mt_loose_pgs) {
2061 np = txn->mt_loose_pgs;
2062 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2063 txn->mt_loose_count--;
2064 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2072 /* If our dirty list is already full, we can't do anything */
2073 if (txn->mt_dirty_room == 0) {
2078 for (op = MDB_FIRST;; op = MDB_NEXT) {
2083 /* Seek a big enough contiguous page range. Prefer
2084 * pages at the tail, just truncating the list.
2090 if (mop[i-n2] == pgno+n2)
2097 if (op == MDB_FIRST) { /* 1st iteration */
2098 /* Prepare to fetch more and coalesce */
2099 last = env->me_pglast;
2100 oldest = env->me_pgoldest;
2101 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2104 key.mv_data = &last; /* will look up last+1 */
2105 key.mv_size = sizeof(last);
2107 if (Paranoid && mc->mc_dbi == FREE_DBI)
2110 if (Paranoid && retry < 0 && mop_len)
2114 /* Do not fetch more if the record will be too recent */
2115 if (oldest <= last) {
2117 oldest = mdb_find_oldest(txn);
2118 env->me_pgoldest = oldest;
2124 rc = mdb_cursor_get(&m2, &key, NULL, op);
2126 if (rc == MDB_NOTFOUND)
2130 last = *(txnid_t*)key.mv_data;
2131 if (oldest <= last) {
2133 oldest = mdb_find_oldest(txn);
2134 env->me_pgoldest = oldest;
2140 np = m2.mc_pg[m2.mc_top];
2141 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2142 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2145 idl = (MDB_ID *) data.mv_data;
2148 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2153 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2155 mop = env->me_pghead;
2157 env->me_pglast = last;
2159 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2160 last, txn->mt_dbs[FREE_DBI].md_root, i));
2162 DPRINTF(("IDL %"Z"u", idl[j]));
2164 /* Merge in descending sorted order */
2165 mdb_midl_xmerge(mop, idl);
2169 /* Use new pages from the map when nothing suitable in the freeDB */
2171 pgno = txn->mt_next_pgno;
2172 if (pgno + num >= env->me_maxpg) {
2173 DPUTS("DB size maxed out");
2179 if (env->me_flags & MDB_WRITEMAP) {
2180 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2182 if (!(np = mdb_page_malloc(txn, num))) {
2188 mop[0] = mop_len -= num;
2189 /* Move any stragglers down */
2190 for (j = i-num; j < mop_len; )
2191 mop[++j] = mop[++i];
2193 txn->mt_next_pgno = pgno + num;
2196 mdb_page_dirty(txn, np);
2202 txn->mt_flags |= MDB_TXN_ERROR;
2206 /** Copy the used portions of a non-overflow page.
2207 * @param[in] dst page to copy into
2208 * @param[in] src page to copy from
2209 * @param[in] psize size of a page
2212 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2214 enum { Align = sizeof(pgno_t) };
2215 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2217 /* If page isn't full, just copy the used portion. Adjust
2218 * alignment so memcpy may copy words instead of bytes.
2220 if ((unused &= -Align) && !IS_LEAF2(src)) {
2221 upper = (upper + PAGEBASE) & -Align;
2222 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2223 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2226 memcpy(dst, src, psize - unused);
2230 /** Pull a page off the txn's spill list, if present.
2231 * If a page being referenced was spilled to disk in this txn, bring
2232 * it back and make it dirty/writable again.
2233 * @param[in] txn the transaction handle.
2234 * @param[in] mp the page being referenced. It must not be dirty.
2235 * @param[out] ret the writable page, if any. ret is unchanged if
2236 * mp wasn't spilled.
2239 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2241 MDB_env *env = txn->mt_env;
2244 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2246 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2247 if (!tx2->mt_spill_pgs)
2249 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2250 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2253 if (txn->mt_dirty_room == 0)
2254 return MDB_TXN_FULL;
2255 if (IS_OVERFLOW(mp))
2259 if (env->me_flags & MDB_WRITEMAP) {
2262 np = mdb_page_malloc(txn, num);
2266 memcpy(np, mp, num * env->me_psize);
2268 mdb_page_copy(np, mp, env->me_psize);
2271 /* If in current txn, this page is no longer spilled.
2272 * If it happens to be the last page, truncate the spill list.
2273 * Otherwise mark it as deleted by setting the LSB.
2275 if (x == txn->mt_spill_pgs[0])
2276 txn->mt_spill_pgs[0]--;
2278 txn->mt_spill_pgs[x] |= 1;
2279 } /* otherwise, if belonging to a parent txn, the
2280 * page remains spilled until child commits
2283 mdb_page_dirty(txn, np);
2284 np->mp_flags |= P_DIRTY;
2292 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2293 * @param[in] mc cursor pointing to the page to be touched
2294 * @return 0 on success, non-zero on failure.
2297 mdb_page_touch(MDB_cursor *mc)
2299 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2300 MDB_txn *txn = mc->mc_txn;
2301 MDB_cursor *m2, *m3;
2305 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2306 if (txn->mt_flags & MDB_TXN_SPILLS) {
2308 rc = mdb_page_unspill(txn, mp, &np);
2314 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2315 (rc = mdb_page_alloc(mc, 1, &np)))
2318 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2319 mp->mp_pgno, pgno));
2320 mdb_cassert(mc, mp->mp_pgno != pgno);
2321 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2322 /* Update the parent page, if any, to point to the new page */
2324 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2325 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2326 SETPGNO(node, pgno);
2328 mc->mc_db->md_root = pgno;
2330 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2331 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2333 /* If txn has a parent, make sure the page is in our
2337 unsigned x = mdb_mid2l_search(dl, pgno);
2338 if (x <= dl[0].mid && dl[x].mid == pgno) {
2339 if (mp != dl[x].mptr) { /* bad cursor? */
2340 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2341 txn->mt_flags |= MDB_TXN_ERROR;
2342 return MDB_CORRUPTED;
2347 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2349 np = mdb_page_malloc(txn, 1);
2354 rc = mdb_mid2l_insert(dl, &mid);
2355 mdb_cassert(mc, rc == 0);
2360 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2362 np->mp_flags |= P_DIRTY;
2365 /* Adjust cursors pointing to mp */
2366 mc->mc_pg[mc->mc_top] = np;
2367 m2 = txn->mt_cursors[mc->mc_dbi];
2368 if (mc->mc_flags & C_SUB) {
2369 for (; m2; m2=m2->mc_next) {
2370 m3 = &m2->mc_xcursor->mx_cursor;
2371 if (m3->mc_snum < mc->mc_snum) continue;
2372 if (m3->mc_pg[mc->mc_top] == mp)
2373 m3->mc_pg[mc->mc_top] = np;
2376 for (; m2; m2=m2->mc_next) {
2377 if (m2->mc_snum < mc->mc_snum) continue;
2378 if (m2->mc_pg[mc->mc_top] == mp) {
2379 m2->mc_pg[mc->mc_top] = np;
2380 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2382 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2384 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2385 if (!(leaf->mn_flags & F_SUBDATA))
2386 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2394 txn->mt_flags |= MDB_TXN_ERROR;
2399 mdb_env_sync(MDB_env *env, int force)
2402 if (env->me_flags & MDB_RDONLY)
2404 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2405 if (env->me_flags & MDB_WRITEMAP) {
2406 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2407 ? MS_ASYNC : MS_SYNC;
2408 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2411 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2415 #ifdef BROKEN_FDATASYNC
2416 if (env->me_flags & MDB_FSYNCONLY) {
2417 if (fsync(env->me_fd))
2421 if (MDB_FDATASYNC(env->me_fd))
2428 /** Back up parent txn's cursors, then grab the originals for tracking */
2430 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2432 MDB_cursor *mc, *bk;
2437 for (i = src->mt_numdbs; --i >= 0; ) {
2438 if ((mc = src->mt_cursors[i]) != NULL) {
2439 size = sizeof(MDB_cursor);
2441 size += sizeof(MDB_xcursor);
2442 for (; mc; mc = bk->mc_next) {
2448 mc->mc_db = &dst->mt_dbs[i];
2449 /* Kill pointers into src - and dst to reduce abuse: The
2450 * user may not use mc until dst ends. Otherwise we'd...
2452 mc->mc_txn = NULL; /* ...set this to dst */
2453 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2454 if ((mx = mc->mc_xcursor) != NULL) {
2455 *(MDB_xcursor *)(bk+1) = *mx;
2456 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2458 mc->mc_next = dst->mt_cursors[i];
2459 dst->mt_cursors[i] = mc;
2466 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2467 * @param[in] txn the transaction handle.
2468 * @param[in] merge true to keep changes to parent cursors, false to revert.
2469 * @return 0 on success, non-zero on failure.
2472 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2474 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2478 for (i = txn->mt_numdbs; --i >= 0; ) {
2479 for (mc = cursors[i]; mc; mc = next) {
2481 if ((bk = mc->mc_backup) != NULL) {
2483 /* Commit changes to parent txn */
2484 mc->mc_next = bk->mc_next;
2485 mc->mc_backup = bk->mc_backup;
2486 mc->mc_txn = bk->mc_txn;
2487 mc->mc_db = bk->mc_db;
2488 mc->mc_dbflag = bk->mc_dbflag;
2489 if ((mx = mc->mc_xcursor) != NULL)
2490 mx->mx_cursor.mc_txn = bk->mc_txn;
2492 /* Abort nested txn */
2494 if ((mx = mc->mc_xcursor) != NULL)
2495 *mx = *(MDB_xcursor *)(bk+1);
2499 /* Only malloced cursors are permanently tracked. */
2506 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2512 Pidset = F_SETLK, Pidcheck = F_GETLK
2516 /** Set or check a pid lock. Set returns 0 on success.
2517 * Check returns 0 if the process is certainly dead, nonzero if it may
2518 * be alive (the lock exists or an error happened so we do not know).
2520 * On Windows Pidset is a no-op, we merely check for the existence
2521 * of the process with the given pid. On POSIX we use a single byte
2522 * lock on the lockfile, set at an offset equal to the pid.
2525 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2527 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2530 if (op == Pidcheck) {
2531 h = OpenProcess(env->me_pidquery, FALSE, pid);
2532 /* No documented "no such process" code, but other program use this: */
2534 return ErrCode() != ERROR_INVALID_PARAMETER;
2535 /* A process exists until all handles to it close. Has it exited? */
2536 ret = WaitForSingleObject(h, 0) != 0;
2543 struct flock lock_info;
2544 memset(&lock_info, 0, sizeof(lock_info));
2545 lock_info.l_type = F_WRLCK;
2546 lock_info.l_whence = SEEK_SET;
2547 lock_info.l_start = pid;
2548 lock_info.l_len = 1;
2549 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2550 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2552 } else if ((rc = ErrCode()) == EINTR) {
2560 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2561 * @param[in] txn the transaction handle to initialize
2562 * @return 0 on success, non-zero on failure.
2565 mdb_txn_renew0(MDB_txn *txn)
2567 MDB_env *env = txn->mt_env;
2568 MDB_txninfo *ti = env->me_txns;
2570 unsigned int i, nr, flags = txn->mt_flags;
2572 int rc, new_notls = 0;
2574 if ((flags &= MDB_TXN_RDONLY) != 0) {
2576 meta = mdb_env_pick_meta(env);
2577 txn->mt_txnid = meta->mm_txnid;
2578 txn->mt_u.reader = NULL;
2580 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2581 pthread_getspecific(env->me_txkey);
2583 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2584 return MDB_BAD_RSLOT;
2586 MDB_PID_T pid = env->me_pid;
2587 MDB_THR_T tid = pthread_self();
2588 mdb_mutexref_t rmutex = env->me_rmutex;
2590 if (!env->me_live_reader) {
2591 rc = mdb_reader_pid(env, Pidset, pid);
2594 env->me_live_reader = 1;
2597 if (LOCK_MUTEX(rc, env, rmutex))
2599 nr = ti->mti_numreaders;
2600 for (i=0; i<nr; i++)
2601 if (ti->mti_readers[i].mr_pid == 0)
2603 if (i == env->me_maxreaders) {
2604 UNLOCK_MUTEX(rmutex);
2605 return MDB_READERS_FULL;
2607 r = &ti->mti_readers[i];
2608 /* Claim the reader slot, carefully since other code
2609 * uses the reader table un-mutexed: First reset the
2610 * slot, next publish it in mti_numreaders. After
2611 * that, it is safe for mdb_env_close() to touch it.
2612 * When it will be closed, we can finally claim it.
2615 r->mr_txnid = (txnid_t)-1;
2618 ti->mti_numreaders = ++nr;
2619 env->me_close_readers = nr;
2621 UNLOCK_MUTEX(rmutex);
2623 new_notls = (env->me_flags & MDB_NOTLS);
2624 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2629 do /* LY: Retry on a race, ITS#7970. */
2630 r->mr_txnid = ti->mti_txnid;
2631 while(r->mr_txnid != ti->mti_txnid);
2632 txn->mt_txnid = r->mr_txnid;
2633 txn->mt_u.reader = r;
2634 meta = env->me_metas[txn->mt_txnid & 1];
2636 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2638 /* Not yet touching txn == env->me_txn0, it may be active */
2640 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2642 txn->mt_txnid = ti->mti_txnid;
2643 meta = env->me_metas[txn->mt_txnid & 1];
2645 meta = mdb_env_pick_meta(env);
2646 txn->mt_txnid = meta->mm_txnid;
2650 if (txn->mt_txnid == mdb_debug_start)
2653 txn->mt_child = NULL;
2654 txn->mt_loose_pgs = NULL;
2655 txn->mt_loose_count = 0;
2656 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2657 txn->mt_u.dirty_list = env->me_dirty_list;
2658 txn->mt_u.dirty_list[0].mid = 0;
2659 txn->mt_free_pgs = env->me_free_pgs;
2660 txn->mt_free_pgs[0] = 0;
2661 txn->mt_spill_pgs = NULL;
2663 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2666 /* Copy the DB info and flags */
2667 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2669 /* Moved to here to avoid a data race in read TXNs */
2670 txn->mt_next_pgno = meta->mm_last_pg+1;
2672 txn->mt_flags = flags;
2675 txn->mt_numdbs = env->me_numdbs;
2676 for (i=2; i<txn->mt_numdbs; i++) {
2677 x = env->me_dbflags[i];
2678 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2679 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2681 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2682 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2684 if (env->me_maxpg < txn->mt_next_pgno) {
2685 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2686 return MDB_MAP_RESIZED;
2693 mdb_txn_renew(MDB_txn *txn)
2697 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2700 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2701 DPUTS("environment had fatal error, must shutdown!");
2705 rc = mdb_txn_renew0(txn);
2706 if (rc == MDB_SUCCESS) {
2707 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2708 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2709 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2715 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2719 int rc, size, tsize;
2721 flags &= MDB_TXN_BEGIN_FLAGS;
2722 flags |= env->me_flags & MDB_WRITEMAP;
2724 if (env->me_flags & MDB_FATAL_ERROR) {
2725 DPUTS("environment had fatal error, must shutdown!");
2728 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2732 /* Nested transactions: Max 1 child, write txns only, no writemap */
2733 flags |= parent->mt_flags;
2734 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2735 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2737 /* Child txns save MDB_pgstate and use own copy of cursors */
2738 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2739 size += tsize = sizeof(MDB_ntxn);
2740 } else if (flags & MDB_RDONLY) {
2741 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2742 size += tsize = sizeof(MDB_txn);
2744 /* Reuse preallocated write txn. However, do not touch it until
2745 * mdb_txn_renew0() succeeds, since it currently may be active.
2750 if ((txn = calloc(1, size)) == NULL) {
2751 DPRINTF(("calloc: %s", strerror(errno)));
2754 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2755 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2756 txn->mt_flags = flags;
2761 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2762 txn->mt_dbiseqs = parent->mt_dbiseqs;
2763 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2764 if (!txn->mt_u.dirty_list ||
2765 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2767 free(txn->mt_u.dirty_list);
2771 txn->mt_txnid = parent->mt_txnid;
2772 txn->mt_dirty_room = parent->mt_dirty_room;
2773 txn->mt_u.dirty_list[0].mid = 0;
2774 txn->mt_spill_pgs = NULL;
2775 txn->mt_next_pgno = parent->mt_next_pgno;
2776 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2777 parent->mt_child = txn;
2778 txn->mt_parent = parent;
2779 txn->mt_numdbs = parent->mt_numdbs;
2780 txn->mt_dbxs = parent->mt_dbxs;
2781 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2782 /* Copy parent's mt_dbflags, but clear DB_NEW */
2783 for (i=0; i<txn->mt_numdbs; i++)
2784 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2786 ntxn = (MDB_ntxn *)txn;
2787 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2788 if (env->me_pghead) {
2789 size = MDB_IDL_SIZEOF(env->me_pghead);
2790 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2792 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2797 rc = mdb_cursor_shadow(parent, txn);
2799 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2800 } else { /* MDB_RDONLY */
2801 txn->mt_dbiseqs = env->me_dbiseqs;
2803 rc = mdb_txn_renew0(txn);
2806 if (txn != env->me_txn0)
2809 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2811 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2812 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2813 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2820 mdb_txn_env(MDB_txn *txn)
2822 if(!txn) return NULL;
2827 mdb_txn_id(MDB_txn *txn)
2830 return txn->mt_txnid;
2833 /** Export or close DBI handles opened in this txn. */
2835 mdb_dbis_update(MDB_txn *txn, int keep)
2838 MDB_dbi n = txn->mt_numdbs;
2839 MDB_env *env = txn->mt_env;
2840 unsigned char *tdbflags = txn->mt_dbflags;
2842 for (i = n; --i >= 2;) {
2843 if (tdbflags[i] & DB_NEW) {
2845 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2847 char *ptr = env->me_dbxs[i].md_name.mv_data;
2849 env->me_dbxs[i].md_name.mv_data = NULL;
2850 env->me_dbxs[i].md_name.mv_size = 0;
2851 env->me_dbflags[i] = 0;
2852 env->me_dbiseqs[i]++;
2858 if (keep && env->me_numdbs < n)
2862 /** End a transaction, except successful commit of a nested transaction.
2863 * May be called twice for readonly txns: First reset it, then abort.
2864 * @param[in] txn the transaction handle to end
2865 * @param[in] mode why and how to end the transaction
2868 mdb_txn_end(MDB_txn *txn, unsigned mode)
2870 MDB_env *env = txn->mt_env;
2872 static const char *const names[] = MDB_END_NAMES;
2875 /* Export or close DBI handles opened in this txn */
2876 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2878 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2879 names[mode & MDB_END_OPMASK],
2880 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2881 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2883 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2884 if (txn->mt_u.reader) {
2885 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2886 if (!(env->me_flags & MDB_NOTLS)) {
2887 txn->mt_u.reader = NULL; /* txn does not own reader */
2888 } else if (mode & MDB_END_SLOT) {
2889 txn->mt_u.reader->mr_pid = 0;
2890 txn->mt_u.reader = NULL;
2891 } /* else txn owns the slot until it does MDB_END_SLOT */
2893 txn->mt_numdbs = 0; /* prevent further DBI activity */
2894 txn->mt_flags |= MDB_TXN_FINISHED;
2895 txn->mt_dbxs = NULL; /* mark txn as reset */
2896 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2897 pgno_t *pghead = env->me_pghead;
2899 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2900 mdb_cursors_close(txn, 0);
2901 if (!(env->me_flags & MDB_WRITEMAP)) {
2902 mdb_dlist_free(txn);
2906 txn->mt_flags = MDB_TXN_FINISHED;
2908 if (!txn->mt_parent) {
2909 mdb_midl_shrink(&txn->mt_free_pgs);
2910 env->me_free_pgs = txn->mt_free_pgs;
2912 env->me_pghead = NULL;
2916 mode = 0; /* txn == env->me_txn0, do not free() it */
2918 /* The writer mutex was locked in mdb_txn_begin. */
2920 UNLOCK_MUTEX(env->me_wmutex);
2922 txn->mt_parent->mt_child = NULL;
2923 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2924 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2925 mdb_midl_free(txn->mt_free_pgs);
2926 mdb_midl_free(txn->mt_spill_pgs);
2927 free(txn->mt_u.dirty_list);
2930 mdb_midl_free(pghead);
2933 if (mode & MDB_END_FREE)
2938 mdb_txn_reset(MDB_txn *txn)
2943 /* This call is only valid for read-only txns */
2944 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2947 mdb_txn_end(txn, MDB_END_RESET);
2951 mdb_txn_abort(MDB_txn *txn)
2957 mdb_txn_abort(txn->mt_child);
2959 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
2962 /** Save the freelist as of this transaction to the freeDB.
2963 * This changes the freelist. Keep trying until it stabilizes.
2966 mdb_freelist_save(MDB_txn *txn)
2968 /* env->me_pghead[] can grow and shrink during this call.
2969 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2970 * Page numbers cannot disappear from txn->mt_free_pgs[].
2973 MDB_env *env = txn->mt_env;
2974 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2975 txnid_t pglast = 0, head_id = 0;
2976 pgno_t freecnt = 0, *free_pgs, *mop;
2977 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2979 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2981 if (env->me_pghead) {
2982 /* Make sure first page of freeDB is touched and on freelist */
2983 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2984 if (rc && rc != MDB_NOTFOUND)
2988 if (!env->me_pghead && txn->mt_loose_pgs) {
2989 /* Put loose page numbers in mt_free_pgs, since
2990 * we may be unable to return them to me_pghead.
2992 MDB_page *mp = txn->mt_loose_pgs;
2993 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2995 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2996 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2997 txn->mt_loose_pgs = NULL;
2998 txn->mt_loose_count = 0;
3001 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3002 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3003 ? SSIZE_MAX : maxfree_1pg;
3006 /* Come back here after each Put() in case freelist changed */
3011 /* If using records from freeDB which we have not yet
3012 * deleted, delete them and any we reserved for me_pghead.
3014 while (pglast < env->me_pglast) {
3015 rc = mdb_cursor_first(&mc, &key, NULL);
3018 pglast = head_id = *(txnid_t *)key.mv_data;
3019 total_room = head_room = 0;
3020 mdb_tassert(txn, pglast <= env->me_pglast);
3021 rc = mdb_cursor_del(&mc, 0);
3026 /* Save the IDL of pages freed by this txn, to a single record */
3027 if (freecnt < txn->mt_free_pgs[0]) {
3029 /* Make sure last page of freeDB is touched and on freelist */
3030 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3031 if (rc && rc != MDB_NOTFOUND)
3034 free_pgs = txn->mt_free_pgs;
3035 /* Write to last page of freeDB */
3036 key.mv_size = sizeof(txn->mt_txnid);
3037 key.mv_data = &txn->mt_txnid;
3039 freecnt = free_pgs[0];
3040 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3041 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3044 /* Retry if mt_free_pgs[] grew during the Put() */
3045 free_pgs = txn->mt_free_pgs;
3046 } while (freecnt < free_pgs[0]);
3047 mdb_midl_sort(free_pgs);
3048 memcpy(data.mv_data, free_pgs, data.mv_size);
3051 unsigned int i = free_pgs[0];
3052 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3053 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3055 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3061 mop = env->me_pghead;
3062 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3064 /* Reserve records for me_pghead[]. Split it if multi-page,
3065 * to avoid searching freeDB for a page range. Use keys in
3066 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3068 if (total_room >= mop_len) {
3069 if (total_room == mop_len || --more < 0)
3071 } else if (head_room >= maxfree_1pg && head_id > 1) {
3072 /* Keep current record (overflow page), add a new one */
3076 /* (Re)write {key = head_id, IDL length = head_room} */
3077 total_room -= head_room;
3078 head_room = mop_len - total_room;
3079 if (head_room > maxfree_1pg && head_id > 1) {
3080 /* Overflow multi-page for part of me_pghead */
3081 head_room /= head_id; /* amortize page sizes */
3082 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3083 } else if (head_room < 0) {
3084 /* Rare case, not bothering to delete this record */
3087 key.mv_size = sizeof(head_id);
3088 key.mv_data = &head_id;
3089 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3090 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3093 /* IDL is initially empty, zero out at least the length */
3094 pgs = (pgno_t *)data.mv_data;
3095 j = head_room > clean_limit ? head_room : 0;
3099 total_room += head_room;
3102 /* Return loose page numbers to me_pghead, though usually none are
3103 * left at this point. The pages themselves remain in dirty_list.
3105 if (txn->mt_loose_pgs) {
3106 MDB_page *mp = txn->mt_loose_pgs;
3107 unsigned count = txn->mt_loose_count;
3109 /* Room for loose pages + temp IDL with same */
3110 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3112 mop = env->me_pghead;
3113 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3114 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3115 loose[ ++count ] = mp->mp_pgno;
3117 mdb_midl_sort(loose);
3118 mdb_midl_xmerge(mop, loose);
3119 txn->mt_loose_pgs = NULL;
3120 txn->mt_loose_count = 0;
3124 /* Fill in the reserved me_pghead records */
3130 rc = mdb_cursor_first(&mc, &key, &data);
3131 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3132 txnid_t id = *(txnid_t *)key.mv_data;
3133 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3136 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3138 if (len > mop_len) {
3140 data.mv_size = (len + 1) * sizeof(MDB_ID);
3142 data.mv_data = mop -= len;
3145 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3147 if (rc || !(mop_len -= len))
3154 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3155 * @param[in] txn the transaction that's being committed
3156 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3157 * @return 0 on success, non-zero on failure.
3160 mdb_page_flush(MDB_txn *txn, int keep)
3162 MDB_env *env = txn->mt_env;
3163 MDB_ID2L dl = txn->mt_u.dirty_list;
3164 unsigned psize = env->me_psize, j;
3165 int i, pagecount = dl[0].mid, rc;
3166 size_t size = 0, pos = 0;
3168 MDB_page *dp = NULL;
3172 struct iovec iov[MDB_COMMIT_PAGES];
3173 ssize_t wpos = 0, wsize = 0, wres;
3174 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3180 if (env->me_flags & MDB_WRITEMAP) {
3181 /* Clear dirty flags */
3182 while (++i <= pagecount) {
3184 /* Don't flush this page yet */
3185 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3186 dp->mp_flags &= ~P_KEEP;
3190 dp->mp_flags &= ~P_DIRTY;
3195 /* Write the pages */
3197 if (++i <= pagecount) {
3199 /* Don't flush this page yet */
3200 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3201 dp->mp_flags &= ~P_KEEP;
3206 /* clear dirty flag */
3207 dp->mp_flags &= ~P_DIRTY;
3210 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3215 /* Windows actually supports scatter/gather I/O, but only on
3216 * unbuffered file handles. Since we're relying on the OS page
3217 * cache for all our data, that's self-defeating. So we just
3218 * write pages one at a time. We use the ov structure to set
3219 * the write offset, to at least save the overhead of a Seek
3222 DPRINTF(("committing page %"Z"u", pgno));
3223 memset(&ov, 0, sizeof(ov));
3224 ov.Offset = pos & 0xffffffff;
3225 ov.OffsetHigh = pos >> 16 >> 16;
3226 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3228 DPRINTF(("WriteFile: %d", rc));
3232 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3233 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3236 /* Write previous page(s) */
3237 #ifdef MDB_USE_PWRITEV
3238 wres = pwritev(env->me_fd, iov, n, wpos);
3241 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3244 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3248 DPRINTF(("lseek: %s", strerror(rc)));
3251 wres = writev(env->me_fd, iov, n);
3254 if (wres != wsize) {
3259 DPRINTF(("Write error: %s", strerror(rc)));
3261 rc = EIO; /* TODO: Use which error code? */
3262 DPUTS("short write, filesystem full?");
3273 DPRINTF(("committing page %"Z"u", pgno));
3274 next_pos = pos + size;
3275 iov[n].iov_len = size;
3276 iov[n].iov_base = (char *)dp;
3282 /* MIPS has cache coherency issues, this is a no-op everywhere else
3283 * Note: for any size >= on-chip cache size, entire on-chip cache is
3286 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3288 for (i = keep; ++i <= pagecount; ) {
3290 /* This is a page we skipped above */
3293 dl[j].mid = dp->mp_pgno;
3296 mdb_dpage_free(env, dp);
3301 txn->mt_dirty_room += i - j;
3307 mdb_txn_commit(MDB_txn *txn)
3310 unsigned int i, end_mode;
3316 /* mdb_txn_end() mode for a commit which writes nothing */
3317 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3319 if (txn->mt_child) {
3320 rc = mdb_txn_commit(txn->mt_child);
3327 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3331 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3332 DPUTS("txn has failed/finished, can't commit");
3334 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3339 if (txn->mt_parent) {
3340 MDB_txn *parent = txn->mt_parent;
3344 unsigned x, y, len, ps_len;
3346 /* Append our free list to parent's */
3347 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3350 mdb_midl_free(txn->mt_free_pgs);
3351 /* Failures after this must either undo the changes
3352 * to the parent or set MDB_TXN_ERROR in the parent.
3355 parent->mt_next_pgno = txn->mt_next_pgno;
3356 parent->mt_flags = txn->mt_flags;
3358 /* Merge our cursors into parent's and close them */
3359 mdb_cursors_close(txn, 1);
3361 /* Update parent's DB table. */
3362 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3363 parent->mt_numdbs = txn->mt_numdbs;
3364 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3365 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3366 for (i=2; i<txn->mt_numdbs; i++) {
3367 /* preserve parent's DB_NEW status */
3368 x = parent->mt_dbflags[i] & DB_NEW;
3369 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3372 dst = parent->mt_u.dirty_list;
3373 src = txn->mt_u.dirty_list;
3374 /* Remove anything in our dirty list from parent's spill list */
3375 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3377 pspill[0] = (pgno_t)-1;
3378 /* Mark our dirty pages as deleted in parent spill list */
3379 for (i=0, len=src[0].mid; ++i <= len; ) {
3380 MDB_ID pn = src[i].mid << 1;
3381 while (pn > pspill[x])
3383 if (pn == pspill[x]) {
3388 /* Squash deleted pagenums if we deleted any */
3389 for (x=y; ++x <= ps_len; )
3390 if (!(pspill[x] & 1))
3391 pspill[++y] = pspill[x];
3395 /* Find len = length of merging our dirty list with parent's */
3397 dst[0].mid = 0; /* simplify loops */
3398 if (parent->mt_parent) {
3399 len = x + src[0].mid;
3400 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3401 for (i = x; y && i; y--) {
3402 pgno_t yp = src[y].mid;
3403 while (yp < dst[i].mid)
3405 if (yp == dst[i].mid) {
3410 } else { /* Simplify the above for single-ancestor case */
3411 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3413 /* Merge our dirty list with parent's */
3415 for (i = len; y; dst[i--] = src[y--]) {
3416 pgno_t yp = src[y].mid;
3417 while (yp < dst[x].mid)
3418 dst[i--] = dst[x--];
3419 if (yp == dst[x].mid)
3420 free(dst[x--].mptr);
3422 mdb_tassert(txn, i == x);
3424 free(txn->mt_u.dirty_list);
3425 parent->mt_dirty_room = txn->mt_dirty_room;
3426 if (txn->mt_spill_pgs) {
3427 if (parent->mt_spill_pgs) {
3428 /* TODO: Prevent failure here, so parent does not fail */
3429 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3431 parent->mt_flags |= MDB_TXN_ERROR;
3432 mdb_midl_free(txn->mt_spill_pgs);
3433 mdb_midl_sort(parent->mt_spill_pgs);
3435 parent->mt_spill_pgs = txn->mt_spill_pgs;
3439 /* Append our loose page list to parent's */
3440 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3442 *lp = txn->mt_loose_pgs;
3443 parent->mt_loose_count += txn->mt_loose_count;
3445 parent->mt_child = NULL;
3446 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3451 if (txn != env->me_txn) {
3452 DPUTS("attempt to commit unknown transaction");
3457 mdb_cursors_close(txn, 0);
3459 if (!txn->mt_u.dirty_list[0].mid &&
3460 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3463 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3464 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3466 /* Update DB root pointers */
3467 if (txn->mt_numdbs > 2) {
3471 data.mv_size = sizeof(MDB_db);
3473 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3474 for (i = 2; i < txn->mt_numdbs; i++) {
3475 if (txn->mt_dbflags[i] & DB_DIRTY) {
3476 if (TXN_DBI_CHANGED(txn, i)) {
3480 data.mv_data = &txn->mt_dbs[i];
3481 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3489 rc = mdb_freelist_save(txn);
3493 mdb_midl_free(env->me_pghead);
3494 env->me_pghead = NULL;
3495 mdb_midl_shrink(&txn->mt_free_pgs);
3501 if ((rc = mdb_page_flush(txn, 0)) ||
3502 (rc = mdb_env_sync(env, 0)) ||
3503 (rc = mdb_env_write_meta(txn)))
3505 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3508 mdb_txn_end(txn, end_mode);
3516 /** Read the environment parameters of a DB environment before
3517 * mapping it into memory.
3518 * @param[in] env the environment handle
3519 * @param[out] meta address of where to store the meta information
3520 * @return 0 on success, non-zero on failure.
3523 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3529 enum { Size = sizeof(pbuf) };
3531 /* We don't know the page size yet, so use a minimum value.
3532 * Read both meta pages so we can use the latest one.
3535 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3539 memset(&ov, 0, sizeof(ov));
3541 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3542 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3545 rc = pread(env->me_fd, &pbuf, Size, off);
3548 if (rc == 0 && off == 0)
3550 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3551 DPRINTF(("read: %s", mdb_strerror(rc)));
3555 p = (MDB_page *)&pbuf;
3557 if (!F_ISSET(p->mp_flags, P_META)) {
3558 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3563 if (m->mm_magic != MDB_MAGIC) {
3564 DPUTS("meta has invalid magic");
3568 if (m->mm_version != MDB_DATA_VERSION) {
3569 DPRINTF(("database is version %u, expected version %u",
3570 m->mm_version, MDB_DATA_VERSION));
3571 return MDB_VERSION_MISMATCH;
3574 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3580 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3582 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3584 meta->mm_magic = MDB_MAGIC;
3585 meta->mm_version = MDB_DATA_VERSION;
3586 meta->mm_mapsize = env->me_mapsize;
3587 meta->mm_psize = env->me_psize;
3588 meta->mm_last_pg = 1;
3589 meta->mm_flags = env->me_flags & 0xffff;
3590 meta->mm_flags |= MDB_INTEGERKEY;
3591 meta->mm_dbs[0].md_root = P_INVALID;
3592 meta->mm_dbs[1].md_root = P_INVALID;
3595 /** Write the environment parameters of a freshly created DB environment.
3596 * @param[in] env the environment handle
3597 * @param[in] meta the #MDB_meta to write
3598 * @return 0 on success, non-zero on failure.
3601 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3609 memset(&ov, 0, sizeof(ov));
3610 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3612 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3615 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3616 len = pwrite(fd, ptr, size, pos); \
3617 if (len == -1 && ErrCode() == EINTR) continue; \
3618 rc = (len >= 0); break; } while(1)
3621 DPUTS("writing new meta page");
3623 psize = env->me_psize;
3625 p = calloc(2, psize);
3630 p->mp_flags = P_META;
3631 *(MDB_meta *)METADATA(p) = *meta;
3633 q = (MDB_page *)((char *)p + psize);
3635 q->mp_flags = P_META;
3636 *(MDB_meta *)METADATA(q) = *meta;
3638 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3641 else if ((unsigned) len == psize * 2)
3649 /** Update the environment info to commit a transaction.
3650 * @param[in] txn the transaction that's being committed
3651 * @return 0 on success, non-zero on failure.
3654 mdb_env_write_meta(MDB_txn *txn)
3657 MDB_meta meta, metab, *mp;
3661 int rc, len, toggle;
3670 toggle = txn->mt_txnid & 1;
3671 DPRINTF(("writing meta page %d for root page %"Z"u",
3672 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3675 flags = env->me_flags;
3676 mp = env->me_metas[toggle];
3677 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3678 /* Persist any increases of mapsize config */
3679 if (mapsize < env->me_mapsize)
3680 mapsize = env->me_mapsize;
3682 if (flags & MDB_WRITEMAP) {
3683 mp->mm_mapsize = mapsize;
3684 mp->mm_dbs[0] = txn->mt_dbs[0];
3685 mp->mm_dbs[1] = txn->mt_dbs[1];
3686 mp->mm_last_pg = txn->mt_next_pgno - 1;
3687 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3688 !(defined(__i386__) || defined(__x86_64__))
3689 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3690 __sync_synchronize();
3692 mp->mm_txnid = txn->mt_txnid;
3693 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3694 unsigned meta_size = env->me_psize;
3695 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3698 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3699 if (meta_size < env->me_os_psize)
3700 meta_size += meta_size;
3705 if (MDB_MSYNC(ptr, meta_size, rc)) {
3712 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3713 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3715 meta.mm_mapsize = mapsize;
3716 meta.mm_dbs[0] = txn->mt_dbs[0];
3717 meta.mm_dbs[1] = txn->mt_dbs[1];
3718 meta.mm_last_pg = txn->mt_next_pgno - 1;
3719 meta.mm_txnid = txn->mt_txnid;
3721 off = offsetof(MDB_meta, mm_mapsize);
3722 ptr = (char *)&meta + off;
3723 len = sizeof(MDB_meta) - off;
3725 off += env->me_psize;
3728 /* Write to the SYNC fd */
3729 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3732 memset(&ov, 0, sizeof(ov));
3734 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3739 rc = pwrite(mfd, ptr, len, off);
3742 rc = rc < 0 ? ErrCode() : EIO;
3747 DPUTS("write failed, disk error?");
3748 /* On a failure, the pagecache still contains the new data.
3749 * Write some old data back, to prevent it from being used.
3750 * Use the non-SYNC fd; we know it will fail anyway.
3752 meta.mm_last_pg = metab.mm_last_pg;
3753 meta.mm_txnid = metab.mm_txnid;
3755 memset(&ov, 0, sizeof(ov));
3757 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3759 r2 = pwrite(env->me_fd, ptr, len, off);
3760 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3763 env->me_flags |= MDB_FATAL_ERROR;
3766 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3767 CACHEFLUSH(env->me_map + off, len, DCACHE);
3769 /* Memory ordering issues are irrelevant; since the entire writer
3770 * is wrapped by wmutex, all of these changes will become visible
3771 * after the wmutex is unlocked. Since the DB is multi-version,
3772 * readers will get consistent data regardless of how fresh or
3773 * how stale their view of these values is.
3776 env->me_txns->mti_txnid = txn->mt_txnid;
3781 /** Check both meta pages to see which one is newer.
3782 * @param[in] env the environment handle
3783 * @return newest #MDB_meta.
3786 mdb_env_pick_meta(const MDB_env *env)
3788 MDB_meta *const *metas = env->me_metas;
3789 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3793 mdb_env_create(MDB_env **env)
3797 e = calloc(1, sizeof(MDB_env));
3801 e->me_maxreaders = DEFAULT_READERS;
3802 e->me_maxdbs = e->me_numdbs = 2;
3803 e->me_fd = INVALID_HANDLE_VALUE;
3804 e->me_lfd = INVALID_HANDLE_VALUE;
3805 e->me_mfd = INVALID_HANDLE_VALUE;
3806 #ifdef MDB_USE_POSIX_SEM
3807 e->me_rmutex = SEM_FAILED;
3808 e->me_wmutex = SEM_FAILED;
3810 e->me_pid = getpid();
3811 GET_PAGESIZE(e->me_os_psize);
3812 VGMEMP_CREATE(e,0,0);
3818 mdb_env_map(MDB_env *env, void *addr)
3821 unsigned int flags = env->me_flags;
3825 LONG sizelo, sizehi;
3828 if (flags & MDB_RDONLY) {
3829 /* Don't set explicit map size, use whatever exists */
3834 msize = env->me_mapsize;
3835 sizelo = msize & 0xffffffff;
3836 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3838 /* Windows won't create mappings for zero length files.
3839 * and won't map more than the file size.
3840 * Just set the maxsize right now.
3842 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3843 || !SetEndOfFile(env->me_fd)
3844 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3848 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3849 PAGE_READWRITE : PAGE_READONLY,
3850 sizehi, sizelo, NULL);
3853 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3854 FILE_MAP_WRITE : FILE_MAP_READ,
3856 rc = env->me_map ? 0 : ErrCode();
3861 int prot = PROT_READ;
3862 if (flags & MDB_WRITEMAP) {
3864 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3867 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3869 if (env->me_map == MAP_FAILED) {
3874 if (flags & MDB_NORDAHEAD) {
3875 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3877 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3879 #ifdef POSIX_MADV_RANDOM
3880 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3881 #endif /* POSIX_MADV_RANDOM */
3882 #endif /* MADV_RANDOM */
3886 /* Can happen because the address argument to mmap() is just a
3887 * hint. mmap() can pick another, e.g. if the range is in use.
3888 * The MAP_FIXED flag would prevent that, but then mmap could
3889 * instead unmap existing pages to make room for the new map.
3891 if (addr && env->me_map != addr)
3892 return EBUSY; /* TODO: Make a new MDB_* error code? */
3894 p = (MDB_page *)env->me_map;
3895 env->me_metas[0] = METADATA(p);
3896 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3902 mdb_env_set_mapsize(MDB_env *env, size_t size)
3904 /* If env is already open, caller is responsible for making
3905 * sure there are no active txns.
3913 meta = mdb_env_pick_meta(env);
3915 size = meta->mm_mapsize;
3917 /* Silently round up to minimum if the size is too small */
3918 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3922 munmap(env->me_map, env->me_mapsize);
3923 env->me_mapsize = size;
3924 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3925 rc = mdb_env_map(env, old);
3929 env->me_mapsize = size;
3931 env->me_maxpg = env->me_mapsize / env->me_psize;
3936 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3940 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3945 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3947 if (env->me_map || readers < 1)
3949 env->me_maxreaders = readers;
3954 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3956 if (!env || !readers)
3958 *readers = env->me_maxreaders;
3963 mdb_fsize(HANDLE fd, size_t *size)
3966 LARGE_INTEGER fsize;
3968 if (!GetFileSizeEx(fd, &fsize))
3971 *size = fsize.QuadPart;
3983 #ifdef BROKEN_FDATASYNC
3984 #include <sys/utsname.h>
3985 #include <sys/vfs.h>
3988 /** Further setup required for opening an LMDB environment
3991 mdb_env_open2(MDB_env *env)
3993 unsigned int flags = env->me_flags;
3994 int i, newenv = 0, rc;
3998 /* See if we should use QueryLimited */
4000 if ((rc & 0xff) > 5)
4001 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4003 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4006 #ifdef BROKEN_FDATASYNC
4007 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4008 * https://lkml.org/lkml/2012/9/3/83
4009 * Kernels after 3.6-rc6 are known good.
4010 * https://lkml.org/lkml/2012/9/10/556
4011 * See if the DB is on ext3/ext4, then check for new enough kernel
4012 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4017 fstatfs(env->me_fd, &st);
4018 while (st.f_type == 0xEF53) {
4022 if (uts.release[0] < '3') {
4023 if (!strncmp(uts.release, "2.6.32.", 7)) {
4024 i = atoi(uts.release+7);
4026 break; /* 2.6.32.60 and newer is OK */
4027 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4028 i = atoi(uts.release+7);
4030 break; /* 2.6.34.15 and newer is OK */
4032 } else if (uts.release[0] == '3') {
4033 i = atoi(uts.release+2);
4035 break; /* 3.6 and newer is OK */
4037 i = atoi(uts.release+4);
4039 break; /* 3.5.4 and newer is OK */
4040 } else if (i == 2) {
4041 i = atoi(uts.release+4);
4043 break; /* 3.2.30 and newer is OK */
4045 } else { /* 4.x and newer is OK */
4048 env->me_flags |= MDB_FSYNCONLY;
4054 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4057 DPUTS("new mdbenv");
4059 env->me_psize = env->me_os_psize;
4060 if (env->me_psize > MAX_PAGESIZE)
4061 env->me_psize = MAX_PAGESIZE;
4062 memset(&meta, 0, sizeof(meta));
4063 mdb_env_init_meta0(env, &meta);
4064 meta.mm_mapsize = DEFAULT_MAPSIZE;
4066 env->me_psize = meta.mm_psize;
4069 /* Was a mapsize configured? */
4070 if (!env->me_mapsize) {
4071 env->me_mapsize = meta.mm_mapsize;
4074 /* Make sure mapsize >= committed data size. Even when using
4075 * mm_mapsize, which could be broken in old files (ITS#7789).
4077 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4078 if (env->me_mapsize < minsize)
4079 env->me_mapsize = minsize;
4081 meta.mm_mapsize = env->me_mapsize;
4083 if (newenv && !(flags & MDB_FIXEDMAP)) {
4084 /* mdb_env_map() may grow the datafile. Write the metapages
4085 * first, so the file will be valid if initialization fails.
4086 * Except with FIXEDMAP, since we do not yet know mm_address.
4087 * We could fill in mm_address later, but then a different
4088 * program might end up doing that - one with a memory layout
4089 * and map address which does not suit the main program.
4091 rc = mdb_env_init_meta(env, &meta);
4097 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4102 if (flags & MDB_FIXEDMAP)
4103 meta.mm_address = env->me_map;
4104 i = mdb_env_init_meta(env, &meta);
4105 if (i != MDB_SUCCESS) {
4110 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4111 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4113 #if !(MDB_MAXKEYSIZE)
4114 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4116 env->me_maxpg = env->me_mapsize / env->me_psize;
4120 MDB_meta *meta = mdb_env_pick_meta(env);
4121 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4123 DPRINTF(("opened database version %u, pagesize %u",
4124 meta->mm_version, env->me_psize));
4125 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4126 DPRINTF(("depth: %u", db->md_depth));
4127 DPRINTF(("entries: %"Z"u", db->md_entries));
4128 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4129 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4130 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4131 DPRINTF(("root: %"Z"u", db->md_root));
4139 /** Release a reader thread's slot in the reader lock table.
4140 * This function is called automatically when a thread exits.
4141 * @param[in] ptr This points to the slot in the reader lock table.
4144 mdb_env_reader_dest(void *ptr)
4146 MDB_reader *reader = ptr;
4152 /** Junk for arranging thread-specific callbacks on Windows. This is
4153 * necessarily platform and compiler-specific. Windows supports up
4154 * to 1088 keys. Let's assume nobody opens more than 64 environments
4155 * in a single process, for now. They can override this if needed.
4157 #ifndef MAX_TLS_KEYS
4158 #define MAX_TLS_KEYS 64
4160 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4161 static int mdb_tls_nkeys;
4163 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4167 case DLL_PROCESS_ATTACH: break;
4168 case DLL_THREAD_ATTACH: break;
4169 case DLL_THREAD_DETACH:
4170 for (i=0; i<mdb_tls_nkeys; i++) {
4171 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4173 mdb_env_reader_dest(r);
4177 case DLL_PROCESS_DETACH: break;
4182 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4184 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4188 /* Force some symbol references.
4189 * _tls_used forces the linker to create the TLS directory if not already done
4190 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4192 #pragma comment(linker, "/INCLUDE:_tls_used")
4193 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4194 #pragma const_seg(".CRT$XLB")
4195 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4196 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4199 #pragma comment(linker, "/INCLUDE:__tls_used")
4200 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4201 #pragma data_seg(".CRT$XLB")
4202 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4204 #endif /* WIN 32/64 */
4205 #endif /* !__GNUC__ */
4208 /** Downgrade the exclusive lock on the region back to shared */
4210 mdb_env_share_locks(MDB_env *env, int *excl)
4213 MDB_meta *meta = mdb_env_pick_meta(env);
4215 env->me_txns->mti_txnid = meta->mm_txnid;
4220 /* First acquire a shared lock. The Unlock will
4221 * then release the existing exclusive lock.
4223 memset(&ov, 0, sizeof(ov));
4224 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4227 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4233 struct flock lock_info;
4234 /* The shared lock replaces the existing lock */
4235 memset((void *)&lock_info, 0, sizeof(lock_info));
4236 lock_info.l_type = F_RDLCK;
4237 lock_info.l_whence = SEEK_SET;
4238 lock_info.l_start = 0;
4239 lock_info.l_len = 1;
4240 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4241 (rc = ErrCode()) == EINTR) ;
4242 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4249 /** Try to get exclusive lock, otherwise shared.
4250 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4253 mdb_env_excl_lock(MDB_env *env, int *excl)
4257 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4261 memset(&ov, 0, sizeof(ov));
4262 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4269 struct flock lock_info;
4270 memset((void *)&lock_info, 0, sizeof(lock_info));
4271 lock_info.l_type = F_WRLCK;
4272 lock_info.l_whence = SEEK_SET;
4273 lock_info.l_start = 0;
4274 lock_info.l_len = 1;
4275 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4276 (rc = ErrCode()) == EINTR) ;
4280 # ifndef MDB_USE_POSIX_MUTEX
4281 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4284 lock_info.l_type = F_RDLCK;
4285 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4286 (rc = ErrCode()) == EINTR) ;
4296 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4298 * @(#) $Revision: 5.1 $
4299 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4300 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4302 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4306 * Please do not copyright this code. This code is in the public domain.
4308 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4309 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4310 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4311 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4312 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4313 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4314 * PERFORMANCE OF THIS SOFTWARE.
4317 * chongo <Landon Curt Noll> /\oo/\
4318 * http://www.isthe.com/chongo/
4320 * Share and Enjoy! :-)
4323 typedef unsigned long long mdb_hash_t;
4324 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4326 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4327 * @param[in] val value to hash
4328 * @param[in] hval initial value for hash
4329 * @return 64 bit hash
4331 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4332 * hval arg on the first call.
4335 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4337 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4338 unsigned char *end = s + val->mv_size;
4340 * FNV-1a hash each octet of the string
4343 /* xor the bottom with the current octet */
4344 hval ^= (mdb_hash_t)*s++;
4346 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4347 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4348 (hval << 7) + (hval << 8) + (hval << 40);
4350 /* return our new hash value */
4354 /** Hash the string and output the encoded hash.
4355 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4356 * very short name limits. We don't care about the encoding being reversible,
4357 * we just want to preserve as many bits of the input as possible in a
4358 * small printable string.
4359 * @param[in] str string to hash
4360 * @param[out] encbuf an array of 11 chars to hold the hash
4362 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4365 mdb_pack85(unsigned long l, char *out)
4369 for (i=0; i<5; i++) {
4370 *out++ = mdb_a85[l % 85];
4376 mdb_hash_enc(MDB_val *val, char *encbuf)
4378 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4380 mdb_pack85(h, encbuf);
4381 mdb_pack85(h>>32, encbuf+5);
4386 /** Open and/or initialize the lock region for the environment.
4387 * @param[in] env The LMDB environment.
4388 * @param[in] lpath The pathname of the file used for the lock region.
4389 * @param[in] mode The Unix permissions for the file, if we create it.
4390 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4391 * @return 0 on success, non-zero on failure.
4394 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4397 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4399 # define MDB_ERRCODE_ROFS EROFS
4400 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4401 # define MDB_CLOEXEC O_CLOEXEC
4404 # define MDB_CLOEXEC 0
4411 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4412 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4413 FILE_ATTRIBUTE_NORMAL, NULL);
4415 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4417 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4419 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4424 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4425 /* Lose record locks when exec*() */
4426 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4427 fcntl(env->me_lfd, F_SETFD, fdflags);
4430 if (!(env->me_flags & MDB_NOTLS)) {
4431 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4434 env->me_flags |= MDB_ENV_TXKEY;
4436 /* Windows TLS callbacks need help finding their TLS info. */
4437 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4441 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4445 /* Try to get exclusive lock. If we succeed, then
4446 * nobody is using the lock region and we should initialize it.
4448 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4451 size = GetFileSize(env->me_lfd, NULL);
4453 size = lseek(env->me_lfd, 0, SEEK_END);
4454 if (size == -1) goto fail_errno;
4456 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4457 if (size < rsize && *excl > 0) {
4459 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4460 || !SetEndOfFile(env->me_lfd))
4463 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4467 size = rsize - sizeof(MDB_txninfo);
4468 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4473 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4475 if (!mh) goto fail_errno;
4476 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4478 if (!env->me_txns) goto fail_errno;
4480 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4482 if (m == MAP_FAILED) goto fail_errno;
4488 BY_HANDLE_FILE_INFORMATION stbuf;
4497 if (!mdb_sec_inited) {
4498 InitializeSecurityDescriptor(&mdb_null_sd,
4499 SECURITY_DESCRIPTOR_REVISION);
4500 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4501 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4502 mdb_all_sa.bInheritHandle = FALSE;
4503 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4506 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4507 idbuf.volume = stbuf.dwVolumeSerialNumber;
4508 idbuf.nhigh = stbuf.nFileIndexHigh;
4509 idbuf.nlow = stbuf.nFileIndexLow;
4510 val.mv_data = &idbuf;
4511 val.mv_size = sizeof(idbuf);
4512 mdb_hash_enc(&val, encbuf);
4513 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4514 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4515 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4516 if (!env->me_rmutex) goto fail_errno;
4517 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4518 if (!env->me_wmutex) goto fail_errno;
4519 #elif defined(MDB_USE_POSIX_SEM)
4528 #if defined(__NetBSD__)
4529 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4531 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4532 idbuf.dev = stbuf.st_dev;
4533 idbuf.ino = stbuf.st_ino;
4534 val.mv_data = &idbuf;
4535 val.mv_size = sizeof(idbuf);
4536 mdb_hash_enc(&val, encbuf);
4537 #ifdef MDB_SHORT_SEMNAMES
4538 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4540 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4541 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4542 /* Clean up after a previous run, if needed: Try to
4543 * remove both semaphores before doing anything else.
4545 sem_unlink(env->me_txns->mti_rmname);
4546 sem_unlink(env->me_txns->mti_wmname);
4547 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4548 O_CREAT|O_EXCL, mode, 1);
4549 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4550 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4551 O_CREAT|O_EXCL, mode, 1);
4552 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4553 #else /* MDB_USE_POSIX_MUTEX: */
4554 pthread_mutexattr_t mattr;
4556 if ((rc = pthread_mutexattr_init(&mattr))
4557 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4558 #ifdef MDB_ROBUST_SUPPORTED
4559 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4561 || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
4562 || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
4564 pthread_mutexattr_destroy(&mattr);
4565 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4567 env->me_txns->mti_magic = MDB_MAGIC;
4568 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4569 env->me_txns->mti_txnid = 0;
4570 env->me_txns->mti_numreaders = 0;
4573 if (env->me_txns->mti_magic != MDB_MAGIC) {
4574 DPUTS("lock region has invalid magic");
4578 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4579 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4580 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4581 rc = MDB_VERSION_MISMATCH;
4585 if (rc && rc != EACCES && rc != EAGAIN) {
4589 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4590 if (!env->me_rmutex) goto fail_errno;
4591 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4592 if (!env->me_wmutex) goto fail_errno;
4593 #elif defined(MDB_USE_POSIX_SEM)
4594 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4595 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4596 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4597 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4608 /** The name of the lock file in the DB environment */
4609 #define LOCKNAME "/lock.mdb"
4610 /** The name of the data file in the DB environment */
4611 #define DATANAME "/data.mdb"
4612 /** The suffix of the lock file when no subdir is used */
4613 #define LOCKSUFF "-lock"
4614 /** Only a subset of the @ref mdb_env flags can be changed
4615 * at runtime. Changing other flags requires closing the
4616 * environment and re-opening it with the new flags.
4618 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4619 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4620 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4622 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4623 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4627 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4629 int oflags, rc, len, excl = -1;
4630 char *lpath, *dpath;
4632 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4636 if (flags & MDB_NOSUBDIR) {
4637 rc = len + sizeof(LOCKSUFF) + len + 1;
4639 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4644 if (flags & MDB_NOSUBDIR) {
4645 dpath = lpath + len + sizeof(LOCKSUFF);
4646 sprintf(lpath, "%s" LOCKSUFF, path);
4647 strcpy(dpath, path);
4649 dpath = lpath + len + sizeof(LOCKNAME);
4650 sprintf(lpath, "%s" LOCKNAME, path);
4651 sprintf(dpath, "%s" DATANAME, path);
4655 flags |= env->me_flags;
4656 if (flags & MDB_RDONLY) {
4657 /* silently ignore WRITEMAP when we're only getting read access */
4658 flags &= ~MDB_WRITEMAP;
4660 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4661 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4664 env->me_flags = flags |= MDB_ENV_ACTIVE;
4668 env->me_path = strdup(path);
4669 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4670 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4671 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4672 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4676 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4678 /* For RDONLY, get lockfile after we know datafile exists */
4679 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4680 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4686 if (F_ISSET(flags, MDB_RDONLY)) {
4687 oflags = GENERIC_READ;
4688 len = OPEN_EXISTING;
4690 oflags = GENERIC_READ|GENERIC_WRITE;
4693 mode = FILE_ATTRIBUTE_NORMAL;
4694 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4695 NULL, len, mode, NULL);
4697 if (F_ISSET(flags, MDB_RDONLY))
4700 oflags = O_RDWR | O_CREAT;
4702 env->me_fd = open(dpath, oflags, mode);
4704 if (env->me_fd == INVALID_HANDLE_VALUE) {
4709 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4710 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4715 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4716 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4717 env->me_mfd = env->me_fd;
4719 /* Synchronous fd for meta writes. Needed even with
4720 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4723 len = OPEN_EXISTING;
4724 env->me_mfd = CreateFile(dpath, oflags,
4725 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4726 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4729 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4731 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4736 DPRINTF(("opened dbenv %p", (void *) env));
4738 rc = mdb_env_share_locks(env, &excl);
4742 if (!(flags & MDB_RDONLY)) {
4744 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4745 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4746 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4747 (txn = calloc(1, size)))
4749 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4750 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4751 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4752 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4754 txn->mt_dbxs = env->me_dbxs;
4755 txn->mt_flags = MDB_TXN_FINISHED;
4765 mdb_env_close0(env, excl);
4771 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4773 mdb_env_close0(MDB_env *env, int excl)
4777 if (!(env->me_flags & MDB_ENV_ACTIVE))
4780 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4782 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4783 free(env->me_dbxs[i].md_name.mv_data);
4788 free(env->me_dbiseqs);
4789 free(env->me_dbflags);
4791 free(env->me_dirty_list);
4793 mdb_midl_free(env->me_free_pgs);
4795 if (env->me_flags & MDB_ENV_TXKEY) {
4796 pthread_key_delete(env->me_txkey);
4798 /* Delete our key from the global list */
4799 for (i=0; i<mdb_tls_nkeys; i++)
4800 if (mdb_tls_keys[i] == env->me_txkey) {
4801 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4809 munmap(env->me_map, env->me_mapsize);
4811 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4812 (void) close(env->me_mfd);
4813 if (env->me_fd != INVALID_HANDLE_VALUE)
4814 (void) close(env->me_fd);
4816 MDB_PID_T pid = env->me_pid;
4817 /* Clearing readers is done in this function because
4818 * me_txkey with its destructor must be disabled first.
4820 * We skip the the reader mutex, so we touch only
4821 * data owned by this process (me_close_readers and
4822 * our readers), and clear each reader atomically.
4824 for (i = env->me_close_readers; --i >= 0; )
4825 if (env->me_txns->mti_readers[i].mr_pid == pid)
4826 env->me_txns->mti_readers[i].mr_pid = 0;
4828 if (env->me_rmutex) {
4829 CloseHandle(env->me_rmutex);
4830 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4832 /* Windows automatically destroys the mutexes when
4833 * the last handle closes.
4835 #elif defined(MDB_USE_POSIX_SEM)
4836 if (env->me_rmutex != SEM_FAILED) {
4837 sem_close(env->me_rmutex);
4838 if (env->me_wmutex != SEM_FAILED)
4839 sem_close(env->me_wmutex);
4840 /* If we have the filelock: If we are the
4841 * only remaining user, clean up semaphores.
4844 mdb_env_excl_lock(env, &excl);
4846 sem_unlink(env->me_txns->mti_rmname);
4847 sem_unlink(env->me_txns->mti_wmname);
4851 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4853 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4856 /* Unlock the lockfile. Windows would have unlocked it
4857 * after closing anyway, but not necessarily at once.
4859 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4862 (void) close(env->me_lfd);
4865 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4869 mdb_env_close(MDB_env *env)
4876 VGMEMP_DESTROY(env);
4877 while ((dp = env->me_dpages) != NULL) {
4878 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4879 env->me_dpages = dp->mp_next;
4883 mdb_env_close0(env, 0);
4887 /** Compare two items pointing at aligned size_t's */
4889 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4891 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4892 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4895 /** Compare two items pointing at aligned unsigned int's.
4897 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4898 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4901 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4903 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4904 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4907 /** Compare two items pointing at unsigned ints of unknown alignment.
4908 * Nodes and keys are guaranteed to be 2-byte aligned.
4911 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4913 #if BYTE_ORDER == LITTLE_ENDIAN
4914 unsigned short *u, *c;
4917 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4918 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4921 } while(!x && u > (unsigned short *)a->mv_data);
4924 unsigned short *u, *c, *end;
4927 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4928 u = (unsigned short *)a->mv_data;
4929 c = (unsigned short *)b->mv_data;
4932 } while(!x && u < end);
4937 /** Compare two items lexically */
4939 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4946 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4952 diff = memcmp(a->mv_data, b->mv_data, len);
4953 return diff ? diff : len_diff<0 ? -1 : len_diff;
4956 /** Compare two items in reverse byte order */
4958 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4960 const unsigned char *p1, *p2, *p1_lim;
4964 p1_lim = (const unsigned char *)a->mv_data;
4965 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4966 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4968 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4974 while (p1 > p1_lim) {
4975 diff = *--p1 - *--p2;
4979 return len_diff<0 ? -1 : len_diff;
4982 /** Search for key within a page, using binary search.
4983 * Returns the smallest entry larger or equal to the key.
4984 * If exactp is non-null, stores whether the found entry was an exact match
4985 * in *exactp (1 or 0).
4986 * Updates the cursor index with the index of the found entry.
4987 * If no entry larger or equal to the key is found, returns NULL.
4990 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4992 unsigned int i = 0, nkeys;
4995 MDB_page *mp = mc->mc_pg[mc->mc_top];
4996 MDB_node *node = NULL;
5001 nkeys = NUMKEYS(mp);
5003 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5004 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5007 low = IS_LEAF(mp) ? 0 : 1;
5009 cmp = mc->mc_dbx->md_cmp;
5011 /* Branch pages have no data, so if using integer keys,
5012 * alignment is guaranteed. Use faster mdb_cmp_int.
5014 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5015 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5022 nodekey.mv_size = mc->mc_db->md_pad;
5023 node = NODEPTR(mp, 0); /* fake */
5024 while (low <= high) {
5025 i = (low + high) >> 1;
5026 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5027 rc = cmp(key, &nodekey);
5028 DPRINTF(("found leaf index %u [%s], rc = %i",
5029 i, DKEY(&nodekey), rc));
5038 while (low <= high) {
5039 i = (low + high) >> 1;
5041 node = NODEPTR(mp, i);
5042 nodekey.mv_size = NODEKSZ(node);
5043 nodekey.mv_data = NODEKEY(node);
5045 rc = cmp(key, &nodekey);
5048 DPRINTF(("found leaf index %u [%s], rc = %i",
5049 i, DKEY(&nodekey), rc));
5051 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5052 i, DKEY(&nodekey), NODEPGNO(node), rc));
5063 if (rc > 0) { /* Found entry is less than the key. */
5064 i++; /* Skip to get the smallest entry larger than key. */
5066 node = NODEPTR(mp, i);
5069 *exactp = (rc == 0 && nkeys > 0);
5070 /* store the key index */
5071 mc->mc_ki[mc->mc_top] = i;
5073 /* There is no entry larger or equal to the key. */
5076 /* nodeptr is fake for LEAF2 */
5082 mdb_cursor_adjust(MDB_cursor *mc, func)
5086 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5087 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5094 /** Pop a page off the top of the cursor's stack. */
5096 mdb_cursor_pop(MDB_cursor *mc)
5099 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5100 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5108 /** Push a page onto the top of the cursor's stack. */
5110 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5112 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5113 DDBI(mc), (void *) mc));
5115 if (mc->mc_snum >= CURSOR_STACK) {
5116 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5117 return MDB_CURSOR_FULL;
5120 mc->mc_top = mc->mc_snum++;
5121 mc->mc_pg[mc->mc_top] = mp;
5122 mc->mc_ki[mc->mc_top] = 0;
5127 /** Find the address of the page corresponding to a given page number.
5128 * @param[in] txn the transaction for this access.
5129 * @param[in] pgno the page number for the page to retrieve.
5130 * @param[out] ret address of a pointer where the page's address will be stored.
5131 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5132 * @return 0 on success, non-zero on failure.
5135 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5137 MDB_env *env = txn->mt_env;
5141 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5145 MDB_ID2L dl = tx2->mt_u.dirty_list;
5147 /* Spilled pages were dirtied in this txn and flushed
5148 * because the dirty list got full. Bring this page
5149 * back in from the map (but don't unspill it here,
5150 * leave that unless page_touch happens again).
5152 if (tx2->mt_spill_pgs) {
5153 MDB_ID pn = pgno << 1;
5154 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5155 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5156 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5161 unsigned x = mdb_mid2l_search(dl, pgno);
5162 if (x <= dl[0].mid && dl[x].mid == pgno) {
5168 } while ((tx2 = tx2->mt_parent) != NULL);
5171 if (pgno < txn->mt_next_pgno) {
5173 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5175 DPRINTF(("page %"Z"u not found", pgno));
5176 txn->mt_flags |= MDB_TXN_ERROR;
5177 return MDB_PAGE_NOTFOUND;
5187 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5188 * The cursor is at the root page, set up the rest of it.
5191 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5193 MDB_page *mp = mc->mc_pg[mc->mc_top];
5197 while (IS_BRANCH(mp)) {
5201 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5202 mdb_cassert(mc, NUMKEYS(mp) > 1);
5203 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5205 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5207 if (flags & MDB_PS_LAST)
5208 i = NUMKEYS(mp) - 1;
5211 node = mdb_node_search(mc, key, &exact);
5213 i = NUMKEYS(mp) - 1;
5215 i = mc->mc_ki[mc->mc_top];
5217 mdb_cassert(mc, i > 0);
5221 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5224 mdb_cassert(mc, i < NUMKEYS(mp));
5225 node = NODEPTR(mp, i);
5227 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5230 mc->mc_ki[mc->mc_top] = i;
5231 if ((rc = mdb_cursor_push(mc, mp)))
5234 if (flags & MDB_PS_MODIFY) {
5235 if ((rc = mdb_page_touch(mc)) != 0)
5237 mp = mc->mc_pg[mc->mc_top];
5242 DPRINTF(("internal error, index points to a %02X page!?",
5244 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5245 return MDB_CORRUPTED;
5248 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5249 key ? DKEY(key) : "null"));
5250 mc->mc_flags |= C_INITIALIZED;
5251 mc->mc_flags &= ~C_EOF;
5256 /** Search for the lowest key under the current branch page.
5257 * This just bypasses a NUMKEYS check in the current page
5258 * before calling mdb_page_search_root(), because the callers
5259 * are all in situations where the current page is known to
5263 mdb_page_search_lowest(MDB_cursor *mc)
5265 MDB_page *mp = mc->mc_pg[mc->mc_top];
5266 MDB_node *node = NODEPTR(mp, 0);
5269 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5272 mc->mc_ki[mc->mc_top] = 0;
5273 if ((rc = mdb_cursor_push(mc, mp)))
5275 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5278 /** Search for the page a given key should be in.
5279 * Push it and its parent pages on the cursor stack.
5280 * @param[in,out] mc the cursor for this operation.
5281 * @param[in] key the key to search for, or NULL for first/last page.
5282 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5283 * are touched (updated with new page numbers).
5284 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5285 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5286 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5287 * @return 0 on success, non-zero on failure.
5290 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5295 /* Make sure the txn is still viable, then find the root from
5296 * the txn's db table and set it as the root of the cursor's stack.
5298 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5299 DPUTS("transaction may not be used now");
5302 /* Make sure we're using an up-to-date root */
5303 if (*mc->mc_dbflag & DB_STALE) {
5305 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5307 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5308 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5315 MDB_node *leaf = mdb_node_search(&mc2,
5316 &mc->mc_dbx->md_name, &exact);
5318 return MDB_NOTFOUND;
5319 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5320 return MDB_INCOMPATIBLE; /* not a named DB */
5321 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5324 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5326 /* The txn may not know this DBI, or another process may
5327 * have dropped and recreated the DB with other flags.
5329 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5330 return MDB_INCOMPATIBLE;
5331 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5333 *mc->mc_dbflag &= ~DB_STALE;
5335 root = mc->mc_db->md_root;
5337 if (root == P_INVALID) { /* Tree is empty. */
5338 DPUTS("tree is empty");
5339 return MDB_NOTFOUND;
5343 mdb_cassert(mc, root > 1);
5344 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5345 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5351 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5352 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5354 if (flags & MDB_PS_MODIFY) {
5355 if ((rc = mdb_page_touch(mc)))
5359 if (flags & MDB_PS_ROOTONLY)
5362 return mdb_page_search_root(mc, key, flags);
5366 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5368 MDB_txn *txn = mc->mc_txn;
5369 pgno_t pg = mp->mp_pgno;
5370 unsigned x = 0, ovpages = mp->mp_pages;
5371 MDB_env *env = txn->mt_env;
5372 MDB_IDL sl = txn->mt_spill_pgs;
5373 MDB_ID pn = pg << 1;
5376 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5377 /* If the page is dirty or on the spill list we just acquired it,
5378 * so we should give it back to our current free list, if any.
5379 * Otherwise put it onto the list of pages we freed in this txn.
5381 * Won't create me_pghead: me_pglast must be inited along with it.
5382 * Unsupported in nested txns: They would need to hide the page
5383 * range in ancestor txns' dirty and spilled lists.
5385 if (env->me_pghead &&
5387 ((mp->mp_flags & P_DIRTY) ||
5388 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5392 MDB_ID2 *dl, ix, iy;
5393 rc = mdb_midl_need(&env->me_pghead, ovpages);
5396 if (!(mp->mp_flags & P_DIRTY)) {
5397 /* This page is no longer spilled */
5404 /* Remove from dirty list */
5405 dl = txn->mt_u.dirty_list;
5407 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5413 mdb_cassert(mc, x > 1);
5415 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5416 txn->mt_flags |= MDB_TXN_ERROR;
5417 return MDB_CORRUPTED;
5420 if (!(env->me_flags & MDB_WRITEMAP))
5421 mdb_dpage_free(env, mp);
5423 /* Insert in me_pghead */
5424 mop = env->me_pghead;
5425 j = mop[0] + ovpages;
5426 for (i = mop[0]; i && mop[i] < pg; i--)
5432 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5436 mc->mc_db->md_overflow_pages -= ovpages;
5440 /** Return the data associated with a given node.
5441 * @param[in] txn The transaction for this operation.
5442 * @param[in] leaf The node being read.
5443 * @param[out] data Updated to point to the node's data.
5444 * @return 0 on success, non-zero on failure.
5447 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5449 MDB_page *omp; /* overflow page */
5453 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5454 data->mv_size = NODEDSZ(leaf);
5455 data->mv_data = NODEDATA(leaf);
5459 /* Read overflow data.
5461 data->mv_size = NODEDSZ(leaf);
5462 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5463 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5464 DPRINTF(("read overflow page %"Z"u failed", pgno));
5467 data->mv_data = METADATA(omp);
5473 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5474 MDB_val *key, MDB_val *data)
5481 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5483 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5486 if (txn->mt_flags & MDB_TXN_BLOCKED)
5489 mdb_cursor_init(&mc, txn, dbi, &mx);
5490 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5493 /** Find a sibling for a page.
5494 * Replaces the page at the top of the cursor's stack with the
5495 * specified sibling, if one exists.
5496 * @param[in] mc The cursor for this operation.
5497 * @param[in] move_right Non-zero if the right sibling is requested,
5498 * otherwise the left sibling.
5499 * @return 0 on success, non-zero on failure.
5502 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5508 if (mc->mc_snum < 2) {
5509 return MDB_NOTFOUND; /* root has no siblings */
5513 DPRINTF(("parent page is page %"Z"u, index %u",
5514 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5516 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5517 : (mc->mc_ki[mc->mc_top] == 0)) {
5518 DPRINTF(("no more keys left, moving to %s sibling",
5519 move_right ? "right" : "left"));
5520 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5521 /* undo cursor_pop before returning */
5528 mc->mc_ki[mc->mc_top]++;
5530 mc->mc_ki[mc->mc_top]--;
5531 DPRINTF(("just moving to %s index key %u",
5532 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5534 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5536 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5537 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5538 /* mc will be inconsistent if caller does mc_snum++ as above */
5539 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5543 mdb_cursor_push(mc, mp);
5545 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5550 /** Move the cursor to the next data item. */
5552 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5558 if (mc->mc_flags & C_EOF) {
5559 return MDB_NOTFOUND;
5562 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5564 mp = mc->mc_pg[mc->mc_top];
5566 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5567 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5568 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5569 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5570 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5571 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5572 if (rc == MDB_SUCCESS)
5573 MDB_GET_KEY(leaf, key);
5578 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5579 if (op == MDB_NEXT_DUP)
5580 return MDB_NOTFOUND;
5584 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5585 mdb_dbg_pgno(mp), (void *) mc));
5586 if (mc->mc_flags & C_DEL)
5589 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5590 DPUTS("=====> move to next sibling page");
5591 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5592 mc->mc_flags |= C_EOF;
5595 mp = mc->mc_pg[mc->mc_top];
5596 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5598 mc->mc_ki[mc->mc_top]++;
5601 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5602 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5605 key->mv_size = mc->mc_db->md_pad;
5606 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5610 mdb_cassert(mc, IS_LEAF(mp));
5611 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5613 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5614 mdb_xcursor_init1(mc, leaf);
5617 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5620 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5621 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5622 if (rc != MDB_SUCCESS)
5627 MDB_GET_KEY(leaf, key);
5631 /** Move the cursor to the previous data item. */
5633 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5639 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5641 mp = mc->mc_pg[mc->mc_top];
5643 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5644 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5645 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5646 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5647 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5648 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5649 if (rc == MDB_SUCCESS) {
5650 MDB_GET_KEY(leaf, key);
5651 mc->mc_flags &= ~C_EOF;
5657 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5658 if (op == MDB_PREV_DUP)
5659 return MDB_NOTFOUND;
5663 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5664 mdb_dbg_pgno(mp), (void *) mc));
5666 if (mc->mc_ki[mc->mc_top] == 0) {
5667 DPUTS("=====> move to prev sibling page");
5668 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5671 mp = mc->mc_pg[mc->mc_top];
5672 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5673 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5675 mc->mc_ki[mc->mc_top]--;
5677 mc->mc_flags &= ~C_EOF;
5679 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5680 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5683 key->mv_size = mc->mc_db->md_pad;
5684 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5688 mdb_cassert(mc, IS_LEAF(mp));
5689 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5691 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5692 mdb_xcursor_init1(mc, leaf);
5695 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5698 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5699 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5700 if (rc != MDB_SUCCESS)
5705 MDB_GET_KEY(leaf, key);
5709 /** Set the cursor on a specific data item. */
5711 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5712 MDB_cursor_op op, int *exactp)
5716 MDB_node *leaf = NULL;
5719 if (key->mv_size == 0)
5720 return MDB_BAD_VALSIZE;
5723 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5725 /* See if we're already on the right page */
5726 if (mc->mc_flags & C_INITIALIZED) {
5729 mp = mc->mc_pg[mc->mc_top];
5731 mc->mc_ki[mc->mc_top] = 0;
5732 return MDB_NOTFOUND;
5734 if (mp->mp_flags & P_LEAF2) {
5735 nodekey.mv_size = mc->mc_db->md_pad;
5736 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5738 leaf = NODEPTR(mp, 0);
5739 MDB_GET_KEY2(leaf, nodekey);
5741 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5743 /* Probably happens rarely, but first node on the page
5744 * was the one we wanted.
5746 mc->mc_ki[mc->mc_top] = 0;
5753 unsigned int nkeys = NUMKEYS(mp);
5755 if (mp->mp_flags & P_LEAF2) {
5756 nodekey.mv_data = LEAF2KEY(mp,
5757 nkeys-1, nodekey.mv_size);
5759 leaf = NODEPTR(mp, nkeys-1);
5760 MDB_GET_KEY2(leaf, nodekey);
5762 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5764 /* last node was the one we wanted */
5765 mc->mc_ki[mc->mc_top] = nkeys-1;
5771 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5772 /* This is definitely the right page, skip search_page */
5773 if (mp->mp_flags & P_LEAF2) {
5774 nodekey.mv_data = LEAF2KEY(mp,
5775 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5777 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5778 MDB_GET_KEY2(leaf, nodekey);
5780 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5782 /* current node was the one we wanted */
5792 /* If any parents have right-sibs, search.
5793 * Otherwise, there's nothing further.
5795 for (i=0; i<mc->mc_top; i++)
5797 NUMKEYS(mc->mc_pg[i])-1)
5799 if (i == mc->mc_top) {
5800 /* There are no other pages */
5801 mc->mc_ki[mc->mc_top] = nkeys;
5802 return MDB_NOTFOUND;
5806 /* There are no other pages */
5807 mc->mc_ki[mc->mc_top] = 0;
5808 if (op == MDB_SET_RANGE && !exactp) {
5812 return MDB_NOTFOUND;
5816 rc = mdb_page_search(mc, key, 0);
5817 if (rc != MDB_SUCCESS)
5820 mp = mc->mc_pg[mc->mc_top];
5821 mdb_cassert(mc, IS_LEAF(mp));
5824 leaf = mdb_node_search(mc, key, exactp);
5825 if (exactp != NULL && !*exactp) {
5826 /* MDB_SET specified and not an exact match. */
5827 return MDB_NOTFOUND;
5831 DPUTS("===> inexact leaf not found, goto sibling");
5832 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5833 mc->mc_flags |= C_EOF;
5834 return rc; /* no entries matched */
5836 mp = mc->mc_pg[mc->mc_top];
5837 mdb_cassert(mc, IS_LEAF(mp));
5838 leaf = NODEPTR(mp, 0);
5842 mc->mc_flags |= C_INITIALIZED;
5843 mc->mc_flags &= ~C_EOF;
5846 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5847 key->mv_size = mc->mc_db->md_pad;
5848 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5853 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5854 mdb_xcursor_init1(mc, leaf);
5857 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5858 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5859 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5862 if (op == MDB_GET_BOTH) {
5868 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5869 if (rc != MDB_SUCCESS)
5872 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5875 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5877 dcmp = mc->mc_dbx->md_dcmp;
5878 #if UINT_MAX < SIZE_MAX
5879 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5880 dcmp = mdb_cmp_clong;
5882 rc = dcmp(data, &olddata);
5884 if (op == MDB_GET_BOTH || rc > 0)
5885 return MDB_NOTFOUND;
5892 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5893 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5898 /* The key already matches in all other cases */
5899 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5900 MDB_GET_KEY(leaf, key);
5901 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5906 /** Move the cursor to the first item in the database. */
5908 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5914 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5916 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5917 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5918 if (rc != MDB_SUCCESS)
5921 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5923 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5924 mc->mc_flags |= C_INITIALIZED;
5925 mc->mc_flags &= ~C_EOF;
5927 mc->mc_ki[mc->mc_top] = 0;
5929 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5930 key->mv_size = mc->mc_db->md_pad;
5931 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5936 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5937 mdb_xcursor_init1(mc, leaf);
5938 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5942 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5946 MDB_GET_KEY(leaf, key);
5950 /** Move the cursor to the last item in the database. */
5952 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5958 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5960 if (!(mc->mc_flags & C_EOF)) {
5962 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5963 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5964 if (rc != MDB_SUCCESS)
5967 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5970 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5971 mc->mc_flags |= C_INITIALIZED|C_EOF;
5972 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5974 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5975 key->mv_size = mc->mc_db->md_pad;
5976 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5981 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5982 mdb_xcursor_init1(mc, leaf);
5983 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5987 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5992 MDB_GET_KEY(leaf, key);
5997 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6002 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6007 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6011 case MDB_GET_CURRENT:
6012 if (!(mc->mc_flags & C_INITIALIZED)) {
6015 MDB_page *mp = mc->mc_pg[mc->mc_top];
6016 int nkeys = NUMKEYS(mp);
6017 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6018 mc->mc_ki[mc->mc_top] = nkeys;
6024 key->mv_size = mc->mc_db->md_pad;
6025 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6027 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6028 MDB_GET_KEY(leaf, key);
6030 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6031 if (mc->mc_flags & C_DEL)
6032 mdb_xcursor_init1(mc, leaf);
6033 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6035 rc = mdb_node_read(mc->mc_txn, leaf, data);
6042 case MDB_GET_BOTH_RANGE:
6047 if (mc->mc_xcursor == NULL) {
6048 rc = MDB_INCOMPATIBLE;
6058 rc = mdb_cursor_set(mc, key, data, op,
6059 op == MDB_SET_RANGE ? NULL : &exact);
6062 case MDB_GET_MULTIPLE:
6063 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6067 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6068 rc = MDB_INCOMPATIBLE;
6072 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6073 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6076 case MDB_NEXT_MULTIPLE:
6081 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6082 rc = MDB_INCOMPATIBLE;
6085 if (!(mc->mc_flags & C_INITIALIZED))
6086 rc = mdb_cursor_first(mc, key, data);
6088 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6089 if (rc == MDB_SUCCESS) {
6090 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6093 mx = &mc->mc_xcursor->mx_cursor;
6094 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6096 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6097 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6105 case MDB_NEXT_NODUP:
6106 if (!(mc->mc_flags & C_INITIALIZED))
6107 rc = mdb_cursor_first(mc, key, data);
6109 rc = mdb_cursor_next(mc, key, data, op);
6113 case MDB_PREV_NODUP:
6114 if (!(mc->mc_flags & C_INITIALIZED)) {
6115 rc = mdb_cursor_last(mc, key, data);
6118 mc->mc_flags |= C_INITIALIZED;
6119 mc->mc_ki[mc->mc_top]++;
6121 rc = mdb_cursor_prev(mc, key, data, op);
6124 rc = mdb_cursor_first(mc, key, data);
6127 mfunc = mdb_cursor_first;
6129 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6133 if (mc->mc_xcursor == NULL) {
6134 rc = MDB_INCOMPATIBLE;
6138 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6139 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6140 MDB_GET_KEY(leaf, key);
6141 rc = mdb_node_read(mc->mc_txn, leaf, data);
6145 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6149 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6152 rc = mdb_cursor_last(mc, key, data);
6155 mfunc = mdb_cursor_last;
6158 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6163 if (mc->mc_flags & C_DEL)
6164 mc->mc_flags ^= C_DEL;
6169 /** Touch all the pages in the cursor stack. Set mc_top.
6170 * Makes sure all the pages are writable, before attempting a write operation.
6171 * @param[in] mc The cursor to operate on.
6174 mdb_cursor_touch(MDB_cursor *mc)
6176 int rc = MDB_SUCCESS;
6178 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6181 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6183 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6184 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6187 *mc->mc_dbflag |= DB_DIRTY;
6192 rc = mdb_page_touch(mc);
6193 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6194 mc->mc_top = mc->mc_snum-1;
6199 /** Do not spill pages to disk if txn is getting full, may fail instead */
6200 #define MDB_NOSPILL 0x8000
6203 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6207 MDB_node *leaf = NULL;
6208 MDB_page *fp, *mp, *sub_root = NULL;
6210 MDB_val xdata, *rdata, dkey, olddata;
6212 int do_sub = 0, insert_key, insert_data;
6213 unsigned int mcount = 0, dcount = 0, nospill;
6216 unsigned int nflags;
6219 if (mc == NULL || key == NULL)
6222 env = mc->mc_txn->mt_env;
6224 /* Check this first so counter will always be zero on any
6227 if (flags & MDB_MULTIPLE) {
6228 dcount = data[1].mv_size;
6229 data[1].mv_size = 0;
6230 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6231 return MDB_INCOMPATIBLE;
6234 nospill = flags & MDB_NOSPILL;
6235 flags &= ~MDB_NOSPILL;
6237 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6238 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6240 if (key->mv_size-1 >= ENV_MAXKEY(env))
6241 return MDB_BAD_VALSIZE;
6243 #if SIZE_MAX > MAXDATASIZE
6244 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6245 return MDB_BAD_VALSIZE;
6247 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6248 return MDB_BAD_VALSIZE;
6251 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6252 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6256 if (flags == MDB_CURRENT) {
6257 if (!(mc->mc_flags & C_INITIALIZED))
6260 } else if (mc->mc_db->md_root == P_INVALID) {
6261 /* new database, cursor has nothing to point to */
6264 mc->mc_flags &= ~C_INITIALIZED;
6269 if (flags & MDB_APPEND) {
6271 rc = mdb_cursor_last(mc, &k2, &d2);
6273 rc = mc->mc_dbx->md_cmp(key, &k2);
6276 mc->mc_ki[mc->mc_top]++;
6278 /* new key is <= last key */
6283 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6285 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6286 DPRINTF(("duplicate key [%s]", DKEY(key)));
6288 return MDB_KEYEXIST;
6290 if (rc && rc != MDB_NOTFOUND)
6294 if (mc->mc_flags & C_DEL)
6295 mc->mc_flags ^= C_DEL;
6297 /* Cursor is positioned, check for room in the dirty list */
6299 if (flags & MDB_MULTIPLE) {
6301 xdata.mv_size = data->mv_size * dcount;
6305 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6309 if (rc == MDB_NO_ROOT) {
6311 /* new database, write a root leaf page */
6312 DPUTS("allocating new root leaf page");
6313 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6316 mdb_cursor_push(mc, np);
6317 mc->mc_db->md_root = np->mp_pgno;
6318 mc->mc_db->md_depth++;
6319 *mc->mc_dbflag |= DB_DIRTY;
6320 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6322 np->mp_flags |= P_LEAF2;
6323 mc->mc_flags |= C_INITIALIZED;
6325 /* make sure all cursor pages are writable */
6326 rc2 = mdb_cursor_touch(mc);
6331 insert_key = insert_data = rc;
6333 /* The key does not exist */
6334 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6335 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6336 LEAFSIZE(key, data) > env->me_nodemax)
6338 /* Too big for a node, insert in sub-DB. Set up an empty
6339 * "old sub-page" for prep_subDB to expand to a full page.
6341 fp_flags = P_LEAF|P_DIRTY;
6343 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6344 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6345 olddata.mv_size = PAGEHDRSZ;
6349 /* there's only a key anyway, so this is a no-op */
6350 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6352 unsigned int ksize = mc->mc_db->md_pad;
6353 if (key->mv_size != ksize)
6354 return MDB_BAD_VALSIZE;
6355 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6356 memcpy(ptr, key->mv_data, ksize);
6358 /* if overwriting slot 0 of leaf, need to
6359 * update branch key if there is a parent page
6361 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6362 unsigned short top = mc->mc_top;
6364 /* slot 0 is always an empty key, find real slot */
6365 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6367 if (mc->mc_ki[mc->mc_top])
6368 rc2 = mdb_update_key(mc, key);
6379 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6380 olddata.mv_size = NODEDSZ(leaf);
6381 olddata.mv_data = NODEDATA(leaf);
6384 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6385 /* Prepare (sub-)page/sub-DB to accept the new item,
6386 * if needed. fp: old sub-page or a header faking
6387 * it. mp: new (sub-)page. offset: growth in page
6388 * size. xdata: node data with new page or DB.
6390 unsigned i, offset = 0;
6391 mp = fp = xdata.mv_data = env->me_pbuf;
6392 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6394 /* Was a single item before, must convert now */
6395 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6397 /* Just overwrite the current item */
6398 if (flags == MDB_CURRENT)
6400 dcmp = mc->mc_dbx->md_dcmp;
6401 #if UINT_MAX < SIZE_MAX
6402 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6403 dcmp = mdb_cmp_clong;
6405 /* does data match? */
6406 if (!dcmp(data, &olddata)) {
6407 if (flags & MDB_NODUPDATA)
6408 return MDB_KEYEXIST;
6413 /* Back up original data item */
6414 dkey.mv_size = olddata.mv_size;
6415 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6417 /* Make sub-page header for the dup items, with dummy body */
6418 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6419 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6420 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6421 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6422 fp->mp_flags |= P_LEAF2;
6423 fp->mp_pad = data->mv_size;
6424 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6426 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6427 (dkey.mv_size & 1) + (data->mv_size & 1);
6429 fp->mp_upper = xdata.mv_size - PAGEBASE;
6430 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6431 } else if (leaf->mn_flags & F_SUBDATA) {
6432 /* Data is on sub-DB, just store it */
6433 flags |= F_DUPDATA|F_SUBDATA;
6436 /* Data is on sub-page */
6437 fp = olddata.mv_data;
6440 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6441 offset = EVEN(NODESIZE + sizeof(indx_t) +
6445 offset = fp->mp_pad;
6446 if (SIZELEFT(fp) < offset) {
6447 offset *= 4; /* space for 4 more */
6450 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6452 fp->mp_flags |= P_DIRTY;
6453 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6454 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6458 xdata.mv_size = olddata.mv_size + offset;
6461 fp_flags = fp->mp_flags;
6462 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6463 /* Too big for a sub-page, convert to sub-DB */
6464 fp_flags &= ~P_SUBP;
6466 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6467 fp_flags |= P_LEAF2;
6468 dummy.md_pad = fp->mp_pad;
6469 dummy.md_flags = MDB_DUPFIXED;
6470 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6471 dummy.md_flags |= MDB_INTEGERKEY;
6477 dummy.md_branch_pages = 0;
6478 dummy.md_leaf_pages = 1;
6479 dummy.md_overflow_pages = 0;
6480 dummy.md_entries = NUMKEYS(fp);
6481 xdata.mv_size = sizeof(MDB_db);
6482 xdata.mv_data = &dummy;
6483 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6485 offset = env->me_psize - olddata.mv_size;
6486 flags |= F_DUPDATA|F_SUBDATA;
6487 dummy.md_root = mp->mp_pgno;
6491 mp->mp_flags = fp_flags | P_DIRTY;
6492 mp->mp_pad = fp->mp_pad;
6493 mp->mp_lower = fp->mp_lower;
6494 mp->mp_upper = fp->mp_upper + offset;
6495 if (fp_flags & P_LEAF2) {
6496 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6498 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6499 olddata.mv_size - fp->mp_upper - PAGEBASE);
6500 for (i=0; i<NUMKEYS(fp); i++)
6501 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6509 mdb_node_del(mc, 0);
6513 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6514 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6515 return MDB_INCOMPATIBLE;
6516 /* overflow page overwrites need special handling */
6517 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6520 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6522 memcpy(&pg, olddata.mv_data, sizeof(pg));
6523 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6525 ovpages = omp->mp_pages;
6527 /* Is the ov page large enough? */
6528 if (ovpages >= dpages) {
6529 if (!(omp->mp_flags & P_DIRTY) &&
6530 (level || (env->me_flags & MDB_WRITEMAP)))
6532 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6535 level = 0; /* dirty in this txn or clean */
6538 if (omp->mp_flags & P_DIRTY) {
6539 /* yes, overwrite it. Note in this case we don't
6540 * bother to try shrinking the page if the new data
6541 * is smaller than the overflow threshold.
6544 /* It is writable only in a parent txn */
6545 size_t sz = (size_t) env->me_psize * ovpages, off;
6546 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6552 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6553 mdb_cassert(mc, rc2 == 0);
6554 if (!(flags & MDB_RESERVE)) {
6555 /* Copy end of page, adjusting alignment so
6556 * compiler may copy words instead of bytes.
6558 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6559 memcpy((size_t *)((char *)np + off),
6560 (size_t *)((char *)omp + off), sz - off);
6563 memcpy(np, omp, sz); /* Copy beginning of page */
6566 SETDSZ(leaf, data->mv_size);
6567 if (F_ISSET(flags, MDB_RESERVE))
6568 data->mv_data = METADATA(omp);
6570 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6574 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6576 } else if (data->mv_size == olddata.mv_size) {
6577 /* same size, just replace it. Note that we could
6578 * also reuse this node if the new data is smaller,
6579 * but instead we opt to shrink the node in that case.
6581 if (F_ISSET(flags, MDB_RESERVE))
6582 data->mv_data = olddata.mv_data;
6583 else if (!(mc->mc_flags & C_SUB))
6584 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6586 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6591 mdb_node_del(mc, 0);
6597 nflags = flags & NODE_ADD_FLAGS;
6598 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6599 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6600 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6601 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6603 nflags |= MDB_SPLIT_REPLACE;
6604 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6606 /* There is room already in this leaf page. */
6607 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6608 if (rc == 0 && insert_key) {
6609 /* Adjust other cursors pointing to mp */
6610 MDB_cursor *m2, *m3;
6611 MDB_dbi dbi = mc->mc_dbi;
6612 unsigned i = mc->mc_top;
6613 MDB_page *mp = mc->mc_pg[i];
6615 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6616 if (mc->mc_flags & C_SUB)
6617 m3 = &m2->mc_xcursor->mx_cursor;
6620 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6621 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6628 if (rc == MDB_SUCCESS) {
6629 /* Now store the actual data in the child DB. Note that we're
6630 * storing the user data in the keys field, so there are strict
6631 * size limits on dupdata. The actual data fields of the child
6632 * DB are all zero size.
6635 int xflags, new_dupdata;
6640 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6641 if (flags & MDB_CURRENT) {
6642 xflags = MDB_CURRENT|MDB_NOSPILL;
6644 mdb_xcursor_init1(mc, leaf);
6645 xflags = (flags & MDB_NODUPDATA) ?
6646 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6649 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6650 new_dupdata = (int)dkey.mv_size;
6651 /* converted, write the original data first */
6653 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6656 /* we've done our job */
6659 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6660 /* Adjust other cursors pointing to mp */
6662 MDB_xcursor *mx = mc->mc_xcursor;
6663 unsigned i = mc->mc_top;
6664 MDB_page *mp = mc->mc_pg[i];
6666 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6667 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6668 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6669 if (m2->mc_pg[i] == mp) {
6670 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6671 mdb_xcursor_init2(m2, mx, new_dupdata);
6672 } else if (!insert_key) {
6673 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6674 if (!(n2->mn_flags & F_SUBDATA))
6675 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6680 ecount = mc->mc_xcursor->mx_db.md_entries;
6681 if (flags & MDB_APPENDDUP)
6682 xflags |= MDB_APPEND;
6683 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6684 if (flags & F_SUBDATA) {
6685 void *db = NODEDATA(leaf);
6686 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6688 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6690 /* Increment count unless we just replaced an existing item. */
6692 mc->mc_db->md_entries++;
6694 /* Invalidate txn if we created an empty sub-DB */
6697 /* If we succeeded and the key didn't exist before,
6698 * make sure the cursor is marked valid.
6700 mc->mc_flags |= C_INITIALIZED;
6702 if (flags & MDB_MULTIPLE) {
6705 /* let caller know how many succeeded, if any */
6706 data[1].mv_size = mcount;
6707 if (mcount < dcount) {
6708 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6709 insert_key = insert_data = 0;
6716 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6719 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6724 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6730 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6731 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6733 if (!(mc->mc_flags & C_INITIALIZED))
6736 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6737 return MDB_NOTFOUND;
6739 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6742 rc = mdb_cursor_touch(mc);
6746 mp = mc->mc_pg[mc->mc_top];
6749 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6751 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6752 if (flags & MDB_NODUPDATA) {
6753 /* mdb_cursor_del0() will subtract the final entry */
6754 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6756 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6757 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6759 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6762 /* If sub-DB still has entries, we're done */
6763 if (mc->mc_xcursor->mx_db.md_entries) {
6764 if (leaf->mn_flags & F_SUBDATA) {
6765 /* update subDB info */
6766 void *db = NODEDATA(leaf);
6767 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6770 /* shrink fake page */
6771 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6772 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6773 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6774 /* fix other sub-DB cursors pointed at fake pages on this page */
6775 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6776 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6777 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6778 if (m2->mc_pg[mc->mc_top] == mp) {
6779 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6780 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6782 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6783 if (!(n2->mn_flags & F_SUBDATA))
6784 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6789 mc->mc_db->md_entries--;
6790 mc->mc_flags |= C_DEL;
6793 /* otherwise fall thru and delete the sub-DB */
6796 if (leaf->mn_flags & F_SUBDATA) {
6797 /* add all the child DB's pages to the free list */
6798 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6803 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6804 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6805 rc = MDB_INCOMPATIBLE;
6809 /* add overflow pages to free list */
6810 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6814 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6815 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6816 (rc = mdb_ovpage_free(mc, omp)))
6821 return mdb_cursor_del0(mc);
6824 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6828 /** Allocate and initialize new pages for a database.
6829 * @param[in] mc a cursor on the database being added to.
6830 * @param[in] flags flags defining what type of page is being allocated.
6831 * @param[in] num the number of pages to allocate. This is usually 1,
6832 * unless allocating overflow pages for a large record.
6833 * @param[out] mp Address of a page, or NULL on failure.
6834 * @return 0 on success, non-zero on failure.
6837 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6842 if ((rc = mdb_page_alloc(mc, num, &np)))
6844 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6845 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6846 np->mp_flags = flags | P_DIRTY;
6847 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6848 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6851 mc->mc_db->md_branch_pages++;
6852 else if (IS_LEAF(np))
6853 mc->mc_db->md_leaf_pages++;
6854 else if (IS_OVERFLOW(np)) {
6855 mc->mc_db->md_overflow_pages += num;
6863 /** Calculate the size of a leaf node.
6864 * The size depends on the environment's page size; if a data item
6865 * is too large it will be put onto an overflow page and the node
6866 * size will only include the key and not the data. Sizes are always
6867 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6868 * of the #MDB_node headers.
6869 * @param[in] env The environment handle.
6870 * @param[in] key The key for the node.
6871 * @param[in] data The data for the node.
6872 * @return The number of bytes needed to store the node.
6875 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6879 sz = LEAFSIZE(key, data);
6880 if (sz > env->me_nodemax) {
6881 /* put on overflow page */
6882 sz -= data->mv_size - sizeof(pgno_t);
6885 return EVEN(sz + sizeof(indx_t));
6888 /** Calculate the size of a branch node.
6889 * The size should depend on the environment's page size but since
6890 * we currently don't support spilling large keys onto overflow
6891 * pages, it's simply the size of the #MDB_node header plus the
6892 * size of the key. Sizes are always rounded up to an even number
6893 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6894 * @param[in] env The environment handle.
6895 * @param[in] key The key for the node.
6896 * @return The number of bytes needed to store the node.
6899 mdb_branch_size(MDB_env *env, MDB_val *key)
6904 if (sz > env->me_nodemax) {
6905 /* put on overflow page */
6906 /* not implemented */
6907 /* sz -= key->size - sizeof(pgno_t); */
6910 return sz + sizeof(indx_t);
6913 /** Add a node to the page pointed to by the cursor.
6914 * @param[in] mc The cursor for this operation.
6915 * @param[in] indx The index on the page where the new node should be added.
6916 * @param[in] key The key for the new node.
6917 * @param[in] data The data for the new node, if any.
6918 * @param[in] pgno The page number, if adding a branch node.
6919 * @param[in] flags Flags for the node.
6920 * @return 0 on success, non-zero on failure. Possible errors are:
6922 * <li>ENOMEM - failed to allocate overflow pages for the node.
6923 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6924 * should never happen since all callers already calculate the
6925 * page's free space before calling this function.
6929 mdb_node_add(MDB_cursor *mc, indx_t indx,
6930 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6933 size_t node_size = NODESIZE;
6937 MDB_page *mp = mc->mc_pg[mc->mc_top];
6938 MDB_page *ofp = NULL; /* overflow page */
6942 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6944 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6945 IS_LEAF(mp) ? "leaf" : "branch",
6946 IS_SUBP(mp) ? "sub-" : "",
6947 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6948 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6951 /* Move higher keys up one slot. */
6952 int ksize = mc->mc_db->md_pad, dif;
6953 char *ptr = LEAF2KEY(mp, indx, ksize);
6954 dif = NUMKEYS(mp) - indx;
6956 memmove(ptr+ksize, ptr, dif*ksize);
6957 /* insert new key */
6958 memcpy(ptr, key->mv_data, ksize);
6960 /* Just using these for counting */
6961 mp->mp_lower += sizeof(indx_t);
6962 mp->mp_upper -= ksize - sizeof(indx_t);
6966 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6968 node_size += key->mv_size;
6970 mdb_cassert(mc, key && data);
6971 if (F_ISSET(flags, F_BIGDATA)) {
6972 /* Data already on overflow page. */
6973 node_size += sizeof(pgno_t);
6974 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6975 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6977 /* Put data on overflow page. */
6978 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6979 data->mv_size, node_size+data->mv_size));
6980 node_size = EVEN(node_size + sizeof(pgno_t));
6981 if ((ssize_t)node_size > room)
6983 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6985 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6989 node_size += data->mv_size;
6992 node_size = EVEN(node_size);
6993 if ((ssize_t)node_size > room)
6997 /* Move higher pointers up one slot. */
6998 for (i = NUMKEYS(mp); i > indx; i--)
6999 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7001 /* Adjust free space offsets. */
7002 ofs = mp->mp_upper - node_size;
7003 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7004 mp->mp_ptrs[indx] = ofs;
7006 mp->mp_lower += sizeof(indx_t);
7008 /* Write the node data. */
7009 node = NODEPTR(mp, indx);
7010 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7011 node->mn_flags = flags;
7013 SETDSZ(node,data->mv_size);
7018 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7021 ndata = NODEDATA(node);
7023 if (F_ISSET(flags, F_BIGDATA))
7024 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7025 else if (F_ISSET(flags, MDB_RESERVE))
7026 data->mv_data = ndata;
7028 memcpy(ndata, data->mv_data, data->mv_size);
7030 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7031 ndata = METADATA(ofp);
7032 if (F_ISSET(flags, MDB_RESERVE))
7033 data->mv_data = ndata;
7035 memcpy(ndata, data->mv_data, data->mv_size);
7042 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7043 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7044 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7045 DPRINTF(("node size = %"Z"u", node_size));
7046 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7047 return MDB_PAGE_FULL;
7050 /** Delete the specified node from a page.
7051 * @param[in] mc Cursor pointing to the node to delete.
7052 * @param[in] ksize The size of a node. Only used if the page is
7053 * part of a #MDB_DUPFIXED database.
7056 mdb_node_del(MDB_cursor *mc, int ksize)
7058 MDB_page *mp = mc->mc_pg[mc->mc_top];
7059 indx_t indx = mc->mc_ki[mc->mc_top];
7061 indx_t i, j, numkeys, ptr;
7065 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7066 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7067 numkeys = NUMKEYS(mp);
7068 mdb_cassert(mc, indx < numkeys);
7071 int x = numkeys - 1 - indx;
7072 base = LEAF2KEY(mp, indx, ksize);
7074 memmove(base, base + ksize, x * ksize);
7075 mp->mp_lower -= sizeof(indx_t);
7076 mp->mp_upper += ksize - sizeof(indx_t);
7080 node = NODEPTR(mp, indx);
7081 sz = NODESIZE + node->mn_ksize;
7083 if (F_ISSET(node->mn_flags, F_BIGDATA))
7084 sz += sizeof(pgno_t);
7086 sz += NODEDSZ(node);
7090 ptr = mp->mp_ptrs[indx];
7091 for (i = j = 0; i < numkeys; i++) {
7093 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7094 if (mp->mp_ptrs[i] < ptr)
7095 mp->mp_ptrs[j] += sz;
7100 base = (char *)mp + mp->mp_upper + PAGEBASE;
7101 memmove(base + sz, base, ptr - mp->mp_upper);
7103 mp->mp_lower -= sizeof(indx_t);
7107 /** Compact the main page after deleting a node on a subpage.
7108 * @param[in] mp The main page to operate on.
7109 * @param[in] indx The index of the subpage on the main page.
7112 mdb_node_shrink(MDB_page *mp, indx_t indx)
7117 indx_t delta, nsize, len, ptr;
7120 node = NODEPTR(mp, indx);
7121 sp = (MDB_page *)NODEDATA(node);
7122 delta = SIZELEFT(sp);
7123 nsize = NODEDSZ(node) - delta;
7125 /* Prepare to shift upward, set len = length(subpage part to shift) */
7129 return; /* do not make the node uneven-sized */
7131 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7132 for (i = NUMKEYS(sp); --i >= 0; )
7133 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7136 sp->mp_upper = sp->mp_lower;
7137 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7138 SETDSZ(node, nsize);
7140 /* Shift <lower nodes...initial part of subpage> upward */
7141 base = (char *)mp + mp->mp_upper + PAGEBASE;
7142 memmove(base + delta, base, (char *)sp + len - base);
7144 ptr = mp->mp_ptrs[indx];
7145 for (i = NUMKEYS(mp); --i >= 0; ) {
7146 if (mp->mp_ptrs[i] <= ptr)
7147 mp->mp_ptrs[i] += delta;
7149 mp->mp_upper += delta;
7152 /** Initial setup of a sorted-dups cursor.
7153 * Sorted duplicates are implemented as a sub-database for the given key.
7154 * The duplicate data items are actually keys of the sub-database.
7155 * Operations on the duplicate data items are performed using a sub-cursor
7156 * initialized when the sub-database is first accessed. This function does
7157 * the preliminary setup of the sub-cursor, filling in the fields that
7158 * depend only on the parent DB.
7159 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7162 mdb_xcursor_init0(MDB_cursor *mc)
7164 MDB_xcursor *mx = mc->mc_xcursor;
7166 mx->mx_cursor.mc_xcursor = NULL;
7167 mx->mx_cursor.mc_txn = mc->mc_txn;
7168 mx->mx_cursor.mc_db = &mx->mx_db;
7169 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7170 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7171 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7172 mx->mx_cursor.mc_snum = 0;
7173 mx->mx_cursor.mc_top = 0;
7174 mx->mx_cursor.mc_flags = C_SUB;
7175 mx->mx_dbx.md_name.mv_size = 0;
7176 mx->mx_dbx.md_name.mv_data = NULL;
7177 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7178 mx->mx_dbx.md_dcmp = NULL;
7179 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7182 /** Final setup of a sorted-dups cursor.
7183 * Sets up the fields that depend on the data from the main cursor.
7184 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7185 * @param[in] node The data containing the #MDB_db record for the
7186 * sorted-dup database.
7189 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7191 MDB_xcursor *mx = mc->mc_xcursor;
7193 if (node->mn_flags & F_SUBDATA) {
7194 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7195 mx->mx_cursor.mc_pg[0] = 0;
7196 mx->mx_cursor.mc_snum = 0;
7197 mx->mx_cursor.mc_top = 0;
7198 mx->mx_cursor.mc_flags = C_SUB;
7200 MDB_page *fp = NODEDATA(node);
7201 mx->mx_db.md_pad = 0;
7202 mx->mx_db.md_flags = 0;
7203 mx->mx_db.md_depth = 1;
7204 mx->mx_db.md_branch_pages = 0;
7205 mx->mx_db.md_leaf_pages = 1;
7206 mx->mx_db.md_overflow_pages = 0;
7207 mx->mx_db.md_entries = NUMKEYS(fp);
7208 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7209 mx->mx_cursor.mc_snum = 1;
7210 mx->mx_cursor.mc_top = 0;
7211 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7212 mx->mx_cursor.mc_pg[0] = fp;
7213 mx->mx_cursor.mc_ki[0] = 0;
7214 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7215 mx->mx_db.md_flags = MDB_DUPFIXED;
7216 mx->mx_db.md_pad = fp->mp_pad;
7217 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7218 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7221 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7222 mx->mx_db.md_root));
7223 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7224 #if UINT_MAX < SIZE_MAX
7225 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7226 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7231 /** Fixup a sorted-dups cursor due to underlying update.
7232 * Sets up some fields that depend on the data from the main cursor.
7233 * Almost the same as init1, but skips initialization steps if the
7234 * xcursor had already been used.
7235 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7236 * @param[in] src_mx The xcursor of an up-to-date cursor.
7237 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7240 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7242 MDB_xcursor *mx = mc->mc_xcursor;
7245 mx->mx_cursor.mc_snum = 1;
7246 mx->mx_cursor.mc_top = 0;
7247 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7248 mx->mx_cursor.mc_ki[0] = 0;
7249 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7250 #if UINT_MAX < SIZE_MAX
7251 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7253 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7256 mx->mx_db = src_mx->mx_db;
7257 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7258 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7259 mx->mx_db.md_root));
7262 /** Initialize a cursor for a given transaction and database. */
7264 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7267 mc->mc_backup = NULL;
7270 mc->mc_db = &txn->mt_dbs[dbi];
7271 mc->mc_dbx = &txn->mt_dbxs[dbi];
7272 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7278 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7279 mdb_tassert(txn, mx != NULL);
7280 mc->mc_xcursor = mx;
7281 mdb_xcursor_init0(mc);
7283 mc->mc_xcursor = NULL;
7285 if (*mc->mc_dbflag & DB_STALE) {
7286 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7291 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7294 size_t size = sizeof(MDB_cursor);
7296 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7299 if (txn->mt_flags & MDB_TXN_BLOCKED)
7302 /* Allow read access to the freelist */
7303 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7306 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7307 size += sizeof(MDB_xcursor);
7309 if ((mc = malloc(size)) != NULL) {
7310 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7311 if (txn->mt_cursors) {
7312 mc->mc_next = txn->mt_cursors[dbi];
7313 txn->mt_cursors[dbi] = mc;
7314 mc->mc_flags |= C_UNTRACK;
7326 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7328 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7331 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7334 if (txn->mt_flags & MDB_TXN_BLOCKED)
7337 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7341 /* Return the count of duplicate data items for the current key */
7343 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7347 if (mc == NULL || countp == NULL)
7350 if (mc->mc_xcursor == NULL)
7351 return MDB_INCOMPATIBLE;
7353 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7356 if (!(mc->mc_flags & C_INITIALIZED))
7359 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7360 return MDB_NOTFOUND;
7362 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7363 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7366 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7369 *countp = mc->mc_xcursor->mx_db.md_entries;
7375 mdb_cursor_close(MDB_cursor *mc)
7377 if (mc && !mc->mc_backup) {
7378 /* remove from txn, if tracked */
7379 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7380 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7381 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7383 *prev = mc->mc_next;
7390 mdb_cursor_txn(MDB_cursor *mc)
7392 if (!mc) return NULL;
7397 mdb_cursor_dbi(MDB_cursor *mc)
7402 /** Replace the key for a branch node with a new key.
7403 * @param[in] mc Cursor pointing to the node to operate on.
7404 * @param[in] key The new key to use.
7405 * @return 0 on success, non-zero on failure.
7408 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7414 int delta, ksize, oksize;
7415 indx_t ptr, i, numkeys, indx;
7418 indx = mc->mc_ki[mc->mc_top];
7419 mp = mc->mc_pg[mc->mc_top];
7420 node = NODEPTR(mp, indx);
7421 ptr = mp->mp_ptrs[indx];
7425 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7426 k2.mv_data = NODEKEY(node);
7427 k2.mv_size = node->mn_ksize;
7428 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7430 mdb_dkey(&k2, kbuf2),
7436 /* Sizes must be 2-byte aligned. */
7437 ksize = EVEN(key->mv_size);
7438 oksize = EVEN(node->mn_ksize);
7439 delta = ksize - oksize;
7441 /* Shift node contents if EVEN(key length) changed. */
7443 if (delta > 0 && SIZELEFT(mp) < delta) {
7445 /* not enough space left, do a delete and split */
7446 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7447 pgno = NODEPGNO(node);
7448 mdb_node_del(mc, 0);
7449 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7452 numkeys = NUMKEYS(mp);
7453 for (i = 0; i < numkeys; i++) {
7454 if (mp->mp_ptrs[i] <= ptr)
7455 mp->mp_ptrs[i] -= delta;
7458 base = (char *)mp + mp->mp_upper + PAGEBASE;
7459 len = ptr - mp->mp_upper + NODESIZE;
7460 memmove(base - delta, base, len);
7461 mp->mp_upper -= delta;
7463 node = NODEPTR(mp, indx);
7466 /* But even if no shift was needed, update ksize */
7467 if (node->mn_ksize != key->mv_size)
7468 node->mn_ksize = key->mv_size;
7471 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7477 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7479 /** Move a node from csrc to cdst.
7482 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7489 unsigned short flags;
7493 /* Mark src and dst as dirty. */
7494 if ((rc = mdb_page_touch(csrc)) ||
7495 (rc = mdb_page_touch(cdst)))
7498 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7499 key.mv_size = csrc->mc_db->md_pad;
7500 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7502 data.mv_data = NULL;
7506 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7507 mdb_cassert(csrc, !((size_t)srcnode & 1));
7508 srcpg = NODEPGNO(srcnode);
7509 flags = srcnode->mn_flags;
7510 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7511 unsigned int snum = csrc->mc_snum;
7513 /* must find the lowest key below src */
7514 rc = mdb_page_search_lowest(csrc);
7517 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7518 key.mv_size = csrc->mc_db->md_pad;
7519 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7521 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7522 key.mv_size = NODEKSZ(s2);
7523 key.mv_data = NODEKEY(s2);
7525 csrc->mc_snum = snum--;
7526 csrc->mc_top = snum;
7528 key.mv_size = NODEKSZ(srcnode);
7529 key.mv_data = NODEKEY(srcnode);
7531 data.mv_size = NODEDSZ(srcnode);
7532 data.mv_data = NODEDATA(srcnode);
7534 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7535 unsigned int snum = cdst->mc_snum;
7538 /* must find the lowest key below dst */
7539 mdb_cursor_copy(cdst, &mn);
7540 rc = mdb_page_search_lowest(&mn);
7543 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7544 bkey.mv_size = mn.mc_db->md_pad;
7545 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7547 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7548 bkey.mv_size = NODEKSZ(s2);
7549 bkey.mv_data = NODEKEY(s2);
7551 mn.mc_snum = snum--;
7554 rc = mdb_update_key(&mn, &bkey);
7559 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7560 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7561 csrc->mc_ki[csrc->mc_top],
7563 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7564 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7566 /* Add the node to the destination page.
7568 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7569 if (rc != MDB_SUCCESS)
7572 /* Delete the node from the source page.
7574 mdb_node_del(csrc, key.mv_size);
7577 /* Adjust other cursors pointing to mp */
7578 MDB_cursor *m2, *m3;
7579 MDB_dbi dbi = csrc->mc_dbi;
7582 mp = cdst->mc_pg[csrc->mc_top];
7583 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7584 if (csrc->mc_flags & C_SUB)
7585 m3 = &m2->mc_xcursor->mx_cursor;
7588 if (m3 == cdst) continue;
7589 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] >=
7590 cdst->mc_ki[csrc->mc_top]) {
7591 m3->mc_ki[csrc->mc_top]++;
7595 mp = csrc->mc_pg[csrc->mc_top];
7596 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7597 if (csrc->mc_flags & C_SUB)
7598 m3 = &m2->mc_xcursor->mx_cursor;
7601 if (m3 == csrc) continue;
7602 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7603 csrc->mc_ki[csrc->mc_top]) {
7604 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7605 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7610 /* Update the parent separators.
7612 if (csrc->mc_ki[csrc->mc_top] == 0) {
7613 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7614 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7615 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7617 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7618 key.mv_size = NODEKSZ(srcnode);
7619 key.mv_data = NODEKEY(srcnode);
7621 DPRINTF(("update separator for source page %"Z"u to [%s]",
7622 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7623 mdb_cursor_copy(csrc, &mn);
7626 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7629 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7631 indx_t ix = csrc->mc_ki[csrc->mc_top];
7632 nullkey.mv_size = 0;
7633 csrc->mc_ki[csrc->mc_top] = 0;
7634 rc = mdb_update_key(csrc, &nullkey);
7635 csrc->mc_ki[csrc->mc_top] = ix;
7636 mdb_cassert(csrc, rc == MDB_SUCCESS);
7640 if (cdst->mc_ki[cdst->mc_top] == 0) {
7641 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7642 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7643 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7645 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7646 key.mv_size = NODEKSZ(srcnode);
7647 key.mv_data = NODEKEY(srcnode);
7649 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7650 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7651 mdb_cursor_copy(cdst, &mn);
7654 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7657 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7659 indx_t ix = cdst->mc_ki[cdst->mc_top];
7660 nullkey.mv_size = 0;
7661 cdst->mc_ki[cdst->mc_top] = 0;
7662 rc = mdb_update_key(cdst, &nullkey);
7663 cdst->mc_ki[cdst->mc_top] = ix;
7664 mdb_cassert(cdst, rc == MDB_SUCCESS);
7671 /** Merge one page into another.
7672 * The nodes from the page pointed to by \b csrc will
7673 * be copied to the page pointed to by \b cdst and then
7674 * the \b csrc page will be freed.
7675 * @param[in] csrc Cursor pointing to the source page.
7676 * @param[in] cdst Cursor pointing to the destination page.
7677 * @return 0 on success, non-zero on failure.
7680 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7682 MDB_page *psrc, *pdst;
7689 psrc = csrc->mc_pg[csrc->mc_top];
7690 pdst = cdst->mc_pg[cdst->mc_top];
7692 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7694 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7695 mdb_cassert(csrc, cdst->mc_snum > 1);
7697 /* Mark dst as dirty. */
7698 if ((rc = mdb_page_touch(cdst)))
7701 /* Move all nodes from src to dst.
7703 j = nkeys = NUMKEYS(pdst);
7704 if (IS_LEAF2(psrc)) {
7705 key.mv_size = csrc->mc_db->md_pad;
7706 key.mv_data = METADATA(psrc);
7707 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7708 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7709 if (rc != MDB_SUCCESS)
7711 key.mv_data = (char *)key.mv_data + key.mv_size;
7714 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7715 srcnode = NODEPTR(psrc, i);
7716 if (i == 0 && IS_BRANCH(psrc)) {
7719 mdb_cursor_copy(csrc, &mn);
7720 /* must find the lowest key below src */
7721 rc = mdb_page_search_lowest(&mn);
7724 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7725 key.mv_size = mn.mc_db->md_pad;
7726 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7728 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7729 key.mv_size = NODEKSZ(s2);
7730 key.mv_data = NODEKEY(s2);
7733 key.mv_size = srcnode->mn_ksize;
7734 key.mv_data = NODEKEY(srcnode);
7737 data.mv_size = NODEDSZ(srcnode);
7738 data.mv_data = NODEDATA(srcnode);
7739 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7740 if (rc != MDB_SUCCESS)
7745 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7746 pdst->mp_pgno, NUMKEYS(pdst),
7747 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7749 /* Unlink the src page from parent and add to free list.
7752 mdb_node_del(csrc, 0);
7753 if (csrc->mc_ki[csrc->mc_top] == 0) {
7755 rc = mdb_update_key(csrc, &key);
7763 psrc = csrc->mc_pg[csrc->mc_top];
7764 /* If not operating on FreeDB, allow this page to be reused
7765 * in this txn. Otherwise just add to free list.
7767 rc = mdb_page_loose(csrc, psrc);
7771 csrc->mc_db->md_leaf_pages--;
7773 csrc->mc_db->md_branch_pages--;
7775 /* Adjust other cursors pointing to mp */
7776 MDB_cursor *m2, *m3;
7777 MDB_dbi dbi = csrc->mc_dbi;
7779 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7780 if (csrc->mc_flags & C_SUB)
7781 m3 = &m2->mc_xcursor->mx_cursor;
7784 if (m3 == csrc) continue;
7785 if (m3->mc_snum < csrc->mc_snum) continue;
7786 if (m3->mc_pg[csrc->mc_top] == psrc) {
7787 m3->mc_pg[csrc->mc_top] = pdst;
7788 m3->mc_ki[csrc->mc_top] += nkeys;
7793 unsigned int snum = cdst->mc_snum;
7794 uint16_t depth = cdst->mc_db->md_depth;
7795 mdb_cursor_pop(cdst);
7796 rc = mdb_rebalance(cdst);
7797 /* Did the tree height change? */
7798 if (depth != cdst->mc_db->md_depth)
7799 snum += cdst->mc_db->md_depth - depth;
7800 cdst->mc_snum = snum;
7801 cdst->mc_top = snum-1;
7806 /** Copy the contents of a cursor.
7807 * @param[in] csrc The cursor to copy from.
7808 * @param[out] cdst The cursor to copy to.
7811 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7815 cdst->mc_txn = csrc->mc_txn;
7816 cdst->mc_dbi = csrc->mc_dbi;
7817 cdst->mc_db = csrc->mc_db;
7818 cdst->mc_dbx = csrc->mc_dbx;
7819 cdst->mc_snum = csrc->mc_snum;
7820 cdst->mc_top = csrc->mc_top;
7821 cdst->mc_flags = csrc->mc_flags;
7823 for (i=0; i<csrc->mc_snum; i++) {
7824 cdst->mc_pg[i] = csrc->mc_pg[i];
7825 cdst->mc_ki[i] = csrc->mc_ki[i];
7829 /** Rebalance the tree after a delete operation.
7830 * @param[in] mc Cursor pointing to the page where rebalancing
7832 * @return 0 on success, non-zero on failure.
7835 mdb_rebalance(MDB_cursor *mc)
7839 unsigned int ptop, minkeys, thresh;
7843 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7848 thresh = FILL_THRESHOLD;
7850 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7851 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7852 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7853 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7855 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7856 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7857 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7858 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7862 if (mc->mc_snum < 2) {
7863 MDB_page *mp = mc->mc_pg[0];
7865 DPUTS("Can't rebalance a subpage, ignoring");
7868 if (NUMKEYS(mp) == 0) {
7869 DPUTS("tree is completely empty");
7870 mc->mc_db->md_root = P_INVALID;
7871 mc->mc_db->md_depth = 0;
7872 mc->mc_db->md_leaf_pages = 0;
7873 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7876 /* Adjust cursors pointing to mp */
7879 mc->mc_flags &= ~C_INITIALIZED;
7881 MDB_cursor *m2, *m3;
7882 MDB_dbi dbi = mc->mc_dbi;
7884 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7885 if (mc->mc_flags & C_SUB)
7886 m3 = &m2->mc_xcursor->mx_cursor;
7889 if (m3->mc_snum < mc->mc_snum) continue;
7890 if (m3->mc_pg[0] == mp) {
7893 m3->mc_flags &= ~C_INITIALIZED;
7897 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7899 DPUTS("collapsing root page!");
7900 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7903 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7904 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7907 mc->mc_db->md_depth--;
7908 mc->mc_db->md_branch_pages--;
7909 mc->mc_ki[0] = mc->mc_ki[1];
7910 for (i = 1; i<mc->mc_db->md_depth; i++) {
7911 mc->mc_pg[i] = mc->mc_pg[i+1];
7912 mc->mc_ki[i] = mc->mc_ki[i+1];
7915 /* Adjust other cursors pointing to mp */
7916 MDB_cursor *m2, *m3;
7917 MDB_dbi dbi = mc->mc_dbi;
7919 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7920 if (mc->mc_flags & C_SUB)
7921 m3 = &m2->mc_xcursor->mx_cursor;
7924 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7925 if (m3->mc_pg[0] == mp) {
7926 for (i=0; i<m3->mc_snum; i++) {
7927 m3->mc_pg[i] = m3->mc_pg[i+1];
7928 m3->mc_ki[i] = m3->mc_ki[i+1];
7936 DPUTS("root page doesn't need rebalancing");
7940 /* The parent (branch page) must have at least 2 pointers,
7941 * otherwise the tree is invalid.
7943 ptop = mc->mc_top-1;
7944 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7946 /* Leaf page fill factor is below the threshold.
7947 * Try to move keys from left or right neighbor, or
7948 * merge with a neighbor page.
7953 mdb_cursor_copy(mc, &mn);
7954 mn.mc_xcursor = NULL;
7956 oldki = mc->mc_ki[mc->mc_top];
7957 if (mc->mc_ki[ptop] == 0) {
7958 /* We're the leftmost leaf in our parent.
7960 DPUTS("reading right neighbor");
7962 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7963 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7966 mn.mc_ki[mn.mc_top] = 0;
7967 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7969 /* There is at least one neighbor to the left.
7971 DPUTS("reading left neighbor");
7973 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7974 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7977 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7978 mc->mc_ki[mc->mc_top] = 0;
7981 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7982 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7983 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7985 /* If the neighbor page is above threshold and has enough keys,
7986 * move one key from it. Otherwise we should try to merge them.
7987 * (A branch page must never have less than 2 keys.)
7989 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7990 rc = mdb_node_move(&mn, mc);
7991 if (mc->mc_ki[mc->mc_top-1]) {
7995 if (mc->mc_ki[ptop] == 0) {
7996 rc = mdb_page_merge(&mn, mc);
7999 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8000 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8001 /* We want mdb_rebalance to find mn when doing fixups */
8002 if (mc->mc_flags & C_SUB) {
8003 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8004 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
8005 dummy.mc_xcursor = (MDB_xcursor *)&mn;
8007 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
8008 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
8010 rc = mdb_page_merge(mc, &mn);
8011 if (mc->mc_flags & C_SUB)
8012 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
8014 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
8015 mdb_cursor_copy(&mn, mc);
8017 mc->mc_flags &= ~C_EOF;
8019 mc->mc_ki[mc->mc_top] = oldki;
8023 /** Complete a delete operation started by #mdb_cursor_del(). */
8025 mdb_cursor_del0(MDB_cursor *mc)
8031 MDB_cursor *m2, *m3;
8032 MDB_dbi dbi = mc->mc_dbi;
8034 ki = mc->mc_ki[mc->mc_top];
8035 mp = mc->mc_pg[mc->mc_top];
8036 mdb_node_del(mc, mc->mc_db->md_pad);
8037 mc->mc_db->md_entries--;
8039 /* Adjust other cursors pointing to mp */
8040 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8041 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8042 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8044 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8046 if (m3->mc_pg[mc->mc_top] == mp) {
8047 if (m3->mc_ki[mc->mc_top] >= ki) {
8048 m3->mc_flags |= C_DEL;
8049 if (m3->mc_ki[mc->mc_top] > ki)
8050 m3->mc_ki[mc->mc_top]--;
8051 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8052 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8057 rc = mdb_rebalance(mc);
8059 if (rc == MDB_SUCCESS) {
8060 /* DB is totally empty now, just bail out.
8061 * Other cursors adjustments were already done
8062 * by mdb_rebalance and aren't needed here.
8067 mp = mc->mc_pg[mc->mc_top];
8068 nkeys = NUMKEYS(mp);
8070 /* Adjust other cursors pointing to mp */
8071 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8072 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8073 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8075 if (m3->mc_snum < mc->mc_snum)
8077 if (m3->mc_pg[mc->mc_top] == mp) {
8078 /* if m3 points past last node in page, find next sibling */
8079 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8080 rc = mdb_cursor_sibling(m3, 1);
8081 if (rc == MDB_NOTFOUND) {
8082 m3->mc_flags |= C_EOF;
8088 mc->mc_flags |= C_DEL;
8092 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8097 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8098 MDB_val *key, MDB_val *data)
8100 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8103 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8104 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8106 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8107 /* must ignore any data */
8111 return mdb_del0(txn, dbi, key, data, 0);
8115 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8116 MDB_val *key, MDB_val *data, unsigned flags)
8121 MDB_val rdata, *xdata;
8125 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8127 mdb_cursor_init(&mc, txn, dbi, &mx);
8136 flags |= MDB_NODUPDATA;
8138 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8140 /* let mdb_page_split know about this cursor if needed:
8141 * delete will trigger a rebalance; if it needs to move
8142 * a node from one page to another, it will have to
8143 * update the parent's separator key(s). If the new sepkey
8144 * is larger than the current one, the parent page may
8145 * run out of space, triggering a split. We need this
8146 * cursor to be consistent until the end of the rebalance.
8148 mc.mc_flags |= C_UNTRACK;
8149 mc.mc_next = txn->mt_cursors[dbi];
8150 txn->mt_cursors[dbi] = &mc;
8151 rc = mdb_cursor_del(&mc, flags);
8152 txn->mt_cursors[dbi] = mc.mc_next;
8157 /** Split a page and insert a new node.
8158 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8159 * The cursor will be updated to point to the actual page and index where
8160 * the node got inserted after the split.
8161 * @param[in] newkey The key for the newly inserted node.
8162 * @param[in] newdata The data for the newly inserted node.
8163 * @param[in] newpgno The page number, if the new node is a branch node.
8164 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8165 * @return 0 on success, non-zero on failure.
8168 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8169 unsigned int nflags)
8172 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8175 int i, j, split_indx, nkeys, pmax;
8176 MDB_env *env = mc->mc_txn->mt_env;
8178 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8179 MDB_page *copy = NULL;
8180 MDB_page *mp, *rp, *pp;
8185 mp = mc->mc_pg[mc->mc_top];
8186 newindx = mc->mc_ki[mc->mc_top];
8187 nkeys = NUMKEYS(mp);
8189 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8190 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8191 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8193 /* Create a right sibling. */
8194 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8196 rp->mp_pad = mp->mp_pad;
8197 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8199 if (mc->mc_snum < 2) {
8200 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8202 /* shift current top to make room for new parent */
8203 mc->mc_pg[1] = mc->mc_pg[0];
8204 mc->mc_ki[1] = mc->mc_ki[0];
8207 mc->mc_db->md_root = pp->mp_pgno;
8208 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8209 new_root = mc->mc_db->md_depth++;
8211 /* Add left (implicit) pointer. */
8212 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8213 /* undo the pre-push */
8214 mc->mc_pg[0] = mc->mc_pg[1];
8215 mc->mc_ki[0] = mc->mc_ki[1];
8216 mc->mc_db->md_root = mp->mp_pgno;
8217 mc->mc_db->md_depth--;
8224 ptop = mc->mc_top-1;
8225 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8228 mc->mc_flags |= C_SPLITTING;
8229 mdb_cursor_copy(mc, &mn);
8230 mn.mc_pg[mn.mc_top] = rp;
8231 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8233 if (nflags & MDB_APPEND) {
8234 mn.mc_ki[mn.mc_top] = 0;
8236 split_indx = newindx;
8240 split_indx = (nkeys+1) / 2;
8245 unsigned int lsize, rsize, ksize;
8246 /* Move half of the keys to the right sibling */
8247 x = mc->mc_ki[mc->mc_top] - split_indx;
8248 ksize = mc->mc_db->md_pad;
8249 split = LEAF2KEY(mp, split_indx, ksize);
8250 rsize = (nkeys - split_indx) * ksize;
8251 lsize = (nkeys - split_indx) * sizeof(indx_t);
8252 mp->mp_lower -= lsize;
8253 rp->mp_lower += lsize;
8254 mp->mp_upper += rsize - lsize;
8255 rp->mp_upper -= rsize - lsize;
8256 sepkey.mv_size = ksize;
8257 if (newindx == split_indx) {
8258 sepkey.mv_data = newkey->mv_data;
8260 sepkey.mv_data = split;
8263 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8264 memcpy(rp->mp_ptrs, split, rsize);
8265 sepkey.mv_data = rp->mp_ptrs;
8266 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8267 memcpy(ins, newkey->mv_data, ksize);
8268 mp->mp_lower += sizeof(indx_t);
8269 mp->mp_upper -= ksize - sizeof(indx_t);
8272 memcpy(rp->mp_ptrs, split, x * ksize);
8273 ins = LEAF2KEY(rp, x, ksize);
8274 memcpy(ins, newkey->mv_data, ksize);
8275 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8276 rp->mp_lower += sizeof(indx_t);
8277 rp->mp_upper -= ksize - sizeof(indx_t);
8278 mc->mc_ki[mc->mc_top] = x;
8279 mc->mc_pg[mc->mc_top] = rp;
8282 int psize, nsize, k;
8283 /* Maximum free space in an empty page */
8284 pmax = env->me_psize - PAGEHDRSZ;
8286 nsize = mdb_leaf_size(env, newkey, newdata);
8288 nsize = mdb_branch_size(env, newkey);
8289 nsize = EVEN(nsize);
8291 /* grab a page to hold a temporary copy */
8292 copy = mdb_page_malloc(mc->mc_txn, 1);
8297 copy->mp_pgno = mp->mp_pgno;
8298 copy->mp_flags = mp->mp_flags;
8299 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8300 copy->mp_upper = env->me_psize - PAGEBASE;
8302 /* prepare to insert */
8303 for (i=0, j=0; i<nkeys; i++) {
8305 copy->mp_ptrs[j++] = 0;
8307 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8310 /* When items are relatively large the split point needs
8311 * to be checked, because being off-by-one will make the
8312 * difference between success or failure in mdb_node_add.
8314 * It's also relevant if a page happens to be laid out
8315 * such that one half of its nodes are all "small" and
8316 * the other half of its nodes are "large." If the new
8317 * item is also "large" and falls on the half with
8318 * "large" nodes, it also may not fit.
8320 * As a final tweak, if the new item goes on the last
8321 * spot on the page (and thus, onto the new page), bias
8322 * the split so the new page is emptier than the old page.
8323 * This yields better packing during sequential inserts.
8325 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8326 /* Find split point */
8328 if (newindx <= split_indx || newindx >= nkeys) {
8330 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8335 for (; i!=k; i+=j) {
8340 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8341 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8343 if (F_ISSET(node->mn_flags, F_BIGDATA))
8344 psize += sizeof(pgno_t);
8346 psize += NODEDSZ(node);
8348 psize = EVEN(psize);
8350 if (psize > pmax || i == k-j) {
8351 split_indx = i + (j<0);
8356 if (split_indx == newindx) {
8357 sepkey.mv_size = newkey->mv_size;
8358 sepkey.mv_data = newkey->mv_data;
8360 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8361 sepkey.mv_size = node->mn_ksize;
8362 sepkey.mv_data = NODEKEY(node);
8367 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8369 /* Copy separator key to the parent.
8371 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8375 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8380 if (mn.mc_snum == mc->mc_snum) {
8381 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8382 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8383 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8384 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8389 /* Right page might now have changed parent.
8390 * Check if left page also changed parent.
8392 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8393 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8394 for (i=0; i<ptop; i++) {
8395 mc->mc_pg[i] = mn.mc_pg[i];
8396 mc->mc_ki[i] = mn.mc_ki[i];
8398 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8399 if (mn.mc_ki[ptop]) {
8400 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8402 /* find right page's left sibling */
8403 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8404 mdb_cursor_sibling(mc, 0);
8409 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8412 mc->mc_flags ^= C_SPLITTING;
8413 if (rc != MDB_SUCCESS) {
8416 if (nflags & MDB_APPEND) {
8417 mc->mc_pg[mc->mc_top] = rp;
8418 mc->mc_ki[mc->mc_top] = 0;
8419 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8422 for (i=0; i<mc->mc_top; i++)
8423 mc->mc_ki[i] = mn.mc_ki[i];
8424 } else if (!IS_LEAF2(mp)) {
8426 mc->mc_pg[mc->mc_top] = rp;
8431 rkey.mv_data = newkey->mv_data;
8432 rkey.mv_size = newkey->mv_size;
8438 /* Update index for the new key. */
8439 mc->mc_ki[mc->mc_top] = j;
8441 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8442 rkey.mv_data = NODEKEY(node);
8443 rkey.mv_size = node->mn_ksize;
8445 xdata.mv_data = NODEDATA(node);
8446 xdata.mv_size = NODEDSZ(node);
8449 pgno = NODEPGNO(node);
8450 flags = node->mn_flags;
8453 if (!IS_LEAF(mp) && j == 0) {
8454 /* First branch index doesn't need key data. */
8458 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8464 mc->mc_pg[mc->mc_top] = copy;
8469 } while (i != split_indx);
8471 nkeys = NUMKEYS(copy);
8472 for (i=0; i<nkeys; i++)
8473 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8474 mp->mp_lower = copy->mp_lower;
8475 mp->mp_upper = copy->mp_upper;
8476 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8477 env->me_psize - copy->mp_upper - PAGEBASE);
8479 /* reset back to original page */
8480 if (newindx < split_indx) {
8481 mc->mc_pg[mc->mc_top] = mp;
8482 if (nflags & MDB_RESERVE) {
8483 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8484 if (!(node->mn_flags & F_BIGDATA))
8485 newdata->mv_data = NODEDATA(node);
8488 mc->mc_pg[mc->mc_top] = rp;
8490 /* Make sure mc_ki is still valid.
8492 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8493 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8494 for (i=0; i<=ptop; i++) {
8495 mc->mc_pg[i] = mn.mc_pg[i];
8496 mc->mc_ki[i] = mn.mc_ki[i];
8503 /* Adjust other cursors pointing to mp */
8504 MDB_cursor *m2, *m3;
8505 MDB_dbi dbi = mc->mc_dbi;
8506 int fixup = NUMKEYS(mp);
8508 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8509 if (mc->mc_flags & C_SUB)
8510 m3 = &m2->mc_xcursor->mx_cursor;
8515 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8517 if (m3->mc_flags & C_SPLITTING)
8522 for (k=new_root; k>=0; k--) {
8523 m3->mc_ki[k+1] = m3->mc_ki[k];
8524 m3->mc_pg[k+1] = m3->mc_pg[k];
8526 if (m3->mc_ki[0] >= split_indx) {
8531 m3->mc_pg[0] = mc->mc_pg[0];
8535 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8536 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8537 m3->mc_ki[mc->mc_top]++;
8538 if (m3->mc_ki[mc->mc_top] >= fixup) {
8539 m3->mc_pg[mc->mc_top] = rp;
8540 m3->mc_ki[mc->mc_top] -= fixup;
8541 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8543 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8544 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8549 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8552 if (copy) /* tmp page */
8553 mdb_page_free(env, copy);
8555 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8560 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8561 MDB_val *key, MDB_val *data, unsigned int flags)
8566 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8569 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8572 mdb_cursor_init(&mc, txn, dbi, &mx);
8573 return mdb_cursor_put(&mc, key, data, flags);
8577 #define MDB_WBUF (1024*1024)
8580 /** State needed for a compacting copy. */
8581 typedef struct mdb_copy {
8582 pthread_mutex_t mc_mutex;
8583 pthread_cond_t mc_cond;
8590 pgno_t mc_next_pgno;
8593 volatile int mc_new;
8598 /** Dedicated writer thread for compacting copy. */
8599 static THREAD_RET ESECT CALL_CONV
8600 mdb_env_copythr(void *arg)
8604 int toggle = 0, wsize, rc;
8607 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8610 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8613 pthread_mutex_lock(&my->mc_mutex);
8615 pthread_cond_signal(&my->mc_cond);
8618 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8619 if (my->mc_new < 0) {
8624 wsize = my->mc_wlen[toggle];
8625 ptr = my->mc_wbuf[toggle];
8628 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8632 } else if (len > 0) {
8646 /* If there's an overflow page tail, write it too */
8647 if (my->mc_olen[toggle]) {
8648 wsize = my->mc_olen[toggle];
8649 ptr = my->mc_over[toggle];
8650 my->mc_olen[toggle] = 0;
8653 my->mc_wlen[toggle] = 0;
8655 pthread_cond_signal(&my->mc_cond);
8657 pthread_cond_signal(&my->mc_cond);
8658 pthread_mutex_unlock(&my->mc_mutex);
8659 return (THREAD_RET)0;
8663 /** Tell the writer thread there's a buffer ready to write */
8665 mdb_env_cthr_toggle(mdb_copy *my, int st)
8667 int toggle = my->mc_toggle ^ 1;
8668 pthread_mutex_lock(&my->mc_mutex);
8669 if (my->mc_status) {
8670 pthread_mutex_unlock(&my->mc_mutex);
8671 return my->mc_status;
8673 while (my->mc_new == 1)
8674 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8676 my->mc_toggle = toggle;
8677 pthread_cond_signal(&my->mc_cond);
8678 pthread_mutex_unlock(&my->mc_mutex);
8682 /** Depth-first tree traversal for compacting copy. */
8684 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8687 MDB_txn *txn = my->mc_txn;
8689 MDB_page *mo, *mp, *leaf;
8694 /* Empty DB, nothing to do */
8695 if (*pg == P_INVALID)
8702 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8705 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8709 /* Make cursor pages writable */
8710 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8714 for (i=0; i<mc.mc_top; i++) {
8715 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8716 mc.mc_pg[i] = (MDB_page *)ptr;
8717 ptr += my->mc_env->me_psize;
8720 /* This is writable space for a leaf page. Usually not needed. */
8721 leaf = (MDB_page *)ptr;
8723 toggle = my->mc_toggle;
8724 while (mc.mc_snum > 0) {
8726 mp = mc.mc_pg[mc.mc_top];
8730 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8731 for (i=0; i<n; i++) {
8732 ni = NODEPTR(mp, i);
8733 if (ni->mn_flags & F_BIGDATA) {
8737 /* Need writable leaf */
8739 mc.mc_pg[mc.mc_top] = leaf;
8740 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8742 ni = NODEPTR(mp, i);
8745 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8746 rc = mdb_page_get(txn, pg, &omp, NULL);
8749 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8750 rc = mdb_env_cthr_toggle(my, 1);
8753 toggle = my->mc_toggle;
8755 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8756 memcpy(mo, omp, my->mc_env->me_psize);
8757 mo->mp_pgno = my->mc_next_pgno;
8758 my->mc_next_pgno += omp->mp_pages;
8759 my->mc_wlen[toggle] += my->mc_env->me_psize;
8760 if (omp->mp_pages > 1) {
8761 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8762 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8763 rc = mdb_env_cthr_toggle(my, 1);
8766 toggle = my->mc_toggle;
8768 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8769 } else if (ni->mn_flags & F_SUBDATA) {
8772 /* Need writable leaf */
8774 mc.mc_pg[mc.mc_top] = leaf;
8775 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8777 ni = NODEPTR(mp, i);
8780 memcpy(&db, NODEDATA(ni), sizeof(db));
8781 my->mc_toggle = toggle;
8782 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8785 toggle = my->mc_toggle;
8786 memcpy(NODEDATA(ni), &db, sizeof(db));
8791 mc.mc_ki[mc.mc_top]++;
8792 if (mc.mc_ki[mc.mc_top] < n) {
8795 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8797 rc = mdb_page_get(txn, pg, &mp, NULL);
8802 mc.mc_ki[mc.mc_top] = 0;
8803 if (IS_BRANCH(mp)) {
8804 /* Whenever we advance to a sibling branch page,
8805 * we must proceed all the way down to its first leaf.
8807 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8810 mc.mc_pg[mc.mc_top] = mp;
8814 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8815 rc = mdb_env_cthr_toggle(my, 1);
8818 toggle = my->mc_toggle;
8820 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8821 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8822 mo->mp_pgno = my->mc_next_pgno++;
8823 my->mc_wlen[toggle] += my->mc_env->me_psize;
8825 /* Update parent if there is one */
8826 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8827 SETPGNO(ni, mo->mp_pgno);
8828 mdb_cursor_pop(&mc);
8830 /* Otherwise we're done */
8840 /** Copy environment with compaction. */
8842 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8847 MDB_txn *txn = NULL;
8852 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8853 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8854 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8855 if (my.mc_wbuf[0] == NULL)
8858 pthread_mutex_init(&my.mc_mutex, NULL);
8859 pthread_cond_init(&my.mc_cond, NULL);
8860 #ifdef HAVE_MEMALIGN
8861 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8862 if (my.mc_wbuf[0] == NULL)
8865 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8870 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8871 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8876 my.mc_next_pgno = 2;
8882 THREAD_CREATE(thr, mdb_env_copythr, &my);
8884 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8888 mp = (MDB_page *)my.mc_wbuf[0];
8889 memset(mp, 0, 2*env->me_psize);
8891 mp->mp_flags = P_META;
8892 mm = (MDB_meta *)METADATA(mp);
8893 mdb_env_init_meta0(env, mm);
8894 mm->mm_address = env->me_metas[0]->mm_address;
8896 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8898 mp->mp_flags = P_META;
8899 *(MDB_meta *)METADATA(mp) = *mm;
8900 mm = (MDB_meta *)METADATA(mp);
8902 /* Count the number of free pages, subtract from lastpg to find
8903 * number of active pages
8906 MDB_ID freecount = 0;
8909 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8910 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8911 freecount += *(MDB_ID *)data.mv_data;
8912 freecount += txn->mt_dbs[0].md_branch_pages +
8913 txn->mt_dbs[0].md_leaf_pages +
8914 txn->mt_dbs[0].md_overflow_pages;
8916 /* Set metapage 1 */
8917 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8918 mm->mm_dbs[1] = txn->mt_dbs[1];
8919 if (mm->mm_last_pg > 1) {
8920 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8923 mm->mm_dbs[1].md_root = P_INVALID;
8926 my.mc_wlen[0] = env->me_psize * 2;
8928 pthread_mutex_lock(&my.mc_mutex);
8930 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8931 pthread_mutex_unlock(&my.mc_mutex);
8932 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8933 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8934 rc = mdb_env_cthr_toggle(&my, 1);
8935 mdb_env_cthr_toggle(&my, -1);
8936 pthread_mutex_lock(&my.mc_mutex);
8938 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8939 pthread_mutex_unlock(&my.mc_mutex);
8944 CloseHandle(my.mc_cond);
8945 CloseHandle(my.mc_mutex);
8946 _aligned_free(my.mc_wbuf[0]);
8948 pthread_cond_destroy(&my.mc_cond);
8949 pthread_mutex_destroy(&my.mc_mutex);
8950 free(my.mc_wbuf[0]);
8955 /** Copy environment as-is. */
8957 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8959 MDB_txn *txn = NULL;
8960 mdb_mutexref_t wmutex = NULL;
8966 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8970 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8973 /* Do the lock/unlock of the reader mutex before starting the
8974 * write txn. Otherwise other read txns could block writers.
8976 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8981 /* We must start the actual read txn after blocking writers */
8982 mdb_txn_end(txn, MDB_END_RESET_TMP);
8984 /* Temporarily block writers until we snapshot the meta pages */
8985 wmutex = env->me_wmutex;
8986 if (LOCK_MUTEX(rc, env, wmutex))
8989 rc = mdb_txn_renew0(txn);
8991 UNLOCK_MUTEX(wmutex);
8996 wsize = env->me_psize * 2;
9000 DO_WRITE(rc, fd, ptr, w2, len);
9004 } else if (len > 0) {
9010 /* Non-blocking or async handles are not supported */
9016 UNLOCK_MUTEX(wmutex);
9021 w2 = txn->mt_next_pgno * env->me_psize;
9024 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9031 if (wsize > MAX_WRITE)
9035 DO_WRITE(rc, fd, ptr, w2, len);
9039 } else if (len > 0) {
9056 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9058 if (flags & MDB_CP_COMPACT)
9059 return mdb_env_copyfd1(env, fd);
9061 return mdb_env_copyfd0(env, fd);
9065 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9067 return mdb_env_copyfd2(env, fd, 0);
9071 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9075 HANDLE newfd = INVALID_HANDLE_VALUE;
9077 if (env->me_flags & MDB_NOSUBDIR) {
9078 lpath = (char *)path;
9081 len += sizeof(DATANAME);
9082 lpath = malloc(len);
9085 sprintf(lpath, "%s" DATANAME, path);
9088 /* The destination path must exist, but the destination file must not.
9089 * We don't want the OS to cache the writes, since the source data is
9090 * already in the OS cache.
9093 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9094 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9096 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9098 if (newfd == INVALID_HANDLE_VALUE) {
9103 if (env->me_psize >= env->me_os_psize) {
9105 /* Set O_DIRECT if the file system supports it */
9106 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9107 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9109 #ifdef F_NOCACHE /* __APPLE__ */
9110 rc = fcntl(newfd, F_NOCACHE, 1);
9118 rc = mdb_env_copyfd2(env, newfd, flags);
9121 if (!(env->me_flags & MDB_NOSUBDIR))
9123 if (newfd != INVALID_HANDLE_VALUE)
9124 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9131 mdb_env_copy(MDB_env *env, const char *path)
9133 return mdb_env_copy2(env, path, 0);
9137 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9139 if (flag & ~CHANGEABLE)
9142 env->me_flags |= flag;
9144 env->me_flags &= ~flag;
9149 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9154 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9159 mdb_env_set_userctx(MDB_env *env, void *ctx)
9163 env->me_userctx = ctx;
9168 mdb_env_get_userctx(MDB_env *env)
9170 return env ? env->me_userctx : NULL;
9174 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9179 env->me_assert_func = func;
9185 mdb_env_get_path(MDB_env *env, const char **arg)
9190 *arg = env->me_path;
9195 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9204 /** Common code for #mdb_stat() and #mdb_env_stat().
9205 * @param[in] env the environment to operate in.
9206 * @param[in] db the #MDB_db record containing the stats to return.
9207 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9208 * @return 0, this function always succeeds.
9211 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9213 arg->ms_psize = env->me_psize;
9214 arg->ms_depth = db->md_depth;
9215 arg->ms_branch_pages = db->md_branch_pages;
9216 arg->ms_leaf_pages = db->md_leaf_pages;
9217 arg->ms_overflow_pages = db->md_overflow_pages;
9218 arg->ms_entries = db->md_entries;
9224 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9228 if (env == NULL || arg == NULL)
9231 meta = mdb_env_pick_meta(env);
9233 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9237 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9241 if (env == NULL || arg == NULL)
9244 meta = mdb_env_pick_meta(env);
9245 arg->me_mapaddr = meta->mm_address;
9246 arg->me_last_pgno = meta->mm_last_pg;
9247 arg->me_last_txnid = meta->mm_txnid;
9249 arg->me_mapsize = env->me_mapsize;
9250 arg->me_maxreaders = env->me_maxreaders;
9251 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9255 /** Set the default comparison functions for a database.
9256 * Called immediately after a database is opened to set the defaults.
9257 * The user can then override them with #mdb_set_compare() or
9258 * #mdb_set_dupsort().
9259 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9260 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9263 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9265 uint16_t f = txn->mt_dbs[dbi].md_flags;
9267 txn->mt_dbxs[dbi].md_cmp =
9268 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9269 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9271 txn->mt_dbxs[dbi].md_dcmp =
9272 !(f & MDB_DUPSORT) ? 0 :
9273 ((f & MDB_INTEGERDUP)
9274 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9275 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9278 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9284 int rc, dbflag, exact;
9285 unsigned int unused = 0, seq;
9288 if (flags & ~VALID_FLAGS)
9290 if (txn->mt_flags & MDB_TXN_BLOCKED)
9296 if (flags & PERSISTENT_FLAGS) {
9297 uint16_t f2 = flags & PERSISTENT_FLAGS;
9298 /* make sure flag changes get committed */
9299 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9300 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9301 txn->mt_flags |= MDB_TXN_DIRTY;
9304 mdb_default_cmp(txn, MAIN_DBI);
9308 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9309 mdb_default_cmp(txn, MAIN_DBI);
9312 /* Is the DB already open? */
9314 for (i=2; i<txn->mt_numdbs; i++) {
9315 if (!txn->mt_dbxs[i].md_name.mv_size) {
9316 /* Remember this free slot */
9317 if (!unused) unused = i;
9320 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9321 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9327 /* If no free slot and max hit, fail */
9328 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9329 return MDB_DBS_FULL;
9331 /* Cannot mix named databases with some mainDB flags */
9332 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9333 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9335 /* Find the DB info */
9336 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9339 key.mv_data = (void *)name;
9340 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9341 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9342 if (rc == MDB_SUCCESS) {
9343 /* make sure this is actually a DB */
9344 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9345 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9346 return MDB_INCOMPATIBLE;
9347 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9348 /* Create if requested */
9349 data.mv_size = sizeof(MDB_db);
9350 data.mv_data = &dummy;
9351 memset(&dummy, 0, sizeof(dummy));
9352 dummy.md_root = P_INVALID;
9353 dummy.md_flags = flags & PERSISTENT_FLAGS;
9354 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9358 /* OK, got info, add to table */
9359 if (rc == MDB_SUCCESS) {
9360 unsigned int slot = unused ? unused : txn->mt_numdbs;
9361 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9362 txn->mt_dbxs[slot].md_name.mv_size = len;
9363 txn->mt_dbxs[slot].md_rel = NULL;
9364 txn->mt_dbflags[slot] = dbflag;
9365 /* txn-> and env-> are the same in read txns, use
9366 * tmp variable to avoid undefined assignment
9368 seq = ++txn->mt_env->me_dbiseqs[slot];
9369 txn->mt_dbiseqs[slot] = seq;
9371 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9373 mdb_default_cmp(txn, slot);
9383 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9385 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9388 if (txn->mt_flags & MDB_TXN_BLOCKED)
9391 if (txn->mt_dbflags[dbi] & DB_STALE) {
9394 /* Stale, must read the DB's root. cursor_init does it for us. */
9395 mdb_cursor_init(&mc, txn, dbi, &mx);
9397 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9400 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9403 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9405 ptr = env->me_dbxs[dbi].md_name.mv_data;
9406 /* If there was no name, this was already closed */
9408 env->me_dbxs[dbi].md_name.mv_data = NULL;
9409 env->me_dbxs[dbi].md_name.mv_size = 0;
9410 env->me_dbflags[dbi] = 0;
9411 env->me_dbiseqs[dbi]++;
9416 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9418 /* We could return the flags for the FREE_DBI too but what's the point? */
9419 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9421 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9425 /** Add all the DB's pages to the free list.
9426 * @param[in] mc Cursor on the DB to free.
9427 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9428 * @return 0 on success, non-zero on failure.
9431 mdb_drop0(MDB_cursor *mc, int subs)
9435 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9436 if (rc == MDB_SUCCESS) {
9437 MDB_txn *txn = mc->mc_txn;
9442 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9443 * This also avoids any P_LEAF2 pages, which have no nodes.
9445 if (mc->mc_flags & C_SUB)
9448 mdb_cursor_copy(mc, &mx);
9449 while (mc->mc_snum > 0) {
9450 MDB_page *mp = mc->mc_pg[mc->mc_top];
9451 unsigned n = NUMKEYS(mp);
9453 for (i=0; i<n; i++) {
9454 ni = NODEPTR(mp, i);
9455 if (ni->mn_flags & F_BIGDATA) {
9458 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9459 rc = mdb_page_get(txn, pg, &omp, NULL);
9462 mdb_cassert(mc, IS_OVERFLOW(omp));
9463 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9467 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9468 mdb_xcursor_init1(mc, ni);
9469 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9475 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9477 for (i=0; i<n; i++) {
9479 ni = NODEPTR(mp, i);
9482 mdb_midl_xappend(txn->mt_free_pgs, pg);
9487 mc->mc_ki[mc->mc_top] = i;
9488 rc = mdb_cursor_sibling(mc, 1);
9490 if (rc != MDB_NOTFOUND)
9492 /* no more siblings, go back to beginning
9493 * of previous level.
9497 for (i=1; i<mc->mc_snum; i++) {
9499 mc->mc_pg[i] = mx.mc_pg[i];
9504 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9507 txn->mt_flags |= MDB_TXN_ERROR;
9508 } else if (rc == MDB_NOTFOUND) {
9514 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9516 MDB_cursor *mc, *m2;
9519 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9522 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9525 if (TXN_DBI_CHANGED(txn, dbi))
9528 rc = mdb_cursor_open(txn, dbi, &mc);
9532 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9533 /* Invalidate the dropped DB's cursors */
9534 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9535 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9539 /* Can't delete the main DB */
9540 if (del && dbi > MAIN_DBI) {
9541 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9543 txn->mt_dbflags[dbi] = DB_STALE;
9544 mdb_dbi_close(txn->mt_env, dbi);
9546 txn->mt_flags |= MDB_TXN_ERROR;
9549 /* reset the DB record, mark it dirty */
9550 txn->mt_dbflags[dbi] |= DB_DIRTY;
9551 txn->mt_dbs[dbi].md_depth = 0;
9552 txn->mt_dbs[dbi].md_branch_pages = 0;
9553 txn->mt_dbs[dbi].md_leaf_pages = 0;
9554 txn->mt_dbs[dbi].md_overflow_pages = 0;
9555 txn->mt_dbs[dbi].md_entries = 0;
9556 txn->mt_dbs[dbi].md_root = P_INVALID;
9558 txn->mt_flags |= MDB_TXN_DIRTY;
9561 mdb_cursor_close(mc);
9565 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9567 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9570 txn->mt_dbxs[dbi].md_cmp = cmp;
9574 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9576 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9579 txn->mt_dbxs[dbi].md_dcmp = cmp;
9583 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9585 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9588 txn->mt_dbxs[dbi].md_rel = rel;
9592 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9594 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9597 txn->mt_dbxs[dbi].md_relctx = ctx;
9602 mdb_env_get_maxkeysize(MDB_env *env)
9604 return ENV_MAXKEY(env);
9608 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9610 unsigned int i, rdrs;
9613 int rc = 0, first = 1;
9617 if (!env->me_txns) {
9618 return func("(no reader locks)\n", ctx);
9620 rdrs = env->me_txns->mti_numreaders;
9621 mr = env->me_txns->mti_readers;
9622 for (i=0; i<rdrs; i++) {
9624 txnid_t txnid = mr[i].mr_txnid;
9625 sprintf(buf, txnid == (txnid_t)-1 ?
9626 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9627 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9630 rc = func(" pid thread txnid\n", ctx);
9634 rc = func(buf, ctx);
9640 rc = func("(no active readers)\n", ctx);
9645 /** Insert pid into list if not already present.
9646 * return -1 if already present.
9649 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9651 /* binary search of pid in list */
9653 unsigned cursor = 1;
9655 unsigned n = ids[0];
9658 unsigned pivot = n >> 1;
9659 cursor = base + pivot + 1;
9660 val = pid - ids[cursor];
9665 } else if ( val > 0 ) {
9670 /* found, so it's a duplicate */
9679 for (n = ids[0]; n > cursor; n--)
9686 mdb_reader_check(MDB_env *env, int *dead)
9692 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9695 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9697 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9699 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9700 unsigned int i, j, rdrs;
9702 MDB_PID_T *pids, pid;
9703 int rc = MDB_SUCCESS, count = 0;
9705 rdrs = env->me_txns->mti_numreaders;
9706 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9710 mr = env->me_txns->mti_readers;
9711 for (i=0; i<rdrs; i++) {
9713 if (pid && pid != env->me_pid) {
9714 if (mdb_pid_insert(pids, pid) == 0) {
9715 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9716 /* Stale reader found */
9719 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9720 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9722 rdrs = 0; /* the above checked all readers */
9724 /* Recheck, a new process may have reused pid */
9725 if (mdb_reader_pid(env, Pidcheck, pid))
9730 if (mr[j].mr_pid == pid) {
9731 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9732 (unsigned) pid, mr[j].mr_txnid));
9737 UNLOCK_MUTEX(rmutex);
9748 #ifdef MDB_ROBUST_SUPPORTED
9749 /** Handle #LOCK_MUTEX0() failure.
9750 * Try to repair the lock file if the mutex owner died.
9751 * @param[in] env the environment handle
9752 * @param[in] mutex LOCK_MUTEX0() mutex
9753 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9754 * @return 0 on success with the mutex locked, or an error code on failure.
9757 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
9762 if (rc == MDB_OWNERDEAD) {
9763 /* We own the mutex. Clean up after dead previous owner. */
9765 rlocked = (mutex == env->me_rmutex);
9767 /* Keep mti_txnid updated, otherwise next writer can
9768 * overwrite data which latest meta page refers to.
9770 meta = mdb_env_pick_meta(env);
9771 env->me_txns->mti_txnid = meta->mm_txnid;
9772 /* env is hosed if the dead thread was ours */
9774 env->me_flags |= MDB_FATAL_ERROR;
9779 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9780 (rc ? "this process' env is hosed" : "recovering")));
9781 rc2 = mdb_reader_check0(env, rlocked, NULL);
9783 rc2 = mdb_mutex_consistent(mutex);
9784 if (rc || (rc = rc2)) {
9785 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9786 UNLOCK_MUTEX(mutex);
9792 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9797 #endif /* MDB_ROBUST_SUPPORTED */
projects / openldap / blob