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 2011-2014 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>
128 #include <valgrind/memcheck.h>
129 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
130 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
131 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
132 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
133 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
135 #define VGMEMP_CREATE(h,r,z)
136 #define VGMEMP_ALLOC(h,a,s)
137 #define VGMEMP_FREE(h,a)
138 #define VGMEMP_DESTROY(h)
139 #define VGMEMP_DEFINED(a,s)
143 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
144 /* Solaris just defines one or the other */
145 # define LITTLE_ENDIAN 1234
146 # define BIG_ENDIAN 4321
147 # ifdef _LITTLE_ENDIAN
148 # define BYTE_ORDER LITTLE_ENDIAN
150 # define BYTE_ORDER BIG_ENDIAN
153 # define BYTE_ORDER __BYTE_ORDER
157 #ifndef LITTLE_ENDIAN
158 #define LITTLE_ENDIAN __LITTLE_ENDIAN
161 #define BIG_ENDIAN __BIG_ENDIAN
164 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
165 #define MISALIGNED_OK 1
171 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
172 # error "Unknown or unsupported endianness (BYTE_ORDER)"
173 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
174 # error "Two's complement, reasonably sized integer types, please"
178 /** Put infrequently used env functions in separate section */
180 # define ESECT __attribute__ ((section("__TEXT,text_env")))
182 # define ESECT __attribute__ ((section("text_env")))
188 /** @defgroup internal LMDB Internals
191 /** @defgroup compat Compatibility Macros
192 * A bunch of macros to minimize the amount of platform-specific ifdefs
193 * needed throughout the rest of the code. When the features this library
194 * needs are similar enough to POSIX to be hidden in a one-or-two line
195 * replacement, this macro approach is used.
199 /** Features under development */
204 /** Wrapper around __func__, which is a C99 feature */
205 #if __STDC_VERSION__ >= 199901L
206 # define mdb_func_ __func__
207 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
208 # define mdb_func_ __FUNCTION__
210 /* If a debug message says <mdb_unknown>(), update the #if statements above */
211 # define mdb_func_ "<mdb_unknown>"
215 #define MDB_USE_HASH 1
216 #define MDB_PIDLOCK 0
217 #define THREAD_RET DWORD
218 #define pthread_t HANDLE
219 #define pthread_mutex_t HANDLE
220 #define pthread_cond_t HANDLE
221 #define pthread_key_t DWORD
222 #define pthread_self() GetCurrentThreadId()
223 #define pthread_key_create(x,y) \
224 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
225 #define pthread_key_delete(x) TlsFree(x)
226 #define pthread_getspecific(x) TlsGetValue(x)
227 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
228 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
229 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
230 #define pthread_cond_signal(x) SetEvent(*x)
231 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
232 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
233 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
234 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
235 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
236 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
237 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
238 #define getpid() GetCurrentProcessId()
239 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
240 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
241 #define ErrCode() GetLastError()
242 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
243 #define close(fd) (CloseHandle(fd) ? 0 : -1)
244 #define munmap(ptr,len) UnmapViewOfFile(ptr)
245 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
246 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
248 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
252 #define THREAD_RET void *
253 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
254 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
255 #define Z "z" /**< printf format modifier for size_t */
257 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
258 #define MDB_PIDLOCK 1
260 #ifdef MDB_USE_POSIX_SEM
262 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
263 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
264 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
265 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
268 mdb_sem_wait(sem_t *sem)
271 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
276 /** Lock the reader mutex.
278 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
279 /** Unlock the reader mutex.
281 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
283 /** Lock the writer mutex.
284 * Only a single write transaction is allowed at a time. Other writers
285 * will block waiting for this mutex.
287 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
288 /** Unlock the writer mutex.
290 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
291 #endif /* MDB_USE_POSIX_SEM */
293 /** Get the error code for the last failed system function.
295 #define ErrCode() errno
297 /** An abstraction for a file handle.
298 * On POSIX systems file handles are small integers. On Windows
299 * they're opaque pointers.
303 /** A value for an invalid file handle.
304 * Mainly used to initialize file variables and signify that they are
307 #define INVALID_HANDLE_VALUE (-1)
309 /** Get the size of a memory page for the system.
310 * This is the basic size that the platform's memory manager uses, and is
311 * fundamental to the use of memory-mapped files.
313 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
316 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
319 #define MNAME_LEN (sizeof(pthread_mutex_t))
325 /** A flag for opening a file and requesting synchronous data writes.
326 * This is only used when writing a meta page. It's not strictly needed;
327 * we could just do a normal write and then immediately perform a flush.
328 * But if this flag is available it saves us an extra system call.
330 * @note If O_DSYNC is undefined but exists in /usr/include,
331 * preferably set some compiler flag to get the definition.
332 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
335 # define MDB_DSYNC O_DSYNC
339 /** Function for flushing the data of a file. Define this to fsync
340 * if fdatasync() is not supported.
342 #ifndef MDB_FDATASYNC
343 # define MDB_FDATASYNC fdatasync
347 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
358 /** A page number in the database.
359 * Note that 64 bit page numbers are overkill, since pages themselves
360 * already represent 12-13 bits of addressable memory, and the OS will
361 * always limit applications to a maximum of 63 bits of address space.
363 * @note In the #MDB_node structure, we only store 48 bits of this value,
364 * which thus limits us to only 60 bits of addressable data.
366 typedef MDB_ID pgno_t;
368 /** A transaction ID.
369 * See struct MDB_txn.mt_txnid for details.
371 typedef MDB_ID txnid_t;
373 /** @defgroup debug Debug Macros
377 /** Enable debug output. Needs variable argument macros (a C99 feature).
378 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
379 * read from and written to the database (used for free space management).
385 static int mdb_debug;
386 static txnid_t mdb_debug_start;
388 /** Print a debug message with printf formatting.
389 * Requires double parenthesis around 2 or more args.
391 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
392 # define DPRINTF0(fmt, ...) \
393 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
395 # define DPRINTF(args) ((void) 0)
397 /** Print a debug string.
398 * The string is printed literally, with no format processing.
400 #define DPUTS(arg) DPRINTF(("%s", arg))
401 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
403 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
406 /** @brief The maximum size of a database page.
408 * It is 32k or 64k, since value-PAGEBASE must fit in
409 * #MDB_page.%mp_upper.
411 * LMDB will use database pages < OS pages if needed.
412 * That causes more I/O in write transactions: The OS must
413 * know (read) the whole page before writing a partial page.
415 * Note that we don't currently support Huge pages. On Linux,
416 * regular data files cannot use Huge pages, and in general
417 * Huge pages aren't actually pageable. We rely on the OS
418 * demand-pager to read our data and page it out when memory
419 * pressure from other processes is high. So until OSs have
420 * actual paging support for Huge pages, they're not viable.
422 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
424 /** The minimum number of keys required in a database page.
425 * Setting this to a larger value will place a smaller bound on the
426 * maximum size of a data item. Data items larger than this size will
427 * be pushed into overflow pages instead of being stored directly in
428 * the B-tree node. This value used to default to 4. With a page size
429 * of 4096 bytes that meant that any item larger than 1024 bytes would
430 * go into an overflow page. That also meant that on average 2-3KB of
431 * each overflow page was wasted space. The value cannot be lower than
432 * 2 because then there would no longer be a tree structure. With this
433 * value, items larger than 2KB will go into overflow pages, and on
434 * average only 1KB will be wasted.
436 #define MDB_MINKEYS 2
438 /** A stamp that identifies a file as an LMDB file.
439 * There's nothing special about this value other than that it is easily
440 * recognizable, and it will reflect any byte order mismatches.
442 #define MDB_MAGIC 0xBEEFC0DE
444 /** The version number for a database's datafile format. */
445 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
446 /** The version number for a database's lockfile format. */
447 #define MDB_LOCK_VERSION 1
449 /** @brief The max size of a key we can write, or 0 for dynamic max.
451 * Define this as 0 to compute the max from the page size. 511
452 * is default for backwards compat: liblmdb <= 0.9.10 can break
453 * when modifying a DB with keys/dupsort data bigger than its max.
454 * #MDB_DEVEL sets the default to 0.
456 * Data items in an #MDB_DUPSORT database are also limited to
457 * this size, since they're actually keys of a sub-DB. Keys and
458 * #MDB_DUPSORT data items must fit on a node in a regular page.
460 #ifndef MDB_MAXKEYSIZE
461 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
464 /** The maximum size of a key we can write to the environment. */
466 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
468 #define ENV_MAXKEY(env) ((env)->me_maxkey)
471 /** @brief The maximum size of a data item.
473 * We only store a 32 bit value for node sizes.
475 #define MAXDATASIZE 0xffffffffUL
478 /** Key size which fits in a #DKBUF.
481 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
484 * This is used for printing a hex dump of a key's contents.
486 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
487 /** Display a key in hex.
489 * Invoke a function to display a key in hex.
491 #define DKEY(x) mdb_dkey(x, kbuf)
497 /** An invalid page number.
498 * Mainly used to denote an empty tree.
500 #define P_INVALID (~(pgno_t)0)
502 /** Test if the flags \b f are set in a flag word \b w. */
503 #define F_ISSET(w, f) (((w) & (f)) == (f))
505 /** Round \b n up to an even number. */
506 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
508 /** Used for offsets within a single page.
509 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
512 typedef uint16_t indx_t;
514 /** Default size of memory map.
515 * This is certainly too small for any actual applications. Apps should always set
516 * the size explicitly using #mdb_env_set_mapsize().
518 #define DEFAULT_MAPSIZE 1048576
520 /** @defgroup readers Reader Lock Table
521 * Readers don't acquire any locks for their data access. Instead, they
522 * simply record their transaction ID in the reader table. The reader
523 * mutex is needed just to find an empty slot in the reader table. The
524 * slot's address is saved in thread-specific data so that subsequent read
525 * transactions started by the same thread need no further locking to proceed.
527 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
529 * No reader table is used if the database is on a read-only filesystem, or
530 * if #MDB_NOLOCK is set.
532 * Since the database uses multi-version concurrency control, readers don't
533 * actually need any locking. This table is used to keep track of which
534 * readers are using data from which old transactions, so that we'll know
535 * when a particular old transaction is no longer in use. Old transactions
536 * that have discarded any data pages can then have those pages reclaimed
537 * for use by a later write transaction.
539 * The lock table is constructed such that reader slots are aligned with the
540 * processor's cache line size. Any slot is only ever used by one thread.
541 * This alignment guarantees that there will be no contention or cache
542 * thrashing as threads update their own slot info, and also eliminates
543 * any need for locking when accessing a slot.
545 * A writer thread will scan every slot in the table to determine the oldest
546 * outstanding reader transaction. Any freed pages older than this will be
547 * reclaimed by the writer. The writer doesn't use any locks when scanning
548 * this table. This means that there's no guarantee that the writer will
549 * see the most up-to-date reader info, but that's not required for correct
550 * operation - all we need is to know the upper bound on the oldest reader,
551 * we don't care at all about the newest reader. So the only consequence of
552 * reading stale information here is that old pages might hang around a
553 * while longer before being reclaimed. That's actually good anyway, because
554 * the longer we delay reclaiming old pages, the more likely it is that a
555 * string of contiguous pages can be found after coalescing old pages from
556 * many old transactions together.
559 /** Number of slots in the reader table.
560 * This value was chosen somewhat arbitrarily. 126 readers plus a
561 * couple mutexes fit exactly into 8KB on my development machine.
562 * Applications should set the table size using #mdb_env_set_maxreaders().
564 #define DEFAULT_READERS 126
566 /** The size of a CPU cache line in bytes. We want our lock structures
567 * aligned to this size to avoid false cache line sharing in the
569 * This value works for most CPUs. For Itanium this should be 128.
575 /** The information we store in a single slot of the reader table.
576 * In addition to a transaction ID, we also record the process and
577 * thread ID that owns a slot, so that we can detect stale information,
578 * e.g. threads or processes that went away without cleaning up.
579 * @note We currently don't check for stale records. We simply re-init
580 * the table when we know that we're the only process opening the
583 typedef struct MDB_rxbody {
584 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
585 * Multiple readers that start at the same time will probably have the
586 * same ID here. Again, it's not important to exclude them from
587 * anything; all we need to know is which version of the DB they
588 * started from so we can avoid overwriting any data used in that
589 * particular version.
592 /** The process ID of the process owning this reader txn. */
594 /** The thread ID of the thread owning this txn. */
598 /** The actual reader record, with cacheline padding. */
599 typedef struct MDB_reader {
602 /** shorthand for mrb_txnid */
603 #define mr_txnid mru.mrx.mrb_txnid
604 #define mr_pid mru.mrx.mrb_pid
605 #define mr_tid mru.mrx.mrb_tid
606 /** cache line alignment */
607 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
611 /** The header for the reader table.
612 * The table resides in a memory-mapped file. (This is a different file
613 * than is used for the main database.)
615 * For POSIX the actual mutexes reside in the shared memory of this
616 * mapped file. On Windows, mutexes are named objects allocated by the
617 * kernel; we store the mutex names in this mapped file so that other
618 * processes can grab them. This same approach is also used on
619 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
620 * process-shared POSIX mutexes. For these cases where a named object
621 * is used, the object name is derived from a 64 bit FNV hash of the
622 * environment pathname. As such, naming collisions are extremely
623 * unlikely. If a collision occurs, the results are unpredictable.
625 typedef struct MDB_txbody {
626 /** Stamp identifying this as an LMDB file. It must be set
629 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
631 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
632 char mtb_rmname[MNAME_LEN];
634 /** Mutex protecting access to this table.
635 * This is the reader lock that #LOCK_MUTEX_R acquires.
637 pthread_mutex_t mtb_mutex;
639 /** The ID of the last transaction committed to the database.
640 * This is recorded here only for convenience; the value can always
641 * be determined by reading the main database meta pages.
644 /** The number of slots that have been used in the reader table.
645 * This always records the maximum count, it is not decremented
646 * when readers release their slots.
648 unsigned mtb_numreaders;
651 /** The actual reader table definition. */
652 typedef struct MDB_txninfo {
655 #define mti_magic mt1.mtb.mtb_magic
656 #define mti_format mt1.mtb.mtb_format
657 #define mti_mutex mt1.mtb.mtb_mutex
658 #define mti_rmname mt1.mtb.mtb_rmname
659 #define mti_txnid mt1.mtb.mtb_txnid
660 #define mti_numreaders mt1.mtb.mtb_numreaders
661 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
664 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
665 char mt2_wmname[MNAME_LEN];
666 #define mti_wmname mt2.mt2_wmname
668 pthread_mutex_t mt2_wmutex;
669 #define mti_wmutex mt2.mt2_wmutex
671 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
673 MDB_reader mti_readers[1];
676 /** Lockfile format signature: version, features and field layout */
677 #define MDB_LOCK_FORMAT \
679 ((MDB_LOCK_VERSION) \
680 /* Flags which describe functionality */ \
681 + (((MDB_PIDLOCK) != 0) << 16)))
684 /** Common header for all page types.
685 * Overflow records occupy a number of contiguous pages with no
686 * headers on any page after the first.
688 typedef struct MDB_page {
689 #define mp_pgno mp_p.p_pgno
690 #define mp_next mp_p.p_next
692 pgno_t p_pgno; /**< page number */
693 struct MDB_page *p_next; /**< for in-memory list of freed pages */
696 /** @defgroup mdb_page Page Flags
698 * Flags for the page headers.
701 #define P_BRANCH 0x01 /**< branch page */
702 #define P_LEAF 0x02 /**< leaf page */
703 #define P_OVERFLOW 0x04 /**< overflow page */
704 #define P_META 0x08 /**< meta page */
705 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
706 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
707 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
708 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
709 #define P_KEEP 0x8000 /**< leave this page alone during spill */
711 uint16_t mp_flags; /**< @ref mdb_page */
712 #define mp_lower mp_pb.pb.pb_lower
713 #define mp_upper mp_pb.pb.pb_upper
714 #define mp_pages mp_pb.pb_pages
717 indx_t pb_lower; /**< lower bound of free space */
718 indx_t pb_upper; /**< upper bound of free space */
720 uint32_t pb_pages; /**< number of overflow pages */
722 indx_t mp_ptrs[1]; /**< dynamic size */
725 /** Size of the page header, excluding dynamic data at the end */
726 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
728 /** Address of first usable data byte in a page, after the header */
729 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
731 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
732 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
734 /** Number of nodes on a page */
735 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
737 /** The amount of space remaining in the page */
738 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
740 /** The percentage of space used in the page, in tenths of a percent. */
741 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
742 ((env)->me_psize - PAGEHDRSZ))
743 /** The minimum page fill factor, in tenths of a percent.
744 * Pages emptier than this are candidates for merging.
746 #define FILL_THRESHOLD 250
748 /** Test if a page is a leaf page */
749 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
750 /** Test if a page is a LEAF2 page */
751 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
752 /** Test if a page is a branch page */
753 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
754 /** Test if a page is an overflow page */
755 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
756 /** Test if a page is a sub page */
757 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
759 /** The number of overflow pages needed to store the given size. */
760 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
762 /** Link in #MDB_txn.%mt_loose_pgs list */
763 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
765 /** Header for a single key/data pair within a page.
766 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
767 * We guarantee 2-byte alignment for 'MDB_node's.
769 typedef struct MDB_node {
770 /** lo and hi are used for data size on leaf nodes and for
771 * child pgno on branch nodes. On 64 bit platforms, flags
772 * is also used for pgno. (Branch nodes have no flags).
773 * They are in host byte order in case that lets some
774 * accesses be optimized into a 32-bit word access.
776 #if BYTE_ORDER == LITTLE_ENDIAN
777 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
779 unsigned short mn_hi, mn_lo;
781 /** @defgroup mdb_node Node Flags
783 * Flags for node headers.
786 #define F_BIGDATA 0x01 /**< data put on overflow page */
787 #define F_SUBDATA 0x02 /**< data is a sub-database */
788 #define F_DUPDATA 0x04 /**< data has duplicates */
790 /** valid flags for #mdb_node_add() */
791 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
794 unsigned short mn_flags; /**< @ref mdb_node */
795 unsigned short mn_ksize; /**< key size */
796 char mn_data[1]; /**< key and data are appended here */
799 /** Size of the node header, excluding dynamic data at the end */
800 #define NODESIZE offsetof(MDB_node, mn_data)
802 /** Bit position of top word in page number, for shifting mn_flags */
803 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
805 /** Size of a node in a branch page with a given key.
806 * This is just the node header plus the key, there is no data.
808 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
810 /** Size of a node in a leaf page with a given key and data.
811 * This is node header plus key plus data size.
813 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
815 /** Address of node \b i in page \b p */
816 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
818 /** Address of the key for the node */
819 #define NODEKEY(node) (void *)((node)->mn_data)
821 /** Address of the data for a node */
822 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
824 /** Get the page number pointed to by a branch node */
825 #define NODEPGNO(node) \
826 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
827 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
828 /** Set the page number in a branch node */
829 #define SETPGNO(node,pgno) do { \
830 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
831 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
833 /** Get the size of the data in a leaf node */
834 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
835 /** Set the size of the data for a leaf node */
836 #define SETDSZ(node,size) do { \
837 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
838 /** The size of a key in a node */
839 #define NODEKSZ(node) ((node)->mn_ksize)
841 /** Copy a page number from src to dst */
843 #define COPY_PGNO(dst,src) dst = src
845 #if SIZE_MAX > 4294967295UL
846 #define COPY_PGNO(dst,src) do { \
847 unsigned short *s, *d; \
848 s = (unsigned short *)&(src); \
849 d = (unsigned short *)&(dst); \
856 #define COPY_PGNO(dst,src) do { \
857 unsigned short *s, *d; \
858 s = (unsigned short *)&(src); \
859 d = (unsigned short *)&(dst); \
865 /** The address of a key in a LEAF2 page.
866 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
867 * There are no node headers, keys are stored contiguously.
869 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
871 /** Set the \b node's key into \b keyptr, if requested. */
872 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
873 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
875 /** Set the \b node's key into \b key. */
876 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
878 /** Information about a single database in the environment. */
879 typedef struct MDB_db {
880 uint32_t md_pad; /**< also ksize for LEAF2 pages */
881 uint16_t md_flags; /**< @ref mdb_dbi_open */
882 uint16_t md_depth; /**< depth of this tree */
883 pgno_t md_branch_pages; /**< number of internal pages */
884 pgno_t md_leaf_pages; /**< number of leaf pages */
885 pgno_t md_overflow_pages; /**< number of overflow pages */
886 size_t md_entries; /**< number of data items */
887 pgno_t md_root; /**< the root page of this tree */
890 /** mdb_dbi_open flags */
891 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
892 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
893 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
894 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
896 /** Handle for the DB used to track free pages. */
898 /** Handle for the default DB. */
901 /** Meta page content.
902 * A meta page is the start point for accessing a database snapshot.
903 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
905 typedef struct MDB_meta {
906 /** Stamp identifying this as an LMDB file. It must be set
909 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
911 void *mm_address; /**< address for fixed mapping */
912 size_t mm_mapsize; /**< size of mmap region */
913 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
914 /** The size of pages used in this DB */
915 #define mm_psize mm_dbs[0].md_pad
916 /** Any persistent environment flags. @ref mdb_env */
917 #define mm_flags mm_dbs[0].md_flags
918 pgno_t mm_last_pg; /**< last used page in file */
919 txnid_t mm_txnid; /**< txnid that committed this page */
922 /** Buffer for a stack-allocated meta page.
923 * The members define size and alignment, and silence type
924 * aliasing warnings. They are not used directly; that could
925 * mean incorrectly using several union members in parallel.
927 typedef union MDB_metabuf {
930 char mm_pad[PAGEHDRSZ];
935 /** Auxiliary DB info.
936 * The information here is mostly static/read-only. There is
937 * only a single copy of this record in the environment.
939 typedef struct MDB_dbx {
940 MDB_val md_name; /**< name of the database */
941 MDB_cmp_func *md_cmp; /**< function for comparing keys */
942 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
943 MDB_rel_func *md_rel; /**< user relocate function */
944 void *md_relctx; /**< user-provided context for md_rel */
947 /** A database transaction.
948 * Every operation requires a transaction handle.
951 MDB_txn *mt_parent; /**< parent of a nested txn */
952 MDB_txn *mt_child; /**< nested txn under this txn */
953 pgno_t mt_next_pgno; /**< next unallocated page */
954 /** The ID of this transaction. IDs are integers incrementing from 1.
955 * Only committed write transactions increment the ID. If a transaction
956 * aborts, the ID may be re-used by the next writer.
959 MDB_env *mt_env; /**< the DB environment */
960 /** The list of pages that became unused during this transaction.
963 /** The list of loose pages that became unused and may be reused
964 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
966 MDB_page *mt_loose_pgs;
967 /* #Number of loose pages (#mt_loose_pgs) */
969 /** The sorted list of dirty pages we temporarily wrote to disk
970 * because the dirty list was full. page numbers in here are
971 * shifted left by 1, deleted slots have the LSB set.
973 MDB_IDL mt_spill_pgs;
975 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
977 /** For read txns: This thread/txn's reader table slot, or NULL. */
980 /** Array of records for each DB known in the environment. */
982 /** Array of MDB_db records for each known DB */
984 /** Array of sequence numbers for each DB handle */
985 unsigned int *mt_dbiseqs;
986 /** @defgroup mt_dbflag Transaction DB Flags
990 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
991 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
992 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
993 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
995 /** In write txns, array of cursors for each DB */
996 MDB_cursor **mt_cursors;
997 /** Array of flags for each DB */
998 unsigned char *mt_dbflags;
999 /** Number of DB records in use. This number only ever increments;
1000 * we don't decrement it when individual DB handles are closed.
1004 /** @defgroup mdb_txn Transaction Flags
1008 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1009 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1010 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1011 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1013 unsigned int mt_flags; /**< @ref mdb_txn */
1014 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1015 * Includes ancestor txns' dirty pages not hidden by other txns'
1016 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1017 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1019 unsigned int mt_dirty_room;
1022 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1023 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1024 * raise this on a 64 bit machine.
1026 #define CURSOR_STACK 32
1030 /** Cursors are used for all DB operations.
1031 * A cursor holds a path of (page pointer, key index) from the DB
1032 * root to a position in the DB, plus other state. #MDB_DUPSORT
1033 * cursors include an xcursor to the current data item. Write txns
1034 * track their cursors and keep them up to date when data moves.
1035 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1036 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1039 /** Next cursor on this DB in this txn */
1040 MDB_cursor *mc_next;
1041 /** Backup of the original cursor if this cursor is a shadow */
1042 MDB_cursor *mc_backup;
1043 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1044 struct MDB_xcursor *mc_xcursor;
1045 /** The transaction that owns this cursor */
1047 /** The database handle this cursor operates on */
1049 /** The database record for this cursor */
1051 /** The database auxiliary record for this cursor */
1053 /** The @ref mt_dbflag for this database */
1054 unsigned char *mc_dbflag;
1055 unsigned short mc_snum; /**< number of pushed pages */
1056 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1057 /** @defgroup mdb_cursor Cursor Flags
1059 * Cursor state flags.
1062 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1063 #define C_EOF 0x02 /**< No more data */
1064 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1065 #define C_DEL 0x08 /**< last op was a cursor_del */
1066 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1067 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1069 unsigned int mc_flags; /**< @ref mdb_cursor */
1070 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1071 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1074 /** Context for sorted-dup records.
1075 * We could have gone to a fully recursive design, with arbitrarily
1076 * deep nesting of sub-databases. But for now we only handle these
1077 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1079 typedef struct MDB_xcursor {
1080 /** A sub-cursor for traversing the Dup DB */
1081 MDB_cursor mx_cursor;
1082 /** The database record for this Dup DB */
1084 /** The auxiliary DB record for this Dup DB */
1086 /** The @ref mt_dbflag for this Dup DB */
1087 unsigned char mx_dbflag;
1090 /** State of FreeDB old pages, stored in the MDB_env */
1091 typedef struct MDB_pgstate {
1092 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1093 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1096 /** The database environment. */
1098 HANDLE me_fd; /**< The main data file */
1099 HANDLE me_lfd; /**< The lock file */
1100 HANDLE me_mfd; /**< just for writing the meta pages */
1101 /** Failed to update the meta page. Probably an I/O error. */
1102 #define MDB_FATAL_ERROR 0x80000000U
1103 /** Some fields are initialized. */
1104 #define MDB_ENV_ACTIVE 0x20000000U
1105 /** me_txkey is set */
1106 #define MDB_ENV_TXKEY 0x10000000U
1107 /** fdatasync is unreliable */
1108 #define MDB_FSYNCONLY 0x08000000U
1109 uint32_t me_flags; /**< @ref mdb_env */
1110 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1111 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1112 unsigned int me_maxreaders; /**< size of the reader table */
1113 unsigned int me_numreaders; /**< max numreaders set by this env */
1114 MDB_dbi me_numdbs; /**< number of DBs opened */
1115 MDB_dbi me_maxdbs; /**< size of the DB table */
1116 MDB_PID_T me_pid; /**< process ID of this env */
1117 char *me_path; /**< path to the DB files */
1118 char *me_map; /**< the memory map of the data file */
1119 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1120 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1121 void *me_pbuf; /**< scratch area for DUPSORT put() */
1122 MDB_txn *me_txn; /**< current write transaction */
1123 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1124 size_t me_mapsize; /**< size of the data memory map */
1125 off_t me_size; /**< current file size */
1126 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1127 MDB_dbx *me_dbxs; /**< array of static DB info */
1128 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1129 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1130 pthread_key_t me_txkey; /**< thread-key for readers */
1131 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1132 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1133 # define me_pglast me_pgstate.mf_pglast
1134 # define me_pghead me_pgstate.mf_pghead
1135 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1136 /** IDL of pages that became unused in a write txn */
1137 MDB_IDL me_free_pgs;
1138 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1139 MDB_ID2L me_dirty_list;
1140 /** Max number of freelist items that can fit in a single overflow page */
1142 /** Max size of a node on a page */
1143 unsigned int me_nodemax;
1144 #if !(MDB_MAXKEYSIZE)
1145 unsigned int me_maxkey; /**< max size of a key */
1147 int me_live_reader; /**< have liveness lock in reader table */
1149 int me_pidquery; /**< Used in OpenProcess */
1150 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1152 #elif defined(MDB_USE_POSIX_SEM)
1153 sem_t *me_rmutex; /* Shared mutexes are not supported */
1156 void *me_userctx; /**< User-settable context */
1157 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1160 /** Nested transaction */
1161 typedef struct MDB_ntxn {
1162 MDB_txn mnt_txn; /**< the transaction */
1163 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1166 /** max number of pages to commit in one writev() call */
1167 #define MDB_COMMIT_PAGES 64
1168 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1169 #undef MDB_COMMIT_PAGES
1170 #define MDB_COMMIT_PAGES IOV_MAX
1173 /** max bytes to write in one call */
1174 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1176 /** Check \b txn and \b dbi arguments to a function */
1177 #define TXN_DBI_EXIST(txn, dbi) \
1178 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1180 /** Check for misused \b dbi handles */
1181 #define TXN_DBI_CHANGED(txn, dbi) \
1182 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1184 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1185 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1186 static int mdb_page_touch(MDB_cursor *mc);
1188 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1189 static int mdb_page_search_root(MDB_cursor *mc,
1190 MDB_val *key, int modify);
1191 #define MDB_PS_MODIFY 1
1192 #define MDB_PS_ROOTONLY 2
1193 #define MDB_PS_FIRST 4
1194 #define MDB_PS_LAST 8
1195 static int mdb_page_search(MDB_cursor *mc,
1196 MDB_val *key, int flags);
1197 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1199 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1200 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1201 pgno_t newpgno, unsigned int nflags);
1203 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1204 static int mdb_env_pick_meta(const MDB_env *env);
1205 static int mdb_env_write_meta(MDB_txn *txn);
1206 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1207 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1209 static void mdb_env_close0(MDB_env *env, int excl);
1211 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1212 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1213 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1214 static void mdb_node_del(MDB_cursor *mc, int ksize);
1215 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1216 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1217 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1218 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1219 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1221 static int mdb_rebalance(MDB_cursor *mc);
1222 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1224 static void mdb_cursor_pop(MDB_cursor *mc);
1225 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1227 static int mdb_cursor_del0(MDB_cursor *mc);
1228 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1229 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1230 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1231 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1232 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1234 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1235 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1237 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1238 static void mdb_xcursor_init0(MDB_cursor *mc);
1239 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1241 static int mdb_drop0(MDB_cursor *mc, int subs);
1242 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1245 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1249 static SECURITY_DESCRIPTOR mdb_null_sd;
1250 static SECURITY_ATTRIBUTES mdb_all_sa;
1251 static int mdb_sec_inited;
1254 /** Return the library version info. */
1256 mdb_version(int *major, int *minor, int *patch)
1258 if (major) *major = MDB_VERSION_MAJOR;
1259 if (minor) *minor = MDB_VERSION_MINOR;
1260 if (patch) *patch = MDB_VERSION_PATCH;
1261 return MDB_VERSION_STRING;
1264 /** Table of descriptions for LMDB @ref errors */
1265 static char *const mdb_errstr[] = {
1266 "MDB_KEYEXIST: Key/data pair already exists",
1267 "MDB_NOTFOUND: No matching key/data pair found",
1268 "MDB_PAGE_NOTFOUND: Requested page not found",
1269 "MDB_CORRUPTED: Located page was wrong type",
1270 "MDB_PANIC: Update of meta page failed",
1271 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1272 "MDB_INVALID: File is not an LMDB file",
1273 "MDB_MAP_FULL: Environment mapsize limit reached",
1274 "MDB_DBS_FULL: Environment maxdbs limit reached",
1275 "MDB_READERS_FULL: Environment maxreaders limit reached",
1276 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1277 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1278 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1279 "MDB_PAGE_FULL: Internal error - page has no more space",
1280 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1281 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1282 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1283 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1284 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1285 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1289 mdb_strerror(int err)
1292 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1293 * This works as long as no function between the call to mdb_strerror
1294 * and the actual use of the message uses more than 4K of stack.
1297 char buf[1024], *ptr = buf;
1301 return ("Successful return: 0");
1303 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1304 i = err - MDB_KEYEXIST;
1305 return mdb_errstr[i];
1309 /* These are the C-runtime error codes we use. The comment indicates
1310 * their numeric value, and the Win32 error they would correspond to
1311 * if the error actually came from a Win32 API. A major mess, we should
1312 * have used LMDB-specific error codes for everything.
1315 case ENOENT: /* 2, FILE_NOT_FOUND */
1316 case EIO: /* 5, ACCESS_DENIED */
1317 case ENOMEM: /* 12, INVALID_ACCESS */
1318 case EACCES: /* 13, INVALID_DATA */
1319 case EBUSY: /* 16, CURRENT_DIRECTORY */
1320 case EINVAL: /* 22, BAD_COMMAND */
1321 case ENOSPC: /* 28, OUT_OF_PAPER */
1322 return strerror(err);
1327 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1328 FORMAT_MESSAGE_IGNORE_INSERTS,
1329 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1332 return strerror(err);
1336 /** assert(3) variant in cursor context */
1337 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1338 /** assert(3) variant in transaction context */
1339 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1340 /** assert(3) variant in environment context */
1341 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1344 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1345 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1348 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1349 const char *func, const char *file, int line)
1352 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1353 file, line, expr_txt, func);
1354 if (env->me_assert_func)
1355 env->me_assert_func(env, buf);
1356 fprintf(stderr, "%s\n", buf);
1360 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1364 /** Return the page number of \b mp which may be sub-page, for debug output */
1366 mdb_dbg_pgno(MDB_page *mp)
1369 COPY_PGNO(ret, mp->mp_pgno);
1373 /** Display a key in hexadecimal and return the address of the result.
1374 * @param[in] key the key to display
1375 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1376 * @return The key in hexadecimal form.
1379 mdb_dkey(MDB_val *key, char *buf)
1382 unsigned char *c = key->mv_data;
1388 if (key->mv_size > DKBUF_MAXKEYSIZE)
1389 return "MDB_MAXKEYSIZE";
1390 /* may want to make this a dynamic check: if the key is mostly
1391 * printable characters, print it as-is instead of converting to hex.
1395 for (i=0; i<key->mv_size; i++)
1396 ptr += sprintf(ptr, "%02x", *c++);
1398 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1404 mdb_leafnode_type(MDB_node *n)
1406 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1407 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1408 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1411 /** Display all the keys in the page. */
1413 mdb_page_list(MDB_page *mp)
1415 pgno_t pgno = mdb_dbg_pgno(mp);
1416 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1418 unsigned int i, nkeys, nsize, total = 0;
1422 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1423 case P_BRANCH: type = "Branch page"; break;
1424 case P_LEAF: type = "Leaf page"; break;
1425 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1426 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1427 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1429 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1430 pgno, mp->mp_pages, state);
1433 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1434 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1437 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1441 nkeys = NUMKEYS(mp);
1442 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1444 for (i=0; i<nkeys; i++) {
1445 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1446 key.mv_size = nsize = mp->mp_pad;
1447 key.mv_data = LEAF2KEY(mp, i, nsize);
1449 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1452 node = NODEPTR(mp, i);
1453 key.mv_size = node->mn_ksize;
1454 key.mv_data = node->mn_data;
1455 nsize = NODESIZE + key.mv_size;
1456 if (IS_BRANCH(mp)) {
1457 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1461 if (F_ISSET(node->mn_flags, F_BIGDATA))
1462 nsize += sizeof(pgno_t);
1464 nsize += NODEDSZ(node);
1466 nsize += sizeof(indx_t);
1467 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1468 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1470 total = EVEN(total);
1472 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1473 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1477 mdb_cursor_chk(MDB_cursor *mc)
1483 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1484 for (i=0; i<mc->mc_top; i++) {
1486 node = NODEPTR(mp, mc->mc_ki[i]);
1487 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1490 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1496 /** Count all the pages in each DB and in the freelist
1497 * and make sure it matches the actual number of pages
1499 * All named DBs must be open for a correct count.
1501 static void mdb_audit(MDB_txn *txn)
1505 MDB_ID freecount, count;
1510 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1511 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1512 freecount += *(MDB_ID *)data.mv_data;
1513 mdb_tassert(txn, rc == MDB_NOTFOUND);
1516 for (i = 0; i<txn->mt_numdbs; i++) {
1518 if (!(txn->mt_dbflags[i] & DB_VALID))
1520 mdb_cursor_init(&mc, txn, i, &mx);
1521 if (txn->mt_dbs[i].md_root == P_INVALID)
1523 count += txn->mt_dbs[i].md_branch_pages +
1524 txn->mt_dbs[i].md_leaf_pages +
1525 txn->mt_dbs[i].md_overflow_pages;
1526 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1527 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1528 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1531 mp = mc.mc_pg[mc.mc_top];
1532 for (j=0; j<NUMKEYS(mp); j++) {
1533 MDB_node *leaf = NODEPTR(mp, j);
1534 if (leaf->mn_flags & F_SUBDATA) {
1536 memcpy(&db, NODEDATA(leaf), sizeof(db));
1537 count += db.md_branch_pages + db.md_leaf_pages +
1538 db.md_overflow_pages;
1542 mdb_tassert(txn, rc == MDB_NOTFOUND);
1545 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1546 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1547 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1553 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1555 return txn->mt_dbxs[dbi].md_cmp(a, b);
1559 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1561 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1564 /** Allocate memory for a page.
1565 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1568 mdb_page_malloc(MDB_txn *txn, unsigned num)
1570 MDB_env *env = txn->mt_env;
1571 MDB_page *ret = env->me_dpages;
1572 size_t psize = env->me_psize, sz = psize, off;
1573 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1574 * For a single page alloc, we init everything after the page header.
1575 * For multi-page, we init the final page; if the caller needed that
1576 * many pages they will be filling in at least up to the last page.
1580 VGMEMP_ALLOC(env, ret, sz);
1581 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1582 env->me_dpages = ret->mp_next;
1585 psize -= off = PAGEHDRSZ;
1590 if ((ret = malloc(sz)) != NULL) {
1591 VGMEMP_ALLOC(env, ret, sz);
1592 if (!(env->me_flags & MDB_NOMEMINIT)) {
1593 memset((char *)ret + off, 0, psize);
1597 txn->mt_flags |= MDB_TXN_ERROR;
1601 /** Free a single page.
1602 * Saves single pages to a list, for future reuse.
1603 * (This is not used for multi-page overflow pages.)
1606 mdb_page_free(MDB_env *env, MDB_page *mp)
1608 mp->mp_next = env->me_dpages;
1609 VGMEMP_FREE(env, mp);
1610 env->me_dpages = mp;
1613 /** Free a dirty page */
1615 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1617 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1618 mdb_page_free(env, dp);
1620 /* large pages just get freed directly */
1621 VGMEMP_FREE(env, dp);
1626 /** Return all dirty pages to dpage list */
1628 mdb_dlist_free(MDB_txn *txn)
1630 MDB_env *env = txn->mt_env;
1631 MDB_ID2L dl = txn->mt_u.dirty_list;
1632 unsigned i, n = dl[0].mid;
1634 for (i = 1; i <= n; i++) {
1635 mdb_dpage_free(env, dl[i].mptr);
1640 /** Loosen or free a single page.
1641 * Saves single pages to a list for future reuse
1642 * in this same txn. It has been pulled from the freeDB
1643 * and already resides on the dirty list, but has been
1644 * deleted. Use these pages first before pulling again
1647 * If the page wasn't dirtied in this txn, just add it
1648 * to this txn's free list.
1651 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1654 pgno_t pgno = mp->mp_pgno;
1655 MDB_txn *txn = mc->mc_txn;
1657 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1658 if (txn->mt_parent) {
1659 MDB_ID2 *dl = txn->mt_u.dirty_list;
1660 /* If txn has a parent, make sure the page is in our
1664 unsigned x = mdb_mid2l_search(dl, pgno);
1665 if (x <= dl[0].mid && dl[x].mid == pgno) {
1666 if (mp != dl[x].mptr) { /* bad cursor? */
1667 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1668 txn->mt_flags |= MDB_TXN_ERROR;
1669 return MDB_CORRUPTED;
1676 /* no parent txn, so it's just ours */
1681 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1683 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1684 txn->mt_loose_pgs = mp;
1685 txn->mt_loose_count++;
1686 mp->mp_flags |= P_LOOSE;
1688 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1696 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1697 * @param[in] mc A cursor handle for the current operation.
1698 * @param[in] pflags Flags of the pages to update:
1699 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1700 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1701 * @return 0 on success, non-zero on failure.
1704 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1706 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1707 MDB_txn *txn = mc->mc_txn;
1713 int rc = MDB_SUCCESS, level;
1715 /* Mark pages seen by cursors */
1716 if (mc->mc_flags & C_UNTRACK)
1717 mc = NULL; /* will find mc in mt_cursors */
1718 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1719 for (; mc; mc=mc->mc_next) {
1720 if (!(mc->mc_flags & C_INITIALIZED))
1722 for (m3 = mc;; m3 = &mx->mx_cursor) {
1724 for (j=0; j<m3->mc_snum; j++) {
1726 if ((mp->mp_flags & Mask) == pflags)
1727 mp->mp_flags ^= P_KEEP;
1729 mx = m3->mc_xcursor;
1730 /* Proceed to mx if it is at a sub-database */
1731 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1733 if (! (mp && (mp->mp_flags & P_LEAF)))
1735 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1736 if (!(leaf->mn_flags & F_SUBDATA))
1745 /* Mark dirty root pages */
1746 for (i=0; i<txn->mt_numdbs; i++) {
1747 if (txn->mt_dbflags[i] & DB_DIRTY) {
1748 pgno_t pgno = txn->mt_dbs[i].md_root;
1749 if (pgno == P_INVALID)
1751 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1753 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1754 dp->mp_flags ^= P_KEEP;
1762 static int mdb_page_flush(MDB_txn *txn, int keep);
1764 /** Spill pages from the dirty list back to disk.
1765 * This is intended to prevent running into #MDB_TXN_FULL situations,
1766 * but note that they may still occur in a few cases:
1767 * 1) our estimate of the txn size could be too small. Currently this
1768 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1769 * 2) child txns may run out of space if their parents dirtied a
1770 * lot of pages and never spilled them. TODO: we probably should do
1771 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1772 * the parent's dirty_room is below a given threshold.
1774 * Otherwise, if not using nested txns, it is expected that apps will
1775 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1776 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1777 * If the txn never references them again, they can be left alone.
1778 * If the txn only reads them, they can be used without any fuss.
1779 * If the txn writes them again, they can be dirtied immediately without
1780 * going thru all of the work of #mdb_page_touch(). Such references are
1781 * handled by #mdb_page_unspill().
1783 * Also note, we never spill DB root pages, nor pages of active cursors,
1784 * because we'll need these back again soon anyway. And in nested txns,
1785 * we can't spill a page in a child txn if it was already spilled in a
1786 * parent txn. That would alter the parent txns' data even though
1787 * the child hasn't committed yet, and we'd have no way to undo it if
1788 * the child aborted.
1790 * @param[in] m0 cursor A cursor handle identifying the transaction and
1791 * database for which we are checking space.
1792 * @param[in] key For a put operation, the key being stored.
1793 * @param[in] data For a put operation, the data being stored.
1794 * @return 0 on success, non-zero on failure.
1797 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1799 MDB_txn *txn = m0->mc_txn;
1801 MDB_ID2L dl = txn->mt_u.dirty_list;
1802 unsigned int i, j, need;
1805 if (m0->mc_flags & C_SUB)
1808 /* Estimate how much space this op will take */
1809 i = m0->mc_db->md_depth;
1810 /* Named DBs also dirty the main DB */
1811 if (m0->mc_dbi > MAIN_DBI)
1812 i += txn->mt_dbs[MAIN_DBI].md_depth;
1813 /* For puts, roughly factor in the key+data size */
1815 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1816 i += i; /* double it for good measure */
1819 if (txn->mt_dirty_room > i)
1822 if (!txn->mt_spill_pgs) {
1823 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1824 if (!txn->mt_spill_pgs)
1827 /* purge deleted slots */
1828 MDB_IDL sl = txn->mt_spill_pgs;
1829 unsigned int num = sl[0];
1831 for (i=1; i<=num; i++) {
1838 /* Preserve pages which may soon be dirtied again */
1839 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1842 /* Less aggressive spill - we originally spilled the entire dirty list,
1843 * with a few exceptions for cursor pages and DB root pages. But this
1844 * turns out to be a lot of wasted effort because in a large txn many
1845 * of those pages will need to be used again. So now we spill only 1/8th
1846 * of the dirty pages. Testing revealed this to be a good tradeoff,
1847 * better than 1/2, 1/4, or 1/10.
1849 if (need < MDB_IDL_UM_MAX / 8)
1850 need = MDB_IDL_UM_MAX / 8;
1852 /* Save the page IDs of all the pages we're flushing */
1853 /* flush from the tail forward, this saves a lot of shifting later on. */
1854 for (i=dl[0].mid; i && need; i--) {
1855 MDB_ID pn = dl[i].mid << 1;
1857 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1859 /* Can't spill twice, make sure it's not already in a parent's
1862 if (txn->mt_parent) {
1864 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1865 if (tx2->mt_spill_pgs) {
1866 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1867 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1868 dp->mp_flags |= P_KEEP;
1876 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1880 mdb_midl_sort(txn->mt_spill_pgs);
1882 /* Flush the spilled part of dirty list */
1883 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1886 /* Reset any dirty pages we kept that page_flush didn't see */
1887 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1890 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1894 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1896 mdb_find_oldest(MDB_txn *txn)
1899 txnid_t mr, oldest = txn->mt_txnid - 1;
1900 if (txn->mt_env->me_txns) {
1901 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1902 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1913 /** Add a page to the txn's dirty list */
1915 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1918 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1920 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1921 insert = mdb_mid2l_append;
1923 insert = mdb_mid2l_insert;
1925 mid.mid = mp->mp_pgno;
1927 rc = insert(txn->mt_u.dirty_list, &mid);
1928 mdb_tassert(txn, rc == 0);
1929 txn->mt_dirty_room--;
1932 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1933 * me_pghead and mt_next_pgno.
1935 * If there are free pages available from older transactions, they
1936 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1937 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1938 * and move me_pglast to say which records were consumed. Only this
1939 * function can create me_pghead and move me_pglast/mt_next_pgno.
1940 * @param[in] mc cursor A cursor handle identifying the transaction and
1941 * database for which we are allocating.
1942 * @param[in] num the number of pages to allocate.
1943 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1944 * will always be satisfied by a single contiguous chunk of memory.
1945 * @return 0 on success, non-zero on failure.
1948 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1950 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1951 /* Get at most <Max_retries> more freeDB records once me_pghead
1952 * has enough pages. If not enough, use new pages from the map.
1953 * If <Paranoid> and mc is updating the freeDB, only get new
1954 * records if me_pghead is empty. Then the freelist cannot play
1955 * catch-up with itself by growing while trying to save it.
1957 enum { Paranoid = 1, Max_retries = 500 };
1959 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1961 int rc, retry = num * 60;
1962 MDB_txn *txn = mc->mc_txn;
1963 MDB_env *env = txn->mt_env;
1964 pgno_t pgno, *mop = env->me_pghead;
1965 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1967 txnid_t oldest = 0, last;
1972 /* If there are any loose pages, just use them */
1973 if (num == 1 && txn->mt_loose_pgs) {
1974 np = txn->mt_loose_pgs;
1975 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1976 txn->mt_loose_count--;
1977 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1985 /* If our dirty list is already full, we can't do anything */
1986 if (txn->mt_dirty_room == 0) {
1991 for (op = MDB_FIRST;; op = MDB_NEXT) {
1996 /* Seek a big enough contiguous page range. Prefer
1997 * pages at the tail, just truncating the list.
2003 if (mop[i-n2] == pgno+n2)
2010 if (op == MDB_FIRST) { /* 1st iteration */
2011 /* Prepare to fetch more and coalesce */
2012 last = env->me_pglast;
2013 oldest = env->me_pgoldest;
2014 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2017 key.mv_data = &last; /* will look up last+1 */
2018 key.mv_size = sizeof(last);
2020 if (Paranoid && mc->mc_dbi == FREE_DBI)
2023 if (Paranoid && retry < 0 && mop_len)
2027 /* Do not fetch more if the record will be too recent */
2028 if (oldest <= last) {
2030 oldest = mdb_find_oldest(txn);
2031 env->me_pgoldest = oldest;
2037 rc = mdb_cursor_get(&m2, &key, NULL, op);
2039 if (rc == MDB_NOTFOUND)
2043 last = *(txnid_t*)key.mv_data;
2044 if (oldest <= last) {
2046 oldest = mdb_find_oldest(txn);
2047 env->me_pgoldest = oldest;
2053 np = m2.mc_pg[m2.mc_top];
2054 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2055 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2058 idl = (MDB_ID *) data.mv_data;
2061 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2066 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2068 mop = env->me_pghead;
2070 env->me_pglast = last;
2072 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2073 last, txn->mt_dbs[FREE_DBI].md_root, i));
2075 DPRINTF(("IDL %"Z"u", idl[j]));
2077 /* Merge in descending sorted order */
2078 mdb_midl_xmerge(mop, idl);
2082 /* Use new pages from the map when nothing suitable in the freeDB */
2084 pgno = txn->mt_next_pgno;
2085 if (pgno + num >= env->me_maxpg) {
2086 DPUTS("DB size maxed out");
2092 if (env->me_flags & MDB_WRITEMAP) {
2093 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2095 if (!(np = mdb_page_malloc(txn, num))) {
2101 mop[0] = mop_len -= num;
2102 /* Move any stragglers down */
2103 for (j = i-num; j < mop_len; )
2104 mop[++j] = mop[++i];
2106 txn->mt_next_pgno = pgno + num;
2109 mdb_page_dirty(txn, np);
2115 txn->mt_flags |= MDB_TXN_ERROR;
2119 /** Copy the used portions of a non-overflow page.
2120 * @param[in] dst page to copy into
2121 * @param[in] src page to copy from
2122 * @param[in] psize size of a page
2125 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2127 enum { Align = sizeof(pgno_t) };
2128 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2130 /* If page isn't full, just copy the used portion. Adjust
2131 * alignment so memcpy may copy words instead of bytes.
2133 if ((unused &= -Align) && !IS_LEAF2(src)) {
2134 upper = (upper + PAGEBASE) & -Align;
2135 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2136 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2139 memcpy(dst, src, psize - unused);
2143 /** Pull a page off the txn's spill list, if present.
2144 * If a page being referenced was spilled to disk in this txn, bring
2145 * it back and make it dirty/writable again.
2146 * @param[in] txn the transaction handle.
2147 * @param[in] mp the page being referenced. It must not be dirty.
2148 * @param[out] ret the writable page, if any. ret is unchanged if
2149 * mp wasn't spilled.
2152 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2154 MDB_env *env = txn->mt_env;
2157 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2159 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2160 if (!tx2->mt_spill_pgs)
2162 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2163 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2166 if (txn->mt_dirty_room == 0)
2167 return MDB_TXN_FULL;
2168 if (IS_OVERFLOW(mp))
2172 if (env->me_flags & MDB_WRITEMAP) {
2175 np = mdb_page_malloc(txn, num);
2179 memcpy(np, mp, num * env->me_psize);
2181 mdb_page_copy(np, mp, env->me_psize);
2184 /* If in current txn, this page is no longer spilled.
2185 * If it happens to be the last page, truncate the spill list.
2186 * Otherwise mark it as deleted by setting the LSB.
2188 if (x == txn->mt_spill_pgs[0])
2189 txn->mt_spill_pgs[0]--;
2191 txn->mt_spill_pgs[x] |= 1;
2192 } /* otherwise, if belonging to a parent txn, the
2193 * page remains spilled until child commits
2196 mdb_page_dirty(txn, np);
2197 np->mp_flags |= P_DIRTY;
2205 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2206 * @param[in] mc cursor pointing to the page to be touched
2207 * @return 0 on success, non-zero on failure.
2210 mdb_page_touch(MDB_cursor *mc)
2212 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2213 MDB_txn *txn = mc->mc_txn;
2214 MDB_cursor *m2, *m3;
2218 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2219 if (txn->mt_flags & MDB_TXN_SPILLS) {
2221 rc = mdb_page_unspill(txn, mp, &np);
2227 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2228 (rc = mdb_page_alloc(mc, 1, &np)))
2231 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2232 mp->mp_pgno, pgno));
2233 mdb_cassert(mc, mp->mp_pgno != pgno);
2234 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2235 /* Update the parent page, if any, to point to the new page */
2237 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2238 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2239 SETPGNO(node, pgno);
2241 mc->mc_db->md_root = pgno;
2243 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2244 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2246 /* If txn has a parent, make sure the page is in our
2250 unsigned x = mdb_mid2l_search(dl, pgno);
2251 if (x <= dl[0].mid && dl[x].mid == pgno) {
2252 if (mp != dl[x].mptr) { /* bad cursor? */
2253 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2254 txn->mt_flags |= MDB_TXN_ERROR;
2255 return MDB_CORRUPTED;
2260 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2262 np = mdb_page_malloc(txn, 1);
2267 rc = mdb_mid2l_insert(dl, &mid);
2268 mdb_cassert(mc, rc == 0);
2273 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2275 np->mp_flags |= P_DIRTY;
2278 /* Adjust cursors pointing to mp */
2279 mc->mc_pg[mc->mc_top] = np;
2280 m2 = txn->mt_cursors[mc->mc_dbi];
2281 if (mc->mc_flags & C_SUB) {
2282 for (; m2; m2=m2->mc_next) {
2283 m3 = &m2->mc_xcursor->mx_cursor;
2284 if (m3->mc_snum < mc->mc_snum) continue;
2285 if (m3->mc_pg[mc->mc_top] == mp)
2286 m3->mc_pg[mc->mc_top] = np;
2289 for (; m2; m2=m2->mc_next) {
2290 if (m2->mc_snum < mc->mc_snum) continue;
2291 if (m2->mc_pg[mc->mc_top] == mp) {
2292 m2->mc_pg[mc->mc_top] = np;
2293 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2295 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2297 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2298 if (!(leaf->mn_flags & F_SUBDATA))
2299 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2307 txn->mt_flags |= MDB_TXN_ERROR;
2312 mdb_env_sync(MDB_env *env, int force)
2315 if (env->me_flags & MDB_RDONLY)
2317 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2318 if (env->me_flags & MDB_WRITEMAP) {
2319 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2320 ? MS_ASYNC : MS_SYNC;
2321 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2324 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2328 #ifdef BROKEN_FDATASYNC
2329 if (env->me_flags & MDB_FSYNCONLY) {
2330 if (fsync(env->me_fd))
2334 if (MDB_FDATASYNC(env->me_fd))
2341 /** Back up parent txn's cursors, then grab the originals for tracking */
2343 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2345 MDB_cursor *mc, *bk;
2350 for (i = src->mt_numdbs; --i >= 0; ) {
2351 if ((mc = src->mt_cursors[i]) != NULL) {
2352 size = sizeof(MDB_cursor);
2354 size += sizeof(MDB_xcursor);
2355 for (; mc; mc = bk->mc_next) {
2361 mc->mc_db = &dst->mt_dbs[i];
2362 /* Kill pointers into src - and dst to reduce abuse: The
2363 * user may not use mc until dst ends. Otherwise we'd...
2365 mc->mc_txn = NULL; /* ...set this to dst */
2366 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2367 if ((mx = mc->mc_xcursor) != NULL) {
2368 *(MDB_xcursor *)(bk+1) = *mx;
2369 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2371 mc->mc_next = dst->mt_cursors[i];
2372 dst->mt_cursors[i] = mc;
2379 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2380 * @param[in] txn the transaction handle.
2381 * @param[in] merge true to keep changes to parent cursors, false to revert.
2382 * @return 0 on success, non-zero on failure.
2385 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2387 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2391 for (i = txn->mt_numdbs; --i >= 0; ) {
2392 for (mc = cursors[i]; mc; mc = next) {
2394 if ((bk = mc->mc_backup) != NULL) {
2396 /* Commit changes to parent txn */
2397 mc->mc_next = bk->mc_next;
2398 mc->mc_backup = bk->mc_backup;
2399 mc->mc_txn = bk->mc_txn;
2400 mc->mc_db = bk->mc_db;
2401 mc->mc_dbflag = bk->mc_dbflag;
2402 if ((mx = mc->mc_xcursor) != NULL)
2403 mx->mx_cursor.mc_txn = bk->mc_txn;
2405 /* Abort nested txn */
2407 if ((mx = mc->mc_xcursor) != NULL)
2408 *mx = *(MDB_xcursor *)(bk+1);
2412 /* Only malloced cursors are permanently tracked. */
2420 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2423 mdb_txn_reset0(MDB_txn *txn, const char *act);
2425 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2431 Pidset = F_SETLK, Pidcheck = F_GETLK
2435 /** Set or check a pid lock. Set returns 0 on success.
2436 * Check returns 0 if the process is certainly dead, nonzero if it may
2437 * be alive (the lock exists or an error happened so we do not know).
2439 * On Windows Pidset is a no-op, we merely check for the existence
2440 * of the process with the given pid. On POSIX we use a single byte
2441 * lock on the lockfile, set at an offset equal to the pid.
2444 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2446 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2449 if (op == Pidcheck) {
2450 h = OpenProcess(env->me_pidquery, FALSE, pid);
2451 /* No documented "no such process" code, but other program use this: */
2453 return ErrCode() != ERROR_INVALID_PARAMETER;
2454 /* A process exists until all handles to it close. Has it exited? */
2455 ret = WaitForSingleObject(h, 0) != 0;
2462 struct flock lock_info;
2463 memset(&lock_info, 0, sizeof(lock_info));
2464 lock_info.l_type = F_WRLCK;
2465 lock_info.l_whence = SEEK_SET;
2466 lock_info.l_start = pid;
2467 lock_info.l_len = 1;
2468 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2469 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2471 } else if ((rc = ErrCode()) == EINTR) {
2479 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2480 * @param[in] txn the transaction handle to initialize
2481 * @return 0 on success, non-zero on failure.
2484 mdb_txn_renew0(MDB_txn *txn)
2486 MDB_env *env = txn->mt_env;
2487 MDB_txninfo *ti = env->me_txns;
2491 int rc, new_notls = 0;
2493 if (txn->mt_flags & MDB_TXN_RDONLY) {
2495 txn->mt_numdbs = env->me_numdbs;
2496 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2498 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2499 txn->mt_txnid = meta->mm_txnid;
2500 txn->mt_u.reader = NULL;
2502 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2503 pthread_getspecific(env->me_txkey);
2505 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2506 return MDB_BAD_RSLOT;
2508 MDB_PID_T pid = env->me_pid;
2509 MDB_THR_T tid = pthread_self();
2511 if (!env->me_live_reader) {
2512 rc = mdb_reader_pid(env, Pidset, pid);
2515 env->me_live_reader = 1;
2519 nr = ti->mti_numreaders;
2520 for (i=0; i<nr; i++)
2521 if (ti->mti_readers[i].mr_pid == 0)
2523 if (i == env->me_maxreaders) {
2524 UNLOCK_MUTEX_R(env);
2525 return MDB_READERS_FULL;
2527 ti->mti_readers[i].mr_pid = pid;
2528 ti->mti_readers[i].mr_tid = tid;
2530 ti->mti_numreaders = ++nr;
2531 /* Save numreaders for un-mutexed mdb_env_close() */
2532 env->me_numreaders = nr;
2533 UNLOCK_MUTEX_R(env);
2535 r = &ti->mti_readers[i];
2536 new_notls = (env->me_flags & MDB_NOTLS);
2537 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2542 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2543 txn->mt_u.reader = r;
2544 meta = env->me_metas[txn->mt_txnid & 1];
2550 txn->mt_txnid = ti->mti_txnid;
2551 meta = env->me_metas[txn->mt_txnid & 1];
2553 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2554 txn->mt_txnid = meta->mm_txnid;
2557 txn->mt_numdbs = env->me_numdbs;
2560 if (txn->mt_txnid == mdb_debug_start)
2564 txn->mt_child = NULL;
2565 txn->mt_loose_pgs = NULL;
2566 txn->mt_loose_count = 0;
2567 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2568 txn->mt_u.dirty_list = env->me_dirty_list;
2569 txn->mt_u.dirty_list[0].mid = 0;
2570 txn->mt_free_pgs = env->me_free_pgs;
2571 txn->mt_free_pgs[0] = 0;
2572 txn->mt_spill_pgs = NULL;
2574 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2577 /* Copy the DB info and flags */
2578 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2580 /* Moved to here to avoid a data race in read TXNs */
2581 txn->mt_next_pgno = meta->mm_last_pg+1;
2583 for (i=2; i<txn->mt_numdbs; i++) {
2584 x = env->me_dbflags[i];
2585 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2586 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2588 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2590 if (env->me_maxpg < txn->mt_next_pgno) {
2591 mdb_txn_reset0(txn, "renew0-mapfail");
2593 txn->mt_u.reader->mr_pid = 0;
2594 txn->mt_u.reader = NULL;
2596 return MDB_MAP_RESIZED;
2603 mdb_txn_renew(MDB_txn *txn)
2607 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2610 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2611 DPUTS("environment had fatal error, must shutdown!");
2615 rc = mdb_txn_renew0(txn);
2616 if (rc == MDB_SUCCESS) {
2617 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2618 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2619 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2625 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2629 int rc, size, tsize = sizeof(MDB_txn);
2631 if (env->me_flags & MDB_FATAL_ERROR) {
2632 DPUTS("environment had fatal error, must shutdown!");
2635 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2638 /* Nested transactions: Max 1 child, write txns only, no writemap */
2639 if (parent->mt_child ||
2640 (flags & MDB_RDONLY) ||
2641 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2642 (env->me_flags & MDB_WRITEMAP))
2644 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2646 tsize = sizeof(MDB_ntxn);
2649 if (!(flags & MDB_RDONLY)) {
2651 txn = env->me_txn0; /* just reuse preallocated write txn */
2654 /* child txns use own copy of cursors */
2655 size += env->me_maxdbs * sizeof(MDB_cursor *);
2657 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2659 if ((txn = calloc(1, size)) == NULL) {
2660 DPRINTF(("calloc: %s", strerror(errno)));
2663 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2664 if (flags & MDB_RDONLY) {
2665 txn->mt_flags |= MDB_TXN_RDONLY;
2666 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2667 txn->mt_dbiseqs = env->me_dbiseqs;
2669 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2671 txn->mt_dbiseqs = parent->mt_dbiseqs;
2672 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2674 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2675 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2683 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2684 if (!txn->mt_u.dirty_list ||
2685 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2687 free(txn->mt_u.dirty_list);
2691 txn->mt_txnid = parent->mt_txnid;
2692 txn->mt_dirty_room = parent->mt_dirty_room;
2693 txn->mt_u.dirty_list[0].mid = 0;
2694 txn->mt_spill_pgs = NULL;
2695 txn->mt_next_pgno = parent->mt_next_pgno;
2696 parent->mt_child = txn;
2697 txn->mt_parent = parent;
2698 txn->mt_numdbs = parent->mt_numdbs;
2699 txn->mt_flags = parent->mt_flags;
2700 txn->mt_dbxs = parent->mt_dbxs;
2701 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2702 /* Copy parent's mt_dbflags, but clear DB_NEW */
2703 for (i=0; i<txn->mt_numdbs; i++)
2704 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2706 ntxn = (MDB_ntxn *)txn;
2707 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2708 if (env->me_pghead) {
2709 size = MDB_IDL_SIZEOF(env->me_pghead);
2710 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2712 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2717 rc = mdb_cursor_shadow(parent, txn);
2719 mdb_txn_reset0(txn, "beginchild-fail");
2721 rc = mdb_txn_renew0(txn);
2724 if (txn != env->me_txn0)
2728 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2729 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2730 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2737 mdb_txn_env(MDB_txn *txn)
2739 if(!txn) return NULL;
2743 /** Export or close DBI handles opened in this txn. */
2745 mdb_dbis_update(MDB_txn *txn, int keep)
2748 MDB_dbi n = txn->mt_numdbs;
2749 MDB_env *env = txn->mt_env;
2750 unsigned char *tdbflags = txn->mt_dbflags;
2752 for (i = n; --i >= 2;) {
2753 if (tdbflags[i] & DB_NEW) {
2755 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2757 char *ptr = env->me_dbxs[i].md_name.mv_data;
2759 env->me_dbxs[i].md_name.mv_data = NULL;
2760 env->me_dbxs[i].md_name.mv_size = 0;
2761 env->me_dbflags[i] = 0;
2762 env->me_dbiseqs[i]++;
2768 if (keep && env->me_numdbs < n)
2772 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2773 * May be called twice for readonly txns: First reset it, then abort.
2774 * @param[in] txn the transaction handle to reset
2775 * @param[in] act why the transaction is being reset
2778 mdb_txn_reset0(MDB_txn *txn, const char *act)
2780 MDB_env *env = txn->mt_env;
2782 /* Close any DBI handles opened in this txn */
2783 mdb_dbis_update(txn, 0);
2785 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2786 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2787 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2789 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2790 if (txn->mt_u.reader) {
2791 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2792 if (!(env->me_flags & MDB_NOTLS))
2793 txn->mt_u.reader = NULL; /* txn does not own reader */
2795 txn->mt_numdbs = 0; /* close nothing if called again */
2796 txn->mt_dbxs = NULL; /* mark txn as reset */
2798 pgno_t *pghead = env->me_pghead;
2800 mdb_cursors_close(txn, 0);
2801 if (!(env->me_flags & MDB_WRITEMAP)) {
2802 mdb_dlist_free(txn);
2805 if (!txn->mt_parent) {
2806 if (mdb_midl_shrink(&txn->mt_free_pgs))
2807 env->me_free_pgs = txn->mt_free_pgs;
2809 env->me_pghead = NULL;
2813 /* The writer mutex was locked in mdb_txn_begin. */
2815 UNLOCK_MUTEX_W(env);
2817 txn->mt_parent->mt_child = NULL;
2818 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2819 mdb_midl_free(txn->mt_free_pgs);
2820 mdb_midl_free(txn->mt_spill_pgs);
2821 free(txn->mt_u.dirty_list);
2824 mdb_midl_free(pghead);
2829 mdb_txn_reset(MDB_txn *txn)
2834 /* This call is only valid for read-only txns */
2835 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2838 mdb_txn_reset0(txn, "reset");
2842 mdb_txn_abort(MDB_txn *txn)
2848 mdb_txn_abort(txn->mt_child);
2850 mdb_txn_reset0(txn, "abort");
2851 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2852 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2853 txn->mt_u.reader->mr_pid = 0;
2855 if (txn != txn->mt_env->me_txn0)
2859 /** Save the freelist as of this transaction to the freeDB.
2860 * This changes the freelist. Keep trying until it stabilizes.
2863 mdb_freelist_save(MDB_txn *txn)
2865 /* env->me_pghead[] can grow and shrink during this call.
2866 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2867 * Page numbers cannot disappear from txn->mt_free_pgs[].
2870 MDB_env *env = txn->mt_env;
2871 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2872 txnid_t pglast = 0, head_id = 0;
2873 pgno_t freecnt = 0, *free_pgs, *mop;
2874 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2876 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2878 if (env->me_pghead) {
2879 /* Make sure first page of freeDB is touched and on freelist */
2880 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2881 if (rc && rc != MDB_NOTFOUND)
2885 if (!env->me_pghead && txn->mt_loose_pgs) {
2886 /* Put loose page numbers in mt_free_pgs, since
2887 * we may be unable to return them to me_pghead.
2889 MDB_page *mp = txn->mt_loose_pgs;
2890 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2892 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2893 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2894 txn->mt_loose_pgs = NULL;
2895 txn->mt_loose_count = 0;
2898 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2899 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2900 ? SSIZE_MAX : maxfree_1pg;
2903 /* Come back here after each Put() in case freelist changed */
2908 /* If using records from freeDB which we have not yet
2909 * deleted, delete them and any we reserved for me_pghead.
2911 while (pglast < env->me_pglast) {
2912 rc = mdb_cursor_first(&mc, &key, NULL);
2915 pglast = head_id = *(txnid_t *)key.mv_data;
2916 total_room = head_room = 0;
2917 mdb_tassert(txn, pglast <= env->me_pglast);
2918 rc = mdb_cursor_del(&mc, 0);
2923 /* Save the IDL of pages freed by this txn, to a single record */
2924 if (freecnt < txn->mt_free_pgs[0]) {
2926 /* Make sure last page of freeDB is touched and on freelist */
2927 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2928 if (rc && rc != MDB_NOTFOUND)
2931 free_pgs = txn->mt_free_pgs;
2932 /* Write to last page of freeDB */
2933 key.mv_size = sizeof(txn->mt_txnid);
2934 key.mv_data = &txn->mt_txnid;
2936 freecnt = free_pgs[0];
2937 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2938 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2941 /* Retry if mt_free_pgs[] grew during the Put() */
2942 free_pgs = txn->mt_free_pgs;
2943 } while (freecnt < free_pgs[0]);
2944 mdb_midl_sort(free_pgs);
2945 memcpy(data.mv_data, free_pgs, data.mv_size);
2948 unsigned int i = free_pgs[0];
2949 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2950 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2952 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2958 mop = env->me_pghead;
2959 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2961 /* Reserve records for me_pghead[]. Split it if multi-page,
2962 * to avoid searching freeDB for a page range. Use keys in
2963 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2965 if (total_room >= mop_len) {
2966 if (total_room == mop_len || --more < 0)
2968 } else if (head_room >= maxfree_1pg && head_id > 1) {
2969 /* Keep current record (overflow page), add a new one */
2973 /* (Re)write {key = head_id, IDL length = head_room} */
2974 total_room -= head_room;
2975 head_room = mop_len - total_room;
2976 if (head_room > maxfree_1pg && head_id > 1) {
2977 /* Overflow multi-page for part of me_pghead */
2978 head_room /= head_id; /* amortize page sizes */
2979 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2980 } else if (head_room < 0) {
2981 /* Rare case, not bothering to delete this record */
2984 key.mv_size = sizeof(head_id);
2985 key.mv_data = &head_id;
2986 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2987 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2990 /* IDL is initially empty, zero out at least the length */
2991 pgs = (pgno_t *)data.mv_data;
2992 j = head_room > clean_limit ? head_room : 0;
2996 total_room += head_room;
2999 /* Return loose page numbers to me_pghead, though usually none are
3000 * left at this point. The pages themselves remain in dirty_list.
3002 if (txn->mt_loose_pgs) {
3003 MDB_page *mp = txn->mt_loose_pgs;
3004 unsigned count = txn->mt_loose_count;
3006 /* Room for loose pages + temp IDL with same */
3007 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3009 mop = env->me_pghead;
3010 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3011 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3012 loose[ ++count ] = mp->mp_pgno;
3014 mdb_midl_sort(loose);
3015 mdb_midl_xmerge(mop, loose);
3016 txn->mt_loose_pgs = NULL;
3017 txn->mt_loose_count = 0;
3021 /* Fill in the reserved me_pghead records */
3027 rc = mdb_cursor_first(&mc, &key, &data);
3028 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3029 txnid_t id = *(txnid_t *)key.mv_data;
3030 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3033 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3035 if (len > mop_len) {
3037 data.mv_size = (len + 1) * sizeof(MDB_ID);
3039 data.mv_data = mop -= len;
3042 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3044 if (rc || !(mop_len -= len))
3051 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3052 * @param[in] txn the transaction that's being committed
3053 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3054 * @return 0 on success, non-zero on failure.
3057 mdb_page_flush(MDB_txn *txn, int keep)
3059 MDB_env *env = txn->mt_env;
3060 MDB_ID2L dl = txn->mt_u.dirty_list;
3061 unsigned psize = env->me_psize, j;
3062 int i, pagecount = dl[0].mid, rc;
3063 size_t size = 0, pos = 0;
3065 MDB_page *dp = NULL;
3069 struct iovec iov[MDB_COMMIT_PAGES];
3070 ssize_t wpos = 0, wsize = 0, wres;
3071 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3077 if (env->me_flags & MDB_WRITEMAP) {
3078 /* Clear dirty flags */
3079 while (++i <= pagecount) {
3081 /* Don't flush this page yet */
3082 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3083 dp->mp_flags &= ~P_KEEP;
3087 dp->mp_flags &= ~P_DIRTY;
3092 /* Write the pages */
3094 if (++i <= pagecount) {
3096 /* Don't flush this page yet */
3097 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3098 dp->mp_flags &= ~P_KEEP;
3103 /* clear dirty flag */
3104 dp->mp_flags &= ~P_DIRTY;
3107 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3112 /* Windows actually supports scatter/gather I/O, but only on
3113 * unbuffered file handles. Since we're relying on the OS page
3114 * cache for all our data, that's self-defeating. So we just
3115 * write pages one at a time. We use the ov structure to set
3116 * the write offset, to at least save the overhead of a Seek
3119 DPRINTF(("committing page %"Z"u", pgno));
3120 memset(&ov, 0, sizeof(ov));
3121 ov.Offset = pos & 0xffffffff;
3122 ov.OffsetHigh = pos >> 16 >> 16;
3123 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3125 DPRINTF(("WriteFile: %d", rc));
3129 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3130 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3132 /* Write previous page(s) */
3133 #ifdef MDB_USE_PWRITEV
3134 wres = pwritev(env->me_fd, iov, n, wpos);
3137 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3139 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3141 DPRINTF(("lseek: %s", strerror(rc)));
3144 wres = writev(env->me_fd, iov, n);
3147 if (wres != wsize) {
3150 DPRINTF(("Write error: %s", strerror(rc)));
3152 rc = EIO; /* TODO: Use which error code? */
3153 DPUTS("short write, filesystem full?");
3164 DPRINTF(("committing page %"Z"u", pgno));
3165 next_pos = pos + size;
3166 iov[n].iov_len = size;
3167 iov[n].iov_base = (char *)dp;
3173 /* MIPS has cache coherency issues, this is a no-op everywhere else
3174 * Note: for any size >= on-chip cache size, entire on-chip cache is
3177 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3179 for (i = keep; ++i <= pagecount; ) {
3181 /* This is a page we skipped above */
3184 dl[j].mid = dp->mp_pgno;
3187 mdb_dpage_free(env, dp);
3192 txn->mt_dirty_room += i - j;
3198 mdb_txn_commit(MDB_txn *txn)
3204 if (txn == NULL || txn->mt_env == NULL)
3207 if (txn->mt_child) {
3208 rc = mdb_txn_commit(txn->mt_child);
3209 txn->mt_child = NULL;
3216 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3217 mdb_dbis_update(txn, 1);
3218 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3223 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3224 DPUTS("error flag is set, can't commit");
3226 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3231 if (txn->mt_parent) {
3232 MDB_txn *parent = txn->mt_parent;
3236 unsigned x, y, len, ps_len;
3238 /* Append our free list to parent's */
3239 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3242 mdb_midl_free(txn->mt_free_pgs);
3243 /* Failures after this must either undo the changes
3244 * to the parent or set MDB_TXN_ERROR in the parent.
3247 parent->mt_next_pgno = txn->mt_next_pgno;
3248 parent->mt_flags = txn->mt_flags;
3250 /* Merge our cursors into parent's and close them */
3251 mdb_cursors_close(txn, 1);
3253 /* Update parent's DB table. */
3254 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3255 parent->mt_numdbs = txn->mt_numdbs;
3256 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3257 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3258 for (i=2; i<txn->mt_numdbs; i++) {
3259 /* preserve parent's DB_NEW status */
3260 x = parent->mt_dbflags[i] & DB_NEW;
3261 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3264 dst = parent->mt_u.dirty_list;
3265 src = txn->mt_u.dirty_list;
3266 /* Remove anything in our dirty list from parent's spill list */
3267 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3269 pspill[0] = (pgno_t)-1;
3270 /* Mark our dirty pages as deleted in parent spill list */
3271 for (i=0, len=src[0].mid; ++i <= len; ) {
3272 MDB_ID pn = src[i].mid << 1;
3273 while (pn > pspill[x])
3275 if (pn == pspill[x]) {
3280 /* Squash deleted pagenums if we deleted any */
3281 for (x=y; ++x <= ps_len; )
3282 if (!(pspill[x] & 1))
3283 pspill[++y] = pspill[x];
3287 /* Find len = length of merging our dirty list with parent's */
3289 dst[0].mid = 0; /* simplify loops */
3290 if (parent->mt_parent) {
3291 len = x + src[0].mid;
3292 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3293 for (i = x; y && i; y--) {
3294 pgno_t yp = src[y].mid;
3295 while (yp < dst[i].mid)
3297 if (yp == dst[i].mid) {
3302 } else { /* Simplify the above for single-ancestor case */
3303 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3305 /* Merge our dirty list with parent's */
3307 for (i = len; y; dst[i--] = src[y--]) {
3308 pgno_t yp = src[y].mid;
3309 while (yp < dst[x].mid)
3310 dst[i--] = dst[x--];
3311 if (yp == dst[x].mid)
3312 free(dst[x--].mptr);
3314 mdb_tassert(txn, i == x);
3316 free(txn->mt_u.dirty_list);
3317 parent->mt_dirty_room = txn->mt_dirty_room;
3318 if (txn->mt_spill_pgs) {
3319 if (parent->mt_spill_pgs) {
3320 /* TODO: Prevent failure here, so parent does not fail */
3321 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3323 parent->mt_flags |= MDB_TXN_ERROR;
3324 mdb_midl_free(txn->mt_spill_pgs);
3325 mdb_midl_sort(parent->mt_spill_pgs);
3327 parent->mt_spill_pgs = txn->mt_spill_pgs;
3331 /* Append our loose page list to parent's */
3332 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3334 *lp = txn->mt_loose_pgs;
3335 parent->mt_loose_count += txn->mt_loose_count;
3337 parent->mt_child = NULL;
3338 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3343 if (txn != env->me_txn) {
3344 DPUTS("attempt to commit unknown transaction");
3349 mdb_cursors_close(txn, 0);
3351 if (!txn->mt_u.dirty_list[0].mid &&
3352 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3355 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3356 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3358 /* Update DB root pointers */
3359 if (txn->mt_numdbs > 2) {
3363 data.mv_size = sizeof(MDB_db);
3365 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3366 for (i = 2; i < txn->mt_numdbs; i++) {
3367 if (txn->mt_dbflags[i] & DB_DIRTY) {
3368 if (TXN_DBI_CHANGED(txn, i)) {
3372 data.mv_data = &txn->mt_dbs[i];
3373 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3380 rc = mdb_freelist_save(txn);
3384 mdb_midl_free(env->me_pghead);
3385 env->me_pghead = NULL;
3386 if (mdb_midl_shrink(&txn->mt_free_pgs))
3387 env->me_free_pgs = txn->mt_free_pgs;
3393 if ((rc = mdb_page_flush(txn, 0)) ||
3394 (rc = mdb_env_sync(env, 0)) ||
3395 (rc = mdb_env_write_meta(txn)))
3398 /* Free P_LOOSE pages left behind in dirty_list */
3399 if (!(env->me_flags & MDB_WRITEMAP))
3400 mdb_dlist_free(txn);
3405 mdb_dbis_update(txn, 1);
3408 UNLOCK_MUTEX_W(env);
3409 if (txn != env->me_txn0)
3419 /** Read the environment parameters of a DB environment before
3420 * mapping it into memory.
3421 * @param[in] env the environment handle
3422 * @param[out] meta address of where to store the meta information
3423 * @return 0 on success, non-zero on failure.
3426 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3432 enum { Size = sizeof(pbuf) };
3434 /* We don't know the page size yet, so use a minimum value.
3435 * Read both meta pages so we can use the latest one.
3438 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3442 memset(&ov, 0, sizeof(ov));
3444 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3445 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3448 rc = pread(env->me_fd, &pbuf, Size, off);
3451 if (rc == 0 && off == 0)
3453 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3454 DPRINTF(("read: %s", mdb_strerror(rc)));
3458 p = (MDB_page *)&pbuf;
3460 if (!F_ISSET(p->mp_flags, P_META)) {
3461 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3466 if (m->mm_magic != MDB_MAGIC) {
3467 DPUTS("meta has invalid magic");
3471 if (m->mm_version != MDB_DATA_VERSION) {
3472 DPRINTF(("database is version %u, expected version %u",
3473 m->mm_version, MDB_DATA_VERSION));
3474 return MDB_VERSION_MISMATCH;
3477 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3484 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3486 meta->mm_magic = MDB_MAGIC;
3487 meta->mm_version = MDB_DATA_VERSION;
3488 meta->mm_mapsize = env->me_mapsize;
3489 meta->mm_psize = env->me_psize;
3490 meta->mm_last_pg = 1;
3491 meta->mm_flags = env->me_flags & 0xffff;
3492 meta->mm_flags |= MDB_INTEGERKEY;
3493 meta->mm_dbs[0].md_root = P_INVALID;
3494 meta->mm_dbs[1].md_root = P_INVALID;
3497 /** Write the environment parameters of a freshly created DB environment.
3498 * @param[in] env the environment handle
3499 * @param[out] meta address of where to store the meta information
3500 * @return 0 on success, non-zero on failure.
3503 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3511 memset(&ov, 0, sizeof(ov));
3512 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3514 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3517 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3518 len = pwrite(fd, ptr, size, pos); \
3519 rc = (len >= 0); } while(0)
3522 DPUTS("writing new meta page");
3524 psize = env->me_psize;
3526 mdb_env_init_meta0(env, meta);
3528 p = calloc(2, psize);
3530 p->mp_flags = P_META;
3531 *(MDB_meta *)METADATA(p) = *meta;
3533 q = (MDB_page *)((char *)p + psize);
3535 q->mp_flags = P_META;
3536 *(MDB_meta *)METADATA(q) = *meta;
3538 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3541 else if ((unsigned) len == psize * 2)
3549 /** Update the environment info to commit a transaction.
3550 * @param[in] txn the transaction that's being committed
3551 * @return 0 on success, non-zero on failure.
3554 mdb_env_write_meta(MDB_txn *txn)
3557 MDB_meta meta, metab, *mp;
3560 int rc, len, toggle;
3569 toggle = txn->mt_txnid & 1;
3570 DPRINTF(("writing meta page %d for root page %"Z"u",
3571 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3574 mp = env->me_metas[toggle];
3575 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3576 /* Persist any increases of mapsize config */
3577 if (mapsize < env->me_mapsize)
3578 mapsize = env->me_mapsize;
3580 if (env->me_flags & MDB_WRITEMAP) {
3581 mp->mm_mapsize = mapsize;
3582 mp->mm_dbs[0] = txn->mt_dbs[0];
3583 mp->mm_dbs[1] = txn->mt_dbs[1];
3584 mp->mm_last_pg = txn->mt_next_pgno - 1;
3585 mp->mm_txnid = txn->mt_txnid;
3586 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3587 unsigned meta_size = env->me_psize;
3588 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3591 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3592 if (meta_size < env->me_os_psize)
3593 meta_size += meta_size;
3598 if (MDB_MSYNC(ptr, meta_size, rc)) {
3605 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3606 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3608 meta.mm_mapsize = mapsize;
3609 meta.mm_dbs[0] = txn->mt_dbs[0];
3610 meta.mm_dbs[1] = txn->mt_dbs[1];
3611 meta.mm_last_pg = txn->mt_next_pgno - 1;
3612 meta.mm_txnid = txn->mt_txnid;
3614 off = offsetof(MDB_meta, mm_mapsize);
3615 ptr = (char *)&meta + off;
3616 len = sizeof(MDB_meta) - off;
3618 off += env->me_psize;
3621 /* Write to the SYNC fd */
3622 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3623 env->me_fd : env->me_mfd;
3626 memset(&ov, 0, sizeof(ov));
3628 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3632 rc = pwrite(mfd, ptr, len, off);
3635 rc = rc < 0 ? ErrCode() : EIO;
3636 DPUTS("write failed, disk error?");
3637 /* On a failure, the pagecache still contains the new data.
3638 * Write some old data back, to prevent it from being used.
3639 * Use the non-SYNC fd; we know it will fail anyway.
3641 meta.mm_last_pg = metab.mm_last_pg;
3642 meta.mm_txnid = metab.mm_txnid;
3644 memset(&ov, 0, sizeof(ov));
3646 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3648 r2 = pwrite(env->me_fd, ptr, len, off);
3649 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3652 env->me_flags |= MDB_FATAL_ERROR;
3655 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3656 CACHEFLUSH(env->me_map + off, len, DCACHE);
3658 /* Memory ordering issues are irrelevant; since the entire writer
3659 * is wrapped by wmutex, all of these changes will become visible
3660 * after the wmutex is unlocked. Since the DB is multi-version,
3661 * readers will get consistent data regardless of how fresh or
3662 * how stale their view of these values is.
3665 env->me_txns->mti_txnid = txn->mt_txnid;
3670 /** Check both meta pages to see which one is newer.
3671 * @param[in] env the environment handle
3672 * @return meta toggle (0 or 1).
3675 mdb_env_pick_meta(const MDB_env *env)
3677 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3681 mdb_env_create(MDB_env **env)
3685 e = calloc(1, sizeof(MDB_env));
3689 e->me_maxreaders = DEFAULT_READERS;
3690 e->me_maxdbs = e->me_numdbs = 2;
3691 e->me_fd = INVALID_HANDLE_VALUE;
3692 e->me_lfd = INVALID_HANDLE_VALUE;
3693 e->me_mfd = INVALID_HANDLE_VALUE;
3694 #ifdef MDB_USE_POSIX_SEM
3695 e->me_rmutex = SEM_FAILED;
3696 e->me_wmutex = SEM_FAILED;
3698 e->me_pid = getpid();
3699 GET_PAGESIZE(e->me_os_psize);
3700 VGMEMP_CREATE(e,0,0);
3706 mdb_env_map(MDB_env *env, void *addr)
3709 unsigned int flags = env->me_flags;
3713 LONG sizelo, sizehi;
3716 if (flags & MDB_RDONLY) {
3717 /* Don't set explicit map size, use whatever exists */
3722 msize = env->me_mapsize;
3723 sizelo = msize & 0xffffffff;
3724 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3726 /* Windows won't create mappings for zero length files.
3727 * and won't map more than the file size.
3728 * Just set the maxsize right now.
3730 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3731 || !SetEndOfFile(env->me_fd)
3732 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3736 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3737 PAGE_READWRITE : PAGE_READONLY,
3738 sizehi, sizelo, NULL);
3741 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3742 FILE_MAP_WRITE : FILE_MAP_READ,
3744 rc = env->me_map ? 0 : ErrCode();
3749 int prot = PROT_READ;
3750 if (flags & MDB_WRITEMAP) {
3752 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3755 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3757 if (env->me_map == MAP_FAILED) {
3762 if (flags & MDB_NORDAHEAD) {
3763 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3765 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3767 #ifdef POSIX_MADV_RANDOM
3768 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3769 #endif /* POSIX_MADV_RANDOM */
3770 #endif /* MADV_RANDOM */
3774 /* Can happen because the address argument to mmap() is just a
3775 * hint. mmap() can pick another, e.g. if the range is in use.
3776 * The MAP_FIXED flag would prevent that, but then mmap could
3777 * instead unmap existing pages to make room for the new map.
3779 if (addr && env->me_map != addr)
3780 return EBUSY; /* TODO: Make a new MDB_* error code? */
3782 p = (MDB_page *)env->me_map;
3783 env->me_metas[0] = METADATA(p);
3784 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3790 mdb_env_set_mapsize(MDB_env *env, size_t size)
3792 /* If env is already open, caller is responsible for making
3793 * sure there are no active txns.
3801 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3802 else if (size < env->me_mapsize) {
3803 /* If the configured size is smaller, make sure it's
3804 * still big enough. Silently round up to minimum if not.
3806 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3810 munmap(env->me_map, env->me_mapsize);
3811 env->me_mapsize = size;
3812 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3813 rc = mdb_env_map(env, old);
3817 env->me_mapsize = size;
3819 env->me_maxpg = env->me_mapsize / env->me_psize;
3824 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3828 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3833 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3835 if (env->me_map || readers < 1)
3837 env->me_maxreaders = readers;
3842 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3844 if (!env || !readers)
3846 *readers = env->me_maxreaders;
3851 mdb_fsize(HANDLE fd, size_t *size)
3854 LARGE_INTEGER fsize;
3856 if (!GetFileSizeEx(fd, &fsize))
3859 *size = fsize.QuadPart;
3871 #ifdef BROKEN_FDATASYNC
3872 #include <sys/utsname.h>
3873 #include <sys/vfs.h>
3876 /** Further setup required for opening an LMDB environment
3879 mdb_env_open2(MDB_env *env)
3881 unsigned int flags = env->me_flags;
3882 int i, newenv = 0, rc;
3886 /* See if we should use QueryLimited */
3888 if ((rc & 0xff) > 5)
3889 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3891 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3893 #ifdef BROKEN_FDATASYNC
3894 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3895 * https://lkml.org/lkml/2012/9/3/83
3896 * Kernels after 3.6-rc6 are known good.
3897 * https://lkml.org/lkml/2012/9/10/556
3898 * See if the DB is on ext3/ext4, then check for new enough kernel
3899 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
3904 fstatfs(env->me_fd, &st);
3905 while (st.f_type == 0xEF53) {
3909 if (uts.release[0] < '3') {
3910 if (!strncmp(uts.release, "2.6.32.", 7)) {
3911 i = atoi(uts.release+7);
3913 break; /* 2.6.32.60 and newer is OK */
3914 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
3915 i = atoi(uts.release+7);
3917 break; /* 2.6.34.15 and newer is OK */
3919 } else if (uts.release[0] == '3') {
3920 i = atoi(uts.release+2);
3922 break; /* 3.6 and newer is OK */
3924 i = atoi(uts.release+4);
3926 break; /* 3.5.4 and newer is OK */
3927 } else if (i == 2) {
3928 i = atoi(uts.release+4);
3930 break; /* 3.2.30 and newer is OK */
3932 } else { /* 4.x and newer is OK */
3935 env->me_flags |= MDB_FSYNCONLY;
3941 memset(&meta, 0, sizeof(meta));
3943 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3946 DPUTS("new mdbenv");
3948 env->me_psize = env->me_os_psize;
3949 if (env->me_psize > MAX_PAGESIZE)
3950 env->me_psize = MAX_PAGESIZE;
3952 env->me_psize = meta.mm_psize;
3955 /* Was a mapsize configured? */
3956 if (!env->me_mapsize) {
3957 /* If this is a new environment, take the default,
3958 * else use the size recorded in the existing env.
3960 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3961 } else if (env->me_mapsize < meta.mm_mapsize) {
3962 /* If the configured size is smaller, make sure it's
3963 * still big enough. Silently round up to minimum if not.
3965 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3966 if (env->me_mapsize < minsize)
3967 env->me_mapsize = minsize;
3970 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3975 if (flags & MDB_FIXEDMAP)
3976 meta.mm_address = env->me_map;
3977 i = mdb_env_init_meta(env, &meta);
3978 if (i != MDB_SUCCESS) {
3983 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3984 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3986 #if !(MDB_MAXKEYSIZE)
3987 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3989 env->me_maxpg = env->me_mapsize / env->me_psize;
3993 int toggle = mdb_env_pick_meta(env);
3994 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3996 DPRINTF(("opened database version %u, pagesize %u",
3997 env->me_metas[0]->mm_version, env->me_psize));
3998 DPRINTF(("using meta page %d", toggle));
3999 DPRINTF(("depth: %u", db->md_depth));
4000 DPRINTF(("entries: %"Z"u", db->md_entries));
4001 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4002 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4003 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4004 DPRINTF(("root: %"Z"u", db->md_root));
4012 /** Release a reader thread's slot in the reader lock table.
4013 * This function is called automatically when a thread exits.
4014 * @param[in] ptr This points to the slot in the reader lock table.
4017 mdb_env_reader_dest(void *ptr)
4019 MDB_reader *reader = ptr;
4025 /** Junk for arranging thread-specific callbacks on Windows. This is
4026 * necessarily platform and compiler-specific. Windows supports up
4027 * to 1088 keys. Let's assume nobody opens more than 64 environments
4028 * in a single process, for now. They can override this if needed.
4030 #ifndef MAX_TLS_KEYS
4031 #define MAX_TLS_KEYS 64
4033 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4034 static int mdb_tls_nkeys;
4036 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4040 case DLL_PROCESS_ATTACH: break;
4041 case DLL_THREAD_ATTACH: break;
4042 case DLL_THREAD_DETACH:
4043 for (i=0; i<mdb_tls_nkeys; i++) {
4044 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4046 mdb_env_reader_dest(r);
4050 case DLL_PROCESS_DETACH: break;
4055 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4057 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4061 /* Force some symbol references.
4062 * _tls_used forces the linker to create the TLS directory if not already done
4063 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4065 #pragma comment(linker, "/INCLUDE:_tls_used")
4066 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4067 #pragma const_seg(".CRT$XLB")
4068 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4069 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4072 #pragma comment(linker, "/INCLUDE:__tls_used")
4073 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4074 #pragma data_seg(".CRT$XLB")
4075 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4077 #endif /* WIN 32/64 */
4078 #endif /* !__GNUC__ */
4081 /** Downgrade the exclusive lock on the region back to shared */
4083 mdb_env_share_locks(MDB_env *env, int *excl)
4085 int rc = 0, toggle = mdb_env_pick_meta(env);
4087 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4092 /* First acquire a shared lock. The Unlock will
4093 * then release the existing exclusive lock.
4095 memset(&ov, 0, sizeof(ov));
4096 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4099 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4105 struct flock lock_info;
4106 /* The shared lock replaces the existing lock */
4107 memset((void *)&lock_info, 0, sizeof(lock_info));
4108 lock_info.l_type = F_RDLCK;
4109 lock_info.l_whence = SEEK_SET;
4110 lock_info.l_start = 0;
4111 lock_info.l_len = 1;
4112 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4113 (rc = ErrCode()) == EINTR) ;
4114 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4121 /** Try to get exclusive lock, otherwise shared.
4122 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4125 mdb_env_excl_lock(MDB_env *env, int *excl)
4129 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4133 memset(&ov, 0, sizeof(ov));
4134 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4141 struct flock lock_info;
4142 memset((void *)&lock_info, 0, sizeof(lock_info));
4143 lock_info.l_type = F_WRLCK;
4144 lock_info.l_whence = SEEK_SET;
4145 lock_info.l_start = 0;
4146 lock_info.l_len = 1;
4147 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4148 (rc = ErrCode()) == EINTR) ;
4152 # ifdef MDB_USE_POSIX_SEM
4153 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4156 lock_info.l_type = F_RDLCK;
4157 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4158 (rc = ErrCode()) == EINTR) ;
4168 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4170 * @(#) $Revision: 5.1 $
4171 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4172 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4174 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4178 * Please do not copyright this code. This code is in the public domain.
4180 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4181 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4182 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4183 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4184 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4185 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4186 * PERFORMANCE OF THIS SOFTWARE.
4189 * chongo <Landon Curt Noll> /\oo/\
4190 * http://www.isthe.com/chongo/
4192 * Share and Enjoy! :-)
4195 typedef unsigned long long mdb_hash_t;
4196 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4198 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4199 * @param[in] val value to hash
4200 * @param[in] hval initial value for hash
4201 * @return 64 bit hash
4203 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4204 * hval arg on the first call.
4207 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4209 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4210 unsigned char *end = s + val->mv_size;
4212 * FNV-1a hash each octet of the string
4215 /* xor the bottom with the current octet */
4216 hval ^= (mdb_hash_t)*s++;
4218 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4219 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4220 (hval << 7) + (hval << 8) + (hval << 40);
4222 /* return our new hash value */
4226 /** Hash the string and output the encoded hash.
4227 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4228 * very short name limits. We don't care about the encoding being reversible,
4229 * we just want to preserve as many bits of the input as possible in a
4230 * small printable string.
4231 * @param[in] str string to hash
4232 * @param[out] encbuf an array of 11 chars to hold the hash
4234 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4237 mdb_pack85(unsigned long l, char *out)
4241 for (i=0; i<5; i++) {
4242 *out++ = mdb_a85[l % 85];
4248 mdb_hash_enc(MDB_val *val, char *encbuf)
4250 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4252 mdb_pack85(h, encbuf);
4253 mdb_pack85(h>>32, encbuf+5);
4258 /** Open and/or initialize the lock region for the environment.
4259 * @param[in] env The LMDB environment.
4260 * @param[in] lpath The pathname of the file used for the lock region.
4261 * @param[in] mode The Unix permissions for the file, if we create it.
4262 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4263 * @return 0 on success, non-zero on failure.
4266 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4269 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4271 # define MDB_ERRCODE_ROFS EROFS
4272 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4273 # define MDB_CLOEXEC O_CLOEXEC
4276 # define MDB_CLOEXEC 0
4283 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4284 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4285 FILE_ATTRIBUTE_NORMAL, NULL);
4287 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4289 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4291 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4296 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4297 /* Lose record locks when exec*() */
4298 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4299 fcntl(env->me_lfd, F_SETFD, fdflags);
4302 if (!(env->me_flags & MDB_NOTLS)) {
4303 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4306 env->me_flags |= MDB_ENV_TXKEY;
4308 /* Windows TLS callbacks need help finding their TLS info. */
4309 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4313 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4317 /* Try to get exclusive lock. If we succeed, then
4318 * nobody is using the lock region and we should initialize it.
4320 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4323 size = GetFileSize(env->me_lfd, NULL);
4325 size = lseek(env->me_lfd, 0, SEEK_END);
4326 if (size == -1) goto fail_errno;
4328 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4329 if (size < rsize && *excl > 0) {
4331 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4332 || !SetEndOfFile(env->me_lfd))
4335 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4339 size = rsize - sizeof(MDB_txninfo);
4340 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4345 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4347 if (!mh) goto fail_errno;
4348 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4350 if (!env->me_txns) goto fail_errno;
4352 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4354 if (m == MAP_FAILED) goto fail_errno;
4360 BY_HANDLE_FILE_INFORMATION stbuf;
4369 if (!mdb_sec_inited) {
4370 InitializeSecurityDescriptor(&mdb_null_sd,
4371 SECURITY_DESCRIPTOR_REVISION);
4372 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4373 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4374 mdb_all_sa.bInheritHandle = FALSE;
4375 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4378 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4379 idbuf.volume = stbuf.dwVolumeSerialNumber;
4380 idbuf.nhigh = stbuf.nFileIndexHigh;
4381 idbuf.nlow = stbuf.nFileIndexLow;
4382 val.mv_data = &idbuf;
4383 val.mv_size = sizeof(idbuf);
4384 mdb_hash_enc(&val, encbuf);
4385 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4386 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4387 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4388 if (!env->me_rmutex) goto fail_errno;
4389 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4390 if (!env->me_wmutex) goto fail_errno;
4391 #elif defined(MDB_USE_POSIX_SEM)
4400 #if defined(__NetBSD__)
4401 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4403 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4404 idbuf.dev = stbuf.st_dev;
4405 idbuf.ino = stbuf.st_ino;
4406 val.mv_data = &idbuf;
4407 val.mv_size = sizeof(idbuf);
4408 mdb_hash_enc(&val, encbuf);
4409 #ifdef MDB_SHORT_SEMNAMES
4410 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4412 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4413 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4414 /* Clean up after a previous run, if needed: Try to
4415 * remove both semaphores before doing anything else.
4417 sem_unlink(env->me_txns->mti_rmname);
4418 sem_unlink(env->me_txns->mti_wmname);
4419 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4420 O_CREAT|O_EXCL, mode, 1);
4421 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4422 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4423 O_CREAT|O_EXCL, mode, 1);
4424 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4425 #else /* MDB_USE_POSIX_SEM */
4426 pthread_mutexattr_t mattr;
4428 if ((rc = pthread_mutexattr_init(&mattr))
4429 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4430 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4431 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4433 pthread_mutexattr_destroy(&mattr);
4434 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4436 env->me_txns->mti_magic = MDB_MAGIC;
4437 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4438 env->me_txns->mti_txnid = 0;
4439 env->me_txns->mti_numreaders = 0;
4442 if (env->me_txns->mti_magic != MDB_MAGIC) {
4443 DPUTS("lock region has invalid magic");
4447 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4448 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4449 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4450 rc = MDB_VERSION_MISMATCH;
4454 if (rc && rc != EACCES && rc != EAGAIN) {
4458 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4459 if (!env->me_rmutex) goto fail_errno;
4460 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4461 if (!env->me_wmutex) goto fail_errno;
4462 #elif defined(MDB_USE_POSIX_SEM)
4463 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4464 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4465 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4466 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4477 /** The name of the lock file in the DB environment */
4478 #define LOCKNAME "/lock.mdb"
4479 /** The name of the data file in the DB environment */
4480 #define DATANAME "/data.mdb"
4481 /** The suffix of the lock file when no subdir is used */
4482 #define LOCKSUFF "-lock"
4483 /** Only a subset of the @ref mdb_env flags can be changed
4484 * at runtime. Changing other flags requires closing the
4485 * environment and re-opening it with the new flags.
4487 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4488 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4489 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4491 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4492 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4496 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4498 int oflags, rc, len, excl = -1;
4499 char *lpath, *dpath;
4501 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4505 if (flags & MDB_NOSUBDIR) {
4506 rc = len + sizeof(LOCKSUFF) + len + 1;
4508 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4513 if (flags & MDB_NOSUBDIR) {
4514 dpath = lpath + len + sizeof(LOCKSUFF);
4515 sprintf(lpath, "%s" LOCKSUFF, path);
4516 strcpy(dpath, path);
4518 dpath = lpath + len + sizeof(LOCKNAME);
4519 sprintf(lpath, "%s" LOCKNAME, path);
4520 sprintf(dpath, "%s" DATANAME, path);
4524 flags |= env->me_flags;
4525 if (flags & MDB_RDONLY) {
4526 /* silently ignore WRITEMAP when we're only getting read access */
4527 flags &= ~MDB_WRITEMAP;
4529 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4530 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4533 env->me_flags = flags |= MDB_ENV_ACTIVE;
4537 env->me_path = strdup(path);
4538 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4539 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4540 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4541 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4546 /* For RDONLY, get lockfile after we know datafile exists */
4547 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4548 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4554 if (F_ISSET(flags, MDB_RDONLY)) {
4555 oflags = GENERIC_READ;
4556 len = OPEN_EXISTING;
4558 oflags = GENERIC_READ|GENERIC_WRITE;
4561 mode = FILE_ATTRIBUTE_NORMAL;
4562 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4563 NULL, len, mode, NULL);
4565 if (F_ISSET(flags, MDB_RDONLY))
4568 oflags = O_RDWR | O_CREAT;
4570 env->me_fd = open(dpath, oflags, mode);
4572 if (env->me_fd == INVALID_HANDLE_VALUE) {
4577 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4578 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4583 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4584 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4585 env->me_mfd = env->me_fd;
4587 /* Synchronous fd for meta writes. Needed even with
4588 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4591 len = OPEN_EXISTING;
4592 env->me_mfd = CreateFile(dpath, oflags,
4593 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4594 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4597 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4599 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4604 DPRINTF(("opened dbenv %p", (void *) env));
4606 rc = mdb_env_share_locks(env, &excl);
4610 if (!((flags & MDB_RDONLY) ||
4611 (env->me_pbuf = calloc(1, env->me_psize))))
4613 if (!(flags & MDB_RDONLY)) {
4615 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4616 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4617 txn = calloc(1, size);
4619 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4620 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4621 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4622 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4624 txn->mt_dbxs = env->me_dbxs;
4634 mdb_env_close0(env, excl);
4640 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4642 mdb_env_close0(MDB_env *env, int excl)
4646 if (!(env->me_flags & MDB_ENV_ACTIVE))
4649 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4650 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4651 free(env->me_dbxs[i].md_name.mv_data);
4654 free(env->me_dbiseqs);
4655 free(env->me_dbflags);
4658 free(env->me_dirty_list);
4660 mdb_midl_free(env->me_free_pgs);
4662 if (env->me_flags & MDB_ENV_TXKEY) {
4663 pthread_key_delete(env->me_txkey);
4665 /* Delete our key from the global list */
4666 for (i=0; i<mdb_tls_nkeys; i++)
4667 if (mdb_tls_keys[i] == env->me_txkey) {
4668 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4676 munmap(env->me_map, env->me_mapsize);
4678 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4679 (void) close(env->me_mfd);
4680 if (env->me_fd != INVALID_HANDLE_VALUE)
4681 (void) close(env->me_fd);
4683 MDB_PID_T pid = env->me_pid;
4684 /* Clearing readers is done in this function because
4685 * me_txkey with its destructor must be disabled first.
4687 for (i = env->me_numreaders; --i >= 0; )
4688 if (env->me_txns->mti_readers[i].mr_pid == pid)
4689 env->me_txns->mti_readers[i].mr_pid = 0;
4691 if (env->me_rmutex) {
4692 CloseHandle(env->me_rmutex);
4693 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4695 /* Windows automatically destroys the mutexes when
4696 * the last handle closes.
4698 #elif defined(MDB_USE_POSIX_SEM)
4699 if (env->me_rmutex != SEM_FAILED) {
4700 sem_close(env->me_rmutex);
4701 if (env->me_wmutex != SEM_FAILED)
4702 sem_close(env->me_wmutex);
4703 /* If we have the filelock: If we are the
4704 * only remaining user, clean up semaphores.
4707 mdb_env_excl_lock(env, &excl);
4709 sem_unlink(env->me_txns->mti_rmname);
4710 sem_unlink(env->me_txns->mti_wmname);
4714 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4716 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4719 /* Unlock the lockfile. Windows would have unlocked it
4720 * after closing anyway, but not necessarily at once.
4722 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4725 (void) close(env->me_lfd);
4728 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4733 mdb_env_close(MDB_env *env)
4740 VGMEMP_DESTROY(env);
4741 while ((dp = env->me_dpages) != NULL) {
4742 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4743 env->me_dpages = dp->mp_next;
4747 mdb_env_close0(env, 0);
4751 /** Compare two items pointing at aligned size_t's */
4753 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4755 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4756 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4759 /** Compare two items pointing at aligned unsigned int's */
4761 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4763 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4764 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4767 /** Compare two items pointing at unsigned ints of unknown alignment.
4768 * Nodes and keys are guaranteed to be 2-byte aligned.
4771 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4773 #if BYTE_ORDER == LITTLE_ENDIAN
4774 unsigned short *u, *c;
4777 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4778 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4781 } while(!x && u > (unsigned short *)a->mv_data);
4784 unsigned short *u, *c, *end;
4787 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4788 u = (unsigned short *)a->mv_data;
4789 c = (unsigned short *)b->mv_data;
4792 } while(!x && u < end);
4797 /** Compare two items pointing at size_t's of unknown alignment. */
4798 #ifdef MISALIGNED_OK
4799 # define mdb_cmp_clong mdb_cmp_long
4801 # define mdb_cmp_clong mdb_cmp_cint
4804 /** Compare two items lexically */
4806 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4813 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4819 diff = memcmp(a->mv_data, b->mv_data, len);
4820 return diff ? diff : len_diff<0 ? -1 : len_diff;
4823 /** Compare two items in reverse byte order */
4825 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4827 const unsigned char *p1, *p2, *p1_lim;
4831 p1_lim = (const unsigned char *)a->mv_data;
4832 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4833 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4835 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4841 while (p1 > p1_lim) {
4842 diff = *--p1 - *--p2;
4846 return len_diff<0 ? -1 : len_diff;
4849 /** Search for key within a page, using binary search.
4850 * Returns the smallest entry larger or equal to the key.
4851 * If exactp is non-null, stores whether the found entry was an exact match
4852 * in *exactp (1 or 0).
4853 * Updates the cursor index with the index of the found entry.
4854 * If no entry larger or equal to the key is found, returns NULL.
4857 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4859 unsigned int i = 0, nkeys;
4862 MDB_page *mp = mc->mc_pg[mc->mc_top];
4863 MDB_node *node = NULL;
4868 nkeys = NUMKEYS(mp);
4870 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4871 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4874 low = IS_LEAF(mp) ? 0 : 1;
4876 cmp = mc->mc_dbx->md_cmp;
4878 /* Branch pages have no data, so if using integer keys,
4879 * alignment is guaranteed. Use faster mdb_cmp_int.
4881 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4882 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4889 nodekey.mv_size = mc->mc_db->md_pad;
4890 node = NODEPTR(mp, 0); /* fake */
4891 while (low <= high) {
4892 i = (low + high) >> 1;
4893 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4894 rc = cmp(key, &nodekey);
4895 DPRINTF(("found leaf index %u [%s], rc = %i",
4896 i, DKEY(&nodekey), rc));
4905 while (low <= high) {
4906 i = (low + high) >> 1;
4908 node = NODEPTR(mp, i);
4909 nodekey.mv_size = NODEKSZ(node);
4910 nodekey.mv_data = NODEKEY(node);
4912 rc = cmp(key, &nodekey);
4915 DPRINTF(("found leaf index %u [%s], rc = %i",
4916 i, DKEY(&nodekey), rc));
4918 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4919 i, DKEY(&nodekey), NODEPGNO(node), rc));
4930 if (rc > 0) { /* Found entry is less than the key. */
4931 i++; /* Skip to get the smallest entry larger than key. */
4933 node = NODEPTR(mp, i);
4936 *exactp = (rc == 0 && nkeys > 0);
4937 /* store the key index */
4938 mc->mc_ki[mc->mc_top] = i;
4940 /* There is no entry larger or equal to the key. */
4943 /* nodeptr is fake for LEAF2 */
4949 mdb_cursor_adjust(MDB_cursor *mc, func)
4953 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4954 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4961 /** Pop a page off the top of the cursor's stack. */
4963 mdb_cursor_pop(MDB_cursor *mc)
4967 MDB_page *top = mc->mc_pg[mc->mc_top];
4973 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4974 DDBI(mc), (void *) mc));
4978 /** Push a page onto the top of the cursor's stack. */
4980 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4982 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4983 DDBI(mc), (void *) mc));
4985 if (mc->mc_snum >= CURSOR_STACK) {
4986 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4987 return MDB_CURSOR_FULL;
4990 mc->mc_top = mc->mc_snum++;
4991 mc->mc_pg[mc->mc_top] = mp;
4992 mc->mc_ki[mc->mc_top] = 0;
4997 /** Find the address of the page corresponding to a given page number.
4998 * @param[in] txn the transaction for this access.
4999 * @param[in] pgno the page number for the page to retrieve.
5000 * @param[out] ret address of a pointer where the page's address will be stored.
5001 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5002 * @return 0 on success, non-zero on failure.
5005 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5007 MDB_env *env = txn->mt_env;
5011 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
5015 MDB_ID2L dl = tx2->mt_u.dirty_list;
5017 /* Spilled pages were dirtied in this txn and flushed
5018 * because the dirty list got full. Bring this page
5019 * back in from the map (but don't unspill it here,
5020 * leave that unless page_touch happens again).
5022 if (tx2->mt_spill_pgs) {
5023 MDB_ID pn = pgno << 1;
5024 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5025 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5026 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5031 unsigned x = mdb_mid2l_search(dl, pgno);
5032 if (x <= dl[0].mid && dl[x].mid == pgno) {
5038 } while ((tx2 = tx2->mt_parent) != NULL);
5041 if (pgno < txn->mt_next_pgno) {
5043 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5045 DPRINTF(("page %"Z"u not found", pgno));
5046 txn->mt_flags |= MDB_TXN_ERROR;
5047 return MDB_PAGE_NOTFOUND;
5057 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5058 * The cursor is at the root page, set up the rest of it.
5061 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5063 MDB_page *mp = mc->mc_pg[mc->mc_top];
5067 while (IS_BRANCH(mp)) {
5071 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5072 mdb_cassert(mc, NUMKEYS(mp) > 1);
5073 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5075 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5077 if (flags & MDB_PS_LAST)
5078 i = NUMKEYS(mp) - 1;
5081 node = mdb_node_search(mc, key, &exact);
5083 i = NUMKEYS(mp) - 1;
5085 i = mc->mc_ki[mc->mc_top];
5087 mdb_cassert(mc, i > 0);
5091 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5094 mdb_cassert(mc, i < NUMKEYS(mp));
5095 node = NODEPTR(mp, i);
5097 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5100 mc->mc_ki[mc->mc_top] = i;
5101 if ((rc = mdb_cursor_push(mc, mp)))
5104 if (flags & MDB_PS_MODIFY) {
5105 if ((rc = mdb_page_touch(mc)) != 0)
5107 mp = mc->mc_pg[mc->mc_top];
5112 DPRINTF(("internal error, index points to a %02X page!?",
5114 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5115 return MDB_CORRUPTED;
5118 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5119 key ? DKEY(key) : "null"));
5120 mc->mc_flags |= C_INITIALIZED;
5121 mc->mc_flags &= ~C_EOF;
5126 /** Search for the lowest key under the current branch page.
5127 * This just bypasses a NUMKEYS check in the current page
5128 * before calling mdb_page_search_root(), because the callers
5129 * are all in situations where the current page is known to
5133 mdb_page_search_lowest(MDB_cursor *mc)
5135 MDB_page *mp = mc->mc_pg[mc->mc_top];
5136 MDB_node *node = NODEPTR(mp, 0);
5139 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5142 mc->mc_ki[mc->mc_top] = 0;
5143 if ((rc = mdb_cursor_push(mc, mp)))
5145 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5148 /** Search for the page a given key should be in.
5149 * Push it and its parent pages on the cursor stack.
5150 * @param[in,out] mc the cursor for this operation.
5151 * @param[in] key the key to search for, or NULL for first/last page.
5152 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5153 * are touched (updated with new page numbers).
5154 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5155 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5156 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5157 * @return 0 on success, non-zero on failure.
5160 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5165 /* Make sure the txn is still viable, then find the root from
5166 * the txn's db table and set it as the root of the cursor's stack.
5168 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5169 DPUTS("transaction has failed, must abort");
5172 /* Make sure we're using an up-to-date root */
5173 if (*mc->mc_dbflag & DB_STALE) {
5175 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5177 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5178 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5185 MDB_node *leaf = mdb_node_search(&mc2,
5186 &mc->mc_dbx->md_name, &exact);
5188 return MDB_NOTFOUND;
5189 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5192 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5194 /* The txn may not know this DBI, or another process may
5195 * have dropped and recreated the DB with other flags.
5197 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5198 return MDB_INCOMPATIBLE;
5199 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5201 *mc->mc_dbflag &= ~DB_STALE;
5203 root = mc->mc_db->md_root;
5205 if (root == P_INVALID) { /* Tree is empty. */
5206 DPUTS("tree is empty");
5207 return MDB_NOTFOUND;
5211 mdb_cassert(mc, root > 1);
5212 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5213 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5219 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5220 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5222 if (flags & MDB_PS_MODIFY) {
5223 if ((rc = mdb_page_touch(mc)))
5227 if (flags & MDB_PS_ROOTONLY)
5230 return mdb_page_search_root(mc, key, flags);
5234 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5236 MDB_txn *txn = mc->mc_txn;
5237 pgno_t pg = mp->mp_pgno;
5238 unsigned x = 0, ovpages = mp->mp_pages;
5239 MDB_env *env = txn->mt_env;
5240 MDB_IDL sl = txn->mt_spill_pgs;
5241 MDB_ID pn = pg << 1;
5244 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5245 /* If the page is dirty or on the spill list we just acquired it,
5246 * so we should give it back to our current free list, if any.
5247 * Otherwise put it onto the list of pages we freed in this txn.
5249 * Won't create me_pghead: me_pglast must be inited along with it.
5250 * Unsupported in nested txns: They would need to hide the page
5251 * range in ancestor txns' dirty and spilled lists.
5253 if (env->me_pghead &&
5255 ((mp->mp_flags & P_DIRTY) ||
5256 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5260 MDB_ID2 *dl, ix, iy;
5261 rc = mdb_midl_need(&env->me_pghead, ovpages);
5264 if (!(mp->mp_flags & P_DIRTY)) {
5265 /* This page is no longer spilled */
5272 /* Remove from dirty list */
5273 dl = txn->mt_u.dirty_list;
5275 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5281 mdb_cassert(mc, x > 1);
5283 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5284 txn->mt_flags |= MDB_TXN_ERROR;
5285 return MDB_CORRUPTED;
5288 if (!(env->me_flags & MDB_WRITEMAP))
5289 mdb_dpage_free(env, mp);
5291 /* Insert in me_pghead */
5292 mop = env->me_pghead;
5293 j = mop[0] + ovpages;
5294 for (i = mop[0]; i && mop[i] < pg; i--)
5300 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5304 mc->mc_db->md_overflow_pages -= ovpages;
5308 /** Return the data associated with a given node.
5309 * @param[in] txn The transaction for this operation.
5310 * @param[in] leaf The node being read.
5311 * @param[out] data Updated to point to the node's data.
5312 * @return 0 on success, non-zero on failure.
5315 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5317 MDB_page *omp; /* overflow page */
5321 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5322 data->mv_size = NODEDSZ(leaf);
5323 data->mv_data = NODEDATA(leaf);
5327 /* Read overflow data.
5329 data->mv_size = NODEDSZ(leaf);
5330 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5331 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5332 DPRINTF(("read overflow page %"Z"u failed", pgno));
5335 data->mv_data = METADATA(omp);
5341 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5342 MDB_val *key, MDB_val *data)
5349 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5351 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5354 if (txn->mt_flags & MDB_TXN_ERROR)
5357 mdb_cursor_init(&mc, txn, dbi, &mx);
5358 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5361 /** Find a sibling for a page.
5362 * Replaces the page at the top of the cursor's stack with the
5363 * specified sibling, if one exists.
5364 * @param[in] mc The cursor for this operation.
5365 * @param[in] move_right Non-zero if the right sibling is requested,
5366 * otherwise the left sibling.
5367 * @return 0 on success, non-zero on failure.
5370 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5376 if (mc->mc_snum < 2) {
5377 return MDB_NOTFOUND; /* root has no siblings */
5381 DPRINTF(("parent page is page %"Z"u, index %u",
5382 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5384 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5385 : (mc->mc_ki[mc->mc_top] == 0)) {
5386 DPRINTF(("no more keys left, moving to %s sibling",
5387 move_right ? "right" : "left"));
5388 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5389 /* undo cursor_pop before returning */
5396 mc->mc_ki[mc->mc_top]++;
5398 mc->mc_ki[mc->mc_top]--;
5399 DPRINTF(("just moving to %s index key %u",
5400 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5402 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5404 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5405 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5406 /* mc will be inconsistent if caller does mc_snum++ as above */
5407 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5411 mdb_cursor_push(mc, mp);
5413 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5418 /** Move the cursor to the next data item. */
5420 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5426 if (mc->mc_flags & C_EOF) {
5427 return MDB_NOTFOUND;
5430 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5432 mp = mc->mc_pg[mc->mc_top];
5434 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5435 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5436 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5437 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5438 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5439 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5440 if (rc == MDB_SUCCESS)
5441 MDB_GET_KEY(leaf, key);
5446 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5447 if (op == MDB_NEXT_DUP)
5448 return MDB_NOTFOUND;
5452 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5453 mdb_dbg_pgno(mp), (void *) mc));
5454 if (mc->mc_flags & C_DEL)
5457 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5458 DPUTS("=====> move to next sibling page");
5459 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5460 mc->mc_flags |= C_EOF;
5463 mp = mc->mc_pg[mc->mc_top];
5464 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5466 mc->mc_ki[mc->mc_top]++;
5469 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5470 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5473 key->mv_size = mc->mc_db->md_pad;
5474 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5478 mdb_cassert(mc, IS_LEAF(mp));
5479 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5481 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5482 mdb_xcursor_init1(mc, leaf);
5485 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5488 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5489 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5490 if (rc != MDB_SUCCESS)
5495 MDB_GET_KEY(leaf, key);
5499 /** Move the cursor to the previous data item. */
5501 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5507 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5509 mp = mc->mc_pg[mc->mc_top];
5511 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5512 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5513 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5514 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5515 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5516 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5517 if (rc == MDB_SUCCESS) {
5518 MDB_GET_KEY(leaf, key);
5519 mc->mc_flags &= ~C_EOF;
5525 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5526 if (op == MDB_PREV_DUP)
5527 return MDB_NOTFOUND;
5531 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5532 mdb_dbg_pgno(mp), (void *) mc));
5534 if (mc->mc_ki[mc->mc_top] == 0) {
5535 DPUTS("=====> move to prev sibling page");
5536 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5539 mp = mc->mc_pg[mc->mc_top];
5540 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5541 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5543 mc->mc_ki[mc->mc_top]--;
5545 mc->mc_flags &= ~C_EOF;
5547 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5548 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5551 key->mv_size = mc->mc_db->md_pad;
5552 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5556 mdb_cassert(mc, IS_LEAF(mp));
5557 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5559 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5560 mdb_xcursor_init1(mc, leaf);
5563 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5566 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5567 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5568 if (rc != MDB_SUCCESS)
5573 MDB_GET_KEY(leaf, key);
5577 /** Set the cursor on a specific data item. */
5579 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5580 MDB_cursor_op op, int *exactp)
5584 MDB_node *leaf = NULL;
5587 if (key->mv_size == 0)
5588 return MDB_BAD_VALSIZE;
5591 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5593 /* See if we're already on the right page */
5594 if (mc->mc_flags & C_INITIALIZED) {
5597 mp = mc->mc_pg[mc->mc_top];
5599 mc->mc_ki[mc->mc_top] = 0;
5600 return MDB_NOTFOUND;
5602 if (mp->mp_flags & P_LEAF2) {
5603 nodekey.mv_size = mc->mc_db->md_pad;
5604 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5606 leaf = NODEPTR(mp, 0);
5607 MDB_GET_KEY2(leaf, nodekey);
5609 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5611 /* Probably happens rarely, but first node on the page
5612 * was the one we wanted.
5614 mc->mc_ki[mc->mc_top] = 0;
5621 unsigned int nkeys = NUMKEYS(mp);
5623 if (mp->mp_flags & P_LEAF2) {
5624 nodekey.mv_data = LEAF2KEY(mp,
5625 nkeys-1, nodekey.mv_size);
5627 leaf = NODEPTR(mp, nkeys-1);
5628 MDB_GET_KEY2(leaf, nodekey);
5630 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5632 /* last node was the one we wanted */
5633 mc->mc_ki[mc->mc_top] = nkeys-1;
5639 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5640 /* This is definitely the right page, skip search_page */
5641 if (mp->mp_flags & P_LEAF2) {
5642 nodekey.mv_data = LEAF2KEY(mp,
5643 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5645 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5646 MDB_GET_KEY2(leaf, nodekey);
5648 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5650 /* current node was the one we wanted */
5660 /* If any parents have right-sibs, search.
5661 * Otherwise, there's nothing further.
5663 for (i=0; i<mc->mc_top; i++)
5665 NUMKEYS(mc->mc_pg[i])-1)
5667 if (i == mc->mc_top) {
5668 /* There are no other pages */
5669 mc->mc_ki[mc->mc_top] = nkeys;
5670 return MDB_NOTFOUND;
5674 /* There are no other pages */
5675 mc->mc_ki[mc->mc_top] = 0;
5676 if (op == MDB_SET_RANGE && !exactp) {
5680 return MDB_NOTFOUND;
5684 rc = mdb_page_search(mc, key, 0);
5685 if (rc != MDB_SUCCESS)
5688 mp = mc->mc_pg[mc->mc_top];
5689 mdb_cassert(mc, IS_LEAF(mp));
5692 leaf = mdb_node_search(mc, key, exactp);
5693 if (exactp != NULL && !*exactp) {
5694 /* MDB_SET specified and not an exact match. */
5695 return MDB_NOTFOUND;
5699 DPUTS("===> inexact leaf not found, goto sibling");
5700 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5701 return rc; /* no entries matched */
5702 mp = mc->mc_pg[mc->mc_top];
5703 mdb_cassert(mc, IS_LEAF(mp));
5704 leaf = NODEPTR(mp, 0);
5708 mc->mc_flags |= C_INITIALIZED;
5709 mc->mc_flags &= ~C_EOF;
5712 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5713 key->mv_size = mc->mc_db->md_pad;
5714 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5719 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5720 mdb_xcursor_init1(mc, leaf);
5723 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5724 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5725 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5728 if (op == MDB_GET_BOTH) {
5734 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5735 if (rc != MDB_SUCCESS)
5738 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5740 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5742 rc = mc->mc_dbx->md_dcmp(data, &d2);
5744 if (op == MDB_GET_BOTH || rc > 0)
5745 return MDB_NOTFOUND;
5752 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5753 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5758 /* The key already matches in all other cases */
5759 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5760 MDB_GET_KEY(leaf, key);
5761 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5766 /** Move the cursor to the first item in the database. */
5768 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5774 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5776 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5777 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5778 if (rc != MDB_SUCCESS)
5781 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5783 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5784 mc->mc_flags |= C_INITIALIZED;
5785 mc->mc_flags &= ~C_EOF;
5787 mc->mc_ki[mc->mc_top] = 0;
5789 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5790 key->mv_size = mc->mc_db->md_pad;
5791 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5796 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5797 mdb_xcursor_init1(mc, leaf);
5798 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5802 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5806 MDB_GET_KEY(leaf, key);
5810 /** Move the cursor to the last item in the database. */
5812 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5818 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5820 if (!(mc->mc_flags & C_EOF)) {
5822 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5823 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5824 if (rc != MDB_SUCCESS)
5827 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5830 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5831 mc->mc_flags |= C_INITIALIZED|C_EOF;
5832 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5834 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5835 key->mv_size = mc->mc_db->md_pad;
5836 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5841 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5842 mdb_xcursor_init1(mc, leaf);
5843 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5847 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5852 MDB_GET_KEY(leaf, key);
5857 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5862 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5867 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5871 case MDB_GET_CURRENT:
5872 if (!(mc->mc_flags & C_INITIALIZED)) {
5875 MDB_page *mp = mc->mc_pg[mc->mc_top];
5876 int nkeys = NUMKEYS(mp);
5877 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5878 mc->mc_ki[mc->mc_top] = nkeys;
5884 key->mv_size = mc->mc_db->md_pad;
5885 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5887 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5888 MDB_GET_KEY(leaf, key);
5890 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5891 if (mc->mc_flags & C_DEL)
5892 mdb_xcursor_init1(mc, leaf);
5893 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5895 rc = mdb_node_read(mc->mc_txn, leaf, data);
5902 case MDB_GET_BOTH_RANGE:
5907 if (mc->mc_xcursor == NULL) {
5908 rc = MDB_INCOMPATIBLE;
5918 rc = mdb_cursor_set(mc, key, data, op,
5919 op == MDB_SET_RANGE ? NULL : &exact);
5922 case MDB_GET_MULTIPLE:
5923 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5927 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5928 rc = MDB_INCOMPATIBLE;
5932 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5933 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5936 case MDB_NEXT_MULTIPLE:
5941 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5942 rc = MDB_INCOMPATIBLE;
5945 if (!(mc->mc_flags & C_INITIALIZED))
5946 rc = mdb_cursor_first(mc, key, data);
5948 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5949 if (rc == MDB_SUCCESS) {
5950 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5953 mx = &mc->mc_xcursor->mx_cursor;
5954 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5956 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5957 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5965 case MDB_NEXT_NODUP:
5966 if (!(mc->mc_flags & C_INITIALIZED))
5967 rc = mdb_cursor_first(mc, key, data);
5969 rc = mdb_cursor_next(mc, key, data, op);
5973 case MDB_PREV_NODUP:
5974 if (!(mc->mc_flags & C_INITIALIZED)) {
5975 rc = mdb_cursor_last(mc, key, data);
5978 mc->mc_flags |= C_INITIALIZED;
5979 mc->mc_ki[mc->mc_top]++;
5981 rc = mdb_cursor_prev(mc, key, data, op);
5984 rc = mdb_cursor_first(mc, key, data);
5987 mfunc = mdb_cursor_first;
5989 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5993 if (mc->mc_xcursor == NULL) {
5994 rc = MDB_INCOMPATIBLE;
5998 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5999 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6000 MDB_GET_KEY(leaf, key);
6001 rc = mdb_node_read(mc->mc_txn, leaf, data);
6005 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6009 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6012 rc = mdb_cursor_last(mc, key, data);
6015 mfunc = mdb_cursor_last;
6018 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6023 if (mc->mc_flags & C_DEL)
6024 mc->mc_flags ^= C_DEL;
6029 /** Touch all the pages in the cursor stack. Set mc_top.
6030 * Makes sure all the pages are writable, before attempting a write operation.
6031 * @param[in] mc The cursor to operate on.
6034 mdb_cursor_touch(MDB_cursor *mc)
6036 int rc = MDB_SUCCESS;
6038 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6041 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6043 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6044 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6047 *mc->mc_dbflag |= DB_DIRTY;
6052 rc = mdb_page_touch(mc);
6053 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6054 mc->mc_top = mc->mc_snum-1;
6059 /** Do not spill pages to disk if txn is getting full, may fail instead */
6060 #define MDB_NOSPILL 0x8000
6063 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6066 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6068 MDB_node *leaf = NULL;
6071 MDB_val xdata, *rdata, dkey, olddata;
6073 int do_sub = 0, insert_key, insert_data;
6074 unsigned int mcount = 0, dcount = 0, nospill;
6077 unsigned int nflags;
6080 if (mc == NULL || key == NULL)
6083 env = mc->mc_txn->mt_env;
6085 /* Check this first so counter will always be zero on any
6088 if (flags & MDB_MULTIPLE) {
6089 dcount = data[1].mv_size;
6090 data[1].mv_size = 0;
6091 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6092 return MDB_INCOMPATIBLE;
6095 nospill = flags & MDB_NOSPILL;
6096 flags &= ~MDB_NOSPILL;
6098 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6099 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6101 if (key->mv_size-1 >= ENV_MAXKEY(env))
6102 return MDB_BAD_VALSIZE;
6104 #if SIZE_MAX > MAXDATASIZE
6105 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6106 return MDB_BAD_VALSIZE;
6108 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6109 return MDB_BAD_VALSIZE;
6112 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6113 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6117 if (flags == MDB_CURRENT) {
6118 if (!(mc->mc_flags & C_INITIALIZED))
6121 } else if (mc->mc_db->md_root == P_INVALID) {
6122 /* new database, cursor has nothing to point to */
6125 mc->mc_flags &= ~C_INITIALIZED;
6130 if (flags & MDB_APPEND) {
6132 rc = mdb_cursor_last(mc, &k2, &d2);
6134 rc = mc->mc_dbx->md_cmp(key, &k2);
6137 mc->mc_ki[mc->mc_top]++;
6139 /* new key is <= last key */
6144 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6146 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6147 DPRINTF(("duplicate key [%s]", DKEY(key)));
6149 return MDB_KEYEXIST;
6151 if (rc && rc != MDB_NOTFOUND)
6155 if (mc->mc_flags & C_DEL)
6156 mc->mc_flags ^= C_DEL;
6158 /* Cursor is positioned, check for room in the dirty list */
6160 if (flags & MDB_MULTIPLE) {
6162 xdata.mv_size = data->mv_size * dcount;
6166 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6170 if (rc == MDB_NO_ROOT) {
6172 /* new database, write a root leaf page */
6173 DPUTS("allocating new root leaf page");
6174 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6177 mdb_cursor_push(mc, np);
6178 mc->mc_db->md_root = np->mp_pgno;
6179 mc->mc_db->md_depth++;
6180 *mc->mc_dbflag |= DB_DIRTY;
6181 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6183 np->mp_flags |= P_LEAF2;
6184 mc->mc_flags |= C_INITIALIZED;
6186 /* make sure all cursor pages are writable */
6187 rc2 = mdb_cursor_touch(mc);
6192 insert_key = insert_data = rc;
6194 /* The key does not exist */
6195 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6196 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6197 LEAFSIZE(key, data) > env->me_nodemax)
6199 /* Too big for a node, insert in sub-DB. Set up an empty
6200 * "old sub-page" for prep_subDB to expand to a full page.
6202 fp_flags = P_LEAF|P_DIRTY;
6204 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6205 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6206 olddata.mv_size = PAGEHDRSZ;
6210 /* there's only a key anyway, so this is a no-op */
6211 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6213 unsigned int ksize = mc->mc_db->md_pad;
6214 if (key->mv_size != ksize)
6215 return MDB_BAD_VALSIZE;
6216 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6217 memcpy(ptr, key->mv_data, ksize);
6219 /* if overwriting slot 0 of leaf, need to
6220 * update branch key if there is a parent page
6222 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6223 unsigned short top = mc->mc_top;
6225 /* slot 0 is always an empty key, find real slot */
6226 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6228 if (mc->mc_ki[mc->mc_top])
6229 rc2 = mdb_update_key(mc, key);
6240 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6241 olddata.mv_size = NODEDSZ(leaf);
6242 olddata.mv_data = NODEDATA(leaf);
6245 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6246 /* Prepare (sub-)page/sub-DB to accept the new item,
6247 * if needed. fp: old sub-page or a header faking
6248 * it. mp: new (sub-)page. offset: growth in page
6249 * size. xdata: node data with new page or DB.
6251 unsigned i, offset = 0;
6252 mp = fp = xdata.mv_data = env->me_pbuf;
6253 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6255 /* Was a single item before, must convert now */
6256 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6257 /* Just overwrite the current item */
6258 if (flags == MDB_CURRENT)
6261 #if UINT_MAX < SIZE_MAX
6262 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6263 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6265 /* does data match? */
6266 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6267 if (flags & MDB_NODUPDATA)
6268 return MDB_KEYEXIST;
6273 /* Back up original data item */
6274 dkey.mv_size = olddata.mv_size;
6275 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6277 /* Make sub-page header for the dup items, with dummy body */
6278 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6279 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6280 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6281 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6282 fp->mp_flags |= P_LEAF2;
6283 fp->mp_pad = data->mv_size;
6284 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6286 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6287 (dkey.mv_size & 1) + (data->mv_size & 1);
6289 fp->mp_upper = xdata.mv_size - PAGEBASE;
6290 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6291 } else if (leaf->mn_flags & F_SUBDATA) {
6292 /* Data is on sub-DB, just store it */
6293 flags |= F_DUPDATA|F_SUBDATA;
6296 /* Data is on sub-page */
6297 fp = olddata.mv_data;
6300 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6301 offset = EVEN(NODESIZE + sizeof(indx_t) +
6305 offset = fp->mp_pad;
6306 if (SIZELEFT(fp) < offset) {
6307 offset *= 4; /* space for 4 more */
6310 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6312 fp->mp_flags |= P_DIRTY;
6313 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6314 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6318 xdata.mv_size = olddata.mv_size + offset;
6321 fp_flags = fp->mp_flags;
6322 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6323 /* Too big for a sub-page, convert to sub-DB */
6324 fp_flags &= ~P_SUBP;
6326 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6327 fp_flags |= P_LEAF2;
6328 dummy.md_pad = fp->mp_pad;
6329 dummy.md_flags = MDB_DUPFIXED;
6330 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6331 dummy.md_flags |= MDB_INTEGERKEY;
6337 dummy.md_branch_pages = 0;
6338 dummy.md_leaf_pages = 1;
6339 dummy.md_overflow_pages = 0;
6340 dummy.md_entries = NUMKEYS(fp);
6341 xdata.mv_size = sizeof(MDB_db);
6342 xdata.mv_data = &dummy;
6343 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6345 offset = env->me_psize - olddata.mv_size;
6346 flags |= F_DUPDATA|F_SUBDATA;
6347 dummy.md_root = mp->mp_pgno;
6350 mp->mp_flags = fp_flags | P_DIRTY;
6351 mp->mp_pad = fp->mp_pad;
6352 mp->mp_lower = fp->mp_lower;
6353 mp->mp_upper = fp->mp_upper + offset;
6354 if (fp_flags & P_LEAF2) {
6355 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6357 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6358 olddata.mv_size - fp->mp_upper - PAGEBASE);
6359 for (i=0; i<NUMKEYS(fp); i++)
6360 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6368 mdb_node_del(mc, 0);
6372 /* overflow page overwrites need special handling */
6373 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6376 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6378 memcpy(&pg, olddata.mv_data, sizeof(pg));
6379 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6381 ovpages = omp->mp_pages;
6383 /* Is the ov page large enough? */
6384 if (ovpages >= dpages) {
6385 if (!(omp->mp_flags & P_DIRTY) &&
6386 (level || (env->me_flags & MDB_WRITEMAP)))
6388 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6391 level = 0; /* dirty in this txn or clean */
6394 if (omp->mp_flags & P_DIRTY) {
6395 /* yes, overwrite it. Note in this case we don't
6396 * bother to try shrinking the page if the new data
6397 * is smaller than the overflow threshold.
6400 /* It is writable only in a parent txn */
6401 size_t sz = (size_t) env->me_psize * ovpages, off;
6402 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6408 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6409 mdb_cassert(mc, rc2 == 0);
6410 if (!(flags & MDB_RESERVE)) {
6411 /* Copy end of page, adjusting alignment so
6412 * compiler may copy words instead of bytes.
6414 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6415 memcpy((size_t *)((char *)np + off),
6416 (size_t *)((char *)omp + off), sz - off);
6419 memcpy(np, omp, sz); /* Copy beginning of page */
6422 SETDSZ(leaf, data->mv_size);
6423 if (F_ISSET(flags, MDB_RESERVE))
6424 data->mv_data = METADATA(omp);
6426 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6430 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6432 } else if (data->mv_size == olddata.mv_size) {
6433 /* same size, just replace it. Note that we could
6434 * also reuse this node if the new data is smaller,
6435 * but instead we opt to shrink the node in that case.
6437 if (F_ISSET(flags, MDB_RESERVE))
6438 data->mv_data = olddata.mv_data;
6439 else if (!(mc->mc_flags & C_SUB))
6440 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6442 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6447 mdb_node_del(mc, 0);
6453 nflags = flags & NODE_ADD_FLAGS;
6454 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6455 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6456 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6457 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6459 nflags |= MDB_SPLIT_REPLACE;
6460 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6462 /* There is room already in this leaf page. */
6463 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6464 if (rc == 0 && insert_key) {
6465 /* Adjust other cursors pointing to mp */
6466 MDB_cursor *m2, *m3;
6467 MDB_dbi dbi = mc->mc_dbi;
6468 unsigned i = mc->mc_top;
6469 MDB_page *mp = mc->mc_pg[i];
6471 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6472 if (mc->mc_flags & C_SUB)
6473 m3 = &m2->mc_xcursor->mx_cursor;
6476 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6477 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6484 if (rc == MDB_SUCCESS) {
6485 /* Now store the actual data in the child DB. Note that we're
6486 * storing the user data in the keys field, so there are strict
6487 * size limits on dupdata. The actual data fields of the child
6488 * DB are all zero size.
6496 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6497 if (flags & MDB_CURRENT) {
6498 xflags = MDB_CURRENT|MDB_NOSPILL;
6500 mdb_xcursor_init1(mc, leaf);
6501 xflags = (flags & MDB_NODUPDATA) ?
6502 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6504 /* converted, write the original data first */
6506 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6510 /* Adjust other cursors pointing to mp */
6512 unsigned i = mc->mc_top;
6513 MDB_page *mp = mc->mc_pg[i];
6515 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6516 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6517 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6518 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6519 mdb_xcursor_init1(m2, leaf);
6523 /* we've done our job */
6526 ecount = mc->mc_xcursor->mx_db.md_entries;
6527 if (flags & MDB_APPENDDUP)
6528 xflags |= MDB_APPEND;
6529 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6530 if (flags & F_SUBDATA) {
6531 void *db = NODEDATA(leaf);
6532 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6534 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6536 /* Increment count unless we just replaced an existing item. */
6538 mc->mc_db->md_entries++;
6540 /* Invalidate txn if we created an empty sub-DB */
6543 /* If we succeeded and the key didn't exist before,
6544 * make sure the cursor is marked valid.
6546 mc->mc_flags |= C_INITIALIZED;
6548 if (flags & MDB_MULTIPLE) {
6551 /* let caller know how many succeeded, if any */
6552 data[1].mv_size = mcount;
6553 if (mcount < dcount) {
6554 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6555 insert_key = insert_data = 0;
6562 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6565 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6570 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6576 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6577 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6579 if (!(mc->mc_flags & C_INITIALIZED))
6582 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6583 return MDB_NOTFOUND;
6585 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6588 rc = mdb_cursor_touch(mc);
6592 mp = mc->mc_pg[mc->mc_top];
6595 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6597 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6598 if (flags & MDB_NODUPDATA) {
6599 /* mdb_cursor_del0() will subtract the final entry */
6600 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6602 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6603 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6605 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6608 /* If sub-DB still has entries, we're done */
6609 if (mc->mc_xcursor->mx_db.md_entries) {
6610 if (leaf->mn_flags & F_SUBDATA) {
6611 /* update subDB info */
6612 void *db = NODEDATA(leaf);
6613 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6616 /* shrink fake page */
6617 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6618 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6619 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6620 /* fix other sub-DB cursors pointed at this fake page */
6621 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6622 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6623 if (m2->mc_pg[mc->mc_top] == mp &&
6624 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6625 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6628 mc->mc_db->md_entries--;
6629 mc->mc_flags |= C_DEL;
6632 /* otherwise fall thru and delete the sub-DB */
6635 if (leaf->mn_flags & F_SUBDATA) {
6636 /* add all the child DB's pages to the free list */
6637 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6643 /* add overflow pages to free list */
6644 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6648 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6649 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6650 (rc = mdb_ovpage_free(mc, omp)))
6655 return mdb_cursor_del0(mc);
6658 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6662 /** Allocate and initialize new pages for a database.
6663 * @param[in] mc a cursor on the database being added to.
6664 * @param[in] flags flags defining what type of page is being allocated.
6665 * @param[in] num the number of pages to allocate. This is usually 1,
6666 * unless allocating overflow pages for a large record.
6667 * @param[out] mp Address of a page, or NULL on failure.
6668 * @return 0 on success, non-zero on failure.
6671 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6676 if ((rc = mdb_page_alloc(mc, num, &np)))
6678 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6679 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6680 np->mp_flags = flags | P_DIRTY;
6681 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6682 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6685 mc->mc_db->md_branch_pages++;
6686 else if (IS_LEAF(np))
6687 mc->mc_db->md_leaf_pages++;
6688 else if (IS_OVERFLOW(np)) {
6689 mc->mc_db->md_overflow_pages += num;
6697 /** Calculate the size of a leaf node.
6698 * The size depends on the environment's page size; if a data item
6699 * is too large it will be put onto an overflow page and the node
6700 * size will only include the key and not the data. Sizes are always
6701 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6702 * of the #MDB_node headers.
6703 * @param[in] env The environment handle.
6704 * @param[in] key The key for the node.
6705 * @param[in] data The data for the node.
6706 * @return The number of bytes needed to store the node.
6709 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6713 sz = LEAFSIZE(key, data);
6714 if (sz > env->me_nodemax) {
6715 /* put on overflow page */
6716 sz -= data->mv_size - sizeof(pgno_t);
6719 return EVEN(sz + sizeof(indx_t));
6722 /** Calculate the size of a branch node.
6723 * The size should depend on the environment's page size but since
6724 * we currently don't support spilling large keys onto overflow
6725 * pages, it's simply the size of the #MDB_node header plus the
6726 * size of the key. Sizes are always rounded up to an even number
6727 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6728 * @param[in] env The environment handle.
6729 * @param[in] key The key for the node.
6730 * @return The number of bytes needed to store the node.
6733 mdb_branch_size(MDB_env *env, MDB_val *key)
6738 if (sz > env->me_nodemax) {
6739 /* put on overflow page */
6740 /* not implemented */
6741 /* sz -= key->size - sizeof(pgno_t); */
6744 return sz + sizeof(indx_t);
6747 /** Add a node to the page pointed to by the cursor.
6748 * @param[in] mc The cursor for this operation.
6749 * @param[in] indx The index on the page where the new node should be added.
6750 * @param[in] key The key for the new node.
6751 * @param[in] data The data for the new node, if any.
6752 * @param[in] pgno The page number, if adding a branch node.
6753 * @param[in] flags Flags for the node.
6754 * @return 0 on success, non-zero on failure. Possible errors are:
6756 * <li>ENOMEM - failed to allocate overflow pages for the node.
6757 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6758 * should never happen since all callers already calculate the
6759 * page's free space before calling this function.
6763 mdb_node_add(MDB_cursor *mc, indx_t indx,
6764 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6767 size_t node_size = NODESIZE;
6771 MDB_page *mp = mc->mc_pg[mc->mc_top];
6772 MDB_page *ofp = NULL; /* overflow page */
6775 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6777 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6778 IS_LEAF(mp) ? "leaf" : "branch",
6779 IS_SUBP(mp) ? "sub-" : "",
6780 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6781 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6784 /* Move higher keys up one slot. */
6785 int ksize = mc->mc_db->md_pad, dif;
6786 char *ptr = LEAF2KEY(mp, indx, ksize);
6787 dif = NUMKEYS(mp) - indx;
6789 memmove(ptr+ksize, ptr, dif*ksize);
6790 /* insert new key */
6791 memcpy(ptr, key->mv_data, ksize);
6793 /* Just using these for counting */
6794 mp->mp_lower += sizeof(indx_t);
6795 mp->mp_upper -= ksize - sizeof(indx_t);
6799 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6801 node_size += key->mv_size;
6803 mdb_cassert(mc, data);
6804 if (F_ISSET(flags, F_BIGDATA)) {
6805 /* Data already on overflow page. */
6806 node_size += sizeof(pgno_t);
6807 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6808 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6810 /* Put data on overflow page. */
6811 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6812 data->mv_size, node_size+data->mv_size));
6813 node_size = EVEN(node_size + sizeof(pgno_t));
6814 if ((ssize_t)node_size > room)
6816 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6818 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6822 node_size += data->mv_size;
6825 node_size = EVEN(node_size);
6826 if ((ssize_t)node_size > room)
6830 /* Move higher pointers up one slot. */
6831 for (i = NUMKEYS(mp); i > indx; i--)
6832 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6834 /* Adjust free space offsets. */
6835 ofs = mp->mp_upper - node_size;
6836 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6837 mp->mp_ptrs[indx] = ofs;
6839 mp->mp_lower += sizeof(indx_t);
6841 /* Write the node data. */
6842 node = NODEPTR(mp, indx);
6843 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6844 node->mn_flags = flags;
6846 SETDSZ(node,data->mv_size);
6851 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6854 mdb_cassert(mc, key);
6856 if (F_ISSET(flags, F_BIGDATA))
6857 memcpy(node->mn_data + key->mv_size, data->mv_data,
6859 else if (F_ISSET(flags, MDB_RESERVE))
6860 data->mv_data = node->mn_data + key->mv_size;
6862 memcpy(node->mn_data + key->mv_size, data->mv_data,
6865 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6867 if (F_ISSET(flags, MDB_RESERVE))
6868 data->mv_data = METADATA(ofp);
6870 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6877 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6878 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6879 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6880 DPRINTF(("node size = %"Z"u", node_size));
6881 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6882 return MDB_PAGE_FULL;
6885 /** Delete the specified node from a page.
6886 * @param[in] mc Cursor pointing to the node to delete.
6887 * @param[in] ksize The size of a node. Only used if the page is
6888 * part of a #MDB_DUPFIXED database.
6891 mdb_node_del(MDB_cursor *mc, int ksize)
6893 MDB_page *mp = mc->mc_pg[mc->mc_top];
6894 indx_t indx = mc->mc_ki[mc->mc_top];
6896 indx_t i, j, numkeys, ptr;
6900 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6901 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6902 numkeys = NUMKEYS(mp);
6903 mdb_cassert(mc, indx < numkeys);
6906 int x = numkeys - 1 - indx;
6907 base = LEAF2KEY(mp, indx, ksize);
6909 memmove(base, base + ksize, x * ksize);
6910 mp->mp_lower -= sizeof(indx_t);
6911 mp->mp_upper += ksize - sizeof(indx_t);
6915 node = NODEPTR(mp, indx);
6916 sz = NODESIZE + node->mn_ksize;
6918 if (F_ISSET(node->mn_flags, F_BIGDATA))
6919 sz += sizeof(pgno_t);
6921 sz += NODEDSZ(node);
6925 ptr = mp->mp_ptrs[indx];
6926 for (i = j = 0; i < numkeys; i++) {
6928 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6929 if (mp->mp_ptrs[i] < ptr)
6930 mp->mp_ptrs[j] += sz;
6935 base = (char *)mp + mp->mp_upper + PAGEBASE;
6936 memmove(base + sz, base, ptr - mp->mp_upper);
6938 mp->mp_lower -= sizeof(indx_t);
6942 /** Compact the main page after deleting a node on a subpage.
6943 * @param[in] mp The main page to operate on.
6944 * @param[in] indx The index of the subpage on the main page.
6947 mdb_node_shrink(MDB_page *mp, indx_t indx)
6953 indx_t i, numkeys, ptr;
6955 node = NODEPTR(mp, indx);
6956 sp = (MDB_page *)NODEDATA(node);
6957 delta = SIZELEFT(sp);
6958 xp = (MDB_page *)((char *)sp + delta);
6960 /* shift subpage upward */
6962 nsize = NUMKEYS(sp) * sp->mp_pad;
6964 return; /* do not make the node uneven-sized */
6965 memmove(METADATA(xp), METADATA(sp), nsize);
6968 numkeys = NUMKEYS(sp);
6969 for (i=numkeys-1; i>=0; i--)
6970 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6972 xp->mp_upper = sp->mp_lower;
6973 xp->mp_lower = sp->mp_lower;
6974 xp->mp_flags = sp->mp_flags;
6975 xp->mp_pad = sp->mp_pad;
6976 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6978 nsize = NODEDSZ(node) - delta;
6979 SETDSZ(node, nsize);
6981 /* shift lower nodes upward */
6982 ptr = mp->mp_ptrs[indx];
6983 numkeys = NUMKEYS(mp);
6984 for (i = 0; i < numkeys; i++) {
6985 if (mp->mp_ptrs[i] <= ptr)
6986 mp->mp_ptrs[i] += delta;
6989 base = (char *)mp + mp->mp_upper + PAGEBASE;
6990 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6991 mp->mp_upper += delta;
6994 /** Initial setup of a sorted-dups cursor.
6995 * Sorted duplicates are implemented as a sub-database for the given key.
6996 * The duplicate data items are actually keys of the sub-database.
6997 * Operations on the duplicate data items are performed using a sub-cursor
6998 * initialized when the sub-database is first accessed. This function does
6999 * the preliminary setup of the sub-cursor, filling in the fields that
7000 * depend only on the parent DB.
7001 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7004 mdb_xcursor_init0(MDB_cursor *mc)
7006 MDB_xcursor *mx = mc->mc_xcursor;
7008 mx->mx_cursor.mc_xcursor = NULL;
7009 mx->mx_cursor.mc_txn = mc->mc_txn;
7010 mx->mx_cursor.mc_db = &mx->mx_db;
7011 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7012 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7013 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7014 mx->mx_cursor.mc_snum = 0;
7015 mx->mx_cursor.mc_top = 0;
7016 mx->mx_cursor.mc_flags = C_SUB;
7017 mx->mx_dbx.md_name.mv_size = 0;
7018 mx->mx_dbx.md_name.mv_data = NULL;
7019 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7020 mx->mx_dbx.md_dcmp = NULL;
7021 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7024 /** Final setup of a sorted-dups cursor.
7025 * Sets up the fields that depend on the data from the main cursor.
7026 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7027 * @param[in] node The data containing the #MDB_db record for the
7028 * sorted-dup database.
7031 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7033 MDB_xcursor *mx = mc->mc_xcursor;
7035 if (node->mn_flags & F_SUBDATA) {
7036 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7037 mx->mx_cursor.mc_pg[0] = 0;
7038 mx->mx_cursor.mc_snum = 0;
7039 mx->mx_cursor.mc_top = 0;
7040 mx->mx_cursor.mc_flags = C_SUB;
7042 MDB_page *fp = NODEDATA(node);
7043 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7044 mx->mx_db.md_flags = 0;
7045 mx->mx_db.md_depth = 1;
7046 mx->mx_db.md_branch_pages = 0;
7047 mx->mx_db.md_leaf_pages = 1;
7048 mx->mx_db.md_overflow_pages = 0;
7049 mx->mx_db.md_entries = NUMKEYS(fp);
7050 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7051 mx->mx_cursor.mc_snum = 1;
7052 mx->mx_cursor.mc_top = 0;
7053 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7054 mx->mx_cursor.mc_pg[0] = fp;
7055 mx->mx_cursor.mc_ki[0] = 0;
7056 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7057 mx->mx_db.md_flags = MDB_DUPFIXED;
7058 mx->mx_db.md_pad = fp->mp_pad;
7059 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7060 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7063 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7064 mx->mx_db.md_root));
7065 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7066 #if UINT_MAX < SIZE_MAX
7067 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7068 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7072 /** Initialize a cursor for a given transaction and database. */
7074 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7077 mc->mc_backup = NULL;
7080 mc->mc_db = &txn->mt_dbs[dbi];
7081 mc->mc_dbx = &txn->mt_dbxs[dbi];
7082 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7087 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7088 mdb_tassert(txn, mx != NULL);
7089 mc->mc_xcursor = mx;
7090 mdb_xcursor_init0(mc);
7092 mc->mc_xcursor = NULL;
7094 if (*mc->mc_dbflag & DB_STALE) {
7095 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7100 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7103 size_t size = sizeof(MDB_cursor);
7105 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7108 if (txn->mt_flags & MDB_TXN_ERROR)
7111 /* Allow read access to the freelist */
7112 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7115 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7116 size += sizeof(MDB_xcursor);
7118 if ((mc = malloc(size)) != NULL) {
7119 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7120 if (txn->mt_cursors) {
7121 mc->mc_next = txn->mt_cursors[dbi];
7122 txn->mt_cursors[dbi] = mc;
7123 mc->mc_flags |= C_UNTRACK;
7135 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7137 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7140 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7143 if (txn->mt_flags & MDB_TXN_ERROR)
7146 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7150 /* Return the count of duplicate data items for the current key */
7152 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7156 if (mc == NULL || countp == NULL)
7159 if (mc->mc_xcursor == NULL)
7160 return MDB_INCOMPATIBLE;
7162 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7165 if (!(mc->mc_flags & C_INITIALIZED))
7168 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7169 return MDB_NOTFOUND;
7171 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7172 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7175 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7178 *countp = mc->mc_xcursor->mx_db.md_entries;
7184 mdb_cursor_close(MDB_cursor *mc)
7186 if (mc && !mc->mc_backup) {
7187 /* remove from txn, if tracked */
7188 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7189 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7190 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7192 *prev = mc->mc_next;
7199 mdb_cursor_txn(MDB_cursor *mc)
7201 if (!mc) return NULL;
7206 mdb_cursor_dbi(MDB_cursor *mc)
7211 /** Replace the key for a branch node with a new key.
7212 * @param[in] mc Cursor pointing to the node to operate on.
7213 * @param[in] key The new key to use.
7214 * @return 0 on success, non-zero on failure.
7217 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7223 int delta, ksize, oksize;
7224 indx_t ptr, i, numkeys, indx;
7227 indx = mc->mc_ki[mc->mc_top];
7228 mp = mc->mc_pg[mc->mc_top];
7229 node = NODEPTR(mp, indx);
7230 ptr = mp->mp_ptrs[indx];
7234 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7235 k2.mv_data = NODEKEY(node);
7236 k2.mv_size = node->mn_ksize;
7237 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7239 mdb_dkey(&k2, kbuf2),
7245 /* Sizes must be 2-byte aligned. */
7246 ksize = EVEN(key->mv_size);
7247 oksize = EVEN(node->mn_ksize);
7248 delta = ksize - oksize;
7250 /* Shift node contents if EVEN(key length) changed. */
7252 if (delta > 0 && SIZELEFT(mp) < delta) {
7254 /* not enough space left, do a delete and split */
7255 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7256 pgno = NODEPGNO(node);
7257 mdb_node_del(mc, 0);
7258 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7261 numkeys = NUMKEYS(mp);
7262 for (i = 0; i < numkeys; i++) {
7263 if (mp->mp_ptrs[i] <= ptr)
7264 mp->mp_ptrs[i] -= delta;
7267 base = (char *)mp + mp->mp_upper + PAGEBASE;
7268 len = ptr - mp->mp_upper + NODESIZE;
7269 memmove(base - delta, base, len);
7270 mp->mp_upper -= delta;
7272 node = NODEPTR(mp, indx);
7275 /* But even if no shift was needed, update ksize */
7276 if (node->mn_ksize != key->mv_size)
7277 node->mn_ksize = key->mv_size;
7280 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7286 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7288 /** Move a node from csrc to cdst.
7291 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7298 unsigned short flags;
7302 /* Mark src and dst as dirty. */
7303 if ((rc = mdb_page_touch(csrc)) ||
7304 (rc = mdb_page_touch(cdst)))
7307 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7308 key.mv_size = csrc->mc_db->md_pad;
7309 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7311 data.mv_data = NULL;
7315 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7316 mdb_cassert(csrc, !((size_t)srcnode & 1));
7317 srcpg = NODEPGNO(srcnode);
7318 flags = srcnode->mn_flags;
7319 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7320 unsigned int snum = csrc->mc_snum;
7322 /* must find the lowest key below src */
7323 rc = mdb_page_search_lowest(csrc);
7326 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7327 key.mv_size = csrc->mc_db->md_pad;
7328 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7330 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7331 key.mv_size = NODEKSZ(s2);
7332 key.mv_data = NODEKEY(s2);
7334 csrc->mc_snum = snum--;
7335 csrc->mc_top = snum;
7337 key.mv_size = NODEKSZ(srcnode);
7338 key.mv_data = NODEKEY(srcnode);
7340 data.mv_size = NODEDSZ(srcnode);
7341 data.mv_data = NODEDATA(srcnode);
7343 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7344 unsigned int snum = cdst->mc_snum;
7347 /* must find the lowest key below dst */
7348 mdb_cursor_copy(cdst, &mn);
7349 rc = mdb_page_search_lowest(&mn);
7352 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7353 bkey.mv_size = mn.mc_db->md_pad;
7354 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7356 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7357 bkey.mv_size = NODEKSZ(s2);
7358 bkey.mv_data = NODEKEY(s2);
7360 mn.mc_snum = snum--;
7363 rc = mdb_update_key(&mn, &bkey);
7368 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7369 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7370 csrc->mc_ki[csrc->mc_top],
7372 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7373 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7375 /* Add the node to the destination page.
7377 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7378 if (rc != MDB_SUCCESS)
7381 /* Delete the node from the source page.
7383 mdb_node_del(csrc, key.mv_size);
7386 /* Adjust other cursors pointing to mp */
7387 MDB_cursor *m2, *m3;
7388 MDB_dbi dbi = csrc->mc_dbi;
7389 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7391 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7392 if (csrc->mc_flags & C_SUB)
7393 m3 = &m2->mc_xcursor->mx_cursor;
7396 if (m3 == csrc) continue;
7397 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7398 csrc->mc_ki[csrc->mc_top]) {
7399 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7400 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7405 /* Update the parent separators.
7407 if (csrc->mc_ki[csrc->mc_top] == 0) {
7408 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7409 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7410 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7412 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7413 key.mv_size = NODEKSZ(srcnode);
7414 key.mv_data = NODEKEY(srcnode);
7416 DPRINTF(("update separator for source page %"Z"u to [%s]",
7417 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7418 mdb_cursor_copy(csrc, &mn);
7421 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7424 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7426 indx_t ix = csrc->mc_ki[csrc->mc_top];
7427 nullkey.mv_size = 0;
7428 csrc->mc_ki[csrc->mc_top] = 0;
7429 rc = mdb_update_key(csrc, &nullkey);
7430 csrc->mc_ki[csrc->mc_top] = ix;
7431 mdb_cassert(csrc, rc == MDB_SUCCESS);
7435 if (cdst->mc_ki[cdst->mc_top] == 0) {
7436 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7437 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7438 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7440 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7441 key.mv_size = NODEKSZ(srcnode);
7442 key.mv_data = NODEKEY(srcnode);
7444 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7445 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7446 mdb_cursor_copy(cdst, &mn);
7449 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7452 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7454 indx_t ix = cdst->mc_ki[cdst->mc_top];
7455 nullkey.mv_size = 0;
7456 cdst->mc_ki[cdst->mc_top] = 0;
7457 rc = mdb_update_key(cdst, &nullkey);
7458 cdst->mc_ki[cdst->mc_top] = ix;
7459 mdb_cassert(csrc, rc == MDB_SUCCESS);
7466 /** Merge one page into another.
7467 * The nodes from the page pointed to by \b csrc will
7468 * be copied to the page pointed to by \b cdst and then
7469 * the \b csrc page will be freed.
7470 * @param[in] csrc Cursor pointing to the source page.
7471 * @param[in] cdst Cursor pointing to the destination page.
7472 * @return 0 on success, non-zero on failure.
7475 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7477 MDB_page *psrc, *pdst;
7484 psrc = csrc->mc_pg[csrc->mc_top];
7485 pdst = cdst->mc_pg[cdst->mc_top];
7487 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7489 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7490 mdb_cassert(csrc, cdst->mc_snum > 1);
7492 /* Mark dst as dirty. */
7493 if ((rc = mdb_page_touch(cdst)))
7496 /* Move all nodes from src to dst.
7498 j = nkeys = NUMKEYS(pdst);
7499 if (IS_LEAF2(psrc)) {
7500 key.mv_size = csrc->mc_db->md_pad;
7501 key.mv_data = METADATA(psrc);
7502 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7503 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7504 if (rc != MDB_SUCCESS)
7506 key.mv_data = (char *)key.mv_data + key.mv_size;
7509 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7510 srcnode = NODEPTR(psrc, i);
7511 if (i == 0 && IS_BRANCH(psrc)) {
7514 mdb_cursor_copy(csrc, &mn);
7515 /* must find the lowest key below src */
7516 rc = mdb_page_search_lowest(&mn);
7519 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7520 key.mv_size = mn.mc_db->md_pad;
7521 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7523 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7524 key.mv_size = NODEKSZ(s2);
7525 key.mv_data = NODEKEY(s2);
7528 key.mv_size = srcnode->mn_ksize;
7529 key.mv_data = NODEKEY(srcnode);
7532 data.mv_size = NODEDSZ(srcnode);
7533 data.mv_data = NODEDATA(srcnode);
7534 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7535 if (rc != MDB_SUCCESS)
7540 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7541 pdst->mp_pgno, NUMKEYS(pdst),
7542 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7544 /* Unlink the src page from parent and add to free list.
7547 mdb_node_del(csrc, 0);
7548 if (csrc->mc_ki[csrc->mc_top] == 0) {
7550 rc = mdb_update_key(csrc, &key);
7558 psrc = csrc->mc_pg[csrc->mc_top];
7559 /* If not operating on FreeDB, allow this page to be reused
7560 * in this txn. Otherwise just add to free list.
7562 rc = mdb_page_loose(csrc, psrc);
7566 csrc->mc_db->md_leaf_pages--;
7568 csrc->mc_db->md_branch_pages--;
7570 /* Adjust other cursors pointing to mp */
7571 MDB_cursor *m2, *m3;
7572 MDB_dbi dbi = csrc->mc_dbi;
7574 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7575 if (csrc->mc_flags & C_SUB)
7576 m3 = &m2->mc_xcursor->mx_cursor;
7579 if (m3 == csrc) continue;
7580 if (m3->mc_snum < csrc->mc_snum) continue;
7581 if (m3->mc_pg[csrc->mc_top] == psrc) {
7582 m3->mc_pg[csrc->mc_top] = pdst;
7583 m3->mc_ki[csrc->mc_top] += nkeys;
7588 unsigned int snum = cdst->mc_snum;
7589 uint16_t depth = cdst->mc_db->md_depth;
7590 mdb_cursor_pop(cdst);
7591 rc = mdb_rebalance(cdst);
7592 /* Did the tree shrink? */
7593 if (depth > cdst->mc_db->md_depth)
7595 cdst->mc_snum = snum;
7596 cdst->mc_top = snum-1;
7601 /** Copy the contents of a cursor.
7602 * @param[in] csrc The cursor to copy from.
7603 * @param[out] cdst The cursor to copy to.
7606 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7610 cdst->mc_txn = csrc->mc_txn;
7611 cdst->mc_dbi = csrc->mc_dbi;
7612 cdst->mc_db = csrc->mc_db;
7613 cdst->mc_dbx = csrc->mc_dbx;
7614 cdst->mc_snum = csrc->mc_snum;
7615 cdst->mc_top = csrc->mc_top;
7616 cdst->mc_flags = csrc->mc_flags;
7618 for (i=0; i<csrc->mc_snum; i++) {
7619 cdst->mc_pg[i] = csrc->mc_pg[i];
7620 cdst->mc_ki[i] = csrc->mc_ki[i];
7624 /** Rebalance the tree after a delete operation.
7625 * @param[in] mc Cursor pointing to the page where rebalancing
7627 * @return 0 on success, non-zero on failure.
7630 mdb_rebalance(MDB_cursor *mc)
7634 unsigned int ptop, minkeys;
7638 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7639 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7640 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7641 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7642 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7644 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7645 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7646 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7647 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7651 if (mc->mc_snum < 2) {
7652 MDB_page *mp = mc->mc_pg[0];
7654 DPUTS("Can't rebalance a subpage, ignoring");
7657 if (NUMKEYS(mp) == 0) {
7658 DPUTS("tree is completely empty");
7659 mc->mc_db->md_root = P_INVALID;
7660 mc->mc_db->md_depth = 0;
7661 mc->mc_db->md_leaf_pages = 0;
7662 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7665 /* Adjust cursors pointing to mp */
7668 mc->mc_flags &= ~C_INITIALIZED;
7670 MDB_cursor *m2, *m3;
7671 MDB_dbi dbi = mc->mc_dbi;
7673 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7674 if (mc->mc_flags & C_SUB)
7675 m3 = &m2->mc_xcursor->mx_cursor;
7678 if (m3->mc_snum < mc->mc_snum) continue;
7679 if (m3->mc_pg[0] == mp) {
7682 m3->mc_flags &= ~C_INITIALIZED;
7686 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7688 DPUTS("collapsing root page!");
7689 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7692 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7693 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7696 mc->mc_db->md_depth--;
7697 mc->mc_db->md_branch_pages--;
7698 mc->mc_ki[0] = mc->mc_ki[1];
7699 for (i = 1; i<mc->mc_db->md_depth; i++) {
7700 mc->mc_pg[i] = mc->mc_pg[i+1];
7701 mc->mc_ki[i] = mc->mc_ki[i+1];
7704 /* Adjust other cursors pointing to mp */
7705 MDB_cursor *m2, *m3;
7706 MDB_dbi dbi = mc->mc_dbi;
7708 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7709 if (mc->mc_flags & C_SUB)
7710 m3 = &m2->mc_xcursor->mx_cursor;
7713 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7714 if (m3->mc_pg[0] == mp) {
7717 for (i=0; i<m3->mc_snum; i++) {
7718 m3->mc_pg[i] = m3->mc_pg[i+1];
7719 m3->mc_ki[i] = m3->mc_ki[i+1];
7725 DPUTS("root page doesn't need rebalancing");
7729 /* The parent (branch page) must have at least 2 pointers,
7730 * otherwise the tree is invalid.
7732 ptop = mc->mc_top-1;
7733 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7735 /* Leaf page fill factor is below the threshold.
7736 * Try to move keys from left or right neighbor, or
7737 * merge with a neighbor page.
7742 mdb_cursor_copy(mc, &mn);
7743 mn.mc_xcursor = NULL;
7745 oldki = mc->mc_ki[mc->mc_top];
7746 if (mc->mc_ki[ptop] == 0) {
7747 /* We're the leftmost leaf in our parent.
7749 DPUTS("reading right neighbor");
7751 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7752 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7755 mn.mc_ki[mn.mc_top] = 0;
7756 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7758 /* There is at least one neighbor to the left.
7760 DPUTS("reading left neighbor");
7762 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7763 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7766 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7767 mc->mc_ki[mc->mc_top] = 0;
7770 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7771 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7772 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7774 /* If the neighbor page is above threshold and has enough keys,
7775 * move one key from it. Otherwise we should try to merge them.
7776 * (A branch page must never have less than 2 keys.)
7778 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7779 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7780 rc = mdb_node_move(&mn, mc);
7781 if (mc->mc_ki[ptop]) {
7785 if (mc->mc_ki[ptop] == 0) {
7786 rc = mdb_page_merge(&mn, mc);
7788 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7789 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7790 rc = mdb_page_merge(mc, &mn);
7791 mdb_cursor_copy(&mn, mc);
7793 mc->mc_flags &= ~C_EOF;
7795 mc->mc_ki[mc->mc_top] = oldki;
7799 /** Complete a delete operation started by #mdb_cursor_del(). */
7801 mdb_cursor_del0(MDB_cursor *mc)
7808 ki = mc->mc_ki[mc->mc_top];
7809 mdb_node_del(mc, mc->mc_db->md_pad);
7810 mc->mc_db->md_entries--;
7811 rc = mdb_rebalance(mc);
7813 if (rc == MDB_SUCCESS) {
7814 MDB_cursor *m2, *m3;
7815 MDB_dbi dbi = mc->mc_dbi;
7817 mp = mc->mc_pg[mc->mc_top];
7818 nkeys = NUMKEYS(mp);
7820 /* if mc points past last node in page, find next sibling */
7821 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7822 rc = mdb_cursor_sibling(mc, 1);
7823 if (rc == MDB_NOTFOUND) {
7824 mc->mc_flags |= C_EOF;
7829 /* Adjust other cursors pointing to mp */
7830 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7831 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7832 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7834 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7836 if (m3->mc_pg[mc->mc_top] == mp) {
7837 if (m3->mc_ki[mc->mc_top] >= ki) {
7838 m3->mc_flags |= C_DEL;
7839 if (m3->mc_ki[mc->mc_top] > ki)
7840 m3->mc_ki[mc->mc_top]--;
7841 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7842 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7844 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7845 rc = mdb_cursor_sibling(m3, 1);
7846 if (rc == MDB_NOTFOUND) {
7847 m3->mc_flags |= C_EOF;
7853 mc->mc_flags |= C_DEL;
7857 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7862 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7863 MDB_val *key, MDB_val *data)
7865 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7868 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7869 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7871 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7872 /* must ignore any data */
7876 return mdb_del0(txn, dbi, key, data, 0);
7880 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7881 MDB_val *key, MDB_val *data, unsigned flags)
7886 MDB_val rdata, *xdata;
7890 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7892 mdb_cursor_init(&mc, txn, dbi, &mx);
7901 flags |= MDB_NODUPDATA;
7903 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7905 /* let mdb_page_split know about this cursor if needed:
7906 * delete will trigger a rebalance; if it needs to move
7907 * a node from one page to another, it will have to
7908 * update the parent's separator key(s). If the new sepkey
7909 * is larger than the current one, the parent page may
7910 * run out of space, triggering a split. We need this
7911 * cursor to be consistent until the end of the rebalance.
7913 mc.mc_flags |= C_UNTRACK;
7914 mc.mc_next = txn->mt_cursors[dbi];
7915 txn->mt_cursors[dbi] = &mc;
7916 rc = mdb_cursor_del(&mc, flags);
7917 txn->mt_cursors[dbi] = mc.mc_next;
7922 /** Split a page and insert a new node.
7923 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7924 * The cursor will be updated to point to the actual page and index where
7925 * the node got inserted after the split.
7926 * @param[in] newkey The key for the newly inserted node.
7927 * @param[in] newdata The data for the newly inserted node.
7928 * @param[in] newpgno The page number, if the new node is a branch node.
7929 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7930 * @return 0 on success, non-zero on failure.
7933 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7934 unsigned int nflags)
7937 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7940 int i, j, split_indx, nkeys, pmax;
7941 MDB_env *env = mc->mc_txn->mt_env;
7943 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7944 MDB_page *copy = NULL;
7945 MDB_page *mp, *rp, *pp;
7950 mp = mc->mc_pg[mc->mc_top];
7951 newindx = mc->mc_ki[mc->mc_top];
7952 nkeys = NUMKEYS(mp);
7954 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7955 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7956 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7958 /* Create a right sibling. */
7959 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7961 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7963 if (mc->mc_snum < 2) {
7964 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7966 /* shift current top to make room for new parent */
7967 mc->mc_pg[1] = mc->mc_pg[0];
7968 mc->mc_ki[1] = mc->mc_ki[0];
7971 mc->mc_db->md_root = pp->mp_pgno;
7972 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7973 mc->mc_db->md_depth++;
7976 /* Add left (implicit) pointer. */
7977 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7978 /* undo the pre-push */
7979 mc->mc_pg[0] = mc->mc_pg[1];
7980 mc->mc_ki[0] = mc->mc_ki[1];
7981 mc->mc_db->md_root = mp->mp_pgno;
7982 mc->mc_db->md_depth--;
7989 ptop = mc->mc_top-1;
7990 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7993 mc->mc_flags |= C_SPLITTING;
7994 mdb_cursor_copy(mc, &mn);
7995 mn.mc_pg[mn.mc_top] = rp;
7996 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7998 if (nflags & MDB_APPEND) {
7999 mn.mc_ki[mn.mc_top] = 0;
8001 split_indx = newindx;
8005 split_indx = (nkeys+1) / 2;
8010 unsigned int lsize, rsize, ksize;
8011 /* Move half of the keys to the right sibling */
8012 x = mc->mc_ki[mc->mc_top] - split_indx;
8013 ksize = mc->mc_db->md_pad;
8014 split = LEAF2KEY(mp, split_indx, ksize);
8015 rsize = (nkeys - split_indx) * ksize;
8016 lsize = (nkeys - split_indx) * sizeof(indx_t);
8017 mp->mp_lower -= lsize;
8018 rp->mp_lower += lsize;
8019 mp->mp_upper += rsize - lsize;
8020 rp->mp_upper -= rsize - lsize;
8021 sepkey.mv_size = ksize;
8022 if (newindx == split_indx) {
8023 sepkey.mv_data = newkey->mv_data;
8025 sepkey.mv_data = split;
8028 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8029 memcpy(rp->mp_ptrs, split, rsize);
8030 sepkey.mv_data = rp->mp_ptrs;
8031 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8032 memcpy(ins, newkey->mv_data, ksize);
8033 mp->mp_lower += sizeof(indx_t);
8034 mp->mp_upper -= ksize - sizeof(indx_t);
8037 memcpy(rp->mp_ptrs, split, x * ksize);
8038 ins = LEAF2KEY(rp, x, ksize);
8039 memcpy(ins, newkey->mv_data, ksize);
8040 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8041 rp->mp_lower += sizeof(indx_t);
8042 rp->mp_upper -= ksize - sizeof(indx_t);
8043 mc->mc_ki[mc->mc_top] = x;
8044 mc->mc_pg[mc->mc_top] = rp;
8047 int psize, nsize, k;
8048 /* Maximum free space in an empty page */
8049 pmax = env->me_psize - PAGEHDRSZ;
8051 nsize = mdb_leaf_size(env, newkey, newdata);
8053 nsize = mdb_branch_size(env, newkey);
8054 nsize = EVEN(nsize);
8056 /* grab a page to hold a temporary copy */
8057 copy = mdb_page_malloc(mc->mc_txn, 1);
8062 copy->mp_pgno = mp->mp_pgno;
8063 copy->mp_flags = mp->mp_flags;
8064 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8065 copy->mp_upper = env->me_psize - PAGEBASE;
8067 /* prepare to insert */
8068 for (i=0, j=0; i<nkeys; i++) {
8070 copy->mp_ptrs[j++] = 0;
8072 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8075 /* When items are relatively large the split point needs
8076 * to be checked, because being off-by-one will make the
8077 * difference between success or failure in mdb_node_add.
8079 * It's also relevant if a page happens to be laid out
8080 * such that one half of its nodes are all "small" and
8081 * the other half of its nodes are "large." If the new
8082 * item is also "large" and falls on the half with
8083 * "large" nodes, it also may not fit.
8085 * As a final tweak, if the new item goes on the last
8086 * spot on the page (and thus, onto the new page), bias
8087 * the split so the new page is emptier than the old page.
8088 * This yields better packing during sequential inserts.
8090 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8091 /* Find split point */
8093 if (newindx <= split_indx || newindx >= nkeys) {
8095 k = newindx >= nkeys ? nkeys : split_indx+2;
8100 for (; i!=k; i+=j) {
8105 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8106 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8108 if (F_ISSET(node->mn_flags, F_BIGDATA))
8109 psize += sizeof(pgno_t);
8111 psize += NODEDSZ(node);
8113 psize = EVEN(psize);
8115 if (psize > pmax || i == k-j) {
8116 split_indx = i + (j<0);
8121 if (split_indx == newindx) {
8122 sepkey.mv_size = newkey->mv_size;
8123 sepkey.mv_data = newkey->mv_data;
8125 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8126 sepkey.mv_size = node->mn_ksize;
8127 sepkey.mv_data = NODEKEY(node);
8132 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8134 /* Copy separator key to the parent.
8136 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8140 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8145 if (mn.mc_snum == mc->mc_snum) {
8146 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8147 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8148 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8149 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8154 /* Right page might now have changed parent.
8155 * Check if left page also changed parent.
8157 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8158 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8159 for (i=0; i<ptop; i++) {
8160 mc->mc_pg[i] = mn.mc_pg[i];
8161 mc->mc_ki[i] = mn.mc_ki[i];
8163 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8164 if (mn.mc_ki[ptop]) {
8165 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8167 /* find right page's left sibling */
8168 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8169 mdb_cursor_sibling(mc, 0);
8174 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8177 mc->mc_flags ^= C_SPLITTING;
8178 if (rc != MDB_SUCCESS) {
8181 if (nflags & MDB_APPEND) {
8182 mc->mc_pg[mc->mc_top] = rp;
8183 mc->mc_ki[mc->mc_top] = 0;
8184 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8187 for (i=0; i<mc->mc_top; i++)
8188 mc->mc_ki[i] = mn.mc_ki[i];
8189 } else if (!IS_LEAF2(mp)) {
8191 mc->mc_pg[mc->mc_top] = rp;
8196 rkey.mv_data = newkey->mv_data;
8197 rkey.mv_size = newkey->mv_size;
8203 /* Update index for the new key. */
8204 mc->mc_ki[mc->mc_top] = j;
8206 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8207 rkey.mv_data = NODEKEY(node);
8208 rkey.mv_size = node->mn_ksize;
8210 xdata.mv_data = NODEDATA(node);
8211 xdata.mv_size = NODEDSZ(node);
8214 pgno = NODEPGNO(node);
8215 flags = node->mn_flags;
8218 if (!IS_LEAF(mp) && j == 0) {
8219 /* First branch index doesn't need key data. */
8223 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8229 mc->mc_pg[mc->mc_top] = copy;
8234 } while (i != split_indx);
8236 nkeys = NUMKEYS(copy);
8237 for (i=0; i<nkeys; i++)
8238 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8239 mp->mp_lower = copy->mp_lower;
8240 mp->mp_upper = copy->mp_upper;
8241 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8242 env->me_psize - copy->mp_upper - PAGEBASE);
8244 /* reset back to original page */
8245 if (newindx < split_indx) {
8246 mc->mc_pg[mc->mc_top] = mp;
8247 if (nflags & MDB_RESERVE) {
8248 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8249 if (!(node->mn_flags & F_BIGDATA))
8250 newdata->mv_data = NODEDATA(node);
8253 mc->mc_pg[mc->mc_top] = rp;
8255 /* Make sure mc_ki is still valid.
8257 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8258 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8259 for (i=0; i<=ptop; i++) {
8260 mc->mc_pg[i] = mn.mc_pg[i];
8261 mc->mc_ki[i] = mn.mc_ki[i];
8268 /* Adjust other cursors pointing to mp */
8269 MDB_cursor *m2, *m3;
8270 MDB_dbi dbi = mc->mc_dbi;
8271 int fixup = NUMKEYS(mp);
8273 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8274 if (mc->mc_flags & C_SUB)
8275 m3 = &m2->mc_xcursor->mx_cursor;
8280 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8282 if (m3->mc_flags & C_SPLITTING)
8287 for (k=m3->mc_top; k>=0; k--) {
8288 m3->mc_ki[k+1] = m3->mc_ki[k];
8289 m3->mc_pg[k+1] = m3->mc_pg[k];
8291 if (m3->mc_ki[0] >= split_indx) {
8296 m3->mc_pg[0] = mc->mc_pg[0];
8300 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8301 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8302 m3->mc_ki[mc->mc_top]++;
8303 if (m3->mc_ki[mc->mc_top] >= fixup) {
8304 m3->mc_pg[mc->mc_top] = rp;
8305 m3->mc_ki[mc->mc_top] -= fixup;
8306 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8308 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8309 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8314 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8317 if (copy) /* tmp page */
8318 mdb_page_free(env, copy);
8320 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8325 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8326 MDB_val *key, MDB_val *data, unsigned int flags)
8331 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8334 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8337 mdb_cursor_init(&mc, txn, dbi, &mx);
8338 return mdb_cursor_put(&mc, key, data, flags);
8342 #define MDB_WBUF (1024*1024)
8345 /** State needed for a compacting copy. */
8346 typedef struct mdb_copy {
8347 pthread_mutex_t mc_mutex;
8348 pthread_cond_t mc_cond;
8355 pgno_t mc_next_pgno;
8358 volatile int mc_new;
8363 /** Dedicated writer thread for compacting copy. */
8364 static THREAD_RET ESECT
8365 mdb_env_copythr(void *arg)
8369 int toggle = 0, wsize, rc;
8372 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8375 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8378 pthread_mutex_lock(&my->mc_mutex);
8380 pthread_cond_signal(&my->mc_cond);
8383 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8384 if (my->mc_new < 0) {
8389 wsize = my->mc_wlen[toggle];
8390 ptr = my->mc_wbuf[toggle];
8393 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8397 } else if (len > 0) {
8411 /* If there's an overflow page tail, write it too */
8412 if (my->mc_olen[toggle]) {
8413 wsize = my->mc_olen[toggle];
8414 ptr = my->mc_over[toggle];
8415 my->mc_olen[toggle] = 0;
8418 my->mc_wlen[toggle] = 0;
8420 pthread_cond_signal(&my->mc_cond);
8422 pthread_cond_signal(&my->mc_cond);
8423 pthread_mutex_unlock(&my->mc_mutex);
8424 return (THREAD_RET)0;
8428 /** Tell the writer thread there's a buffer ready to write */
8430 mdb_env_cthr_toggle(mdb_copy *my, int st)
8432 int toggle = my->mc_toggle ^ 1;
8433 pthread_mutex_lock(&my->mc_mutex);
8434 if (my->mc_status) {
8435 pthread_mutex_unlock(&my->mc_mutex);
8436 return my->mc_status;
8438 while (my->mc_new == 1)
8439 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8441 my->mc_toggle = toggle;
8442 pthread_cond_signal(&my->mc_cond);
8443 pthread_mutex_unlock(&my->mc_mutex);
8447 /** Depth-first tree traversal for compacting copy. */
8449 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8452 MDB_txn *txn = my->mc_txn;
8454 MDB_page *mo, *mp, *leaf;
8459 /* Empty DB, nothing to do */
8460 if (*pg == P_INVALID)
8467 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8470 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8474 /* Make cursor pages writable */
8475 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8479 for (i=0; i<mc.mc_top; i++) {
8480 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8481 mc.mc_pg[i] = (MDB_page *)ptr;
8482 ptr += my->mc_env->me_psize;
8485 /* This is writable space for a leaf page. Usually not needed. */
8486 leaf = (MDB_page *)ptr;
8488 toggle = my->mc_toggle;
8489 while (mc.mc_snum > 0) {
8491 mp = mc.mc_pg[mc.mc_top];
8495 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8496 for (i=0; i<n; i++) {
8497 ni = NODEPTR(mp, i);
8498 if (ni->mn_flags & F_BIGDATA) {
8502 /* Need writable leaf */
8504 mc.mc_pg[mc.mc_top] = leaf;
8505 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8507 ni = NODEPTR(mp, i);
8510 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8511 rc = mdb_page_get(txn, pg, &omp, NULL);
8514 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8515 rc = mdb_env_cthr_toggle(my, 1);
8518 toggle = my->mc_toggle;
8520 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8521 memcpy(mo, omp, my->mc_env->me_psize);
8522 mo->mp_pgno = my->mc_next_pgno;
8523 my->mc_next_pgno += omp->mp_pages;
8524 my->mc_wlen[toggle] += my->mc_env->me_psize;
8525 if (omp->mp_pages > 1) {
8526 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8527 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8528 rc = mdb_env_cthr_toggle(my, 1);
8531 toggle = my->mc_toggle;
8533 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8534 } else if (ni->mn_flags & F_SUBDATA) {
8537 /* Need writable leaf */
8539 mc.mc_pg[mc.mc_top] = leaf;
8540 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8542 ni = NODEPTR(mp, i);
8545 memcpy(&db, NODEDATA(ni), sizeof(db));
8546 my->mc_toggle = toggle;
8547 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8550 toggle = my->mc_toggle;
8551 memcpy(NODEDATA(ni), &db, sizeof(db));
8556 mc.mc_ki[mc.mc_top]++;
8557 if (mc.mc_ki[mc.mc_top] < n) {
8560 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8562 rc = mdb_page_get(txn, pg, &mp, NULL);
8567 mc.mc_ki[mc.mc_top] = 0;
8568 if (IS_BRANCH(mp)) {
8569 /* Whenever we advance to a sibling branch page,
8570 * we must proceed all the way down to its first leaf.
8572 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8575 mc.mc_pg[mc.mc_top] = mp;
8579 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8580 rc = mdb_env_cthr_toggle(my, 1);
8583 toggle = my->mc_toggle;
8585 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8586 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8587 mo->mp_pgno = my->mc_next_pgno++;
8588 my->mc_wlen[toggle] += my->mc_env->me_psize;
8590 /* Update parent if there is one */
8591 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8592 SETPGNO(ni, mo->mp_pgno);
8593 mdb_cursor_pop(&mc);
8595 /* Otherwise we're done */
8605 /** Copy environment with compaction. */
8607 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8612 MDB_txn *txn = NULL;
8617 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8618 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8619 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8620 if (my.mc_wbuf[0] == NULL)
8623 pthread_mutex_init(&my.mc_mutex, NULL);
8624 pthread_cond_init(&my.mc_cond, NULL);
8625 #ifdef HAVE_MEMALIGN
8626 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8627 if (my.mc_wbuf[0] == NULL)
8630 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8635 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8636 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8641 my.mc_next_pgno = 2;
8647 THREAD_CREATE(thr, mdb_env_copythr, &my);
8649 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8653 mp = (MDB_page *)my.mc_wbuf[0];
8654 memset(mp, 0, 2*env->me_psize);
8656 mp->mp_flags = P_META;
8657 mm = (MDB_meta *)METADATA(mp);
8658 mdb_env_init_meta0(env, mm);
8659 mm->mm_address = env->me_metas[0]->mm_address;
8661 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8663 mp->mp_flags = P_META;
8664 *(MDB_meta *)METADATA(mp) = *mm;
8665 mm = (MDB_meta *)METADATA(mp);
8667 /* Count the number of free pages, subtract from lastpg to find
8668 * number of active pages
8671 MDB_ID freecount = 0;
8674 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8675 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8676 freecount += *(MDB_ID *)data.mv_data;
8677 freecount += txn->mt_dbs[0].md_branch_pages +
8678 txn->mt_dbs[0].md_leaf_pages +
8679 txn->mt_dbs[0].md_overflow_pages;
8681 /* Set metapage 1 */
8682 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8683 mm->mm_dbs[1] = txn->mt_dbs[1];
8684 if (mm->mm_last_pg > 1) {
8685 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8688 mm->mm_dbs[1].md_root = P_INVALID;
8691 my.mc_wlen[0] = env->me_psize * 2;
8693 pthread_mutex_lock(&my.mc_mutex);
8695 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8696 pthread_mutex_unlock(&my.mc_mutex);
8697 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8698 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8699 rc = mdb_env_cthr_toggle(&my, 1);
8700 mdb_env_cthr_toggle(&my, -1);
8701 pthread_mutex_lock(&my.mc_mutex);
8703 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8704 pthread_mutex_unlock(&my.mc_mutex);
8709 CloseHandle(my.mc_cond);
8710 CloseHandle(my.mc_mutex);
8711 _aligned_free(my.mc_wbuf[0]);
8713 pthread_cond_destroy(&my.mc_cond);
8714 pthread_mutex_destroy(&my.mc_mutex);
8715 free(my.mc_wbuf[0]);
8720 /** Copy environment as-is. */
8722 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8724 MDB_txn *txn = NULL;
8730 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8734 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8737 /* Do the lock/unlock of the reader mutex before starting the
8738 * write txn. Otherwise other read txns could block writers.
8740 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8745 /* We must start the actual read txn after blocking writers */
8746 mdb_txn_reset0(txn, "reset-stage1");
8748 /* Temporarily block writers until we snapshot the meta pages */
8751 rc = mdb_txn_renew0(txn);
8753 UNLOCK_MUTEX_W(env);
8758 wsize = env->me_psize * 2;
8762 DO_WRITE(rc, fd, ptr, w2, len);
8766 } else if (len > 0) {
8772 /* Non-blocking or async handles are not supported */
8778 UNLOCK_MUTEX_W(env);
8783 w2 = txn->mt_next_pgno * env->me_psize;
8786 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8793 if (wsize > MAX_WRITE)
8797 DO_WRITE(rc, fd, ptr, w2, len);
8801 } else if (len > 0) {
8818 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8820 if (flags & MDB_CP_COMPACT)
8821 return mdb_env_copyfd1(env, fd);
8823 return mdb_env_copyfd0(env, fd);
8827 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8829 return mdb_env_copyfd2(env, fd, 0);
8833 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8837 HANDLE newfd = INVALID_HANDLE_VALUE;
8839 if (env->me_flags & MDB_NOSUBDIR) {
8840 lpath = (char *)path;
8843 len += sizeof(DATANAME);
8844 lpath = malloc(len);
8847 sprintf(lpath, "%s" DATANAME, path);
8850 /* The destination path must exist, but the destination file must not.
8851 * We don't want the OS to cache the writes, since the source data is
8852 * already in the OS cache.
8855 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8856 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8858 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8860 if (newfd == INVALID_HANDLE_VALUE) {
8865 if (env->me_psize >= env->me_os_psize) {
8867 /* Set O_DIRECT if the file system supports it */
8868 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8869 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8871 #ifdef F_NOCACHE /* __APPLE__ */
8872 rc = fcntl(newfd, F_NOCACHE, 1);
8880 rc = mdb_env_copyfd2(env, newfd, flags);
8883 if (!(env->me_flags & MDB_NOSUBDIR))
8885 if (newfd != INVALID_HANDLE_VALUE)
8886 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8893 mdb_env_copy(MDB_env *env, const char *path)
8895 return mdb_env_copy2(env, path, 0);
8899 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8901 if ((flag & CHANGEABLE) != flag)
8904 env->me_flags |= flag;
8906 env->me_flags &= ~flag;
8911 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8916 *arg = env->me_flags;
8921 mdb_env_set_userctx(MDB_env *env, void *ctx)
8925 env->me_userctx = ctx;
8930 mdb_env_get_userctx(MDB_env *env)
8932 return env ? env->me_userctx : NULL;
8936 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8941 env->me_assert_func = func;
8947 mdb_env_get_path(MDB_env *env, const char **arg)
8952 *arg = env->me_path;
8957 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8966 /** Common code for #mdb_stat() and #mdb_env_stat().
8967 * @param[in] env the environment to operate in.
8968 * @param[in] db the #MDB_db record containing the stats to return.
8969 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8970 * @return 0, this function always succeeds.
8973 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8975 arg->ms_psize = env->me_psize;
8976 arg->ms_depth = db->md_depth;
8977 arg->ms_branch_pages = db->md_branch_pages;
8978 arg->ms_leaf_pages = db->md_leaf_pages;
8979 arg->ms_overflow_pages = db->md_overflow_pages;
8980 arg->ms_entries = db->md_entries;
8986 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8990 if (env == NULL || arg == NULL)
8993 toggle = mdb_env_pick_meta(env);
8995 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8999 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9003 if (env == NULL || arg == NULL)
9006 toggle = mdb_env_pick_meta(env);
9007 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9008 arg->me_mapsize = env->me_mapsize;
9009 arg->me_maxreaders = env->me_maxreaders;
9011 /* me_numreaders may be zero if this process never used any readers. Use
9012 * the shared numreader count if it exists.
9014 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
9016 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9017 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9021 /** Set the default comparison functions for a database.
9022 * Called immediately after a database is opened to set the defaults.
9023 * The user can then override them with #mdb_set_compare() or
9024 * #mdb_set_dupsort().
9025 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9026 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9029 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9031 uint16_t f = txn->mt_dbs[dbi].md_flags;
9033 txn->mt_dbxs[dbi].md_cmp =
9034 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9035 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9037 txn->mt_dbxs[dbi].md_dcmp =
9038 !(f & MDB_DUPSORT) ? 0 :
9039 ((f & MDB_INTEGERDUP)
9040 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9041 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9044 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9050 int rc, dbflag, exact;
9051 unsigned int unused = 0, seq;
9054 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9055 mdb_default_cmp(txn, FREE_DBI);
9058 if ((flags & VALID_FLAGS) != flags)
9060 if (txn->mt_flags & MDB_TXN_ERROR)
9066 if (flags & PERSISTENT_FLAGS) {
9067 uint16_t f2 = flags & PERSISTENT_FLAGS;
9068 /* make sure flag changes get committed */
9069 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9070 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9071 txn->mt_flags |= MDB_TXN_DIRTY;
9074 mdb_default_cmp(txn, MAIN_DBI);
9078 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9079 mdb_default_cmp(txn, MAIN_DBI);
9082 /* Is the DB already open? */
9084 for (i=2; i<txn->mt_numdbs; i++) {
9085 if (!txn->mt_dbxs[i].md_name.mv_size) {
9086 /* Remember this free slot */
9087 if (!unused) unused = i;
9090 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9091 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9097 /* If no free slot and max hit, fail */
9098 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9099 return MDB_DBS_FULL;
9101 /* Cannot mix named databases with some mainDB flags */
9102 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9103 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9105 /* Find the DB info */
9106 dbflag = DB_NEW|DB_VALID;
9109 key.mv_data = (void *)name;
9110 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9111 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9112 if (rc == MDB_SUCCESS) {
9113 /* make sure this is actually a DB */
9114 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9115 if (!(node->mn_flags & F_SUBDATA))
9116 return MDB_INCOMPATIBLE;
9117 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9118 /* Create if requested */
9119 data.mv_size = sizeof(MDB_db);
9120 data.mv_data = &dummy;
9121 memset(&dummy, 0, sizeof(dummy));
9122 dummy.md_root = P_INVALID;
9123 dummy.md_flags = flags & PERSISTENT_FLAGS;
9124 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9128 /* OK, got info, add to table */
9129 if (rc == MDB_SUCCESS) {
9130 unsigned int slot = unused ? unused : txn->mt_numdbs;
9131 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9132 txn->mt_dbxs[slot].md_name.mv_size = len;
9133 txn->mt_dbxs[slot].md_rel = NULL;
9134 txn->mt_dbflags[slot] = dbflag;
9135 /* txn-> and env-> are the same in read txns, use
9136 * tmp variable to avoid undefined assignment
9138 seq = ++txn->mt_env->me_dbiseqs[slot];
9139 txn->mt_dbiseqs[slot] = seq;
9141 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9143 mdb_default_cmp(txn, slot);
9152 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9154 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9157 if (txn->mt_flags & MDB_TXN_ERROR)
9160 if (txn->mt_dbflags[dbi] & DB_STALE) {
9163 /* Stale, must read the DB's root. cursor_init does it for us. */
9164 mdb_cursor_init(&mc, txn, dbi, &mx);
9166 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9169 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9172 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9174 ptr = env->me_dbxs[dbi].md_name.mv_data;
9175 /* If there was no name, this was already closed */
9177 env->me_dbxs[dbi].md_name.mv_data = NULL;
9178 env->me_dbxs[dbi].md_name.mv_size = 0;
9179 env->me_dbflags[dbi] = 0;
9180 env->me_dbiseqs[dbi]++;
9185 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9187 /* We could return the flags for the FREE_DBI too but what's the point? */
9188 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9190 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9194 /** Add all the DB's pages to the free list.
9195 * @param[in] mc Cursor on the DB to free.
9196 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9197 * @return 0 on success, non-zero on failure.
9200 mdb_drop0(MDB_cursor *mc, int subs)
9204 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9205 if (rc == MDB_SUCCESS) {
9206 MDB_txn *txn = mc->mc_txn;
9211 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9212 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9215 mdb_cursor_copy(mc, &mx);
9216 while (mc->mc_snum > 0) {
9217 MDB_page *mp = mc->mc_pg[mc->mc_top];
9218 unsigned n = NUMKEYS(mp);
9220 for (i=0; i<n; i++) {
9221 ni = NODEPTR(mp, i);
9222 if (ni->mn_flags & F_BIGDATA) {
9225 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9226 rc = mdb_page_get(txn, pg, &omp, NULL);
9229 mdb_cassert(mc, IS_OVERFLOW(omp));
9230 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9234 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9235 mdb_xcursor_init1(mc, ni);
9236 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9242 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9244 for (i=0; i<n; i++) {
9246 ni = NODEPTR(mp, i);
9249 mdb_midl_xappend(txn->mt_free_pgs, pg);
9254 mc->mc_ki[mc->mc_top] = i;
9255 rc = mdb_cursor_sibling(mc, 1);
9257 if (rc != MDB_NOTFOUND)
9259 /* no more siblings, go back to beginning
9260 * of previous level.
9264 for (i=1; i<mc->mc_snum; i++) {
9266 mc->mc_pg[i] = mx.mc_pg[i];
9271 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9274 txn->mt_flags |= MDB_TXN_ERROR;
9275 } else if (rc == MDB_NOTFOUND) {
9281 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9283 MDB_cursor *mc, *m2;
9286 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9289 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9292 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9295 rc = mdb_cursor_open(txn, dbi, &mc);
9299 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9300 /* Invalidate the dropped DB's cursors */
9301 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9302 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9306 /* Can't delete the main DB */
9307 if (del && dbi > MAIN_DBI) {
9308 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9310 txn->mt_dbflags[dbi] = DB_STALE;
9311 mdb_dbi_close(txn->mt_env, dbi);
9313 txn->mt_flags |= MDB_TXN_ERROR;
9316 /* reset the DB record, mark it dirty */
9317 txn->mt_dbflags[dbi] |= DB_DIRTY;
9318 txn->mt_dbs[dbi].md_depth = 0;
9319 txn->mt_dbs[dbi].md_branch_pages = 0;
9320 txn->mt_dbs[dbi].md_leaf_pages = 0;
9321 txn->mt_dbs[dbi].md_overflow_pages = 0;
9322 txn->mt_dbs[dbi].md_entries = 0;
9323 txn->mt_dbs[dbi].md_root = P_INVALID;
9325 txn->mt_flags |= MDB_TXN_DIRTY;
9328 mdb_cursor_close(mc);
9332 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9334 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9337 txn->mt_dbxs[dbi].md_cmp = cmp;
9341 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9343 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9346 txn->mt_dbxs[dbi].md_dcmp = cmp;
9350 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9352 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9355 txn->mt_dbxs[dbi].md_rel = rel;
9359 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9361 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9364 txn->mt_dbxs[dbi].md_relctx = ctx;
9369 mdb_env_get_maxkeysize(MDB_env *env)
9371 return ENV_MAXKEY(env);
9375 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9377 unsigned int i, rdrs;
9380 int rc = 0, first = 1;
9384 if (!env->me_txns) {
9385 return func("(no reader locks)\n", ctx);
9387 rdrs = env->me_txns->mti_numreaders;
9388 mr = env->me_txns->mti_readers;
9389 for (i=0; i<rdrs; i++) {
9391 txnid_t txnid = mr[i].mr_txnid;
9392 sprintf(buf, txnid == (txnid_t)-1 ?
9393 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9394 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9397 rc = func(" pid thread txnid\n", ctx);
9401 rc = func(buf, ctx);
9407 rc = func("(no active readers)\n", ctx);
9412 /** Insert pid into list if not already present.
9413 * return -1 if already present.
9416 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9418 /* binary search of pid in list */
9420 unsigned cursor = 1;
9422 unsigned n = ids[0];
9425 unsigned pivot = n >> 1;
9426 cursor = base + pivot + 1;
9427 val = pid - ids[cursor];
9432 } else if ( val > 0 ) {
9437 /* found, so it's a duplicate */
9446 for (n = ids[0]; n > cursor; n--)
9453 mdb_reader_check(MDB_env *env, int *dead)
9455 unsigned int i, j, rdrs;
9457 MDB_PID_T *pids, pid;
9466 rdrs = env->me_txns->mti_numreaders;
9467 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9471 mr = env->me_txns->mti_readers;
9472 for (i=0; i<rdrs; i++) {
9473 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9475 if (mdb_pid_insert(pids, pid) == 0) {
9476 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9478 /* Recheck, a new process may have reused pid */
9479 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9480 for (j=i; j<rdrs; j++)
9481 if (mr[j].mr_pid == pid) {
9482 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9483 (unsigned) pid, mr[j].mr_txnid));
9488 UNLOCK_MUTEX_R(env);