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 2011-2016 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
102 typedef SSIZE_T ssize_t;
107 #if defined(__sun) || defined(ANDROID)
108 /* Most platforms have posix_memalign, older may only have memalign */
109 #define HAVE_MEMALIGN 1
113 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
114 #include <netinet/in.h>
115 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
118 #if defined(__APPLE__) || defined (BSD)
119 # define MDB_USE_POSIX_SEM 1
120 # define MDB_FDATASYNC fsync
121 #elif defined(ANDROID)
122 # define MDB_FDATASYNC fsync
127 #ifdef MDB_USE_POSIX_SEM
128 # define MDB_USE_HASH 1
129 #include <semaphore.h>
131 #define MDB_USE_POSIX_MUTEX 1
135 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \
136 + defined(MDB_USE_POSIX_MUTEX) != 1
137 # error "Ambiguous shared-lock implementation"
141 #include <valgrind/memcheck.h>
142 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
143 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
144 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
145 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
146 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
148 #define VGMEMP_CREATE(h,r,z)
149 #define VGMEMP_ALLOC(h,a,s)
150 #define VGMEMP_FREE(h,a)
151 #define VGMEMP_DESTROY(h)
152 #define VGMEMP_DEFINED(a,s)
156 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
157 /* Solaris just defines one or the other */
158 # define LITTLE_ENDIAN 1234
159 # define BIG_ENDIAN 4321
160 # ifdef _LITTLE_ENDIAN
161 # define BYTE_ORDER LITTLE_ENDIAN
163 # define BYTE_ORDER BIG_ENDIAN
166 # define BYTE_ORDER __BYTE_ORDER
170 #ifndef LITTLE_ENDIAN
171 #define LITTLE_ENDIAN __LITTLE_ENDIAN
174 #define BIG_ENDIAN __BIG_ENDIAN
177 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
178 #define MISALIGNED_OK 1
184 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
185 # error "Unknown or unsupported endianness (BYTE_ORDER)"
186 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
187 # error "Two's complement, reasonably sized integer types, please"
191 /** Put infrequently used env functions in separate section */
193 # define ESECT __attribute__ ((section("__TEXT,text_env")))
195 # define ESECT __attribute__ ((section("text_env")))
202 #define CALL_CONV WINAPI
207 /** @defgroup internal LMDB Internals
210 /** @defgroup compat Compatibility Macros
211 * A bunch of macros to minimize the amount of platform-specific ifdefs
212 * needed throughout the rest of the code. When the features this library
213 * needs are similar enough to POSIX to be hidden in a one-or-two line
214 * replacement, this macro approach is used.
218 /** Features under development */
223 /** Wrapper around __func__, which is a C99 feature */
224 #if __STDC_VERSION__ >= 199901L
225 # define mdb_func_ __func__
226 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
227 # define mdb_func_ __FUNCTION__
229 /* If a debug message says <mdb_unknown>(), update the #if statements above */
230 # define mdb_func_ "<mdb_unknown>"
233 /* Internal error codes, not exposed outside liblmdb */
234 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
236 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
237 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
238 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
242 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
244 /** Some platforms define the EOWNERDEAD error code
245 * even though they don't support Robust Mutexes.
246 * Compile with -DMDB_USE_ROBUST=0, or use some other
247 * mechanism like -DMDB_USE_SYSV_SEM instead of
248 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
249 * also Robust, but some systems don't support them
252 #ifndef MDB_USE_ROBUST
253 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
254 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
255 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
256 # define MDB_USE_ROBUST 0
258 # define MDB_USE_ROBUST 1
259 /* glibc < 2.12 only provided _np API */
260 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
261 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
262 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
263 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
264 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
267 #endif /* MDB_USE_ROBUST */
269 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
270 #define MDB_ROBUST_SUPPORTED 1
274 #define MDB_USE_HASH 1
275 #define MDB_PIDLOCK 0
276 #define THREAD_RET DWORD
277 #define pthread_t HANDLE
278 #define pthread_mutex_t HANDLE
279 #define pthread_cond_t HANDLE
280 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
281 #define pthread_key_t DWORD
282 #define pthread_self() GetCurrentThreadId()
283 #define pthread_key_create(x,y) \
284 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
285 #define pthread_key_delete(x) TlsFree(x)
286 #define pthread_getspecific(x) TlsGetValue(x)
287 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
288 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
289 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
290 #define pthread_cond_signal(x) SetEvent(*x)
291 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
292 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
293 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
294 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
295 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
296 #define mdb_mutex_consistent(mutex) 0
297 #define getpid() GetCurrentProcessId()
298 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
299 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
300 #define ErrCode() GetLastError()
301 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
302 #define close(fd) (CloseHandle(fd) ? 0 : -1)
303 #define munmap(ptr,len) UnmapViewOfFile(ptr)
304 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
305 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
307 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
311 #define THREAD_RET void *
312 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
313 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
314 #define Z "z" /**< printf format modifier for size_t */
316 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
317 #define MDB_PIDLOCK 1
319 #ifdef MDB_USE_POSIX_SEM
321 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
322 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
323 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
326 mdb_sem_wait(sem_t *sem)
329 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
333 #else /* MDB_USE_POSIX_MUTEX: */
334 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
335 * local variables keep it (mdb_mutexref_t).
337 * When #mdb_mutexref_t is a pointer declaration and #mdb_mutex_t is
338 * not, then it is array[size 1] so it can be assigned to a pointer.
341 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
343 /** Lock the reader or writer mutex.
344 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
346 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
347 /** Unlock the reader or writer mutex.
349 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
350 /** Mark mutex-protected data as repaired, after death of previous owner.
352 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
353 #endif /* MDB_USE_POSIX_SEM */
355 /** Get the error code for the last failed system function.
357 #define ErrCode() errno
359 /** An abstraction for a file handle.
360 * On POSIX systems file handles are small integers. On Windows
361 * they're opaque pointers.
365 /** A value for an invalid file handle.
366 * Mainly used to initialize file variables and signify that they are
369 #define INVALID_HANDLE_VALUE (-1)
371 /** Get the size of a memory page for the system.
372 * This is the basic size that the platform's memory manager uses, and is
373 * fundamental to the use of memory-mapped files.
375 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
378 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
381 #define MNAME_LEN (sizeof(pthread_mutex_t))
386 #ifdef MDB_ROBUST_SUPPORTED
387 /** Lock mutex, handle any error, set rc = result.
388 * Return 0 on success, nonzero (not rc) on error.
390 #define LOCK_MUTEX(rc, env, mutex) \
391 (((rc) = LOCK_MUTEX0(mutex)) && \
392 ((rc) = mdb_mutex_failed(env, mutex, rc)))
393 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
395 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
396 #define mdb_mutex_failed(env, mutex, rc) (rc)
400 /** A flag for opening a file and requesting synchronous data writes.
401 * This is only used when writing a meta page. It's not strictly needed;
402 * we could just do a normal write and then immediately perform a flush.
403 * But if this flag is available it saves us an extra system call.
405 * @note If O_DSYNC is undefined but exists in /usr/include,
406 * preferably set some compiler flag to get the definition.
410 # define MDB_DSYNC O_DSYNC
412 # define MDB_DSYNC O_SYNC
417 /** Function for flushing the data of a file. Define this to fsync
418 * if fdatasync() is not supported.
420 #ifndef MDB_FDATASYNC
421 # define MDB_FDATASYNC fdatasync
425 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
436 /** A page number in the database.
437 * Note that 64 bit page numbers are overkill, since pages themselves
438 * already represent 12-13 bits of addressable memory, and the OS will
439 * always limit applications to a maximum of 63 bits of address space.
441 * @note In the #MDB_node structure, we only store 48 bits of this value,
442 * which thus limits us to only 60 bits of addressable data.
444 typedef MDB_ID pgno_t;
446 /** A transaction ID.
447 * See struct MDB_txn.mt_txnid for details.
449 typedef MDB_ID txnid_t;
451 /** @defgroup debug Debug Macros
455 /** Enable debug output. Needs variable argument macros (a C99 feature).
456 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
457 * read from and written to the database (used for free space management).
463 static int mdb_debug;
464 static txnid_t mdb_debug_start;
466 /** Print a debug message with printf formatting.
467 * Requires double parenthesis around 2 or more args.
469 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
470 # define DPRINTF0(fmt, ...) \
471 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
473 # define DPRINTF(args) ((void) 0)
475 /** Print a debug string.
476 * The string is printed literally, with no format processing.
478 #define DPUTS(arg) DPRINTF(("%s", arg))
479 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
481 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
484 /** @brief The maximum size of a database page.
486 * It is 32k or 64k, since value-PAGEBASE must fit in
487 * #MDB_page.%mp_upper.
489 * LMDB will use database pages < OS pages if needed.
490 * That causes more I/O in write transactions: The OS must
491 * know (read) the whole page before writing a partial page.
493 * Note that we don't currently support Huge pages. On Linux,
494 * regular data files cannot use Huge pages, and in general
495 * Huge pages aren't actually pageable. We rely on the OS
496 * demand-pager to read our data and page it out when memory
497 * pressure from other processes is high. So until OSs have
498 * actual paging support for Huge pages, they're not viable.
500 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
502 /** The minimum number of keys required in a database page.
503 * Setting this to a larger value will place a smaller bound on the
504 * maximum size of a data item. Data items larger than this size will
505 * be pushed into overflow pages instead of being stored directly in
506 * the B-tree node. This value used to default to 4. With a page size
507 * of 4096 bytes that meant that any item larger than 1024 bytes would
508 * go into an overflow page. That also meant that on average 2-3KB of
509 * each overflow page was wasted space. The value cannot be lower than
510 * 2 because then there would no longer be a tree structure. With this
511 * value, items larger than 2KB will go into overflow pages, and on
512 * average only 1KB will be wasted.
514 #define MDB_MINKEYS 2
516 /** A stamp that identifies a file as an LMDB file.
517 * There's nothing special about this value other than that it is easily
518 * recognizable, and it will reflect any byte order mismatches.
520 #define MDB_MAGIC 0xBEEFC0DE
522 /** The version number for a database's datafile format. */
523 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
524 /** The version number for a database's lockfile format. */
525 #define MDB_LOCK_VERSION 1
527 /** @brief The max size of a key we can write, or 0 for computed max.
529 * This macro should normally be left alone or set to 0.
530 * Note that a database with big keys or dupsort data cannot be
531 * reliably modified by a liblmdb which uses a smaller max.
532 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
534 * Other values are allowed, for backwards compat. However:
535 * A value bigger than the computed max can break if you do not
536 * know what you are doing, and liblmdb <= 0.9.10 can break when
537 * modifying a DB with keys/dupsort data bigger than its max.
539 * Data items in an #MDB_DUPSORT database are also limited to
540 * this size, since they're actually keys of a sub-DB. Keys and
541 * #MDB_DUPSORT data items must fit on a node in a regular page.
543 #ifndef MDB_MAXKEYSIZE
544 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
547 /** The maximum size of a key we can write to the environment. */
549 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
551 #define ENV_MAXKEY(env) ((env)->me_maxkey)
554 /** @brief The maximum size of a data item.
556 * We only store a 32 bit value for node sizes.
558 #define MAXDATASIZE 0xffffffffUL
561 /** Key size which fits in a #DKBUF.
564 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
567 * This is used for printing a hex dump of a key's contents.
569 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
570 /** Display a key in hex.
572 * Invoke a function to display a key in hex.
574 #define DKEY(x) mdb_dkey(x, kbuf)
580 /** An invalid page number.
581 * Mainly used to denote an empty tree.
583 #define P_INVALID (~(pgno_t)0)
585 /** Test if the flags \b f are set in a flag word \b w. */
586 #define F_ISSET(w, f) (((w) & (f)) == (f))
588 /** Round \b n up to an even number. */
589 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
591 /** Used for offsets within a single page.
592 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
595 typedef uint16_t indx_t;
597 /** Default size of memory map.
598 * This is certainly too small for any actual applications. Apps should always set
599 * the size explicitly using #mdb_env_set_mapsize().
601 #define DEFAULT_MAPSIZE 1048576
603 /** @defgroup readers Reader Lock Table
604 * Readers don't acquire any locks for their data access. Instead, they
605 * simply record their transaction ID in the reader table. The reader
606 * mutex is needed just to find an empty slot in the reader table. The
607 * slot's address is saved in thread-specific data so that subsequent read
608 * transactions started by the same thread need no further locking to proceed.
610 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
612 * No reader table is used if the database is on a read-only filesystem, or
613 * if #MDB_NOLOCK is set.
615 * Since the database uses multi-version concurrency control, readers don't
616 * actually need any locking. This table is used to keep track of which
617 * readers are using data from which old transactions, so that we'll know
618 * when a particular old transaction is no longer in use. Old transactions
619 * that have discarded any data pages can then have those pages reclaimed
620 * for use by a later write transaction.
622 * The lock table is constructed such that reader slots are aligned with the
623 * processor's cache line size. Any slot is only ever used by one thread.
624 * This alignment guarantees that there will be no contention or cache
625 * thrashing as threads update their own slot info, and also eliminates
626 * any need for locking when accessing a slot.
628 * A writer thread will scan every slot in the table to determine the oldest
629 * outstanding reader transaction. Any freed pages older than this will be
630 * reclaimed by the writer. The writer doesn't use any locks when scanning
631 * this table. This means that there's no guarantee that the writer will
632 * see the most up-to-date reader info, but that's not required for correct
633 * operation - all we need is to know the upper bound on the oldest reader,
634 * we don't care at all about the newest reader. So the only consequence of
635 * reading stale information here is that old pages might hang around a
636 * while longer before being reclaimed. That's actually good anyway, because
637 * the longer we delay reclaiming old pages, the more likely it is that a
638 * string of contiguous pages can be found after coalescing old pages from
639 * many old transactions together.
642 /** Number of slots in the reader table.
643 * This value was chosen somewhat arbitrarily. 126 readers plus a
644 * couple mutexes fit exactly into 8KB on my development machine.
645 * Applications should set the table size using #mdb_env_set_maxreaders().
647 #define DEFAULT_READERS 126
649 /** The size of a CPU cache line in bytes. We want our lock structures
650 * aligned to this size to avoid false cache line sharing in the
652 * This value works for most CPUs. For Itanium this should be 128.
658 /** The information we store in a single slot of the reader table.
659 * In addition to a transaction ID, we also record the process and
660 * thread ID that owns a slot, so that we can detect stale information,
661 * e.g. threads or processes that went away without cleaning up.
662 * @note We currently don't check for stale records. We simply re-init
663 * the table when we know that we're the only process opening the
666 typedef struct MDB_rxbody {
667 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
668 * Multiple readers that start at the same time will probably have the
669 * same ID here. Again, it's not important to exclude them from
670 * anything; all we need to know is which version of the DB they
671 * started from so we can avoid overwriting any data used in that
672 * particular version.
674 volatile txnid_t mrb_txnid;
675 /** The process ID of the process owning this reader txn. */
676 volatile MDB_PID_T mrb_pid;
677 /** The thread ID of the thread owning this txn. */
678 volatile MDB_THR_T mrb_tid;
681 /** The actual reader record, with cacheline padding. */
682 typedef struct MDB_reader {
685 /** shorthand for mrb_txnid */
686 #define mr_txnid mru.mrx.mrb_txnid
687 #define mr_pid mru.mrx.mrb_pid
688 #define mr_tid mru.mrx.mrb_tid
689 /** cache line alignment */
690 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
694 /** The header for the reader table.
695 * The table resides in a memory-mapped file. (This is a different file
696 * than is used for the main database.)
698 * For POSIX the actual mutexes reside in the shared memory of this
699 * mapped file. On Windows, mutexes are named objects allocated by the
700 * kernel; we store the mutex names in this mapped file so that other
701 * processes can grab them. This same approach is also used on
702 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
703 * process-shared POSIX mutexes. For these cases where a named object
704 * is used, the object name is derived from a 64 bit FNV hash of the
705 * environment pathname. As such, naming collisions are extremely
706 * unlikely. If a collision occurs, the results are unpredictable.
708 typedef struct MDB_txbody {
709 /** Stamp identifying this as an LMDB file. It must be set
712 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
714 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
715 char mtb_rmname[MNAME_LEN];
717 /** Mutex protecting access to this table.
718 * This is the reader table lock used with LOCK_MUTEX().
720 mdb_mutex_t mtb_rmutex;
722 /** The ID of the last transaction committed to the database.
723 * This is recorded here only for convenience; the value can always
724 * be determined by reading the main database meta pages.
726 volatile txnid_t mtb_txnid;
727 /** The number of slots that have been used in the reader table.
728 * This always records the maximum count, it is not decremented
729 * when readers release their slots.
731 volatile unsigned mtb_numreaders;
734 /** The actual reader table definition. */
735 typedef struct MDB_txninfo {
738 #define mti_magic mt1.mtb.mtb_magic
739 #define mti_format mt1.mtb.mtb_format
740 #define mti_rmutex mt1.mtb.mtb_rmutex
741 #define mti_rmname mt1.mtb.mtb_rmname
742 #define mti_txnid mt1.mtb.mtb_txnid
743 #define mti_numreaders mt1.mtb.mtb_numreaders
744 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
747 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
748 char mt2_wmname[MNAME_LEN];
749 #define mti_wmname mt2.mt2_wmname
751 mdb_mutex_t mt2_wmutex;
752 #define mti_wmutex mt2.mt2_wmutex
754 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
756 MDB_reader mti_readers[1];
759 /** Lockfile format signature: version, features and field layout */
760 #define MDB_LOCK_FORMAT \
762 ((MDB_LOCK_VERSION) \
763 /* Flags which describe functionality */ \
764 + (((MDB_PIDLOCK) != 0) << 16)))
767 /** Common header for all page types.
768 * Overflow records occupy a number of contiguous pages with no
769 * headers on any page after the first.
771 typedef struct MDB_page {
772 #define mp_pgno mp_p.p_pgno
773 #define mp_next mp_p.p_next
775 pgno_t p_pgno; /**< page number */
776 struct MDB_page *p_next; /**< for in-memory list of freed pages */
779 /** @defgroup mdb_page Page Flags
781 * Flags for the page headers.
784 #define P_BRANCH 0x01 /**< branch page */
785 #define P_LEAF 0x02 /**< leaf page */
786 #define P_OVERFLOW 0x04 /**< overflow page */
787 #define P_META 0x08 /**< meta page */
788 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
789 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
790 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
791 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
792 #define P_KEEP 0x8000 /**< leave this page alone during spill */
794 uint16_t mp_flags; /**< @ref mdb_page */
795 #define mp_lower mp_pb.pb.pb_lower
796 #define mp_upper mp_pb.pb.pb_upper
797 #define mp_pages mp_pb.pb_pages
800 indx_t pb_lower; /**< lower bound of free space */
801 indx_t pb_upper; /**< upper bound of free space */
803 uint32_t pb_pages; /**< number of overflow pages */
805 indx_t mp_ptrs[1]; /**< dynamic size */
808 /** Size of the page header, excluding dynamic data at the end */
809 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
811 /** Address of first usable data byte in a page, after the header */
812 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
814 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
815 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
817 /** Number of nodes on a page */
818 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
820 /** The amount of space remaining in the page */
821 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
823 /** The percentage of space used in the page, in tenths of a percent. */
824 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
825 ((env)->me_psize - PAGEHDRSZ))
826 /** The minimum page fill factor, in tenths of a percent.
827 * Pages emptier than this are candidates for merging.
829 #define FILL_THRESHOLD 250
831 /** Test if a page is a leaf page */
832 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
833 /** Test if a page is a LEAF2 page */
834 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
835 /** Test if a page is a branch page */
836 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
837 /** Test if a page is an overflow page */
838 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
839 /** Test if a page is a sub page */
840 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
842 /** The number of overflow pages needed to store the given size. */
843 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
845 /** Link in #MDB_txn.%mt_loose_pgs list */
846 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
848 /** Header for a single key/data pair within a page.
849 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
850 * We guarantee 2-byte alignment for 'MDB_node's.
852 typedef struct MDB_node {
853 /** lo and hi are used for data size on leaf nodes and for
854 * child pgno on branch nodes. On 64 bit platforms, flags
855 * is also used for pgno. (Branch nodes have no flags).
856 * They are in host byte order in case that lets some
857 * accesses be optimized into a 32-bit word access.
859 #if BYTE_ORDER == LITTLE_ENDIAN
860 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
862 unsigned short mn_hi, mn_lo;
864 /** @defgroup mdb_node Node Flags
866 * Flags for node headers.
869 #define F_BIGDATA 0x01 /**< data put on overflow page */
870 #define F_SUBDATA 0x02 /**< data is a sub-database */
871 #define F_DUPDATA 0x04 /**< data has duplicates */
873 /** valid flags for #mdb_node_add() */
874 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
877 unsigned short mn_flags; /**< @ref mdb_node */
878 unsigned short mn_ksize; /**< key size */
879 char mn_data[1]; /**< key and data are appended here */
882 /** Size of the node header, excluding dynamic data at the end */
883 #define NODESIZE offsetof(MDB_node, mn_data)
885 /** Bit position of top word in page number, for shifting mn_flags */
886 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
888 /** Size of a node in a branch page with a given key.
889 * This is just the node header plus the key, there is no data.
891 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
893 /** Size of a node in a leaf page with a given key and data.
894 * This is node header plus key plus data size.
896 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
898 /** Address of node \b i in page \b p */
899 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
901 /** Address of the key for the node */
902 #define NODEKEY(node) (void *)((node)->mn_data)
904 /** Address of the data for a node */
905 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
907 /** Get the page number pointed to by a branch node */
908 #define NODEPGNO(node) \
909 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
910 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
911 /** Set the page number in a branch node */
912 #define SETPGNO(node,pgno) do { \
913 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
914 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
916 /** Get the size of the data in a leaf node */
917 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
918 /** Set the size of the data for a leaf node */
919 #define SETDSZ(node,size) do { \
920 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
921 /** The size of a key in a node */
922 #define NODEKSZ(node) ((node)->mn_ksize)
924 /** Copy a page number from src to dst */
926 #define COPY_PGNO(dst,src) dst = src
928 #if SIZE_MAX > 4294967295UL
929 #define COPY_PGNO(dst,src) do { \
930 unsigned short *s, *d; \
931 s = (unsigned short *)&(src); \
932 d = (unsigned short *)&(dst); \
939 #define COPY_PGNO(dst,src) do { \
940 unsigned short *s, *d; \
941 s = (unsigned short *)&(src); \
942 d = (unsigned short *)&(dst); \
948 /** The address of a key in a LEAF2 page.
949 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
950 * There are no node headers, keys are stored contiguously.
952 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
954 /** Set the \b node's key into \b keyptr, if requested. */
955 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
956 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
958 /** Set the \b node's key into \b key. */
959 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
961 /** Information about a single database in the environment. */
962 typedef struct MDB_db {
963 uint32_t md_pad; /**< also ksize for LEAF2 pages */
964 uint16_t md_flags; /**< @ref mdb_dbi_open */
965 uint16_t md_depth; /**< depth of this tree */
966 pgno_t md_branch_pages; /**< number of internal pages */
967 pgno_t md_leaf_pages; /**< number of leaf pages */
968 pgno_t md_overflow_pages; /**< number of overflow pages */
969 size_t md_entries; /**< number of data items */
970 pgno_t md_root; /**< the root page of this tree */
973 /** mdb_dbi_open flags */
974 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
975 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
976 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
977 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
979 /** Handle for the DB used to track free pages. */
981 /** Handle for the default DB. */
983 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
986 /** Number of meta pages - also hardcoded elsewhere */
989 /** Meta page content.
990 * A meta page is the start point for accessing a database snapshot.
991 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
993 typedef struct MDB_meta {
994 /** Stamp identifying this as an LMDB file. It must be set
997 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
999 void *mm_address; /**< address for fixed mapping */
1000 size_t mm_mapsize; /**< size of mmap region */
1001 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1002 /** The size of pages used in this DB */
1003 #define mm_psize mm_dbs[FREE_DBI].md_pad
1004 /** Any persistent environment flags. @ref mdb_env */
1005 #define mm_flags mm_dbs[FREE_DBI].md_flags
1006 pgno_t mm_last_pg; /**< last used page in file */
1007 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1010 /** Buffer for a stack-allocated meta page.
1011 * The members define size and alignment, and silence type
1012 * aliasing warnings. They are not used directly; that could
1013 * mean incorrectly using several union members in parallel.
1015 typedef union MDB_metabuf {
1018 char mm_pad[PAGEHDRSZ];
1023 /** Auxiliary DB info.
1024 * The information here is mostly static/read-only. There is
1025 * only a single copy of this record in the environment.
1027 typedef struct MDB_dbx {
1028 MDB_val md_name; /**< name of the database */
1029 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1030 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1031 MDB_rel_func *md_rel; /**< user relocate function */
1032 void *md_relctx; /**< user-provided context for md_rel */
1035 /** A database transaction.
1036 * Every operation requires a transaction handle.
1039 MDB_txn *mt_parent; /**< parent of a nested txn */
1040 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1042 pgno_t mt_next_pgno; /**< next unallocated page */
1043 /** The ID of this transaction. IDs are integers incrementing from 1.
1044 * Only committed write transactions increment the ID. If a transaction
1045 * aborts, the ID may be re-used by the next writer.
1048 MDB_env *mt_env; /**< the DB environment */
1049 /** The list of pages that became unused during this transaction.
1051 MDB_IDL mt_free_pgs;
1052 /** The list of loose pages that became unused and may be reused
1053 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1055 MDB_page *mt_loose_pgs;
1056 /* #Number of loose pages (#mt_loose_pgs) */
1058 /** The sorted list of dirty pages we temporarily wrote to disk
1059 * because the dirty list was full. page numbers in here are
1060 * shifted left by 1, deleted slots have the LSB set.
1062 MDB_IDL mt_spill_pgs;
1064 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1065 MDB_ID2L dirty_list;
1066 /** For read txns: This thread/txn's reader table slot, or NULL. */
1069 /** Array of records for each DB known in the environment. */
1071 /** Array of MDB_db records for each known DB */
1073 /** Array of sequence numbers for each DB handle */
1074 unsigned int *mt_dbiseqs;
1075 /** @defgroup mt_dbflag Transaction DB Flags
1079 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1080 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1081 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1082 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1083 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1085 /** In write txns, array of cursors for each DB */
1086 MDB_cursor **mt_cursors;
1087 /** Array of flags for each DB */
1088 unsigned char *mt_dbflags;
1089 /** Number of DB records in use, or 0 when the txn is finished.
1090 * This number only ever increments until the txn finishes; we
1091 * don't decrement it when individual DB handles are closed.
1095 /** @defgroup mdb_txn Transaction Flags
1099 /** #mdb_txn_begin() flags */
1100 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1101 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1102 /* internal txn flags */
1103 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1104 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1105 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1106 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1107 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1108 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1109 /** most operations on the txn are currently illegal */
1110 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1112 unsigned int mt_flags; /**< @ref mdb_txn */
1113 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1114 * Includes ancestor txns' dirty pages not hidden by other txns'
1115 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1116 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1118 unsigned int mt_dirty_room;
1121 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1122 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1123 * raise this on a 64 bit machine.
1125 #define CURSOR_STACK 32
1129 /** Cursors are used for all DB operations.
1130 * A cursor holds a path of (page pointer, key index) from the DB
1131 * root to a position in the DB, plus other state. #MDB_DUPSORT
1132 * cursors include an xcursor to the current data item. Write txns
1133 * track their cursors and keep them up to date when data moves.
1134 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1135 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1138 /** Next cursor on this DB in this txn */
1139 MDB_cursor *mc_next;
1140 /** Backup of the original cursor if this cursor is a shadow */
1141 MDB_cursor *mc_backup;
1142 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1143 struct MDB_xcursor *mc_xcursor;
1144 /** The transaction that owns this cursor */
1146 /** The database handle this cursor operates on */
1148 /** The database record for this cursor */
1150 /** The database auxiliary record for this cursor */
1152 /** The @ref mt_dbflag for this database */
1153 unsigned char *mc_dbflag;
1154 unsigned short mc_snum; /**< number of pushed pages */
1155 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1156 /** @defgroup mdb_cursor Cursor Flags
1158 * Cursor state flags.
1161 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1162 #define C_EOF 0x02 /**< No more data */
1163 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1164 #define C_DEL 0x08 /**< last op was a cursor_del */
1165 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1167 unsigned int mc_flags; /**< @ref mdb_cursor */
1168 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1169 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1172 /** Context for sorted-dup records.
1173 * We could have gone to a fully recursive design, with arbitrarily
1174 * deep nesting of sub-databases. But for now we only handle these
1175 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1177 typedef struct MDB_xcursor {
1178 /** A sub-cursor for traversing the Dup DB */
1179 MDB_cursor mx_cursor;
1180 /** The database record for this Dup DB */
1182 /** The auxiliary DB record for this Dup DB */
1184 /** The @ref mt_dbflag for this Dup DB */
1185 unsigned char mx_dbflag;
1188 /** State of FreeDB old pages, stored in the MDB_env */
1189 typedef struct MDB_pgstate {
1190 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1191 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1194 /** The database environment. */
1196 HANDLE me_fd; /**< The main data file */
1197 HANDLE me_lfd; /**< The lock file */
1198 HANDLE me_mfd; /**< just for writing the meta pages */
1199 /** Failed to update the meta page. Probably an I/O error. */
1200 #define MDB_FATAL_ERROR 0x80000000U
1201 /** Some fields are initialized. */
1202 #define MDB_ENV_ACTIVE 0x20000000U
1203 /** me_txkey is set */
1204 #define MDB_ENV_TXKEY 0x10000000U
1205 /** fdatasync is unreliable */
1206 #define MDB_FSYNCONLY 0x08000000U
1207 uint32_t me_flags; /**< @ref mdb_env */
1208 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1209 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1210 unsigned int me_maxreaders; /**< size of the reader table */
1211 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1212 volatile int me_close_readers;
1213 MDB_dbi me_numdbs; /**< number of DBs opened */
1214 MDB_dbi me_maxdbs; /**< size of the DB table */
1215 MDB_PID_T me_pid; /**< process ID of this env */
1216 char *me_path; /**< path to the DB files */
1217 char *me_map; /**< the memory map of the data file */
1218 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1219 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1220 void *me_pbuf; /**< scratch area for DUPSORT put() */
1221 MDB_txn *me_txn; /**< current write transaction */
1222 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1223 size_t me_mapsize; /**< size of the data memory map */
1224 off_t me_size; /**< current file size */
1225 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1226 MDB_dbx *me_dbxs; /**< array of static DB info */
1227 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1228 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1229 pthread_key_t me_txkey; /**< thread-key for readers */
1230 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1231 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1232 # define me_pglast me_pgstate.mf_pglast
1233 # define me_pghead me_pgstate.mf_pghead
1234 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1235 /** IDL of pages that became unused in a write txn */
1236 MDB_IDL me_free_pgs;
1237 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1238 MDB_ID2L me_dirty_list;
1239 /** Max number of freelist items that can fit in a single overflow page */
1241 /** Max size of a node on a page */
1242 unsigned int me_nodemax;
1243 #if !(MDB_MAXKEYSIZE)
1244 unsigned int me_maxkey; /**< max size of a key */
1246 int me_live_reader; /**< have liveness lock in reader table */
1248 int me_pidquery; /**< Used in OpenProcess */
1250 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1251 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1252 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1254 mdb_mutex_t me_rmutex;
1255 mdb_mutex_t me_wmutex;
1257 void *me_userctx; /**< User-settable context */
1258 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1261 /** Nested transaction */
1262 typedef struct MDB_ntxn {
1263 MDB_txn mnt_txn; /**< the transaction */
1264 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1267 /** max number of pages to commit in one writev() call */
1268 #define MDB_COMMIT_PAGES 64
1269 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1270 #undef MDB_COMMIT_PAGES
1271 #define MDB_COMMIT_PAGES IOV_MAX
1274 /** max bytes to write in one call */
1275 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1277 /** Check \b txn and \b dbi arguments to a function */
1278 #define TXN_DBI_EXIST(txn, dbi, validity) \
1279 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1281 /** Check for misused \b dbi handles */
1282 #define TXN_DBI_CHANGED(txn, dbi) \
1283 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1285 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1286 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1287 static int mdb_page_touch(MDB_cursor *mc);
1289 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1290 "reset-tmp", "fail-begin", "fail-beginchild"}
1292 /* mdb_txn_end operation number, for logging */
1293 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1294 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1296 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1297 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1298 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1299 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1300 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1302 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1303 static int mdb_page_search_root(MDB_cursor *mc,
1304 MDB_val *key, int modify);
1305 #define MDB_PS_MODIFY 1
1306 #define MDB_PS_ROOTONLY 2
1307 #define MDB_PS_FIRST 4
1308 #define MDB_PS_LAST 8
1309 static int mdb_page_search(MDB_cursor *mc,
1310 MDB_val *key, int flags);
1311 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1313 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1314 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1315 pgno_t newpgno, unsigned int nflags);
1317 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1318 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1319 static int mdb_env_write_meta(MDB_txn *txn);
1320 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1321 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1323 static void mdb_env_close0(MDB_env *env, int excl);
1325 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1326 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1327 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1328 static void mdb_node_del(MDB_cursor *mc, int ksize);
1329 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1330 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1331 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1332 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1333 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1335 static int mdb_rebalance(MDB_cursor *mc);
1336 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1338 static void mdb_cursor_pop(MDB_cursor *mc);
1339 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1341 static int mdb_cursor_del0(MDB_cursor *mc);
1342 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1343 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1344 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1345 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1346 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1348 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1349 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1351 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1352 static void mdb_xcursor_init0(MDB_cursor *mc);
1353 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1354 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1356 static int mdb_drop0(MDB_cursor *mc, int subs);
1357 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1358 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1361 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1364 /** Compare two items pointing at size_t's of unknown alignment. */
1365 #ifdef MISALIGNED_OK
1366 # define mdb_cmp_clong mdb_cmp_long
1368 # define mdb_cmp_clong mdb_cmp_cint
1372 static SECURITY_DESCRIPTOR mdb_null_sd;
1373 static SECURITY_ATTRIBUTES mdb_all_sa;
1374 static int mdb_sec_inited;
1376 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
1379 /** Return the library version info. */
1381 mdb_version(int *major, int *minor, int *patch)
1383 if (major) *major = MDB_VERSION_MAJOR;
1384 if (minor) *minor = MDB_VERSION_MINOR;
1385 if (patch) *patch = MDB_VERSION_PATCH;
1386 return MDB_VERSION_STRING;
1389 /** Table of descriptions for LMDB @ref errors */
1390 static char *const mdb_errstr[] = {
1391 "MDB_KEYEXIST: Key/data pair already exists",
1392 "MDB_NOTFOUND: No matching key/data pair found",
1393 "MDB_PAGE_NOTFOUND: Requested page not found",
1394 "MDB_CORRUPTED: Located page was wrong type",
1395 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1396 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1397 "MDB_INVALID: File is not an LMDB file",
1398 "MDB_MAP_FULL: Environment mapsize limit reached",
1399 "MDB_DBS_FULL: Environment maxdbs limit reached",
1400 "MDB_READERS_FULL: Environment maxreaders limit reached",
1401 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1402 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1403 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1404 "MDB_PAGE_FULL: Internal error - page has no more space",
1405 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1406 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1407 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1408 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1409 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1410 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1414 mdb_strerror(int err)
1417 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1418 * This works as long as no function between the call to mdb_strerror
1419 * and the actual use of the message uses more than 4K of stack.
1421 #define MSGSIZE 1024
1422 #define PADSIZE 4096
1423 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1427 return ("Successful return: 0");
1429 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1430 i = err - MDB_KEYEXIST;
1431 return mdb_errstr[i];
1435 /* These are the C-runtime error codes we use. The comment indicates
1436 * their numeric value, and the Win32 error they would correspond to
1437 * if the error actually came from a Win32 API. A major mess, we should
1438 * have used LMDB-specific error codes for everything.
1441 case ENOENT: /* 2, FILE_NOT_FOUND */
1442 case EIO: /* 5, ACCESS_DENIED */
1443 case ENOMEM: /* 12, INVALID_ACCESS */
1444 case EACCES: /* 13, INVALID_DATA */
1445 case EBUSY: /* 16, CURRENT_DIRECTORY */
1446 case EINVAL: /* 22, BAD_COMMAND */
1447 case ENOSPC: /* 28, OUT_OF_PAPER */
1448 return strerror(err);
1453 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1454 FORMAT_MESSAGE_IGNORE_INSERTS,
1455 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1458 return strerror(err);
1462 /** assert(3) variant in cursor context */
1463 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1464 /** assert(3) variant in transaction context */
1465 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1466 /** assert(3) variant in environment context */
1467 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1470 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1471 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1474 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1475 const char *func, const char *file, int line)
1478 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1479 file, line, expr_txt, func);
1480 if (env->me_assert_func)
1481 env->me_assert_func(env, buf);
1482 fprintf(stderr, "%s\n", buf);
1486 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1490 /** Return the page number of \b mp which may be sub-page, for debug output */
1492 mdb_dbg_pgno(MDB_page *mp)
1495 COPY_PGNO(ret, mp->mp_pgno);
1499 /** Display a key in hexadecimal and return the address of the result.
1500 * @param[in] key the key to display
1501 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1502 * @return The key in hexadecimal form.
1505 mdb_dkey(MDB_val *key, char *buf)
1508 unsigned char *c = key->mv_data;
1514 if (key->mv_size > DKBUF_MAXKEYSIZE)
1515 return "MDB_MAXKEYSIZE";
1516 /* may want to make this a dynamic check: if the key is mostly
1517 * printable characters, print it as-is instead of converting to hex.
1521 for (i=0; i<key->mv_size; i++)
1522 ptr += sprintf(ptr, "%02x", *c++);
1524 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1530 mdb_leafnode_type(MDB_node *n)
1532 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1533 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1534 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1537 /** Display all the keys in the page. */
1539 mdb_page_list(MDB_page *mp)
1541 pgno_t pgno = mdb_dbg_pgno(mp);
1542 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1544 unsigned int i, nkeys, nsize, total = 0;
1548 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1549 case P_BRANCH: type = "Branch page"; break;
1550 case P_LEAF: type = "Leaf page"; break;
1551 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1552 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1553 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1555 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1556 pgno, mp->mp_pages, state);
1559 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1560 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1563 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1567 nkeys = NUMKEYS(mp);
1568 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1570 for (i=0; i<nkeys; i++) {
1571 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1572 key.mv_size = nsize = mp->mp_pad;
1573 key.mv_data = LEAF2KEY(mp, i, nsize);
1575 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1578 node = NODEPTR(mp, i);
1579 key.mv_size = node->mn_ksize;
1580 key.mv_data = node->mn_data;
1581 nsize = NODESIZE + key.mv_size;
1582 if (IS_BRANCH(mp)) {
1583 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1587 if (F_ISSET(node->mn_flags, F_BIGDATA))
1588 nsize += sizeof(pgno_t);
1590 nsize += NODEDSZ(node);
1592 nsize += sizeof(indx_t);
1593 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1594 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1596 total = EVEN(total);
1598 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1599 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1603 mdb_cursor_chk(MDB_cursor *mc)
1609 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1610 for (i=0; i<mc->mc_top; i++) {
1612 node = NODEPTR(mp, mc->mc_ki[i]);
1613 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1616 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1618 if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
1619 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1620 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1621 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1629 /** Count all the pages in each DB and in the freelist
1630 * and make sure it matches the actual number of pages
1632 * All named DBs must be open for a correct count.
1634 static void mdb_audit(MDB_txn *txn)
1638 MDB_ID freecount, count;
1643 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1644 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1645 freecount += *(MDB_ID *)data.mv_data;
1646 mdb_tassert(txn, rc == MDB_NOTFOUND);
1649 for (i = 0; i<txn->mt_numdbs; i++) {
1651 if (!(txn->mt_dbflags[i] & DB_VALID))
1653 mdb_cursor_init(&mc, txn, i, &mx);
1654 if (txn->mt_dbs[i].md_root == P_INVALID)
1656 count += txn->mt_dbs[i].md_branch_pages +
1657 txn->mt_dbs[i].md_leaf_pages +
1658 txn->mt_dbs[i].md_overflow_pages;
1659 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1660 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1661 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1664 mp = mc.mc_pg[mc.mc_top];
1665 for (j=0; j<NUMKEYS(mp); j++) {
1666 MDB_node *leaf = NODEPTR(mp, j);
1667 if (leaf->mn_flags & F_SUBDATA) {
1669 memcpy(&db, NODEDATA(leaf), sizeof(db));
1670 count += db.md_branch_pages + db.md_leaf_pages +
1671 db.md_overflow_pages;
1675 mdb_tassert(txn, rc == MDB_NOTFOUND);
1678 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1679 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1680 txn->mt_txnid, freecount, count+NUM_METAS,
1681 freecount+count+NUM_METAS, txn->mt_next_pgno);
1687 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1689 return txn->mt_dbxs[dbi].md_cmp(a, b);
1693 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1695 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1696 #if UINT_MAX < SIZE_MAX
1697 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1698 dcmp = mdb_cmp_clong;
1703 /** Allocate memory for a page.
1704 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1707 mdb_page_malloc(MDB_txn *txn, unsigned num)
1709 MDB_env *env = txn->mt_env;
1710 MDB_page *ret = env->me_dpages;
1711 size_t psize = env->me_psize, sz = psize, off;
1712 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1713 * For a single page alloc, we init everything after the page header.
1714 * For multi-page, we init the final page; if the caller needed that
1715 * many pages they will be filling in at least up to the last page.
1719 VGMEMP_ALLOC(env, ret, sz);
1720 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1721 env->me_dpages = ret->mp_next;
1724 psize -= off = PAGEHDRSZ;
1729 if ((ret = malloc(sz)) != NULL) {
1730 VGMEMP_ALLOC(env, ret, sz);
1731 if (!(env->me_flags & MDB_NOMEMINIT)) {
1732 memset((char *)ret + off, 0, psize);
1736 txn->mt_flags |= MDB_TXN_ERROR;
1740 /** Free a single page.
1741 * Saves single pages to a list, for future reuse.
1742 * (This is not used for multi-page overflow pages.)
1745 mdb_page_free(MDB_env *env, MDB_page *mp)
1747 mp->mp_next = env->me_dpages;
1748 VGMEMP_FREE(env, mp);
1749 env->me_dpages = mp;
1752 /** Free a dirty page */
1754 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1756 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1757 mdb_page_free(env, dp);
1759 /* large pages just get freed directly */
1760 VGMEMP_FREE(env, dp);
1765 /** Return all dirty pages to dpage list */
1767 mdb_dlist_free(MDB_txn *txn)
1769 MDB_env *env = txn->mt_env;
1770 MDB_ID2L dl = txn->mt_u.dirty_list;
1771 unsigned i, n = dl[0].mid;
1773 for (i = 1; i <= n; i++) {
1774 mdb_dpage_free(env, dl[i].mptr);
1779 /** Loosen or free a single page.
1780 * Saves single pages to a list for future reuse
1781 * in this same txn. It has been pulled from the freeDB
1782 * and already resides on the dirty list, but has been
1783 * deleted. Use these pages first before pulling again
1786 * If the page wasn't dirtied in this txn, just add it
1787 * to this txn's free list.
1790 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1793 pgno_t pgno = mp->mp_pgno;
1794 MDB_txn *txn = mc->mc_txn;
1796 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1797 if (txn->mt_parent) {
1798 MDB_ID2 *dl = txn->mt_u.dirty_list;
1799 /* If txn has a parent, make sure the page is in our
1803 unsigned x = mdb_mid2l_search(dl, pgno);
1804 if (x <= dl[0].mid && dl[x].mid == pgno) {
1805 if (mp != dl[x].mptr) { /* bad cursor? */
1806 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1807 txn->mt_flags |= MDB_TXN_ERROR;
1808 return MDB_CORRUPTED;
1815 /* no parent txn, so it's just ours */
1820 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1822 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1823 txn->mt_loose_pgs = mp;
1824 txn->mt_loose_count++;
1825 mp->mp_flags |= P_LOOSE;
1827 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1835 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1836 * @param[in] mc A cursor handle for the current operation.
1837 * @param[in] pflags Flags of the pages to update:
1838 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1839 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1840 * @return 0 on success, non-zero on failure.
1843 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1845 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1846 MDB_txn *txn = mc->mc_txn;
1852 int rc = MDB_SUCCESS, level;
1854 /* Mark pages seen by cursors */
1855 if (mc->mc_flags & C_UNTRACK)
1856 mc = NULL; /* will find mc in mt_cursors */
1857 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1858 for (; mc; mc=mc->mc_next) {
1859 if (!(mc->mc_flags & C_INITIALIZED))
1861 for (m3 = mc;; m3 = &mx->mx_cursor) {
1863 for (j=0; j<m3->mc_snum; j++) {
1865 if ((mp->mp_flags & Mask) == pflags)
1866 mp->mp_flags ^= P_KEEP;
1868 mx = m3->mc_xcursor;
1869 /* Proceed to mx if it is at a sub-database */
1870 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1872 if (! (mp && (mp->mp_flags & P_LEAF)))
1874 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1875 if (!(leaf->mn_flags & F_SUBDATA))
1884 /* Mark dirty root pages */
1885 for (i=0; i<txn->mt_numdbs; i++) {
1886 if (txn->mt_dbflags[i] & DB_DIRTY) {
1887 pgno_t pgno = txn->mt_dbs[i].md_root;
1888 if (pgno == P_INVALID)
1890 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1892 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1893 dp->mp_flags ^= P_KEEP;
1901 static int mdb_page_flush(MDB_txn *txn, int keep);
1903 /** Spill pages from the dirty list back to disk.
1904 * This is intended to prevent running into #MDB_TXN_FULL situations,
1905 * but note that they may still occur in a few cases:
1906 * 1) our estimate of the txn size could be too small. Currently this
1907 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1908 * 2) child txns may run out of space if their parents dirtied a
1909 * lot of pages and never spilled them. TODO: we probably should do
1910 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1911 * the parent's dirty_room is below a given threshold.
1913 * Otherwise, if not using nested txns, it is expected that apps will
1914 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1915 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1916 * If the txn never references them again, they can be left alone.
1917 * If the txn only reads them, they can be used without any fuss.
1918 * If the txn writes them again, they can be dirtied immediately without
1919 * going thru all of the work of #mdb_page_touch(). Such references are
1920 * handled by #mdb_page_unspill().
1922 * Also note, we never spill DB root pages, nor pages of active cursors,
1923 * because we'll need these back again soon anyway. And in nested txns,
1924 * we can't spill a page in a child txn if it was already spilled in a
1925 * parent txn. That would alter the parent txns' data even though
1926 * the child hasn't committed yet, and we'd have no way to undo it if
1927 * the child aborted.
1929 * @param[in] m0 cursor A cursor handle identifying the transaction and
1930 * database for which we are checking space.
1931 * @param[in] key For a put operation, the key being stored.
1932 * @param[in] data For a put operation, the data being stored.
1933 * @return 0 on success, non-zero on failure.
1936 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1938 MDB_txn *txn = m0->mc_txn;
1940 MDB_ID2L dl = txn->mt_u.dirty_list;
1941 unsigned int i, j, need;
1944 if (m0->mc_flags & C_SUB)
1947 /* Estimate how much space this op will take */
1948 i = m0->mc_db->md_depth;
1949 /* Named DBs also dirty the main DB */
1950 if (m0->mc_dbi >= CORE_DBS)
1951 i += txn->mt_dbs[MAIN_DBI].md_depth;
1952 /* For puts, roughly factor in the key+data size */
1954 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1955 i += i; /* double it for good measure */
1958 if (txn->mt_dirty_room > i)
1961 if (!txn->mt_spill_pgs) {
1962 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1963 if (!txn->mt_spill_pgs)
1966 /* purge deleted slots */
1967 MDB_IDL sl = txn->mt_spill_pgs;
1968 unsigned int num = sl[0];
1970 for (i=1; i<=num; i++) {
1977 /* Preserve pages which may soon be dirtied again */
1978 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1981 /* Less aggressive spill - we originally spilled the entire dirty list,
1982 * with a few exceptions for cursor pages and DB root pages. But this
1983 * turns out to be a lot of wasted effort because in a large txn many
1984 * of those pages will need to be used again. So now we spill only 1/8th
1985 * of the dirty pages. Testing revealed this to be a good tradeoff,
1986 * better than 1/2, 1/4, or 1/10.
1988 if (need < MDB_IDL_UM_MAX / 8)
1989 need = MDB_IDL_UM_MAX / 8;
1991 /* Save the page IDs of all the pages we're flushing */
1992 /* flush from the tail forward, this saves a lot of shifting later on. */
1993 for (i=dl[0].mid; i && need; i--) {
1994 MDB_ID pn = dl[i].mid << 1;
1996 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1998 /* Can't spill twice, make sure it's not already in a parent's
2001 if (txn->mt_parent) {
2003 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2004 if (tx2->mt_spill_pgs) {
2005 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2006 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2007 dp->mp_flags |= P_KEEP;
2015 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2019 mdb_midl_sort(txn->mt_spill_pgs);
2021 /* Flush the spilled part of dirty list */
2022 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2025 /* Reset any dirty pages we kept that page_flush didn't see */
2026 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2029 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2033 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2035 mdb_find_oldest(MDB_txn *txn)
2038 txnid_t mr, oldest = txn->mt_txnid - 1;
2039 if (txn->mt_env->me_txns) {
2040 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2041 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2052 /** Add a page to the txn's dirty list */
2054 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2057 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2059 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2060 insert = mdb_mid2l_append;
2062 insert = mdb_mid2l_insert;
2064 mid.mid = mp->mp_pgno;
2066 rc = insert(txn->mt_u.dirty_list, &mid);
2067 mdb_tassert(txn, rc == 0);
2068 txn->mt_dirty_room--;
2071 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2072 * me_pghead and mt_next_pgno.
2074 * If there are free pages available from older transactions, they
2075 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2076 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2077 * and move me_pglast to say which records were consumed. Only this
2078 * function can create me_pghead and move me_pglast/mt_next_pgno.
2079 * @param[in] mc cursor A cursor handle identifying the transaction and
2080 * database for which we are allocating.
2081 * @param[in] num the number of pages to allocate.
2082 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2083 * will always be satisfied by a single contiguous chunk of memory.
2084 * @return 0 on success, non-zero on failure.
2087 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2089 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2090 /* Get at most <Max_retries> more freeDB records once me_pghead
2091 * has enough pages. If not enough, use new pages from the map.
2092 * If <Paranoid> and mc is updating the freeDB, only get new
2093 * records if me_pghead is empty. Then the freelist cannot play
2094 * catch-up with itself by growing while trying to save it.
2096 enum { Paranoid = 1, Max_retries = 500 };
2098 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2100 int rc, retry = num * 60;
2101 MDB_txn *txn = mc->mc_txn;
2102 MDB_env *env = txn->mt_env;
2103 pgno_t pgno, *mop = env->me_pghead;
2104 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2106 txnid_t oldest = 0, last;
2111 /* If there are any loose pages, just use them */
2112 if (num == 1 && txn->mt_loose_pgs) {
2113 np = txn->mt_loose_pgs;
2114 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2115 txn->mt_loose_count--;
2116 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2124 /* If our dirty list is already full, we can't do anything */
2125 if (txn->mt_dirty_room == 0) {
2130 for (op = MDB_FIRST;; op = MDB_NEXT) {
2135 /* Seek a big enough contiguous page range. Prefer
2136 * pages at the tail, just truncating the list.
2142 if (mop[i-n2] == pgno+n2)
2149 if (op == MDB_FIRST) { /* 1st iteration */
2150 /* Prepare to fetch more and coalesce */
2151 last = env->me_pglast;
2152 oldest = env->me_pgoldest;
2153 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2156 key.mv_data = &last; /* will look up last+1 */
2157 key.mv_size = sizeof(last);
2159 if (Paranoid && mc->mc_dbi == FREE_DBI)
2162 if (Paranoid && retry < 0 && mop_len)
2166 /* Do not fetch more if the record will be too recent */
2167 if (oldest <= last) {
2169 oldest = mdb_find_oldest(txn);
2170 env->me_pgoldest = oldest;
2176 rc = mdb_cursor_get(&m2, &key, NULL, op);
2178 if (rc == MDB_NOTFOUND)
2182 last = *(txnid_t*)key.mv_data;
2183 if (oldest <= last) {
2185 oldest = mdb_find_oldest(txn);
2186 env->me_pgoldest = oldest;
2192 np = m2.mc_pg[m2.mc_top];
2193 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2194 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2197 idl = (MDB_ID *) data.mv_data;
2200 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2205 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2207 mop = env->me_pghead;
2209 env->me_pglast = last;
2211 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2212 last, txn->mt_dbs[FREE_DBI].md_root, i));
2214 DPRINTF(("IDL %"Z"u", idl[j]));
2216 /* Merge in descending sorted order */
2217 mdb_midl_xmerge(mop, idl);
2221 /* Use new pages from the map when nothing suitable in the freeDB */
2223 pgno = txn->mt_next_pgno;
2224 if (pgno + num >= env->me_maxpg) {
2225 DPUTS("DB size maxed out");
2231 if (env->me_flags & MDB_WRITEMAP) {
2232 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2234 if (!(np = mdb_page_malloc(txn, num))) {
2240 mop[0] = mop_len -= num;
2241 /* Move any stragglers down */
2242 for (j = i-num; j < mop_len; )
2243 mop[++j] = mop[++i];
2245 txn->mt_next_pgno = pgno + num;
2248 mdb_page_dirty(txn, np);
2254 txn->mt_flags |= MDB_TXN_ERROR;
2258 /** Copy the used portions of a non-overflow page.
2259 * @param[in] dst page to copy into
2260 * @param[in] src page to copy from
2261 * @param[in] psize size of a page
2264 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2266 enum { Align = sizeof(pgno_t) };
2267 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2269 /* If page isn't full, just copy the used portion. Adjust
2270 * alignment so memcpy may copy words instead of bytes.
2272 if ((unused &= -Align) && !IS_LEAF2(src)) {
2273 upper = (upper + PAGEBASE) & -Align;
2274 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2275 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2278 memcpy(dst, src, psize - unused);
2282 /** Pull a page off the txn's spill list, if present.
2283 * If a page being referenced was spilled to disk in this txn, bring
2284 * it back and make it dirty/writable again.
2285 * @param[in] txn the transaction handle.
2286 * @param[in] mp the page being referenced. It must not be dirty.
2287 * @param[out] ret the writable page, if any. ret is unchanged if
2288 * mp wasn't spilled.
2291 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2293 MDB_env *env = txn->mt_env;
2296 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2298 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2299 if (!tx2->mt_spill_pgs)
2301 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2302 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2305 if (txn->mt_dirty_room == 0)
2306 return MDB_TXN_FULL;
2307 if (IS_OVERFLOW(mp))
2311 if (env->me_flags & MDB_WRITEMAP) {
2314 np = mdb_page_malloc(txn, num);
2318 memcpy(np, mp, num * env->me_psize);
2320 mdb_page_copy(np, mp, env->me_psize);
2323 /* If in current txn, this page is no longer spilled.
2324 * If it happens to be the last page, truncate the spill list.
2325 * Otherwise mark it as deleted by setting the LSB.
2327 if (x == txn->mt_spill_pgs[0])
2328 txn->mt_spill_pgs[0]--;
2330 txn->mt_spill_pgs[x] |= 1;
2331 } /* otherwise, if belonging to a parent txn, the
2332 * page remains spilled until child commits
2335 mdb_page_dirty(txn, np);
2336 np->mp_flags |= P_DIRTY;
2344 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2345 * @param[in] mc cursor pointing to the page to be touched
2346 * @return 0 on success, non-zero on failure.
2349 mdb_page_touch(MDB_cursor *mc)
2351 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2352 MDB_txn *txn = mc->mc_txn;
2353 MDB_cursor *m2, *m3;
2357 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2358 if (txn->mt_flags & MDB_TXN_SPILLS) {
2360 rc = mdb_page_unspill(txn, mp, &np);
2366 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2367 (rc = mdb_page_alloc(mc, 1, &np)))
2370 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2371 mp->mp_pgno, pgno));
2372 mdb_cassert(mc, mp->mp_pgno != pgno);
2373 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2374 /* Update the parent page, if any, to point to the new page */
2376 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2377 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2378 SETPGNO(node, pgno);
2380 mc->mc_db->md_root = pgno;
2382 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2383 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2385 /* If txn has a parent, make sure the page is in our
2389 unsigned x = mdb_mid2l_search(dl, pgno);
2390 if (x <= dl[0].mid && dl[x].mid == pgno) {
2391 if (mp != dl[x].mptr) { /* bad cursor? */
2392 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2393 txn->mt_flags |= MDB_TXN_ERROR;
2394 return MDB_CORRUPTED;
2399 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2401 np = mdb_page_malloc(txn, 1);
2406 rc = mdb_mid2l_insert(dl, &mid);
2407 mdb_cassert(mc, rc == 0);
2412 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2414 np->mp_flags |= P_DIRTY;
2417 /* Adjust cursors pointing to mp */
2418 mc->mc_pg[mc->mc_top] = np;
2419 m2 = txn->mt_cursors[mc->mc_dbi];
2420 if (mc->mc_flags & C_SUB) {
2421 for (; m2; m2=m2->mc_next) {
2422 m3 = &m2->mc_xcursor->mx_cursor;
2423 if (m3->mc_snum < mc->mc_snum) continue;
2424 if (m3->mc_pg[mc->mc_top] == mp)
2425 m3->mc_pg[mc->mc_top] = np;
2428 for (; m2; m2=m2->mc_next) {
2429 if (m2->mc_snum < mc->mc_snum) continue;
2430 if (m2 == mc) continue;
2431 if (m2->mc_pg[mc->mc_top] == mp) {
2432 m2->mc_pg[mc->mc_top] = np;
2433 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2435 (m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
2437 MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
2438 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
2439 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2447 txn->mt_flags |= MDB_TXN_ERROR;
2452 mdb_env_sync(MDB_env *env, int force)
2455 if (env->me_flags & MDB_RDONLY)
2457 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2458 if (env->me_flags & MDB_WRITEMAP) {
2459 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2460 ? MS_ASYNC : MS_SYNC;
2461 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2464 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2468 #ifdef BROKEN_FDATASYNC
2469 if (env->me_flags & MDB_FSYNCONLY) {
2470 if (fsync(env->me_fd))
2474 if (MDB_FDATASYNC(env->me_fd))
2481 /** Back up parent txn's cursors, then grab the originals for tracking */
2483 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2485 MDB_cursor *mc, *bk;
2490 for (i = src->mt_numdbs; --i >= 0; ) {
2491 if ((mc = src->mt_cursors[i]) != NULL) {
2492 size = sizeof(MDB_cursor);
2494 size += sizeof(MDB_xcursor);
2495 for (; mc; mc = bk->mc_next) {
2501 mc->mc_db = &dst->mt_dbs[i];
2502 /* Kill pointers into src to reduce abuse: The
2503 * user may not use mc until dst ends. But we need a valid
2504 * txn pointer here for cursor fixups to keep working.
2507 mc->mc_dbflag = &dst->mt_dbflags[i];
2508 if ((mx = mc->mc_xcursor) != NULL) {
2509 *(MDB_xcursor *)(bk+1) = *mx;
2510 mx->mx_cursor.mc_txn = dst;
2512 mc->mc_next = dst->mt_cursors[i];
2513 dst->mt_cursors[i] = mc;
2520 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2521 * @param[in] txn the transaction handle.
2522 * @param[in] merge true to keep changes to parent cursors, false to revert.
2523 * @return 0 on success, non-zero on failure.
2526 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2528 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2532 for (i = txn->mt_numdbs; --i >= 0; ) {
2533 for (mc = cursors[i]; mc; mc = next) {
2535 if ((bk = mc->mc_backup) != NULL) {
2537 /* Commit changes to parent txn */
2538 mc->mc_next = bk->mc_next;
2539 mc->mc_backup = bk->mc_backup;
2540 mc->mc_txn = bk->mc_txn;
2541 mc->mc_db = bk->mc_db;
2542 mc->mc_dbflag = bk->mc_dbflag;
2543 if ((mx = mc->mc_xcursor) != NULL)
2544 mx->mx_cursor.mc_txn = bk->mc_txn;
2546 /* Abort nested txn */
2548 if ((mx = mc->mc_xcursor) != NULL)
2549 *mx = *(MDB_xcursor *)(bk+1);
2553 /* Only malloced cursors are permanently tracked. */
2560 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2566 Pidset = F_SETLK, Pidcheck = F_GETLK
2570 /** Set or check a pid lock. Set returns 0 on success.
2571 * Check returns 0 if the process is certainly dead, nonzero if it may
2572 * be alive (the lock exists or an error happened so we do not know).
2574 * On Windows Pidset is a no-op, we merely check for the existence
2575 * of the process with the given pid. On POSIX we use a single byte
2576 * lock on the lockfile, set at an offset equal to the pid.
2579 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2581 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2584 if (op == Pidcheck) {
2585 h = OpenProcess(env->me_pidquery, FALSE, pid);
2586 /* No documented "no such process" code, but other program use this: */
2588 return ErrCode() != ERROR_INVALID_PARAMETER;
2589 /* A process exists until all handles to it close. Has it exited? */
2590 ret = WaitForSingleObject(h, 0) != 0;
2597 struct flock lock_info;
2598 memset(&lock_info, 0, sizeof(lock_info));
2599 lock_info.l_type = F_WRLCK;
2600 lock_info.l_whence = SEEK_SET;
2601 lock_info.l_start = pid;
2602 lock_info.l_len = 1;
2603 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2604 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2606 } else if ((rc = ErrCode()) == EINTR) {
2614 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2615 * @param[in] txn the transaction handle to initialize
2616 * @return 0 on success, non-zero on failure.
2619 mdb_txn_renew0(MDB_txn *txn)
2621 MDB_env *env = txn->mt_env;
2622 MDB_txninfo *ti = env->me_txns;
2624 unsigned int i, nr, flags = txn->mt_flags;
2626 int rc, new_notls = 0;
2628 if ((flags &= MDB_TXN_RDONLY) != 0) {
2630 meta = mdb_env_pick_meta(env);
2631 txn->mt_txnid = meta->mm_txnid;
2632 txn->mt_u.reader = NULL;
2634 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2635 pthread_getspecific(env->me_txkey);
2637 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2638 return MDB_BAD_RSLOT;
2640 MDB_PID_T pid = env->me_pid;
2641 MDB_THR_T tid = pthread_self();
2642 mdb_mutexref_t rmutex = env->me_rmutex;
2644 if (!env->me_live_reader) {
2645 rc = mdb_reader_pid(env, Pidset, pid);
2648 env->me_live_reader = 1;
2651 if (LOCK_MUTEX(rc, env, rmutex))
2653 nr = ti->mti_numreaders;
2654 for (i=0; i<nr; i++)
2655 if (ti->mti_readers[i].mr_pid == 0)
2657 if (i == env->me_maxreaders) {
2658 UNLOCK_MUTEX(rmutex);
2659 return MDB_READERS_FULL;
2661 r = &ti->mti_readers[i];
2662 /* Claim the reader slot, carefully since other code
2663 * uses the reader table un-mutexed: First reset the
2664 * slot, next publish it in mti_numreaders. After
2665 * that, it is safe for mdb_env_close() to touch it.
2666 * When it will be closed, we can finally claim it.
2669 r->mr_txnid = (txnid_t)-1;
2672 ti->mti_numreaders = ++nr;
2673 env->me_close_readers = nr;
2675 UNLOCK_MUTEX(rmutex);
2677 new_notls = (env->me_flags & MDB_NOTLS);
2678 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2683 do /* LY: Retry on a race, ITS#7970. */
2684 r->mr_txnid = ti->mti_txnid;
2685 while(r->mr_txnid != ti->mti_txnid);
2686 txn->mt_txnid = r->mr_txnid;
2687 txn->mt_u.reader = r;
2688 meta = env->me_metas[txn->mt_txnid & 1];
2692 /* Not yet touching txn == env->me_txn0, it may be active */
2694 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2696 txn->mt_txnid = ti->mti_txnid;
2697 meta = env->me_metas[txn->mt_txnid & 1];
2699 meta = mdb_env_pick_meta(env);
2700 txn->mt_txnid = meta->mm_txnid;
2704 if (txn->mt_txnid == mdb_debug_start)
2707 txn->mt_child = NULL;
2708 txn->mt_loose_pgs = NULL;
2709 txn->mt_loose_count = 0;
2710 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2711 txn->mt_u.dirty_list = env->me_dirty_list;
2712 txn->mt_u.dirty_list[0].mid = 0;
2713 txn->mt_free_pgs = env->me_free_pgs;
2714 txn->mt_free_pgs[0] = 0;
2715 txn->mt_spill_pgs = NULL;
2717 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2720 /* Copy the DB info and flags */
2721 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2723 /* Moved to here to avoid a data race in read TXNs */
2724 txn->mt_next_pgno = meta->mm_last_pg+1;
2726 txn->mt_flags = flags;
2729 txn->mt_numdbs = env->me_numdbs;
2730 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2731 x = env->me_dbflags[i];
2732 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2733 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2735 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2736 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2738 if (env->me_flags & MDB_FATAL_ERROR) {
2739 DPUTS("environment had fatal error, must shutdown!");
2741 } else if (env->me_maxpg < txn->mt_next_pgno) {
2742 rc = MDB_MAP_RESIZED;
2746 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2751 mdb_txn_renew(MDB_txn *txn)
2755 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2758 rc = mdb_txn_renew0(txn);
2759 if (rc == MDB_SUCCESS) {
2760 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2761 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2762 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2768 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2772 int rc, size, tsize;
2774 flags &= MDB_TXN_BEGIN_FLAGS;
2775 flags |= env->me_flags & MDB_WRITEMAP;
2777 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2781 /* Nested transactions: Max 1 child, write txns only, no writemap */
2782 flags |= parent->mt_flags;
2783 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2784 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2786 /* Child txns save MDB_pgstate and use own copy of cursors */
2787 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2788 size += tsize = sizeof(MDB_ntxn);
2789 } else if (flags & MDB_RDONLY) {
2790 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2791 size += tsize = sizeof(MDB_txn);
2793 /* Reuse preallocated write txn. However, do not touch it until
2794 * mdb_txn_renew0() succeeds, since it currently may be active.
2799 if ((txn = calloc(1, size)) == NULL) {
2800 DPRINTF(("calloc: %s", strerror(errno)));
2803 txn->mt_dbxs = env->me_dbxs; /* static */
2804 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2805 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2806 txn->mt_flags = flags;
2811 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2812 txn->mt_dbiseqs = parent->mt_dbiseqs;
2813 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2814 if (!txn->mt_u.dirty_list ||
2815 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2817 free(txn->mt_u.dirty_list);
2821 txn->mt_txnid = parent->mt_txnid;
2822 txn->mt_dirty_room = parent->mt_dirty_room;
2823 txn->mt_u.dirty_list[0].mid = 0;
2824 txn->mt_spill_pgs = NULL;
2825 txn->mt_next_pgno = parent->mt_next_pgno;
2826 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2827 parent->mt_child = txn;
2828 txn->mt_parent = parent;
2829 txn->mt_numdbs = parent->mt_numdbs;
2830 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2831 /* Copy parent's mt_dbflags, but clear DB_NEW */
2832 for (i=0; i<txn->mt_numdbs; i++)
2833 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2835 ntxn = (MDB_ntxn *)txn;
2836 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2837 if (env->me_pghead) {
2838 size = MDB_IDL_SIZEOF(env->me_pghead);
2839 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2841 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2846 rc = mdb_cursor_shadow(parent, txn);
2848 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2849 } else { /* MDB_RDONLY */
2850 txn->mt_dbiseqs = env->me_dbiseqs;
2852 rc = mdb_txn_renew0(txn);
2855 if (txn != env->me_txn0)
2858 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2860 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2861 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2862 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2869 mdb_txn_env(MDB_txn *txn)
2871 if(!txn) return NULL;
2876 mdb_txn_id(MDB_txn *txn)
2879 return txn->mt_txnid;
2882 /** Export or close DBI handles opened in this txn. */
2884 mdb_dbis_update(MDB_txn *txn, int keep)
2887 MDB_dbi n = txn->mt_numdbs;
2888 MDB_env *env = txn->mt_env;
2889 unsigned char *tdbflags = txn->mt_dbflags;
2891 for (i = n; --i >= CORE_DBS;) {
2892 if (tdbflags[i] & DB_NEW) {
2894 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2896 char *ptr = env->me_dbxs[i].md_name.mv_data;
2898 env->me_dbxs[i].md_name.mv_data = NULL;
2899 env->me_dbxs[i].md_name.mv_size = 0;
2900 env->me_dbflags[i] = 0;
2901 env->me_dbiseqs[i]++;
2907 if (keep && env->me_numdbs < n)
2911 /** End a transaction, except successful commit of a nested transaction.
2912 * May be called twice for readonly txns: First reset it, then abort.
2913 * @param[in] txn the transaction handle to end
2914 * @param[in] mode why and how to end the transaction
2917 mdb_txn_end(MDB_txn *txn, unsigned mode)
2919 MDB_env *env = txn->mt_env;
2921 static const char *const names[] = MDB_END_NAMES;
2924 /* Export or close DBI handles opened in this txn */
2925 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2927 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2928 names[mode & MDB_END_OPMASK],
2929 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2930 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2932 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2933 if (txn->mt_u.reader) {
2934 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2935 if (!(env->me_flags & MDB_NOTLS)) {
2936 txn->mt_u.reader = NULL; /* txn does not own reader */
2937 } else if (mode & MDB_END_SLOT) {
2938 txn->mt_u.reader->mr_pid = 0;
2939 txn->mt_u.reader = NULL;
2940 } /* else txn owns the slot until it does MDB_END_SLOT */
2942 txn->mt_numdbs = 0; /* prevent further DBI activity */
2943 txn->mt_flags |= MDB_TXN_FINISHED;
2945 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2946 pgno_t *pghead = env->me_pghead;
2948 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2949 mdb_cursors_close(txn, 0);
2950 if (!(env->me_flags & MDB_WRITEMAP)) {
2951 mdb_dlist_free(txn);
2955 txn->mt_flags = MDB_TXN_FINISHED;
2957 if (!txn->mt_parent) {
2958 mdb_midl_shrink(&txn->mt_free_pgs);
2959 env->me_free_pgs = txn->mt_free_pgs;
2961 env->me_pghead = NULL;
2965 mode = 0; /* txn == env->me_txn0, do not free() it */
2967 /* The writer mutex was locked in mdb_txn_begin. */
2969 UNLOCK_MUTEX(env->me_wmutex);
2971 txn->mt_parent->mt_child = NULL;
2972 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
2973 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2974 mdb_midl_free(txn->mt_free_pgs);
2975 mdb_midl_free(txn->mt_spill_pgs);
2976 free(txn->mt_u.dirty_list);
2979 mdb_midl_free(pghead);
2982 if (mode & MDB_END_FREE)
2987 mdb_txn_reset(MDB_txn *txn)
2992 /* This call is only valid for read-only txns */
2993 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2996 mdb_txn_end(txn, MDB_END_RESET);
3000 mdb_txn_abort(MDB_txn *txn)
3006 mdb_txn_abort(txn->mt_child);
3008 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3011 /** Save the freelist as of this transaction to the freeDB.
3012 * This changes the freelist. Keep trying until it stabilizes.
3015 mdb_freelist_save(MDB_txn *txn)
3017 /* env->me_pghead[] can grow and shrink during this call.
3018 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3019 * Page numbers cannot disappear from txn->mt_free_pgs[].
3022 MDB_env *env = txn->mt_env;
3023 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3024 txnid_t pglast = 0, head_id = 0;
3025 pgno_t freecnt = 0, *free_pgs, *mop;
3026 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3028 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3030 if (env->me_pghead) {
3031 /* Make sure first page of freeDB is touched and on freelist */
3032 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3033 if (rc && rc != MDB_NOTFOUND)
3037 if (!env->me_pghead && txn->mt_loose_pgs) {
3038 /* Put loose page numbers in mt_free_pgs, since
3039 * we may be unable to return them to me_pghead.
3041 MDB_page *mp = txn->mt_loose_pgs;
3042 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3044 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3045 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3046 txn->mt_loose_pgs = NULL;
3047 txn->mt_loose_count = 0;
3050 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3051 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3052 ? SSIZE_MAX : maxfree_1pg;
3055 /* Come back here after each Put() in case freelist changed */
3060 /* If using records from freeDB which we have not yet
3061 * deleted, delete them and any we reserved for me_pghead.
3063 while (pglast < env->me_pglast) {
3064 rc = mdb_cursor_first(&mc, &key, NULL);
3067 pglast = head_id = *(txnid_t *)key.mv_data;
3068 total_room = head_room = 0;
3069 mdb_tassert(txn, pglast <= env->me_pglast);
3070 rc = mdb_cursor_del(&mc, 0);
3075 /* Save the IDL of pages freed by this txn, to a single record */
3076 if (freecnt < txn->mt_free_pgs[0]) {
3078 /* Make sure last page of freeDB is touched and on freelist */
3079 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3080 if (rc && rc != MDB_NOTFOUND)
3083 free_pgs = txn->mt_free_pgs;
3084 /* Write to last page of freeDB */
3085 key.mv_size = sizeof(txn->mt_txnid);
3086 key.mv_data = &txn->mt_txnid;
3088 freecnt = free_pgs[0];
3089 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3090 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3093 /* Retry if mt_free_pgs[] grew during the Put() */
3094 free_pgs = txn->mt_free_pgs;
3095 } while (freecnt < free_pgs[0]);
3096 mdb_midl_sort(free_pgs);
3097 memcpy(data.mv_data, free_pgs, data.mv_size);
3100 unsigned int i = free_pgs[0];
3101 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3102 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3104 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3110 mop = env->me_pghead;
3111 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3113 /* Reserve records for me_pghead[]. Split it if multi-page,
3114 * to avoid searching freeDB for a page range. Use keys in
3115 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3117 if (total_room >= mop_len) {
3118 if (total_room == mop_len || --more < 0)
3120 } else if (head_room >= maxfree_1pg && head_id > 1) {
3121 /* Keep current record (overflow page), add a new one */
3125 /* (Re)write {key = head_id, IDL length = head_room} */
3126 total_room -= head_room;
3127 head_room = mop_len - total_room;
3128 if (head_room > maxfree_1pg && head_id > 1) {
3129 /* Overflow multi-page for part of me_pghead */
3130 head_room /= head_id; /* amortize page sizes */
3131 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3132 } else if (head_room < 0) {
3133 /* Rare case, not bothering to delete this record */
3136 key.mv_size = sizeof(head_id);
3137 key.mv_data = &head_id;
3138 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3139 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3142 /* IDL is initially empty, zero out at least the length */
3143 pgs = (pgno_t *)data.mv_data;
3144 j = head_room > clean_limit ? head_room : 0;
3148 total_room += head_room;
3151 /* Return loose page numbers to me_pghead, though usually none are
3152 * left at this point. The pages themselves remain in dirty_list.
3154 if (txn->mt_loose_pgs) {
3155 MDB_page *mp = txn->mt_loose_pgs;
3156 unsigned count = txn->mt_loose_count;
3158 /* Room for loose pages + temp IDL with same */
3159 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3161 mop = env->me_pghead;
3162 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3163 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3164 loose[ ++count ] = mp->mp_pgno;
3166 mdb_midl_sort(loose);
3167 mdb_midl_xmerge(mop, loose);
3168 txn->mt_loose_pgs = NULL;
3169 txn->mt_loose_count = 0;
3173 /* Fill in the reserved me_pghead records */
3179 rc = mdb_cursor_first(&mc, &key, &data);
3180 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3181 txnid_t id = *(txnid_t *)key.mv_data;
3182 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3185 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3187 if (len > mop_len) {
3189 data.mv_size = (len + 1) * sizeof(MDB_ID);
3191 data.mv_data = mop -= len;
3194 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3196 if (rc || !(mop_len -= len))
3203 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3204 * @param[in] txn the transaction that's being committed
3205 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3206 * @return 0 on success, non-zero on failure.
3209 mdb_page_flush(MDB_txn *txn, int keep)
3211 MDB_env *env = txn->mt_env;
3212 MDB_ID2L dl = txn->mt_u.dirty_list;
3213 unsigned psize = env->me_psize, j;
3214 int i, pagecount = dl[0].mid, rc;
3215 size_t size = 0, pos = 0;
3217 MDB_page *dp = NULL;
3221 struct iovec iov[MDB_COMMIT_PAGES];
3222 ssize_t wpos = 0, wsize = 0, wres;
3223 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3229 if (env->me_flags & MDB_WRITEMAP) {
3230 /* Clear dirty flags */
3231 while (++i <= pagecount) {
3233 /* Don't flush this page yet */
3234 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3235 dp->mp_flags &= ~P_KEEP;
3239 dp->mp_flags &= ~P_DIRTY;
3244 /* Write the pages */
3246 if (++i <= pagecount) {
3248 /* Don't flush this page yet */
3249 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3250 dp->mp_flags &= ~P_KEEP;
3255 /* clear dirty flag */
3256 dp->mp_flags &= ~P_DIRTY;
3259 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3264 /* Windows actually supports scatter/gather I/O, but only on
3265 * unbuffered file handles. Since we're relying on the OS page
3266 * cache for all our data, that's self-defeating. So we just
3267 * write pages one at a time. We use the ov structure to set
3268 * the write offset, to at least save the overhead of a Seek
3271 DPRINTF(("committing page %"Z"u", pgno));
3272 memset(&ov, 0, sizeof(ov));
3273 ov.Offset = pos & 0xffffffff;
3274 ov.OffsetHigh = pos >> 16 >> 16;
3275 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3277 DPRINTF(("WriteFile: %d", rc));
3281 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3282 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3285 /* Write previous page(s) */
3286 #ifdef MDB_USE_PWRITEV
3287 wres = pwritev(env->me_fd, iov, n, wpos);
3290 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3293 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3297 DPRINTF(("lseek: %s", strerror(rc)));
3300 wres = writev(env->me_fd, iov, n);
3303 if (wres != wsize) {
3308 DPRINTF(("Write error: %s", strerror(rc)));
3310 rc = EIO; /* TODO: Use which error code? */
3311 DPUTS("short write, filesystem full?");
3322 DPRINTF(("committing page %"Z"u", pgno));
3323 next_pos = pos + size;
3324 iov[n].iov_len = size;
3325 iov[n].iov_base = (char *)dp;
3331 /* MIPS has cache coherency issues, this is a no-op everywhere else
3332 * Note: for any size >= on-chip cache size, entire on-chip cache is
3335 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3337 for (i = keep; ++i <= pagecount; ) {
3339 /* This is a page we skipped above */
3342 dl[j].mid = dp->mp_pgno;
3345 mdb_dpage_free(env, dp);
3350 txn->mt_dirty_room += i - j;
3356 mdb_txn_commit(MDB_txn *txn)
3359 unsigned int i, end_mode;
3365 /* mdb_txn_end() mode for a commit which writes nothing */
3366 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3368 if (txn->mt_child) {
3369 rc = mdb_txn_commit(txn->mt_child);
3376 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3380 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3381 DPUTS("txn has failed/finished, can't commit");
3383 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3388 if (txn->mt_parent) {
3389 MDB_txn *parent = txn->mt_parent;
3393 unsigned x, y, len, ps_len;
3395 /* Append our free list to parent's */
3396 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3399 mdb_midl_free(txn->mt_free_pgs);
3400 /* Failures after this must either undo the changes
3401 * to the parent or set MDB_TXN_ERROR in the parent.
3404 parent->mt_next_pgno = txn->mt_next_pgno;
3405 parent->mt_flags = txn->mt_flags;
3407 /* Merge our cursors into parent's and close them */
3408 mdb_cursors_close(txn, 1);
3410 /* Update parent's DB table. */
3411 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3412 parent->mt_numdbs = txn->mt_numdbs;
3413 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3414 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3415 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3416 /* preserve parent's DB_NEW status */
3417 x = parent->mt_dbflags[i] & DB_NEW;
3418 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3421 dst = parent->mt_u.dirty_list;
3422 src = txn->mt_u.dirty_list;
3423 /* Remove anything in our dirty list from parent's spill list */
3424 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3426 pspill[0] = (pgno_t)-1;
3427 /* Mark our dirty pages as deleted in parent spill list */
3428 for (i=0, len=src[0].mid; ++i <= len; ) {
3429 MDB_ID pn = src[i].mid << 1;
3430 while (pn > pspill[x])
3432 if (pn == pspill[x]) {
3437 /* Squash deleted pagenums if we deleted any */
3438 for (x=y; ++x <= ps_len; )
3439 if (!(pspill[x] & 1))
3440 pspill[++y] = pspill[x];
3444 /* Remove anything in our spill list from parent's dirty list */
3445 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3446 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3447 MDB_ID pn = txn->mt_spill_pgs[i];
3449 continue; /* deleted spillpg */
3451 y = mdb_mid2l_search(dst, pn);
3452 if (y <= dst[0].mid && dst[y].mid == pn) {
3454 while (y < dst[0].mid) {
3463 /* Find len = length of merging our dirty list with parent's */
3465 dst[0].mid = 0; /* simplify loops */
3466 if (parent->mt_parent) {
3467 len = x + src[0].mid;
3468 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3469 for (i = x; y && i; y--) {
3470 pgno_t yp = src[y].mid;
3471 while (yp < dst[i].mid)
3473 if (yp == dst[i].mid) {
3478 } else { /* Simplify the above for single-ancestor case */
3479 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3481 /* Merge our dirty list with parent's */
3483 for (i = len; y; dst[i--] = src[y--]) {
3484 pgno_t yp = src[y].mid;
3485 while (yp < dst[x].mid)
3486 dst[i--] = dst[x--];
3487 if (yp == dst[x].mid)
3488 free(dst[x--].mptr);
3490 mdb_tassert(txn, i == x);
3492 free(txn->mt_u.dirty_list);
3493 parent->mt_dirty_room = txn->mt_dirty_room;
3494 if (txn->mt_spill_pgs) {
3495 if (parent->mt_spill_pgs) {
3496 /* TODO: Prevent failure here, so parent does not fail */
3497 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3499 parent->mt_flags |= MDB_TXN_ERROR;
3500 mdb_midl_free(txn->mt_spill_pgs);
3501 mdb_midl_sort(parent->mt_spill_pgs);
3503 parent->mt_spill_pgs = txn->mt_spill_pgs;
3507 /* Append our loose page list to parent's */
3508 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3510 *lp = txn->mt_loose_pgs;
3511 parent->mt_loose_count += txn->mt_loose_count;
3513 parent->mt_child = NULL;
3514 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3519 if (txn != env->me_txn) {
3520 DPUTS("attempt to commit unknown transaction");
3525 mdb_cursors_close(txn, 0);
3527 if (!txn->mt_u.dirty_list[0].mid &&
3528 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3531 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3532 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3534 /* Update DB root pointers */
3535 if (txn->mt_numdbs > CORE_DBS) {
3539 data.mv_size = sizeof(MDB_db);
3541 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3542 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3543 if (txn->mt_dbflags[i] & DB_DIRTY) {
3544 if (TXN_DBI_CHANGED(txn, i)) {
3548 data.mv_data = &txn->mt_dbs[i];
3549 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3557 rc = mdb_freelist_save(txn);
3561 mdb_midl_free(env->me_pghead);
3562 env->me_pghead = NULL;
3563 mdb_midl_shrink(&txn->mt_free_pgs);
3569 if ((rc = mdb_page_flush(txn, 0)) ||
3570 (rc = mdb_env_sync(env, 0)) ||
3571 (rc = mdb_env_write_meta(txn)))
3573 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3576 mdb_txn_end(txn, end_mode);
3584 /** Read the environment parameters of a DB environment before
3585 * mapping it into memory.
3586 * @param[in] env the environment handle
3587 * @param[out] meta address of where to store the meta information
3588 * @return 0 on success, non-zero on failure.
3591 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3597 enum { Size = sizeof(pbuf) };
3599 /* We don't know the page size yet, so use a minimum value.
3600 * Read both meta pages so we can use the latest one.
3603 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3607 memset(&ov, 0, sizeof(ov));
3609 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3610 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3613 rc = pread(env->me_fd, &pbuf, Size, off);
3616 if (rc == 0 && off == 0)
3618 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3619 DPRINTF(("read: %s", mdb_strerror(rc)));
3623 p = (MDB_page *)&pbuf;
3625 if (!F_ISSET(p->mp_flags, P_META)) {
3626 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3631 if (m->mm_magic != MDB_MAGIC) {
3632 DPUTS("meta has invalid magic");
3636 if (m->mm_version != MDB_DATA_VERSION) {
3637 DPRINTF(("database is version %u, expected version %u",
3638 m->mm_version, MDB_DATA_VERSION));
3639 return MDB_VERSION_MISMATCH;
3642 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3648 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3650 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3652 meta->mm_magic = MDB_MAGIC;
3653 meta->mm_version = MDB_DATA_VERSION;
3654 meta->mm_mapsize = env->me_mapsize;
3655 meta->mm_psize = env->me_psize;
3656 meta->mm_last_pg = NUM_METAS-1;
3657 meta->mm_flags = env->me_flags & 0xffff;
3658 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3659 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3660 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3663 /** Write the environment parameters of a freshly created DB environment.
3664 * @param[in] env the environment handle
3665 * @param[in] meta the #MDB_meta to write
3666 * @return 0 on success, non-zero on failure.
3669 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3677 memset(&ov, 0, sizeof(ov));
3678 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3680 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3683 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3684 len = pwrite(fd, ptr, size, pos); \
3685 if (len == -1 && ErrCode() == EINTR) continue; \
3686 rc = (len >= 0); break; } while(1)
3689 DPUTS("writing new meta page");
3691 psize = env->me_psize;
3693 p = calloc(NUM_METAS, psize);
3698 p->mp_flags = P_META;
3699 *(MDB_meta *)METADATA(p) = *meta;
3701 q = (MDB_page *)((char *)p + psize);
3703 q->mp_flags = P_META;
3704 *(MDB_meta *)METADATA(q) = *meta;
3706 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3709 else if ((unsigned) len == psize * NUM_METAS)
3717 /** Update the environment info to commit a transaction.
3718 * @param[in] txn the transaction that's being committed
3719 * @return 0 on success, non-zero on failure.
3722 mdb_env_write_meta(MDB_txn *txn)
3725 MDB_meta meta, metab, *mp;
3729 int rc, len, toggle;
3738 toggle = txn->mt_txnid & 1;
3739 DPRINTF(("writing meta page %d for root page %"Z"u",
3740 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3743 flags = env->me_flags;
3744 mp = env->me_metas[toggle];
3745 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3746 /* Persist any increases of mapsize config */
3747 if (mapsize < env->me_mapsize)
3748 mapsize = env->me_mapsize;
3750 if (flags & MDB_WRITEMAP) {
3751 mp->mm_mapsize = mapsize;
3752 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3753 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3754 mp->mm_last_pg = txn->mt_next_pgno - 1;
3755 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3756 !(defined(__i386__) || defined(__x86_64__))
3757 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3758 __sync_synchronize();
3760 mp->mm_txnid = txn->mt_txnid;
3761 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3762 unsigned meta_size = env->me_psize;
3763 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3764 ptr = (char *)mp - PAGEHDRSZ;
3765 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3766 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3770 if (MDB_MSYNC(ptr, meta_size, rc)) {
3777 metab.mm_txnid = mp->mm_txnid;
3778 metab.mm_last_pg = mp->mm_last_pg;
3780 meta.mm_mapsize = mapsize;
3781 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3782 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3783 meta.mm_last_pg = txn->mt_next_pgno - 1;
3784 meta.mm_txnid = txn->mt_txnid;
3786 off = offsetof(MDB_meta, mm_mapsize);
3787 ptr = (char *)&meta + off;
3788 len = sizeof(MDB_meta) - off;
3789 off += (char *)mp - env->me_map;
3791 /* Write to the SYNC fd */
3792 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3795 memset(&ov, 0, sizeof(ov));
3797 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3802 rc = pwrite(mfd, ptr, len, off);
3805 rc = rc < 0 ? ErrCode() : EIO;
3810 DPUTS("write failed, disk error?");
3811 /* On a failure, the pagecache still contains the new data.
3812 * Write some old data back, to prevent it from being used.
3813 * Use the non-SYNC fd; we know it will fail anyway.
3815 meta.mm_last_pg = metab.mm_last_pg;
3816 meta.mm_txnid = metab.mm_txnid;
3818 memset(&ov, 0, sizeof(ov));
3820 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3822 r2 = pwrite(env->me_fd, ptr, len, off);
3823 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3826 env->me_flags |= MDB_FATAL_ERROR;
3829 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3830 CACHEFLUSH(env->me_map + off, len, DCACHE);
3832 /* Memory ordering issues are irrelevant; since the entire writer
3833 * is wrapped by wmutex, all of these changes will become visible
3834 * after the wmutex is unlocked. Since the DB is multi-version,
3835 * readers will get consistent data regardless of how fresh or
3836 * how stale their view of these values is.
3839 env->me_txns->mti_txnid = txn->mt_txnid;
3844 /** Check both meta pages to see which one is newer.
3845 * @param[in] env the environment handle
3846 * @return newest #MDB_meta.
3849 mdb_env_pick_meta(const MDB_env *env)
3851 MDB_meta *const *metas = env->me_metas;
3852 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3856 mdb_env_create(MDB_env **env)
3860 e = calloc(1, sizeof(MDB_env));
3864 e->me_maxreaders = DEFAULT_READERS;
3865 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3866 e->me_fd = INVALID_HANDLE_VALUE;
3867 e->me_lfd = INVALID_HANDLE_VALUE;
3868 e->me_mfd = INVALID_HANDLE_VALUE;
3869 #ifdef MDB_USE_POSIX_SEM
3870 e->me_rmutex = SEM_FAILED;
3871 e->me_wmutex = SEM_FAILED;
3873 e->me_pid = getpid();
3874 GET_PAGESIZE(e->me_os_psize);
3875 VGMEMP_CREATE(e,0,0);
3881 mdb_env_map(MDB_env *env, void *addr)
3884 unsigned int flags = env->me_flags;
3888 LONG sizelo, sizehi;
3891 if (flags & MDB_RDONLY) {
3892 /* Don't set explicit map size, use whatever exists */
3897 msize = env->me_mapsize;
3898 sizelo = msize & 0xffffffff;
3899 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3901 /* Windows won't create mappings for zero length files.
3902 * and won't map more than the file size.
3903 * Just set the maxsize right now.
3905 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3906 || !SetEndOfFile(env->me_fd)
3907 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3911 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3912 PAGE_READWRITE : PAGE_READONLY,
3913 sizehi, sizelo, NULL);
3916 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3917 FILE_MAP_WRITE : FILE_MAP_READ,
3919 rc = env->me_map ? 0 : ErrCode();
3924 int prot = PROT_READ;
3925 if (flags & MDB_WRITEMAP) {
3927 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3930 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3932 if (env->me_map == MAP_FAILED) {
3937 if (flags & MDB_NORDAHEAD) {
3938 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3940 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3942 #ifdef POSIX_MADV_RANDOM
3943 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3944 #endif /* POSIX_MADV_RANDOM */
3945 #endif /* MADV_RANDOM */
3949 /* Can happen because the address argument to mmap() is just a
3950 * hint. mmap() can pick another, e.g. if the range is in use.
3951 * The MAP_FIXED flag would prevent that, but then mmap could
3952 * instead unmap existing pages to make room for the new map.
3954 if (addr && env->me_map != addr)
3955 return EBUSY; /* TODO: Make a new MDB_* error code? */
3957 p = (MDB_page *)env->me_map;
3958 env->me_metas[0] = METADATA(p);
3959 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3965 mdb_env_set_mapsize(MDB_env *env, size_t size)
3967 /* If env is already open, caller is responsible for making
3968 * sure there are no active txns.
3976 meta = mdb_env_pick_meta(env);
3978 size = meta->mm_mapsize;
3980 /* Silently round up to minimum if the size is too small */
3981 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3985 munmap(env->me_map, env->me_mapsize);
3986 env->me_mapsize = size;
3987 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3988 rc = mdb_env_map(env, old);
3992 env->me_mapsize = size;
3994 env->me_maxpg = env->me_mapsize / env->me_psize;
3999 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4003 env->me_maxdbs = dbs + CORE_DBS;
4008 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4010 if (env->me_map || readers < 1)
4012 env->me_maxreaders = readers;
4017 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4019 if (!env || !readers)
4021 *readers = env->me_maxreaders;
4026 mdb_fsize(HANDLE fd, size_t *size)
4029 LARGE_INTEGER fsize;
4031 if (!GetFileSizeEx(fd, &fsize))
4034 *size = fsize.QuadPart;
4046 #ifdef BROKEN_FDATASYNC
4047 #include <sys/utsname.h>
4048 #include <sys/vfs.h>
4051 /** Further setup required for opening an LMDB environment
4054 mdb_env_open2(MDB_env *env)
4056 unsigned int flags = env->me_flags;
4057 int i, newenv = 0, rc;
4061 /* See if we should use QueryLimited */
4063 if ((rc & 0xff) > 5)
4064 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4066 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4069 #ifdef BROKEN_FDATASYNC
4070 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4071 * https://lkml.org/lkml/2012/9/3/83
4072 * Kernels after 3.6-rc6 are known good.
4073 * https://lkml.org/lkml/2012/9/10/556
4074 * See if the DB is on ext3/ext4, then check for new enough kernel
4075 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4080 fstatfs(env->me_fd, &st);
4081 while (st.f_type == 0xEF53) {
4085 if (uts.release[0] < '3') {
4086 if (!strncmp(uts.release, "2.6.32.", 7)) {
4087 i = atoi(uts.release+7);
4089 break; /* 2.6.32.60 and newer is OK */
4090 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4091 i = atoi(uts.release+7);
4093 break; /* 2.6.34.15 and newer is OK */
4095 } else if (uts.release[0] == '3') {
4096 i = atoi(uts.release+2);
4098 break; /* 3.6 and newer is OK */
4100 i = atoi(uts.release+4);
4102 break; /* 3.5.4 and newer is OK */
4103 } else if (i == 2) {
4104 i = atoi(uts.release+4);
4106 break; /* 3.2.30 and newer is OK */
4108 } else { /* 4.x and newer is OK */
4111 env->me_flags |= MDB_FSYNCONLY;
4117 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4120 DPUTS("new mdbenv");
4122 env->me_psize = env->me_os_psize;
4123 if (env->me_psize > MAX_PAGESIZE)
4124 env->me_psize = MAX_PAGESIZE;
4125 memset(&meta, 0, sizeof(meta));
4126 mdb_env_init_meta0(env, &meta);
4127 meta.mm_mapsize = DEFAULT_MAPSIZE;
4129 env->me_psize = meta.mm_psize;
4132 /* Was a mapsize configured? */
4133 if (!env->me_mapsize) {
4134 env->me_mapsize = meta.mm_mapsize;
4137 /* Make sure mapsize >= committed data size. Even when using
4138 * mm_mapsize, which could be broken in old files (ITS#7789).
4140 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4141 if (env->me_mapsize < minsize)
4142 env->me_mapsize = minsize;
4144 meta.mm_mapsize = env->me_mapsize;
4146 if (newenv && !(flags & MDB_FIXEDMAP)) {
4147 /* mdb_env_map() may grow the datafile. Write the metapages
4148 * first, so the file will be valid if initialization fails.
4149 * Except with FIXEDMAP, since we do not yet know mm_address.
4150 * We could fill in mm_address later, but then a different
4151 * program might end up doing that - one with a memory layout
4152 * and map address which does not suit the main program.
4154 rc = mdb_env_init_meta(env, &meta);
4160 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4165 if (flags & MDB_FIXEDMAP)
4166 meta.mm_address = env->me_map;
4167 i = mdb_env_init_meta(env, &meta);
4168 if (i != MDB_SUCCESS) {
4173 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4174 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4176 #if !(MDB_MAXKEYSIZE)
4177 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4179 env->me_maxpg = env->me_mapsize / env->me_psize;
4183 MDB_meta *meta = mdb_env_pick_meta(env);
4184 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4186 DPRINTF(("opened database version %u, pagesize %u",
4187 meta->mm_version, env->me_psize));
4188 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4189 DPRINTF(("depth: %u", db->md_depth));
4190 DPRINTF(("entries: %"Z"u", db->md_entries));
4191 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4192 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4193 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4194 DPRINTF(("root: %"Z"u", db->md_root));
4202 /** Release a reader thread's slot in the reader lock table.
4203 * This function is called automatically when a thread exits.
4204 * @param[in] ptr This points to the slot in the reader lock table.
4207 mdb_env_reader_dest(void *ptr)
4209 MDB_reader *reader = ptr;
4215 /** Junk for arranging thread-specific callbacks on Windows. This is
4216 * necessarily platform and compiler-specific. Windows supports up
4217 * to 1088 keys. Let's assume nobody opens more than 64 environments
4218 * in a single process, for now. They can override this if needed.
4220 #ifndef MAX_TLS_KEYS
4221 #define MAX_TLS_KEYS 64
4223 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4224 static int mdb_tls_nkeys;
4226 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4230 case DLL_PROCESS_ATTACH: break;
4231 case DLL_THREAD_ATTACH: break;
4232 case DLL_THREAD_DETACH:
4233 for (i=0; i<mdb_tls_nkeys; i++) {
4234 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4236 mdb_env_reader_dest(r);
4240 case DLL_PROCESS_DETACH: break;
4245 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4247 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4251 /* Force some symbol references.
4252 * _tls_used forces the linker to create the TLS directory if not already done
4253 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4255 #pragma comment(linker, "/INCLUDE:_tls_used")
4256 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4257 #pragma const_seg(".CRT$XLB")
4258 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4259 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4262 #pragma comment(linker, "/INCLUDE:__tls_used")
4263 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4264 #pragma data_seg(".CRT$XLB")
4265 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4267 #endif /* WIN 32/64 */
4268 #endif /* !__GNUC__ */
4271 /** Downgrade the exclusive lock on the region back to shared */
4273 mdb_env_share_locks(MDB_env *env, int *excl)
4276 MDB_meta *meta = mdb_env_pick_meta(env);
4278 env->me_txns->mti_txnid = meta->mm_txnid;
4283 /* First acquire a shared lock. The Unlock will
4284 * then release the existing exclusive lock.
4286 memset(&ov, 0, sizeof(ov));
4287 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4290 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4296 struct flock lock_info;
4297 /* The shared lock replaces the existing lock */
4298 memset((void *)&lock_info, 0, sizeof(lock_info));
4299 lock_info.l_type = F_RDLCK;
4300 lock_info.l_whence = SEEK_SET;
4301 lock_info.l_start = 0;
4302 lock_info.l_len = 1;
4303 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4304 (rc = ErrCode()) == EINTR) ;
4305 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4312 /** Try to get exclusive lock, otherwise shared.
4313 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4316 mdb_env_excl_lock(MDB_env *env, int *excl)
4320 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4324 memset(&ov, 0, sizeof(ov));
4325 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4332 struct flock lock_info;
4333 memset((void *)&lock_info, 0, sizeof(lock_info));
4334 lock_info.l_type = F_WRLCK;
4335 lock_info.l_whence = SEEK_SET;
4336 lock_info.l_start = 0;
4337 lock_info.l_len = 1;
4338 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4339 (rc = ErrCode()) == EINTR) ;
4343 # ifndef MDB_USE_POSIX_MUTEX
4344 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4347 lock_info.l_type = F_RDLCK;
4348 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4349 (rc = ErrCode()) == EINTR) ;
4359 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4361 * @(#) $Revision: 5.1 $
4362 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4363 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4365 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4369 * Please do not copyright this code. This code is in the public domain.
4371 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4372 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4373 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4374 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4375 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4376 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4377 * PERFORMANCE OF THIS SOFTWARE.
4380 * chongo <Landon Curt Noll> /\oo/\
4381 * http://www.isthe.com/chongo/
4383 * Share and Enjoy! :-)
4386 typedef unsigned long long mdb_hash_t;
4387 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4389 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4390 * @param[in] val value to hash
4391 * @param[in] hval initial value for hash
4392 * @return 64 bit hash
4394 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4395 * hval arg on the first call.
4398 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4400 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4401 unsigned char *end = s + val->mv_size;
4403 * FNV-1a hash each octet of the string
4406 /* xor the bottom with the current octet */
4407 hval ^= (mdb_hash_t)*s++;
4409 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4410 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4411 (hval << 7) + (hval << 8) + (hval << 40);
4413 /* return our new hash value */
4417 /** Hash the string and output the encoded hash.
4418 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4419 * very short name limits. We don't care about the encoding being reversible,
4420 * we just want to preserve as many bits of the input as possible in a
4421 * small printable string.
4422 * @param[in] str string to hash
4423 * @param[out] encbuf an array of 11 chars to hold the hash
4425 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4428 mdb_pack85(unsigned long l, char *out)
4432 for (i=0; i<5; i++) {
4433 *out++ = mdb_a85[l % 85];
4439 mdb_hash_enc(MDB_val *val, char *encbuf)
4441 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4443 mdb_pack85(h, encbuf);
4444 mdb_pack85(h>>32, encbuf+5);
4449 /** Open and/or initialize the lock region for the environment.
4450 * @param[in] env The LMDB environment.
4451 * @param[in] lpath The pathname of the file used for the lock region.
4452 * @param[in] mode The Unix permissions for the file, if we create it.
4453 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4454 * @return 0 on success, non-zero on failure.
4457 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4460 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4462 # define MDB_ERRCODE_ROFS EROFS
4463 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4464 # define MDB_CLOEXEC O_CLOEXEC
4467 # define MDB_CLOEXEC 0
4475 rc = utf8_to_utf16(lpath, -1, &wlpath, NULL);
4478 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4479 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4480 FILE_ATTRIBUTE_NORMAL, NULL);
4483 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4485 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4487 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4492 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4493 /* Lose record locks when exec*() */
4494 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4495 fcntl(env->me_lfd, F_SETFD, fdflags);
4498 if (!(env->me_flags & MDB_NOTLS)) {
4499 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4502 env->me_flags |= MDB_ENV_TXKEY;
4504 /* Windows TLS callbacks need help finding their TLS info. */
4505 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4509 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4513 /* Try to get exclusive lock. If we succeed, then
4514 * nobody is using the lock region and we should initialize it.
4516 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4519 size = GetFileSize(env->me_lfd, NULL);
4521 size = lseek(env->me_lfd, 0, SEEK_END);
4522 if (size == -1) goto fail_errno;
4524 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4525 if (size < rsize && *excl > 0) {
4527 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4528 || !SetEndOfFile(env->me_lfd))
4531 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4535 size = rsize - sizeof(MDB_txninfo);
4536 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4541 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4543 if (!mh) goto fail_errno;
4544 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4546 if (!env->me_txns) goto fail_errno;
4548 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4550 if (m == MAP_FAILED) goto fail_errno;
4556 BY_HANDLE_FILE_INFORMATION stbuf;
4565 if (!mdb_sec_inited) {
4566 InitializeSecurityDescriptor(&mdb_null_sd,
4567 SECURITY_DESCRIPTOR_REVISION);
4568 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4569 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4570 mdb_all_sa.bInheritHandle = FALSE;
4571 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4574 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4575 idbuf.volume = stbuf.dwVolumeSerialNumber;
4576 idbuf.nhigh = stbuf.nFileIndexHigh;
4577 idbuf.nlow = stbuf.nFileIndexLow;
4578 val.mv_data = &idbuf;
4579 val.mv_size = sizeof(idbuf);
4580 mdb_hash_enc(&val, encbuf);
4581 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4582 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4583 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4584 if (!env->me_rmutex) goto fail_errno;
4585 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4586 if (!env->me_wmutex) goto fail_errno;
4587 #elif defined(MDB_USE_POSIX_SEM)
4596 #if defined(__NetBSD__)
4597 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4599 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4600 idbuf.dev = stbuf.st_dev;
4601 idbuf.ino = stbuf.st_ino;
4602 val.mv_data = &idbuf;
4603 val.mv_size = sizeof(idbuf);
4604 mdb_hash_enc(&val, encbuf);
4605 #ifdef MDB_SHORT_SEMNAMES
4606 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4608 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4609 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4610 /* Clean up after a previous run, if needed: Try to
4611 * remove both semaphores before doing anything else.
4613 sem_unlink(env->me_txns->mti_rmname);
4614 sem_unlink(env->me_txns->mti_wmname);
4615 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4616 O_CREAT|O_EXCL, mode, 1);
4617 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4618 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4619 O_CREAT|O_EXCL, mode, 1);
4620 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4621 #else /* MDB_USE_POSIX_MUTEX: */
4622 pthread_mutexattr_t mattr;
4624 /* Solaris needs this before initing a robust mutex. Otherwise
4625 * it may skip the init and return EBUSY "seems someone already
4626 * inited" or EINVAL "it was inited differently".
4628 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
4629 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
4631 if ((rc = pthread_mutexattr_init(&mattr)))
4634 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
4635 #ifdef MDB_ROBUST_SUPPORTED
4636 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
4638 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
4639 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
4640 pthread_mutexattr_destroy(&mattr);
4643 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4645 env->me_txns->mti_magic = MDB_MAGIC;
4646 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4647 env->me_txns->mti_txnid = 0;
4648 env->me_txns->mti_numreaders = 0;
4651 if (env->me_txns->mti_magic != MDB_MAGIC) {
4652 DPUTS("lock region has invalid magic");
4656 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4657 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4658 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4659 rc = MDB_VERSION_MISMATCH;
4663 if (rc && rc != EACCES && rc != EAGAIN) {
4667 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4668 if (!env->me_rmutex) goto fail_errno;
4669 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4670 if (!env->me_wmutex) goto fail_errno;
4671 #elif defined(MDB_USE_POSIX_SEM)
4672 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4673 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4674 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4675 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4686 /** The name of the lock file in the DB environment */
4687 #define LOCKNAME "/lock.mdb"
4688 /** The name of the data file in the DB environment */
4689 #define DATANAME "/data.mdb"
4690 /** The suffix of the lock file when no subdir is used */
4691 #define LOCKSUFF "-lock"
4692 /** Only a subset of the @ref mdb_env flags can be changed
4693 * at runtime. Changing other flags requires closing the
4694 * environment and re-opening it with the new flags.
4696 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4697 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4698 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4700 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4701 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4705 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4707 int oflags, rc, len, excl = -1;
4708 char *lpath, *dpath;
4713 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4717 if (flags & MDB_NOSUBDIR) {
4718 rc = len + sizeof(LOCKSUFF) + len + 1;
4720 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4725 if (flags & MDB_NOSUBDIR) {
4726 dpath = lpath + len + sizeof(LOCKSUFF);
4727 sprintf(lpath, "%s" LOCKSUFF, path);
4728 strcpy(dpath, path);
4730 dpath = lpath + len + sizeof(LOCKNAME);
4731 sprintf(lpath, "%s" LOCKNAME, path);
4732 sprintf(dpath, "%s" DATANAME, path);
4736 flags |= env->me_flags;
4737 if (flags & MDB_RDONLY) {
4738 /* silently ignore WRITEMAP when we're only getting read access */
4739 flags &= ~MDB_WRITEMAP;
4741 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4742 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4745 env->me_flags = flags |= MDB_ENV_ACTIVE;
4749 env->me_path = strdup(path);
4750 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4751 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4752 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4753 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4757 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4759 /* For RDONLY, get lockfile after we know datafile exists */
4760 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4761 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4767 if (F_ISSET(flags, MDB_RDONLY)) {
4768 oflags = GENERIC_READ;
4769 len = OPEN_EXISTING;
4771 oflags = GENERIC_READ|GENERIC_WRITE;
4774 mode = FILE_ATTRIBUTE_NORMAL;
4775 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
4778 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4779 NULL, len, mode, NULL);
4782 if (F_ISSET(flags, MDB_RDONLY))
4785 oflags = O_RDWR | O_CREAT;
4787 env->me_fd = open(dpath, oflags, mode);
4789 if (env->me_fd == INVALID_HANDLE_VALUE) {
4794 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4795 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4800 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4801 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4802 env->me_mfd = env->me_fd;
4804 /* Synchronous fd for meta writes. Needed even with
4805 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4808 len = OPEN_EXISTING;
4809 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
4812 env->me_mfd = CreateFileW(wpath, oflags,
4813 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4814 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4818 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4820 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4825 DPRINTF(("opened dbenv %p", (void *) env));
4827 rc = mdb_env_share_locks(env, &excl);
4831 if (!(flags & MDB_RDONLY)) {
4833 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4834 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4835 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4836 (txn = calloc(1, size)))
4838 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4839 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4840 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4841 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4843 txn->mt_dbxs = env->me_dbxs;
4844 txn->mt_flags = MDB_TXN_FINISHED;
4854 mdb_env_close0(env, excl);
4860 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4862 mdb_env_close0(MDB_env *env, int excl)
4866 if (!(env->me_flags & MDB_ENV_ACTIVE))
4869 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4871 for (i = env->me_maxdbs; --i >= CORE_DBS; )
4872 free(env->me_dbxs[i].md_name.mv_data);
4877 free(env->me_dbiseqs);
4878 free(env->me_dbflags);
4880 free(env->me_dirty_list);
4882 mdb_midl_free(env->me_free_pgs);
4884 if (env->me_flags & MDB_ENV_TXKEY) {
4885 pthread_key_delete(env->me_txkey);
4887 /* Delete our key from the global list */
4888 for (i=0; i<mdb_tls_nkeys; i++)
4889 if (mdb_tls_keys[i] == env->me_txkey) {
4890 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4898 munmap(env->me_map, env->me_mapsize);
4900 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4901 (void) close(env->me_mfd);
4902 if (env->me_fd != INVALID_HANDLE_VALUE)
4903 (void) close(env->me_fd);
4905 MDB_PID_T pid = env->me_pid;
4906 /* Clearing readers is done in this function because
4907 * me_txkey with its destructor must be disabled first.
4909 * We skip the the reader mutex, so we touch only
4910 * data owned by this process (me_close_readers and
4911 * our readers), and clear each reader atomically.
4913 for (i = env->me_close_readers; --i >= 0; )
4914 if (env->me_txns->mti_readers[i].mr_pid == pid)
4915 env->me_txns->mti_readers[i].mr_pid = 0;
4917 if (env->me_rmutex) {
4918 CloseHandle(env->me_rmutex);
4919 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4921 /* Windows automatically destroys the mutexes when
4922 * the last handle closes.
4924 #elif defined(MDB_USE_POSIX_SEM)
4925 if (env->me_rmutex != SEM_FAILED) {
4926 sem_close(env->me_rmutex);
4927 if (env->me_wmutex != SEM_FAILED)
4928 sem_close(env->me_wmutex);
4929 /* If we have the filelock: If we are the
4930 * only remaining user, clean up semaphores.
4933 mdb_env_excl_lock(env, &excl);
4935 sem_unlink(env->me_txns->mti_rmname);
4936 sem_unlink(env->me_txns->mti_wmname);
4940 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4942 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4945 /* Unlock the lockfile. Windows would have unlocked it
4946 * after closing anyway, but not necessarily at once.
4948 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4951 (void) close(env->me_lfd);
4954 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4958 mdb_env_close(MDB_env *env)
4965 VGMEMP_DESTROY(env);
4966 while ((dp = env->me_dpages) != NULL) {
4967 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4968 env->me_dpages = dp->mp_next;
4972 mdb_env_close0(env, 0);
4976 /** Compare two items pointing at aligned size_t's */
4978 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4980 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4981 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4984 /** Compare two items pointing at aligned unsigned int's.
4986 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4987 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4990 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4992 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4993 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4996 /** Compare two items pointing at unsigned ints of unknown alignment.
4997 * Nodes and keys are guaranteed to be 2-byte aligned.
5000 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5002 #if BYTE_ORDER == LITTLE_ENDIAN
5003 unsigned short *u, *c;
5006 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5007 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5010 } while(!x && u > (unsigned short *)a->mv_data);
5013 unsigned short *u, *c, *end;
5016 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5017 u = (unsigned short *)a->mv_data;
5018 c = (unsigned short *)b->mv_data;
5021 } while(!x && u < end);
5026 /** Compare two items lexically */
5028 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5035 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5041 diff = memcmp(a->mv_data, b->mv_data, len);
5042 return diff ? diff : len_diff<0 ? -1 : len_diff;
5045 /** Compare two items in reverse byte order */
5047 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5049 const unsigned char *p1, *p2, *p1_lim;
5053 p1_lim = (const unsigned char *)a->mv_data;
5054 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5055 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5057 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5063 while (p1 > p1_lim) {
5064 diff = *--p1 - *--p2;
5068 return len_diff<0 ? -1 : len_diff;
5071 /** Search for key within a page, using binary search.
5072 * Returns the smallest entry larger or equal to the key.
5073 * If exactp is non-null, stores whether the found entry was an exact match
5074 * in *exactp (1 or 0).
5075 * Updates the cursor index with the index of the found entry.
5076 * If no entry larger or equal to the key is found, returns NULL.
5079 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5081 unsigned int i = 0, nkeys;
5084 MDB_page *mp = mc->mc_pg[mc->mc_top];
5085 MDB_node *node = NULL;
5090 nkeys = NUMKEYS(mp);
5092 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5093 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5096 low = IS_LEAF(mp) ? 0 : 1;
5098 cmp = mc->mc_dbx->md_cmp;
5100 /* Branch pages have no data, so if using integer keys,
5101 * alignment is guaranteed. Use faster mdb_cmp_int.
5103 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5104 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5111 nodekey.mv_size = mc->mc_db->md_pad;
5112 node = NODEPTR(mp, 0); /* fake */
5113 while (low <= high) {
5114 i = (low + high) >> 1;
5115 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5116 rc = cmp(key, &nodekey);
5117 DPRINTF(("found leaf index %u [%s], rc = %i",
5118 i, DKEY(&nodekey), rc));
5127 while (low <= high) {
5128 i = (low + high) >> 1;
5130 node = NODEPTR(mp, i);
5131 nodekey.mv_size = NODEKSZ(node);
5132 nodekey.mv_data = NODEKEY(node);
5134 rc = cmp(key, &nodekey);
5137 DPRINTF(("found leaf index %u [%s], rc = %i",
5138 i, DKEY(&nodekey), rc));
5140 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5141 i, DKEY(&nodekey), NODEPGNO(node), rc));
5152 if (rc > 0) { /* Found entry is less than the key. */
5153 i++; /* Skip to get the smallest entry larger than key. */
5155 node = NODEPTR(mp, i);
5158 *exactp = (rc == 0 && nkeys > 0);
5159 /* store the key index */
5160 mc->mc_ki[mc->mc_top] = i;
5162 /* There is no entry larger or equal to the key. */
5165 /* nodeptr is fake for LEAF2 */
5171 mdb_cursor_adjust(MDB_cursor *mc, func)
5175 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5176 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5183 /** Pop a page off the top of the cursor's stack. */
5185 mdb_cursor_pop(MDB_cursor *mc)
5188 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5189 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5195 mc->mc_flags &= ~C_INITIALIZED;
5200 /** Push a page onto the top of the cursor's stack. */
5202 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5204 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5205 DDBI(mc), (void *) mc));
5207 if (mc->mc_snum >= CURSOR_STACK) {
5208 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5209 return MDB_CURSOR_FULL;
5212 mc->mc_top = mc->mc_snum++;
5213 mc->mc_pg[mc->mc_top] = mp;
5214 mc->mc_ki[mc->mc_top] = 0;
5219 /** Find the address of the page corresponding to a given page number.
5220 * @param[in] txn the transaction for this access.
5221 * @param[in] pgno the page number for the page to retrieve.
5222 * @param[out] ret address of a pointer where the page's address will be stored.
5223 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5224 * @return 0 on success, non-zero on failure.
5227 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5229 MDB_env *env = txn->mt_env;
5233 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5237 MDB_ID2L dl = tx2->mt_u.dirty_list;
5239 /* Spilled pages were dirtied in this txn and flushed
5240 * because the dirty list got full. Bring this page
5241 * back in from the map (but don't unspill it here,
5242 * leave that unless page_touch happens again).
5244 if (tx2->mt_spill_pgs) {
5245 MDB_ID pn = pgno << 1;
5246 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5247 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5248 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5253 unsigned x = mdb_mid2l_search(dl, pgno);
5254 if (x <= dl[0].mid && dl[x].mid == pgno) {
5260 } while ((tx2 = tx2->mt_parent) != NULL);
5263 if (pgno < txn->mt_next_pgno) {
5265 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5267 DPRINTF(("page %"Z"u not found", pgno));
5268 txn->mt_flags |= MDB_TXN_ERROR;
5269 return MDB_PAGE_NOTFOUND;
5279 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5280 * The cursor is at the root page, set up the rest of it.
5283 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5285 MDB_page *mp = mc->mc_pg[mc->mc_top];
5289 while (IS_BRANCH(mp)) {
5293 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5294 /* Don't assert on branch pages in the FreeDB. We can get here
5295 * while in the process of rebalancing a FreeDB branch page; we must
5296 * let that proceed. ITS#8336
5298 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
5299 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5301 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5303 if (flags & MDB_PS_LAST)
5304 i = NUMKEYS(mp) - 1;
5307 node = mdb_node_search(mc, key, &exact);
5309 i = NUMKEYS(mp) - 1;
5311 i = mc->mc_ki[mc->mc_top];
5313 mdb_cassert(mc, i > 0);
5317 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5320 mdb_cassert(mc, i < NUMKEYS(mp));
5321 node = NODEPTR(mp, i);
5323 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5326 mc->mc_ki[mc->mc_top] = i;
5327 if ((rc = mdb_cursor_push(mc, mp)))
5330 if (flags & MDB_PS_MODIFY) {
5331 if ((rc = mdb_page_touch(mc)) != 0)
5333 mp = mc->mc_pg[mc->mc_top];
5338 DPRINTF(("internal error, index points to a %02X page!?",
5340 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5341 return MDB_CORRUPTED;
5344 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5345 key ? DKEY(key) : "null"));
5346 mc->mc_flags |= C_INITIALIZED;
5347 mc->mc_flags &= ~C_EOF;
5352 /** Search for the lowest key under the current branch page.
5353 * This just bypasses a NUMKEYS check in the current page
5354 * before calling mdb_page_search_root(), because the callers
5355 * are all in situations where the current page is known to
5359 mdb_page_search_lowest(MDB_cursor *mc)
5361 MDB_page *mp = mc->mc_pg[mc->mc_top];
5362 MDB_node *node = NODEPTR(mp, 0);
5365 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5368 mc->mc_ki[mc->mc_top] = 0;
5369 if ((rc = mdb_cursor_push(mc, mp)))
5371 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5374 /** Search for the page a given key should be in.
5375 * Push it and its parent pages on the cursor stack.
5376 * @param[in,out] mc the cursor for this operation.
5377 * @param[in] key the key to search for, or NULL for first/last page.
5378 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5379 * are touched (updated with new page numbers).
5380 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5381 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5382 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5383 * @return 0 on success, non-zero on failure.
5386 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5391 /* Make sure the txn is still viable, then find the root from
5392 * the txn's db table and set it as the root of the cursor's stack.
5394 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5395 DPUTS("transaction may not be used now");
5398 /* Make sure we're using an up-to-date root */
5399 if (*mc->mc_dbflag & DB_STALE) {
5401 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5403 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5404 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5411 MDB_node *leaf = mdb_node_search(&mc2,
5412 &mc->mc_dbx->md_name, &exact);
5414 return MDB_NOTFOUND;
5415 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5416 return MDB_INCOMPATIBLE; /* not a named DB */
5417 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5420 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5422 /* The txn may not know this DBI, or another process may
5423 * have dropped and recreated the DB with other flags.
5425 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5426 return MDB_INCOMPATIBLE;
5427 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5429 *mc->mc_dbflag &= ~DB_STALE;
5431 root = mc->mc_db->md_root;
5433 if (root == P_INVALID) { /* Tree is empty. */
5434 DPUTS("tree is empty");
5435 return MDB_NOTFOUND;
5439 mdb_cassert(mc, root > 1);
5440 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5441 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5447 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5448 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5450 if (flags & MDB_PS_MODIFY) {
5451 if ((rc = mdb_page_touch(mc)))
5455 if (flags & MDB_PS_ROOTONLY)
5458 return mdb_page_search_root(mc, key, flags);
5462 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5464 MDB_txn *txn = mc->mc_txn;
5465 pgno_t pg = mp->mp_pgno;
5466 unsigned x = 0, ovpages = mp->mp_pages;
5467 MDB_env *env = txn->mt_env;
5468 MDB_IDL sl = txn->mt_spill_pgs;
5469 MDB_ID pn = pg << 1;
5472 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5473 /* If the page is dirty or on the spill list we just acquired it,
5474 * so we should give it back to our current free list, if any.
5475 * Otherwise put it onto the list of pages we freed in this txn.
5477 * Won't create me_pghead: me_pglast must be inited along with it.
5478 * Unsupported in nested txns: They would need to hide the page
5479 * range in ancestor txns' dirty and spilled lists.
5481 if (env->me_pghead &&
5483 ((mp->mp_flags & P_DIRTY) ||
5484 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5488 MDB_ID2 *dl, ix, iy;
5489 rc = mdb_midl_need(&env->me_pghead, ovpages);
5492 if (!(mp->mp_flags & P_DIRTY)) {
5493 /* This page is no longer spilled */
5500 /* Remove from dirty list */
5501 dl = txn->mt_u.dirty_list;
5503 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5509 mdb_cassert(mc, x > 1);
5511 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5512 txn->mt_flags |= MDB_TXN_ERROR;
5513 return MDB_CORRUPTED;
5516 txn->mt_dirty_room++;
5517 if (!(env->me_flags & MDB_WRITEMAP))
5518 mdb_dpage_free(env, mp);
5520 /* Insert in me_pghead */
5521 mop = env->me_pghead;
5522 j = mop[0] + ovpages;
5523 for (i = mop[0]; i && mop[i] < pg; i--)
5529 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5533 mc->mc_db->md_overflow_pages -= ovpages;
5537 /** Return the data associated with a given node.
5538 * @param[in] txn The transaction for this operation.
5539 * @param[in] leaf The node being read.
5540 * @param[out] data Updated to point to the node's data.
5541 * @return 0 on success, non-zero on failure.
5544 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5546 MDB_page *omp; /* overflow page */
5550 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5551 data->mv_size = NODEDSZ(leaf);
5552 data->mv_data = NODEDATA(leaf);
5556 /* Read overflow data.
5558 data->mv_size = NODEDSZ(leaf);
5559 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5560 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5561 DPRINTF(("read overflow page %"Z"u failed", pgno));
5564 data->mv_data = METADATA(omp);
5570 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5571 MDB_val *key, MDB_val *data)
5578 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5580 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5583 if (txn->mt_flags & MDB_TXN_BLOCKED)
5586 mdb_cursor_init(&mc, txn, dbi, &mx);
5587 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5590 /** Find a sibling for a page.
5591 * Replaces the page at the top of the cursor's stack with the
5592 * specified sibling, if one exists.
5593 * @param[in] mc The cursor for this operation.
5594 * @param[in] move_right Non-zero if the right sibling is requested,
5595 * otherwise the left sibling.
5596 * @return 0 on success, non-zero on failure.
5599 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5605 if (mc->mc_snum < 2) {
5606 return MDB_NOTFOUND; /* root has no siblings */
5610 DPRINTF(("parent page is page %"Z"u, index %u",
5611 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5613 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5614 : (mc->mc_ki[mc->mc_top] == 0)) {
5615 DPRINTF(("no more keys left, moving to %s sibling",
5616 move_right ? "right" : "left"));
5617 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5618 /* undo cursor_pop before returning */
5625 mc->mc_ki[mc->mc_top]++;
5627 mc->mc_ki[mc->mc_top]--;
5628 DPRINTF(("just moving to %s index key %u",
5629 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5631 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5633 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5634 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5635 /* mc will be inconsistent if caller does mc_snum++ as above */
5636 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5640 mdb_cursor_push(mc, mp);
5642 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5647 /** Move the cursor to the next data item. */
5649 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5655 if ((mc->mc_flags & C_EOF) ||
5656 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
5657 return MDB_NOTFOUND;
5659 if (!(mc->mc_flags & C_INITIALIZED))
5660 return mdb_cursor_first(mc, key, data);
5662 mp = mc->mc_pg[mc->mc_top];
5664 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5665 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5666 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5667 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5668 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5669 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5670 if (rc == MDB_SUCCESS)
5671 MDB_GET_KEY(leaf, key);
5676 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5677 if (op == MDB_NEXT_DUP)
5678 return MDB_NOTFOUND;
5682 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5683 mdb_dbg_pgno(mp), (void *) mc));
5684 if (mc->mc_flags & C_DEL) {
5685 mc->mc_flags ^= C_DEL;
5689 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5690 DPUTS("=====> move to next sibling page");
5691 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5692 mc->mc_flags |= C_EOF;
5695 mp = mc->mc_pg[mc->mc_top];
5696 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5698 mc->mc_ki[mc->mc_top]++;
5701 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5702 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5705 key->mv_size = mc->mc_db->md_pad;
5706 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5710 mdb_cassert(mc, IS_LEAF(mp));
5711 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5713 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5714 mdb_xcursor_init1(mc, leaf);
5717 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5720 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5721 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5722 if (rc != MDB_SUCCESS)
5727 MDB_GET_KEY(leaf, key);
5731 /** Move the cursor to the previous data item. */
5733 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5739 if (!(mc->mc_flags & C_INITIALIZED)) {
5740 rc = mdb_cursor_last(mc, key, data);
5743 mc->mc_ki[mc->mc_top]++;
5746 mp = mc->mc_pg[mc->mc_top];
5748 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5749 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5750 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5751 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5752 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5753 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5754 if (rc == MDB_SUCCESS) {
5755 MDB_GET_KEY(leaf, key);
5756 mc->mc_flags &= ~C_EOF;
5762 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5763 if (op == MDB_PREV_DUP)
5764 return MDB_NOTFOUND;
5768 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5769 mdb_dbg_pgno(mp), (void *) mc));
5771 mc->mc_flags &= ~(C_EOF|C_DEL);
5773 if (mc->mc_ki[mc->mc_top] == 0) {
5774 DPUTS("=====> move to prev sibling page");
5775 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5778 mp = mc->mc_pg[mc->mc_top];
5779 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5780 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5782 mc->mc_ki[mc->mc_top]--;
5784 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5785 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5788 key->mv_size = mc->mc_db->md_pad;
5789 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5793 mdb_cassert(mc, IS_LEAF(mp));
5794 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5796 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5797 mdb_xcursor_init1(mc, leaf);
5800 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5803 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5804 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5805 if (rc != MDB_SUCCESS)
5810 MDB_GET_KEY(leaf, key);
5814 /** Set the cursor on a specific data item. */
5816 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5817 MDB_cursor_op op, int *exactp)
5821 MDB_node *leaf = NULL;
5824 if (key->mv_size == 0)
5825 return MDB_BAD_VALSIZE;
5828 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5830 /* See if we're already on the right page */
5831 if (mc->mc_flags & C_INITIALIZED) {
5834 mp = mc->mc_pg[mc->mc_top];
5836 mc->mc_ki[mc->mc_top] = 0;
5837 return MDB_NOTFOUND;
5839 if (mp->mp_flags & P_LEAF2) {
5840 nodekey.mv_size = mc->mc_db->md_pad;
5841 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5843 leaf = NODEPTR(mp, 0);
5844 MDB_GET_KEY2(leaf, nodekey);
5846 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5848 /* Probably happens rarely, but first node on the page
5849 * was the one we wanted.
5851 mc->mc_ki[mc->mc_top] = 0;
5858 unsigned int nkeys = NUMKEYS(mp);
5860 if (mp->mp_flags & P_LEAF2) {
5861 nodekey.mv_data = LEAF2KEY(mp,
5862 nkeys-1, nodekey.mv_size);
5864 leaf = NODEPTR(mp, nkeys-1);
5865 MDB_GET_KEY2(leaf, nodekey);
5867 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5869 /* last node was the one we wanted */
5870 mc->mc_ki[mc->mc_top] = nkeys-1;
5876 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5877 /* This is definitely the right page, skip search_page */
5878 if (mp->mp_flags & P_LEAF2) {
5879 nodekey.mv_data = LEAF2KEY(mp,
5880 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5882 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5883 MDB_GET_KEY2(leaf, nodekey);
5885 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5887 /* current node was the one we wanted */
5897 /* If any parents have right-sibs, search.
5898 * Otherwise, there's nothing further.
5900 for (i=0; i<mc->mc_top; i++)
5902 NUMKEYS(mc->mc_pg[i])-1)
5904 if (i == mc->mc_top) {
5905 /* There are no other pages */
5906 mc->mc_ki[mc->mc_top] = nkeys;
5907 return MDB_NOTFOUND;
5911 /* There are no other pages */
5912 mc->mc_ki[mc->mc_top] = 0;
5913 if (op == MDB_SET_RANGE && !exactp) {
5917 return MDB_NOTFOUND;
5923 rc = mdb_page_search(mc, key, 0);
5924 if (rc != MDB_SUCCESS)
5927 mp = mc->mc_pg[mc->mc_top];
5928 mdb_cassert(mc, IS_LEAF(mp));
5931 leaf = mdb_node_search(mc, key, exactp);
5932 if (exactp != NULL && !*exactp) {
5933 /* MDB_SET specified and not an exact match. */
5934 return MDB_NOTFOUND;
5938 DPUTS("===> inexact leaf not found, goto sibling");
5939 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5940 mc->mc_flags |= C_EOF;
5941 return rc; /* no entries matched */
5943 mp = mc->mc_pg[mc->mc_top];
5944 mdb_cassert(mc, IS_LEAF(mp));
5945 leaf = NODEPTR(mp, 0);
5949 mc->mc_flags |= C_INITIALIZED;
5950 mc->mc_flags &= ~C_EOF;
5953 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5954 key->mv_size = mc->mc_db->md_pad;
5955 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5960 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5961 mdb_xcursor_init1(mc, leaf);
5964 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5965 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5966 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5969 if (op == MDB_GET_BOTH) {
5975 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5976 if (rc != MDB_SUCCESS)
5979 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5982 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5984 dcmp = mc->mc_dbx->md_dcmp;
5985 #if UINT_MAX < SIZE_MAX
5986 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5987 dcmp = mdb_cmp_clong;
5989 rc = dcmp(data, &olddata);
5991 if (op == MDB_GET_BOTH || rc > 0)
5992 return MDB_NOTFOUND;
5999 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6000 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6005 /* The key already matches in all other cases */
6006 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6007 MDB_GET_KEY(leaf, key);
6008 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6013 /** Move the cursor to the first item in the database. */
6015 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6021 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6023 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6024 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6025 if (rc != MDB_SUCCESS)
6028 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6030 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6031 mc->mc_flags |= C_INITIALIZED;
6032 mc->mc_flags &= ~C_EOF;
6034 mc->mc_ki[mc->mc_top] = 0;
6036 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6037 key->mv_size = mc->mc_db->md_pad;
6038 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6043 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6044 mdb_xcursor_init1(mc, leaf);
6045 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6049 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6053 MDB_GET_KEY(leaf, key);
6057 /** Move the cursor to the last item in the database. */
6059 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6065 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6067 if (!(mc->mc_flags & C_EOF)) {
6069 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6070 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6071 if (rc != MDB_SUCCESS)
6074 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6077 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6078 mc->mc_flags |= C_INITIALIZED|C_EOF;
6079 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6081 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6082 key->mv_size = mc->mc_db->md_pad;
6083 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6088 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6089 mdb_xcursor_init1(mc, leaf);
6090 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6094 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
6099 MDB_GET_KEY(leaf, key);
6104 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6109 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6114 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6118 case MDB_GET_CURRENT:
6119 if (!(mc->mc_flags & C_INITIALIZED)) {
6122 MDB_page *mp = mc->mc_pg[mc->mc_top];
6123 int nkeys = NUMKEYS(mp);
6124 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6125 mc->mc_ki[mc->mc_top] = nkeys;
6131 key->mv_size = mc->mc_db->md_pad;
6132 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6134 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6135 MDB_GET_KEY(leaf, key);
6137 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6138 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6140 rc = mdb_node_read(mc->mc_txn, leaf, data);
6147 case MDB_GET_BOTH_RANGE:
6152 if (mc->mc_xcursor == NULL) {
6153 rc = MDB_INCOMPATIBLE;
6163 rc = mdb_cursor_set(mc, key, data, op,
6164 op == MDB_SET_RANGE ? NULL : &exact);
6167 case MDB_GET_MULTIPLE:
6168 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6172 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6173 rc = MDB_INCOMPATIBLE;
6177 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6178 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6181 case MDB_NEXT_MULTIPLE:
6186 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6187 rc = MDB_INCOMPATIBLE;
6190 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6191 if (rc == MDB_SUCCESS) {
6192 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6195 mx = &mc->mc_xcursor->mx_cursor;
6196 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6198 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6199 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6207 case MDB_NEXT_NODUP:
6208 rc = mdb_cursor_next(mc, key, data, op);
6212 case MDB_PREV_NODUP:
6213 rc = mdb_cursor_prev(mc, key, data, op);
6216 rc = mdb_cursor_first(mc, key, data);
6219 mfunc = mdb_cursor_first;
6221 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6225 if (mc->mc_xcursor == NULL) {
6226 rc = MDB_INCOMPATIBLE;
6230 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6231 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6232 MDB_GET_KEY(leaf, key);
6233 rc = mdb_node_read(mc->mc_txn, leaf, data);
6237 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6241 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6244 rc = mdb_cursor_last(mc, key, data);
6247 mfunc = mdb_cursor_last;
6250 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6255 if (mc->mc_flags & C_DEL)
6256 mc->mc_flags ^= C_DEL;
6261 /** Touch all the pages in the cursor stack. Set mc_top.
6262 * Makes sure all the pages are writable, before attempting a write operation.
6263 * @param[in] mc The cursor to operate on.
6266 mdb_cursor_touch(MDB_cursor *mc)
6268 int rc = MDB_SUCCESS;
6270 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
6273 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6275 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6276 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6279 *mc->mc_dbflag |= DB_DIRTY;
6284 rc = mdb_page_touch(mc);
6285 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6286 mc->mc_top = mc->mc_snum-1;
6291 /** Do not spill pages to disk if txn is getting full, may fail instead */
6292 #define MDB_NOSPILL 0x8000
6295 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6299 MDB_node *leaf = NULL;
6300 MDB_page *fp, *mp, *sub_root = NULL;
6302 MDB_val xdata, *rdata, dkey, olddata;
6304 int do_sub = 0, insert_key, insert_data;
6305 unsigned int mcount = 0, dcount = 0, nospill;
6308 unsigned int nflags;
6311 if (mc == NULL || key == NULL)
6314 env = mc->mc_txn->mt_env;
6316 /* Check this first so counter will always be zero on any
6319 if (flags & MDB_MULTIPLE) {
6320 dcount = data[1].mv_size;
6321 data[1].mv_size = 0;
6322 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6323 return MDB_INCOMPATIBLE;
6326 nospill = flags & MDB_NOSPILL;
6327 flags &= ~MDB_NOSPILL;
6329 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6330 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6332 if (key->mv_size-1 >= ENV_MAXKEY(env))
6333 return MDB_BAD_VALSIZE;
6335 #if SIZE_MAX > MAXDATASIZE
6336 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6337 return MDB_BAD_VALSIZE;
6339 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6340 return MDB_BAD_VALSIZE;
6343 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6344 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6348 if (flags == MDB_CURRENT) {
6349 if (!(mc->mc_flags & C_INITIALIZED))
6352 } else if (mc->mc_db->md_root == P_INVALID) {
6353 /* new database, cursor has nothing to point to */
6356 mc->mc_flags &= ~C_INITIALIZED;
6361 if (flags & MDB_APPEND) {
6363 rc = mdb_cursor_last(mc, &k2, &d2);
6365 rc = mc->mc_dbx->md_cmp(key, &k2);
6368 mc->mc_ki[mc->mc_top]++;
6370 /* new key is <= last key */
6375 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6377 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6378 DPRINTF(("duplicate key [%s]", DKEY(key)));
6380 return MDB_KEYEXIST;
6382 if (rc && rc != MDB_NOTFOUND)
6386 if (mc->mc_flags & C_DEL)
6387 mc->mc_flags ^= C_DEL;
6389 /* Cursor is positioned, check for room in the dirty list */
6391 if (flags & MDB_MULTIPLE) {
6393 xdata.mv_size = data->mv_size * dcount;
6397 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6401 if (rc == MDB_NO_ROOT) {
6403 /* new database, write a root leaf page */
6404 DPUTS("allocating new root leaf page");
6405 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6408 mdb_cursor_push(mc, np);
6409 mc->mc_db->md_root = np->mp_pgno;
6410 mc->mc_db->md_depth++;
6411 *mc->mc_dbflag |= DB_DIRTY;
6412 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6414 np->mp_flags |= P_LEAF2;
6415 mc->mc_flags |= C_INITIALIZED;
6417 /* make sure all cursor pages are writable */
6418 rc2 = mdb_cursor_touch(mc);
6423 insert_key = insert_data = rc;
6425 /* The key does not exist */
6426 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6427 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6428 LEAFSIZE(key, data) > env->me_nodemax)
6430 /* Too big for a node, insert in sub-DB. Set up an empty
6431 * "old sub-page" for prep_subDB to expand to a full page.
6433 fp_flags = P_LEAF|P_DIRTY;
6435 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6436 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6437 olddata.mv_size = PAGEHDRSZ;
6441 /* there's only a key anyway, so this is a no-op */
6442 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6444 unsigned int ksize = mc->mc_db->md_pad;
6445 if (key->mv_size != ksize)
6446 return MDB_BAD_VALSIZE;
6447 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6448 memcpy(ptr, key->mv_data, ksize);
6450 /* if overwriting slot 0 of leaf, need to
6451 * update branch key if there is a parent page
6453 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6454 unsigned short dtop = 1;
6456 /* slot 0 is always an empty key, find real slot */
6457 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6461 if (mc->mc_ki[mc->mc_top])
6462 rc2 = mdb_update_key(mc, key);
6473 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6474 olddata.mv_size = NODEDSZ(leaf);
6475 olddata.mv_data = NODEDATA(leaf);
6478 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6479 /* Prepare (sub-)page/sub-DB to accept the new item,
6480 * if needed. fp: old sub-page or a header faking
6481 * it. mp: new (sub-)page. offset: growth in page
6482 * size. xdata: node data with new page or DB.
6484 unsigned i, offset = 0;
6485 mp = fp = xdata.mv_data = env->me_pbuf;
6486 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6488 /* Was a single item before, must convert now */
6489 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6491 /* Just overwrite the current item */
6492 if (flags == MDB_CURRENT)
6494 dcmp = mc->mc_dbx->md_dcmp;
6495 #if UINT_MAX < SIZE_MAX
6496 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6497 dcmp = mdb_cmp_clong;
6499 /* does data match? */
6500 if (!dcmp(data, &olddata)) {
6501 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
6502 return MDB_KEYEXIST;
6507 /* Back up original data item */
6508 dkey.mv_size = olddata.mv_size;
6509 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6511 /* Make sub-page header for the dup items, with dummy body */
6512 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6513 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6514 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6515 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6516 fp->mp_flags |= P_LEAF2;
6517 fp->mp_pad = data->mv_size;
6518 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6520 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6521 (dkey.mv_size & 1) + (data->mv_size & 1);
6523 fp->mp_upper = xdata.mv_size - PAGEBASE;
6524 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6525 } else if (leaf->mn_flags & F_SUBDATA) {
6526 /* Data is on sub-DB, just store it */
6527 flags |= F_DUPDATA|F_SUBDATA;
6530 /* Data is on sub-page */
6531 fp = olddata.mv_data;
6534 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6535 offset = EVEN(NODESIZE + sizeof(indx_t) +
6539 offset = fp->mp_pad;
6540 if (SIZELEFT(fp) < offset) {
6541 offset *= 4; /* space for 4 more */
6544 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6546 fp->mp_flags |= P_DIRTY;
6547 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6548 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6552 xdata.mv_size = olddata.mv_size + offset;
6555 fp_flags = fp->mp_flags;
6556 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6557 /* Too big for a sub-page, convert to sub-DB */
6558 fp_flags &= ~P_SUBP;
6560 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6561 fp_flags |= P_LEAF2;
6562 dummy.md_pad = fp->mp_pad;
6563 dummy.md_flags = MDB_DUPFIXED;
6564 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6565 dummy.md_flags |= MDB_INTEGERKEY;
6571 dummy.md_branch_pages = 0;
6572 dummy.md_leaf_pages = 1;
6573 dummy.md_overflow_pages = 0;
6574 dummy.md_entries = NUMKEYS(fp);
6575 xdata.mv_size = sizeof(MDB_db);
6576 xdata.mv_data = &dummy;
6577 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6579 offset = env->me_psize - olddata.mv_size;
6580 flags |= F_DUPDATA|F_SUBDATA;
6581 dummy.md_root = mp->mp_pgno;
6585 mp->mp_flags = fp_flags | P_DIRTY;
6586 mp->mp_pad = fp->mp_pad;
6587 mp->mp_lower = fp->mp_lower;
6588 mp->mp_upper = fp->mp_upper + offset;
6589 if (fp_flags & P_LEAF2) {
6590 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6592 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6593 olddata.mv_size - fp->mp_upper - PAGEBASE);
6594 for (i=0; i<NUMKEYS(fp); i++)
6595 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6603 mdb_node_del(mc, 0);
6607 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6608 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6609 return MDB_INCOMPATIBLE;
6610 /* overflow page overwrites need special handling */
6611 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6614 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6616 memcpy(&pg, olddata.mv_data, sizeof(pg));
6617 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6619 ovpages = omp->mp_pages;
6621 /* Is the ov page large enough? */
6622 if (ovpages >= dpages) {
6623 if (!(omp->mp_flags & P_DIRTY) &&
6624 (level || (env->me_flags & MDB_WRITEMAP)))
6626 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6629 level = 0; /* dirty in this txn or clean */
6632 if (omp->mp_flags & P_DIRTY) {
6633 /* yes, overwrite it. Note in this case we don't
6634 * bother to try shrinking the page if the new data
6635 * is smaller than the overflow threshold.
6638 /* It is writable only in a parent txn */
6639 size_t sz = (size_t) env->me_psize * ovpages, off;
6640 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6646 /* Note - this page is already counted in parent's dirty_room */
6647 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6648 mdb_cassert(mc, rc2 == 0);
6649 /* Currently we make the page look as with put() in the
6650 * parent txn, in case the user peeks at MDB_RESERVEd
6651 * or unused parts. Some users treat ovpages specially.
6653 if (!(flags & MDB_RESERVE)) {
6654 /* Skip the part where LMDB will put *data.
6655 * Copy end of page, adjusting alignment so
6656 * compiler may copy words instead of bytes.
6658 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6659 memcpy((size_t *)((char *)np + off),
6660 (size_t *)((char *)omp + off), sz - off);
6663 memcpy(np, omp, sz); /* Copy beginning of page */
6666 SETDSZ(leaf, data->mv_size);
6667 if (F_ISSET(flags, MDB_RESERVE))
6668 data->mv_data = METADATA(omp);
6670 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6674 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6676 } else if (data->mv_size == olddata.mv_size) {
6677 /* same size, just replace it. Note that we could
6678 * also reuse this node if the new data is smaller,
6679 * but instead we opt to shrink the node in that case.
6681 if (F_ISSET(flags, MDB_RESERVE))
6682 data->mv_data = olddata.mv_data;
6683 else if (!(mc->mc_flags & C_SUB))
6684 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6686 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6691 mdb_node_del(mc, 0);
6697 nflags = flags & NODE_ADD_FLAGS;
6698 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6699 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6700 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6701 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6703 nflags |= MDB_SPLIT_REPLACE;
6704 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6706 /* There is room already in this leaf page. */
6707 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6709 /* Adjust other cursors pointing to mp */
6710 MDB_cursor *m2, *m3;
6711 MDB_dbi dbi = mc->mc_dbi;
6712 unsigned i = mc->mc_top;
6713 MDB_page *mp = mc->mc_pg[i];
6715 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6716 if (mc->mc_flags & C_SUB)
6717 m3 = &m2->mc_xcursor->mx_cursor;
6720 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
6721 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
6724 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6725 MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
6726 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
6727 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6733 if (rc == MDB_SUCCESS) {
6734 /* Now store the actual data in the child DB. Note that we're
6735 * storing the user data in the keys field, so there are strict
6736 * size limits on dupdata. The actual data fields of the child
6737 * DB are all zero size.
6740 int xflags, new_dupdata;
6745 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6746 if (flags & MDB_CURRENT) {
6747 xflags = MDB_CURRENT|MDB_NOSPILL;
6749 mdb_xcursor_init1(mc, leaf);
6750 xflags = (flags & MDB_NODUPDATA) ?
6751 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6754 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6755 new_dupdata = (int)dkey.mv_size;
6756 /* converted, write the original data first */
6758 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6761 /* we've done our job */
6764 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6765 /* Adjust other cursors pointing to mp */
6767 MDB_xcursor *mx = mc->mc_xcursor;
6768 unsigned i = mc->mc_top;
6769 MDB_page *mp = mc->mc_pg[i];
6770 int nkeys = NUMKEYS(mp);
6772 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6773 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6774 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6775 if (m2->mc_pg[i] == mp) {
6776 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6777 mdb_xcursor_init2(m2, mx, new_dupdata);
6778 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6779 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6780 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
6781 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6786 ecount = mc->mc_xcursor->mx_db.md_entries;
6787 if (flags & MDB_APPENDDUP)
6788 xflags |= MDB_APPEND;
6789 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6790 if (flags & F_SUBDATA) {
6791 void *db = NODEDATA(leaf);
6792 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6794 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6796 /* Increment count unless we just replaced an existing item. */
6798 mc->mc_db->md_entries++;
6800 /* Invalidate txn if we created an empty sub-DB */
6803 /* If we succeeded and the key didn't exist before,
6804 * make sure the cursor is marked valid.
6806 mc->mc_flags |= C_INITIALIZED;
6808 if (flags & MDB_MULTIPLE) {
6811 /* let caller know how many succeeded, if any */
6812 data[1].mv_size = mcount;
6813 if (mcount < dcount) {
6814 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6815 insert_key = insert_data = 0;
6822 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6825 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6830 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6836 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6837 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6839 if (!(mc->mc_flags & C_INITIALIZED))
6842 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6843 return MDB_NOTFOUND;
6845 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6848 rc = mdb_cursor_touch(mc);
6852 mp = mc->mc_pg[mc->mc_top];
6855 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6857 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6858 if (flags & MDB_NODUPDATA) {
6859 /* mdb_cursor_del0() will subtract the final entry */
6860 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6861 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
6863 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6864 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6866 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6869 /* If sub-DB still has entries, we're done */
6870 if (mc->mc_xcursor->mx_db.md_entries) {
6871 if (leaf->mn_flags & F_SUBDATA) {
6872 /* update subDB info */
6873 void *db = NODEDATA(leaf);
6874 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6877 /* shrink fake page */
6878 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6879 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6880 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6881 /* fix other sub-DB cursors pointed at fake pages on this page */
6882 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6883 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6884 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6885 if (m2->mc_pg[mc->mc_top] == mp) {
6886 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6887 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6889 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6890 if (!(n2->mn_flags & F_SUBDATA))
6891 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6896 mc->mc_db->md_entries--;
6899 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
6901 /* otherwise fall thru and delete the sub-DB */
6904 if (leaf->mn_flags & F_SUBDATA) {
6905 /* add all the child DB's pages to the free list */
6906 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6911 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6912 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6913 rc = MDB_INCOMPATIBLE;
6917 /* add overflow pages to free list */
6918 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6922 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6923 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6924 (rc = mdb_ovpage_free(mc, omp)))
6929 return mdb_cursor_del0(mc);
6932 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6936 /** Allocate and initialize new pages for a database.
6937 * @param[in] mc a cursor on the database being added to.
6938 * @param[in] flags flags defining what type of page is being allocated.
6939 * @param[in] num the number of pages to allocate. This is usually 1,
6940 * unless allocating overflow pages for a large record.
6941 * @param[out] mp Address of a page, or NULL on failure.
6942 * @return 0 on success, non-zero on failure.
6945 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6950 if ((rc = mdb_page_alloc(mc, num, &np)))
6952 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6953 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6954 np->mp_flags = flags | P_DIRTY;
6955 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6956 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6959 mc->mc_db->md_branch_pages++;
6960 else if (IS_LEAF(np))
6961 mc->mc_db->md_leaf_pages++;
6962 else if (IS_OVERFLOW(np)) {
6963 mc->mc_db->md_overflow_pages += num;
6971 /** Calculate the size of a leaf node.
6972 * The size depends on the environment's page size; if a data item
6973 * is too large it will be put onto an overflow page and the node
6974 * size will only include the key and not the data. Sizes are always
6975 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6976 * of the #MDB_node headers.
6977 * @param[in] env The environment handle.
6978 * @param[in] key The key for the node.
6979 * @param[in] data The data for the node.
6980 * @return The number of bytes needed to store the node.
6983 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6987 sz = LEAFSIZE(key, data);
6988 if (sz > env->me_nodemax) {
6989 /* put on overflow page */
6990 sz -= data->mv_size - sizeof(pgno_t);
6993 return EVEN(sz + sizeof(indx_t));
6996 /** Calculate the size of a branch node.
6997 * The size should depend on the environment's page size but since
6998 * we currently don't support spilling large keys onto overflow
6999 * pages, it's simply the size of the #MDB_node header plus the
7000 * size of the key. Sizes are always rounded up to an even number
7001 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7002 * @param[in] env The environment handle.
7003 * @param[in] key The key for the node.
7004 * @return The number of bytes needed to store the node.
7007 mdb_branch_size(MDB_env *env, MDB_val *key)
7012 if (sz > env->me_nodemax) {
7013 /* put on overflow page */
7014 /* not implemented */
7015 /* sz -= key->size - sizeof(pgno_t); */
7018 return sz + sizeof(indx_t);
7021 /** Add a node to the page pointed to by the cursor.
7022 * @param[in] mc The cursor for this operation.
7023 * @param[in] indx The index on the page where the new node should be added.
7024 * @param[in] key The key for the new node.
7025 * @param[in] data The data for the new node, if any.
7026 * @param[in] pgno The page number, if adding a branch node.
7027 * @param[in] flags Flags for the node.
7028 * @return 0 on success, non-zero on failure. Possible errors are:
7030 * <li>ENOMEM - failed to allocate overflow pages for the node.
7031 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7032 * should never happen since all callers already calculate the
7033 * page's free space before calling this function.
7037 mdb_node_add(MDB_cursor *mc, indx_t indx,
7038 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7041 size_t node_size = NODESIZE;
7045 MDB_page *mp = mc->mc_pg[mc->mc_top];
7046 MDB_page *ofp = NULL; /* overflow page */
7050 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7052 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7053 IS_LEAF(mp) ? "leaf" : "branch",
7054 IS_SUBP(mp) ? "sub-" : "",
7055 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7056 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7059 /* Move higher keys up one slot. */
7060 int ksize = mc->mc_db->md_pad, dif;
7061 char *ptr = LEAF2KEY(mp, indx, ksize);
7062 dif = NUMKEYS(mp) - indx;
7064 memmove(ptr+ksize, ptr, dif*ksize);
7065 /* insert new key */
7066 memcpy(ptr, key->mv_data, ksize);
7068 /* Just using these for counting */
7069 mp->mp_lower += sizeof(indx_t);
7070 mp->mp_upper -= ksize - sizeof(indx_t);
7074 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7076 node_size += key->mv_size;
7078 mdb_cassert(mc, key && data);
7079 if (F_ISSET(flags, F_BIGDATA)) {
7080 /* Data already on overflow page. */
7081 node_size += sizeof(pgno_t);
7082 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7083 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7085 /* Put data on overflow page. */
7086 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7087 data->mv_size, node_size+data->mv_size));
7088 node_size = EVEN(node_size + sizeof(pgno_t));
7089 if ((ssize_t)node_size > room)
7091 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7093 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7097 node_size += data->mv_size;
7100 node_size = EVEN(node_size);
7101 if ((ssize_t)node_size > room)
7105 /* Move higher pointers up one slot. */
7106 for (i = NUMKEYS(mp); i > indx; i--)
7107 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7109 /* Adjust free space offsets. */
7110 ofs = mp->mp_upper - node_size;
7111 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7112 mp->mp_ptrs[indx] = ofs;
7114 mp->mp_lower += sizeof(indx_t);
7116 /* Write the node data. */
7117 node = NODEPTR(mp, indx);
7118 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7119 node->mn_flags = flags;
7121 SETDSZ(node,data->mv_size);
7126 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7129 ndata = NODEDATA(node);
7131 if (F_ISSET(flags, F_BIGDATA))
7132 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7133 else if (F_ISSET(flags, MDB_RESERVE))
7134 data->mv_data = ndata;
7136 memcpy(ndata, data->mv_data, data->mv_size);
7138 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7139 ndata = METADATA(ofp);
7140 if (F_ISSET(flags, MDB_RESERVE))
7141 data->mv_data = ndata;
7143 memcpy(ndata, data->mv_data, data->mv_size);
7150 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7151 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7152 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7153 DPRINTF(("node size = %"Z"u", node_size));
7154 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7155 return MDB_PAGE_FULL;
7158 /** Delete the specified node from a page.
7159 * @param[in] mc Cursor pointing to the node to delete.
7160 * @param[in] ksize The size of a node. Only used if the page is
7161 * part of a #MDB_DUPFIXED database.
7164 mdb_node_del(MDB_cursor *mc, int ksize)
7166 MDB_page *mp = mc->mc_pg[mc->mc_top];
7167 indx_t indx = mc->mc_ki[mc->mc_top];
7169 indx_t i, j, numkeys, ptr;
7173 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7174 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7175 numkeys = NUMKEYS(mp);
7176 mdb_cassert(mc, indx < numkeys);
7179 int x = numkeys - 1 - indx;
7180 base = LEAF2KEY(mp, indx, ksize);
7182 memmove(base, base + ksize, x * ksize);
7183 mp->mp_lower -= sizeof(indx_t);
7184 mp->mp_upper += ksize - sizeof(indx_t);
7188 node = NODEPTR(mp, indx);
7189 sz = NODESIZE + node->mn_ksize;
7191 if (F_ISSET(node->mn_flags, F_BIGDATA))
7192 sz += sizeof(pgno_t);
7194 sz += NODEDSZ(node);
7198 ptr = mp->mp_ptrs[indx];
7199 for (i = j = 0; i < numkeys; i++) {
7201 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7202 if (mp->mp_ptrs[i] < ptr)
7203 mp->mp_ptrs[j] += sz;
7208 base = (char *)mp + mp->mp_upper + PAGEBASE;
7209 memmove(base + sz, base, ptr - mp->mp_upper);
7211 mp->mp_lower -= sizeof(indx_t);
7215 /** Compact the main page after deleting a node on a subpage.
7216 * @param[in] mp The main page to operate on.
7217 * @param[in] indx The index of the subpage on the main page.
7220 mdb_node_shrink(MDB_page *mp, indx_t indx)
7225 indx_t delta, nsize, len, ptr;
7228 node = NODEPTR(mp, indx);
7229 sp = (MDB_page *)NODEDATA(node);
7230 delta = SIZELEFT(sp);
7231 nsize = NODEDSZ(node) - delta;
7233 /* Prepare to shift upward, set len = length(subpage part to shift) */
7237 return; /* do not make the node uneven-sized */
7239 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7240 for (i = NUMKEYS(sp); --i >= 0; )
7241 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7244 sp->mp_upper = sp->mp_lower;
7245 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7246 SETDSZ(node, nsize);
7248 /* Shift <lower nodes...initial part of subpage> upward */
7249 base = (char *)mp + mp->mp_upper + PAGEBASE;
7250 memmove(base + delta, base, (char *)sp + len - base);
7252 ptr = mp->mp_ptrs[indx];
7253 for (i = NUMKEYS(mp); --i >= 0; ) {
7254 if (mp->mp_ptrs[i] <= ptr)
7255 mp->mp_ptrs[i] += delta;
7257 mp->mp_upper += delta;
7260 /** Initial setup of a sorted-dups cursor.
7261 * Sorted duplicates are implemented as a sub-database for the given key.
7262 * The duplicate data items are actually keys of the sub-database.
7263 * Operations on the duplicate data items are performed using a sub-cursor
7264 * initialized when the sub-database is first accessed. This function does
7265 * the preliminary setup of the sub-cursor, filling in the fields that
7266 * depend only on the parent DB.
7267 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7270 mdb_xcursor_init0(MDB_cursor *mc)
7272 MDB_xcursor *mx = mc->mc_xcursor;
7274 mx->mx_cursor.mc_xcursor = NULL;
7275 mx->mx_cursor.mc_txn = mc->mc_txn;
7276 mx->mx_cursor.mc_db = &mx->mx_db;
7277 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7278 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7279 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7280 mx->mx_cursor.mc_snum = 0;
7281 mx->mx_cursor.mc_top = 0;
7282 mx->mx_cursor.mc_flags = C_SUB;
7283 mx->mx_dbx.md_name.mv_size = 0;
7284 mx->mx_dbx.md_name.mv_data = NULL;
7285 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7286 mx->mx_dbx.md_dcmp = NULL;
7287 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7290 /** Final setup of a sorted-dups cursor.
7291 * Sets up the fields that depend on the data from the main cursor.
7292 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7293 * @param[in] node The data containing the #MDB_db record for the
7294 * sorted-dup database.
7297 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7299 MDB_xcursor *mx = mc->mc_xcursor;
7301 if (node->mn_flags & F_SUBDATA) {
7302 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7303 mx->mx_cursor.mc_pg[0] = 0;
7304 mx->mx_cursor.mc_snum = 0;
7305 mx->mx_cursor.mc_top = 0;
7306 mx->mx_cursor.mc_flags = C_SUB;
7308 MDB_page *fp = NODEDATA(node);
7309 mx->mx_db.md_pad = 0;
7310 mx->mx_db.md_flags = 0;
7311 mx->mx_db.md_depth = 1;
7312 mx->mx_db.md_branch_pages = 0;
7313 mx->mx_db.md_leaf_pages = 1;
7314 mx->mx_db.md_overflow_pages = 0;
7315 mx->mx_db.md_entries = NUMKEYS(fp);
7316 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7317 mx->mx_cursor.mc_snum = 1;
7318 mx->mx_cursor.mc_top = 0;
7319 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7320 mx->mx_cursor.mc_pg[0] = fp;
7321 mx->mx_cursor.mc_ki[0] = 0;
7322 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7323 mx->mx_db.md_flags = MDB_DUPFIXED;
7324 mx->mx_db.md_pad = fp->mp_pad;
7325 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7326 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7329 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7330 mx->mx_db.md_root));
7331 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7332 #if UINT_MAX < SIZE_MAX
7333 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7334 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7339 /** Fixup a sorted-dups cursor due to underlying update.
7340 * Sets up some fields that depend on the data from the main cursor.
7341 * Almost the same as init1, but skips initialization steps if the
7342 * xcursor had already been used.
7343 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7344 * @param[in] src_mx The xcursor of an up-to-date cursor.
7345 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7348 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7350 MDB_xcursor *mx = mc->mc_xcursor;
7353 mx->mx_cursor.mc_snum = 1;
7354 mx->mx_cursor.mc_top = 0;
7355 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7356 mx->mx_cursor.mc_ki[0] = 0;
7357 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7358 #if UINT_MAX < SIZE_MAX
7359 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7361 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7364 mx->mx_db = src_mx->mx_db;
7365 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7366 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7367 mx->mx_db.md_root));
7370 /** Initialize a cursor for a given transaction and database. */
7372 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7375 mc->mc_backup = NULL;
7378 mc->mc_db = &txn->mt_dbs[dbi];
7379 mc->mc_dbx = &txn->mt_dbxs[dbi];
7380 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7386 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7387 mdb_tassert(txn, mx != NULL);
7388 mc->mc_xcursor = mx;
7389 mdb_xcursor_init0(mc);
7391 mc->mc_xcursor = NULL;
7393 if (*mc->mc_dbflag & DB_STALE) {
7394 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7399 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7402 size_t size = sizeof(MDB_cursor);
7404 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7407 if (txn->mt_flags & MDB_TXN_BLOCKED)
7410 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7413 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7414 size += sizeof(MDB_xcursor);
7416 if ((mc = malloc(size)) != NULL) {
7417 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7418 if (txn->mt_cursors) {
7419 mc->mc_next = txn->mt_cursors[dbi];
7420 txn->mt_cursors[dbi] = mc;
7421 mc->mc_flags |= C_UNTRACK;
7433 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7435 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7438 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7441 if (txn->mt_flags & MDB_TXN_BLOCKED)
7444 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7448 /* Return the count of duplicate data items for the current key */
7450 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7454 if (mc == NULL || countp == NULL)
7457 if (mc->mc_xcursor == NULL)
7458 return MDB_INCOMPATIBLE;
7460 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7463 if (!(mc->mc_flags & C_INITIALIZED))
7466 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7467 return MDB_NOTFOUND;
7469 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7470 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7473 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7476 *countp = mc->mc_xcursor->mx_db.md_entries;
7482 mdb_cursor_close(MDB_cursor *mc)
7484 if (mc && !mc->mc_backup) {
7485 /* remove from txn, if tracked */
7486 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7487 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7488 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7490 *prev = mc->mc_next;
7497 mdb_cursor_txn(MDB_cursor *mc)
7499 if (!mc) return NULL;
7504 mdb_cursor_dbi(MDB_cursor *mc)
7509 /** Replace the key for a branch node with a new key.
7510 * @param[in] mc Cursor pointing to the node to operate on.
7511 * @param[in] key The new key to use.
7512 * @return 0 on success, non-zero on failure.
7515 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7521 int delta, ksize, oksize;
7522 indx_t ptr, i, numkeys, indx;
7525 indx = mc->mc_ki[mc->mc_top];
7526 mp = mc->mc_pg[mc->mc_top];
7527 node = NODEPTR(mp, indx);
7528 ptr = mp->mp_ptrs[indx];
7532 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7533 k2.mv_data = NODEKEY(node);
7534 k2.mv_size = node->mn_ksize;
7535 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7537 mdb_dkey(&k2, kbuf2),
7543 /* Sizes must be 2-byte aligned. */
7544 ksize = EVEN(key->mv_size);
7545 oksize = EVEN(node->mn_ksize);
7546 delta = ksize - oksize;
7548 /* Shift node contents if EVEN(key length) changed. */
7550 if (delta > 0 && SIZELEFT(mp) < delta) {
7552 /* not enough space left, do a delete and split */
7553 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7554 pgno = NODEPGNO(node);
7555 mdb_node_del(mc, 0);
7556 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7559 numkeys = NUMKEYS(mp);
7560 for (i = 0; i < numkeys; i++) {
7561 if (mp->mp_ptrs[i] <= ptr)
7562 mp->mp_ptrs[i] -= delta;
7565 base = (char *)mp + mp->mp_upper + PAGEBASE;
7566 len = ptr - mp->mp_upper + NODESIZE;
7567 memmove(base - delta, base, len);
7568 mp->mp_upper -= delta;
7570 node = NODEPTR(mp, indx);
7573 /* But even if no shift was needed, update ksize */
7574 if (node->mn_ksize != key->mv_size)
7575 node->mn_ksize = key->mv_size;
7578 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7584 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7586 /** Perform \b act while tracking temporary cursor \b mn */
7587 #define WITH_CURSOR_TRACKING(mn, act) do { \
7588 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
7589 if ((mn).mc_flags & C_SUB) { \
7590 dummy.mc_flags = C_INITIALIZED; \
7591 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
7596 tracked->mc_next = *tp; \
7599 *tp = tracked->mc_next; \
7602 /** Move a node from csrc to cdst.
7605 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7612 unsigned short flags;
7616 /* Mark src and dst as dirty. */
7617 if ((rc = mdb_page_touch(csrc)) ||
7618 (rc = mdb_page_touch(cdst)))
7621 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7622 key.mv_size = csrc->mc_db->md_pad;
7623 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7625 data.mv_data = NULL;
7629 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7630 mdb_cassert(csrc, !((size_t)srcnode & 1));
7631 srcpg = NODEPGNO(srcnode);
7632 flags = srcnode->mn_flags;
7633 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7634 unsigned int snum = csrc->mc_snum;
7636 /* must find the lowest key below src */
7637 rc = mdb_page_search_lowest(csrc);
7640 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7641 key.mv_size = csrc->mc_db->md_pad;
7642 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7644 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7645 key.mv_size = NODEKSZ(s2);
7646 key.mv_data = NODEKEY(s2);
7648 csrc->mc_snum = snum--;
7649 csrc->mc_top = snum;
7651 key.mv_size = NODEKSZ(srcnode);
7652 key.mv_data = NODEKEY(srcnode);
7654 data.mv_size = NODEDSZ(srcnode);
7655 data.mv_data = NODEDATA(srcnode);
7657 mn.mc_xcursor = NULL;
7658 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7659 unsigned int snum = cdst->mc_snum;
7662 /* must find the lowest key below dst */
7663 mdb_cursor_copy(cdst, &mn);
7664 rc = mdb_page_search_lowest(&mn);
7667 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7668 bkey.mv_size = mn.mc_db->md_pad;
7669 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7671 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7672 bkey.mv_size = NODEKSZ(s2);
7673 bkey.mv_data = NODEKEY(s2);
7675 mn.mc_snum = snum--;
7678 rc = mdb_update_key(&mn, &bkey);
7683 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7684 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7685 csrc->mc_ki[csrc->mc_top],
7687 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7688 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7690 /* Add the node to the destination page.
7692 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7693 if (rc != MDB_SUCCESS)
7696 /* Delete the node from the source page.
7698 mdb_node_del(csrc, key.mv_size);
7701 /* Adjust other cursors pointing to mp */
7702 MDB_cursor *m2, *m3;
7703 MDB_dbi dbi = csrc->mc_dbi;
7704 MDB_page *mpd, *mps;
7706 mps = csrc->mc_pg[csrc->mc_top];
7707 /* If we're adding on the left, bump others up */
7709 mpd = cdst->mc_pg[csrc->mc_top];
7710 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7711 if (csrc->mc_flags & C_SUB)
7712 m3 = &m2->mc_xcursor->mx_cursor;
7715 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7718 m3->mc_pg[csrc->mc_top] == mpd &&
7719 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7720 m3->mc_ki[csrc->mc_top]++;
7723 m3->mc_pg[csrc->mc_top] == mps &&
7724 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7725 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7726 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7727 m3->mc_ki[csrc->mc_top-1]++;
7729 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
7731 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7732 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
7733 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
7737 /* Adding on the right, bump others down */
7739 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7740 if (csrc->mc_flags & C_SUB)
7741 m3 = &m2->mc_xcursor->mx_cursor;
7744 if (m3 == csrc) continue;
7745 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7747 if (m3->mc_pg[csrc->mc_top] == mps) {
7748 if (!m3->mc_ki[csrc->mc_top]) {
7749 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7750 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7751 m3->mc_ki[csrc->mc_top-1]--;
7753 m3->mc_ki[csrc->mc_top]--;
7755 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
7757 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7758 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
7759 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
7766 /* Update the parent separators.
7768 if (csrc->mc_ki[csrc->mc_top] == 0) {
7769 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7770 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7771 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7773 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7774 key.mv_size = NODEKSZ(srcnode);
7775 key.mv_data = NODEKEY(srcnode);
7777 DPRINTF(("update separator for source page %"Z"u to [%s]",
7778 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7779 mdb_cursor_copy(csrc, &mn);
7782 /* We want mdb_rebalance to find mn when doing fixups */
7783 WITH_CURSOR_TRACKING(mn,
7784 rc = mdb_update_key(&mn, &key));
7788 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7790 indx_t ix = csrc->mc_ki[csrc->mc_top];
7791 nullkey.mv_size = 0;
7792 csrc->mc_ki[csrc->mc_top] = 0;
7793 rc = mdb_update_key(csrc, &nullkey);
7794 csrc->mc_ki[csrc->mc_top] = ix;
7795 mdb_cassert(csrc, rc == MDB_SUCCESS);
7799 if (cdst->mc_ki[cdst->mc_top] == 0) {
7800 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7801 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7802 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7804 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7805 key.mv_size = NODEKSZ(srcnode);
7806 key.mv_data = NODEKEY(srcnode);
7808 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7809 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7810 mdb_cursor_copy(cdst, &mn);
7813 /* We want mdb_rebalance to find mn when doing fixups */
7814 WITH_CURSOR_TRACKING(mn,
7815 rc = mdb_update_key(&mn, &key));
7819 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7821 indx_t ix = cdst->mc_ki[cdst->mc_top];
7822 nullkey.mv_size = 0;
7823 cdst->mc_ki[cdst->mc_top] = 0;
7824 rc = mdb_update_key(cdst, &nullkey);
7825 cdst->mc_ki[cdst->mc_top] = ix;
7826 mdb_cassert(cdst, rc == MDB_SUCCESS);
7833 /** Merge one page into another.
7834 * The nodes from the page pointed to by \b csrc will
7835 * be copied to the page pointed to by \b cdst and then
7836 * the \b csrc page will be freed.
7837 * @param[in] csrc Cursor pointing to the source page.
7838 * @param[in] cdst Cursor pointing to the destination page.
7839 * @return 0 on success, non-zero on failure.
7842 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7844 MDB_page *psrc, *pdst;
7851 psrc = csrc->mc_pg[csrc->mc_top];
7852 pdst = cdst->mc_pg[cdst->mc_top];
7854 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7856 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7857 mdb_cassert(csrc, cdst->mc_snum > 1);
7859 /* Mark dst as dirty. */
7860 if ((rc = mdb_page_touch(cdst)))
7863 /* get dst page again now that we've touched it. */
7864 pdst = cdst->mc_pg[cdst->mc_top];
7866 /* Move all nodes from src to dst.
7868 j = nkeys = NUMKEYS(pdst);
7869 if (IS_LEAF2(psrc)) {
7870 key.mv_size = csrc->mc_db->md_pad;
7871 key.mv_data = METADATA(psrc);
7872 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7873 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7874 if (rc != MDB_SUCCESS)
7876 key.mv_data = (char *)key.mv_data + key.mv_size;
7879 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7880 srcnode = NODEPTR(psrc, i);
7881 if (i == 0 && IS_BRANCH(psrc)) {
7884 mdb_cursor_copy(csrc, &mn);
7885 mn.mc_xcursor = NULL;
7886 /* must find the lowest key below src */
7887 rc = mdb_page_search_lowest(&mn);
7890 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7891 key.mv_size = mn.mc_db->md_pad;
7892 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7894 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7895 key.mv_size = NODEKSZ(s2);
7896 key.mv_data = NODEKEY(s2);
7899 key.mv_size = srcnode->mn_ksize;
7900 key.mv_data = NODEKEY(srcnode);
7903 data.mv_size = NODEDSZ(srcnode);
7904 data.mv_data = NODEDATA(srcnode);
7905 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7906 if (rc != MDB_SUCCESS)
7911 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7912 pdst->mp_pgno, NUMKEYS(pdst),
7913 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7915 /* Unlink the src page from parent and add to free list.
7918 mdb_node_del(csrc, 0);
7919 if (csrc->mc_ki[csrc->mc_top] == 0) {
7921 rc = mdb_update_key(csrc, &key);
7929 psrc = csrc->mc_pg[csrc->mc_top];
7930 /* If not operating on FreeDB, allow this page to be reused
7931 * in this txn. Otherwise just add to free list.
7933 rc = mdb_page_loose(csrc, psrc);
7937 csrc->mc_db->md_leaf_pages--;
7939 csrc->mc_db->md_branch_pages--;
7941 /* Adjust other cursors pointing to mp */
7942 MDB_cursor *m2, *m3;
7943 MDB_dbi dbi = csrc->mc_dbi;
7944 unsigned int top = csrc->mc_top;
7946 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7947 if (csrc->mc_flags & C_SUB)
7948 m3 = &m2->mc_xcursor->mx_cursor;
7951 if (m3 == csrc) continue;
7952 if (m3->mc_snum < csrc->mc_snum) continue;
7953 if (m3->mc_pg[top] == psrc) {
7954 m3->mc_pg[top] = pdst;
7955 m3->mc_ki[top] += nkeys;
7956 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
7957 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
7958 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
7961 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
7963 MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
7964 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
7965 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
7970 unsigned int snum = cdst->mc_snum;
7971 uint16_t depth = cdst->mc_db->md_depth;
7972 mdb_cursor_pop(cdst);
7973 rc = mdb_rebalance(cdst);
7974 /* Did the tree height change? */
7975 if (depth != cdst->mc_db->md_depth)
7976 snum += cdst->mc_db->md_depth - depth;
7977 cdst->mc_snum = snum;
7978 cdst->mc_top = snum-1;
7983 /** Copy the contents of a cursor.
7984 * @param[in] csrc The cursor to copy from.
7985 * @param[out] cdst The cursor to copy to.
7988 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7992 cdst->mc_txn = csrc->mc_txn;
7993 cdst->mc_dbi = csrc->mc_dbi;
7994 cdst->mc_db = csrc->mc_db;
7995 cdst->mc_dbx = csrc->mc_dbx;
7996 cdst->mc_snum = csrc->mc_snum;
7997 cdst->mc_top = csrc->mc_top;
7998 cdst->mc_flags = csrc->mc_flags;
8000 for (i=0; i<csrc->mc_snum; i++) {
8001 cdst->mc_pg[i] = csrc->mc_pg[i];
8002 cdst->mc_ki[i] = csrc->mc_ki[i];
8006 /** Rebalance the tree after a delete operation.
8007 * @param[in] mc Cursor pointing to the page where rebalancing
8009 * @return 0 on success, non-zero on failure.
8012 mdb_rebalance(MDB_cursor *mc)
8016 unsigned int ptop, minkeys, thresh;
8020 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8025 thresh = FILL_THRESHOLD;
8027 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8028 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8029 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8030 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8032 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8033 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8034 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8035 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8039 if (mc->mc_snum < 2) {
8040 MDB_page *mp = mc->mc_pg[0];
8042 DPUTS("Can't rebalance a subpage, ignoring");
8045 if (NUMKEYS(mp) == 0) {
8046 DPUTS("tree is completely empty");
8047 mc->mc_db->md_root = P_INVALID;
8048 mc->mc_db->md_depth = 0;
8049 mc->mc_db->md_leaf_pages = 0;
8050 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8053 /* Adjust cursors pointing to mp */
8056 mc->mc_flags &= ~C_INITIALIZED;
8058 MDB_cursor *m2, *m3;
8059 MDB_dbi dbi = mc->mc_dbi;
8061 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8062 if (mc->mc_flags & C_SUB)
8063 m3 = &m2->mc_xcursor->mx_cursor;
8066 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8068 if (m3->mc_pg[0] == mp) {
8071 m3->mc_flags &= ~C_INITIALIZED;
8075 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8077 DPUTS("collapsing root page!");
8078 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8081 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8082 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
8085 mc->mc_db->md_depth--;
8086 mc->mc_db->md_branch_pages--;
8087 mc->mc_ki[0] = mc->mc_ki[1];
8088 for (i = 1; i<mc->mc_db->md_depth; i++) {
8089 mc->mc_pg[i] = mc->mc_pg[i+1];
8090 mc->mc_ki[i] = mc->mc_ki[i+1];
8093 /* Adjust other cursors pointing to mp */
8094 MDB_cursor *m2, *m3;
8095 MDB_dbi dbi = mc->mc_dbi;
8097 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8098 if (mc->mc_flags & C_SUB)
8099 m3 = &m2->mc_xcursor->mx_cursor;
8102 if (m3 == mc) continue;
8103 if (!(m3->mc_flags & C_INITIALIZED))
8105 if (m3->mc_pg[0] == mp) {
8106 for (i=0; i<mc->mc_db->md_depth; i++) {
8107 m3->mc_pg[i] = m3->mc_pg[i+1];
8108 m3->mc_ki[i] = m3->mc_ki[i+1];
8116 DPUTS("root page doesn't need rebalancing");
8120 /* The parent (branch page) must have at least 2 pointers,
8121 * otherwise the tree is invalid.
8123 ptop = mc->mc_top-1;
8124 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8126 /* Leaf page fill factor is below the threshold.
8127 * Try to move keys from left or right neighbor, or
8128 * merge with a neighbor page.
8133 mdb_cursor_copy(mc, &mn);
8134 mn.mc_xcursor = NULL;
8136 oldki = mc->mc_ki[mc->mc_top];
8137 if (mc->mc_ki[ptop] == 0) {
8138 /* We're the leftmost leaf in our parent.
8140 DPUTS("reading right neighbor");
8142 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8143 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8146 mn.mc_ki[mn.mc_top] = 0;
8147 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8150 /* There is at least one neighbor to the left.
8152 DPUTS("reading left neighbor");
8154 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8155 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
8158 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8159 mc->mc_ki[mc->mc_top] = 0;
8163 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8164 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8165 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8167 /* If the neighbor page is above threshold and has enough keys,
8168 * move one key from it. Otherwise we should try to merge them.
8169 * (A branch page must never have less than 2 keys.)
8171 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8172 rc = mdb_node_move(&mn, mc, fromleft);
8174 /* if we inserted on left, bump position up */
8179 rc = mdb_page_merge(&mn, mc);
8181 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8182 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8183 /* We want mdb_rebalance to find mn when doing fixups */
8184 WITH_CURSOR_TRACKING(mn,
8185 rc = mdb_page_merge(mc, &mn));
8186 mdb_cursor_copy(&mn, mc);
8188 mc->mc_flags &= ~C_EOF;
8190 mc->mc_ki[mc->mc_top] = oldki;
8194 /** Complete a delete operation started by #mdb_cursor_del(). */
8196 mdb_cursor_del0(MDB_cursor *mc)
8202 MDB_cursor *m2, *m3;
8203 MDB_dbi dbi = mc->mc_dbi;
8205 ki = mc->mc_ki[mc->mc_top];
8206 mp = mc->mc_pg[mc->mc_top];
8207 mdb_node_del(mc, mc->mc_db->md_pad);
8208 mc->mc_db->md_entries--;
8210 /* Adjust other cursors pointing to mp */
8211 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8212 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8213 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8215 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8217 if (m3->mc_pg[mc->mc_top] == mp) {
8218 if (m3->mc_ki[mc->mc_top] == ki) {
8219 m3->mc_flags |= C_DEL;
8220 } else if (m3->mc_ki[mc->mc_top] > ki) {
8221 m3->mc_ki[mc->mc_top]--;
8223 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
8224 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8225 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8226 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8231 rc = mdb_rebalance(mc);
8233 if (rc == MDB_SUCCESS) {
8234 /* DB is totally empty now, just bail out.
8235 * Other cursors adjustments were already done
8236 * by mdb_rebalance and aren't needed here.
8241 mp = mc->mc_pg[mc->mc_top];
8242 nkeys = NUMKEYS(mp);
8244 /* Adjust other cursors pointing to mp */
8245 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8246 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8247 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8249 if (m3->mc_snum < mc->mc_snum)
8251 if (m3->mc_pg[mc->mc_top] == mp) {
8252 /* if m3 points past last node in page, find next sibling */
8253 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
8254 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8255 rc = mdb_cursor_sibling(m3, 1);
8256 if (rc == MDB_NOTFOUND) {
8257 m3->mc_flags |= C_EOF;
8262 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8263 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
8264 if (node->mn_flags & F_DUPDATA) {
8265 mdb_xcursor_init1(m3, node);
8266 m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL;
8272 mc->mc_flags |= C_DEL;
8276 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8281 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8282 MDB_val *key, MDB_val *data)
8284 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8287 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8288 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8290 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8291 /* must ignore any data */
8295 return mdb_del0(txn, dbi, key, data, 0);
8299 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8300 MDB_val *key, MDB_val *data, unsigned flags)
8305 MDB_val rdata, *xdata;
8309 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8311 mdb_cursor_init(&mc, txn, dbi, &mx);
8320 flags |= MDB_NODUPDATA;
8322 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8324 /* let mdb_page_split know about this cursor if needed:
8325 * delete will trigger a rebalance; if it needs to move
8326 * a node from one page to another, it will have to
8327 * update the parent's separator key(s). If the new sepkey
8328 * is larger than the current one, the parent page may
8329 * run out of space, triggering a split. We need this
8330 * cursor to be consistent until the end of the rebalance.
8332 mc.mc_flags |= C_UNTRACK;
8333 mc.mc_next = txn->mt_cursors[dbi];
8334 txn->mt_cursors[dbi] = &mc;
8335 rc = mdb_cursor_del(&mc, flags);
8336 txn->mt_cursors[dbi] = mc.mc_next;
8341 /** Split a page and insert a new node.
8342 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8343 * The cursor will be updated to point to the actual page and index where
8344 * the node got inserted after the split.
8345 * @param[in] newkey The key for the newly inserted node.
8346 * @param[in] newdata The data for the newly inserted node.
8347 * @param[in] newpgno The page number, if the new node is a branch node.
8348 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8349 * @return 0 on success, non-zero on failure.
8352 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8353 unsigned int nflags)
8356 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8359 int i, j, split_indx, nkeys, pmax;
8360 MDB_env *env = mc->mc_txn->mt_env;
8362 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8363 MDB_page *copy = NULL;
8364 MDB_page *mp, *rp, *pp;
8369 mp = mc->mc_pg[mc->mc_top];
8370 newindx = mc->mc_ki[mc->mc_top];
8371 nkeys = NUMKEYS(mp);
8373 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8374 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8375 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8377 /* Create a right sibling. */
8378 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8380 rp->mp_pad = mp->mp_pad;
8381 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8383 /* Usually when splitting the root page, the cursor
8384 * height is 1. But when called from mdb_update_key,
8385 * the cursor height may be greater because it walks
8386 * up the stack while finding the branch slot to update.
8388 if (mc->mc_top < 1) {
8389 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8391 /* shift current top to make room for new parent */
8392 for (i=mc->mc_snum; i>0; i--) {
8393 mc->mc_pg[i] = mc->mc_pg[i-1];
8394 mc->mc_ki[i] = mc->mc_ki[i-1];
8398 mc->mc_db->md_root = pp->mp_pgno;
8399 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8400 new_root = mc->mc_db->md_depth++;
8402 /* Add left (implicit) pointer. */
8403 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8404 /* undo the pre-push */
8405 mc->mc_pg[0] = mc->mc_pg[1];
8406 mc->mc_ki[0] = mc->mc_ki[1];
8407 mc->mc_db->md_root = mp->mp_pgno;
8408 mc->mc_db->md_depth--;
8415 ptop = mc->mc_top-1;
8416 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8419 mdb_cursor_copy(mc, &mn);
8420 mn.mc_xcursor = NULL;
8421 mn.mc_pg[mn.mc_top] = rp;
8422 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8424 if (nflags & MDB_APPEND) {
8425 mn.mc_ki[mn.mc_top] = 0;
8427 split_indx = newindx;
8431 split_indx = (nkeys+1) / 2;
8436 unsigned int lsize, rsize, ksize;
8437 /* Move half of the keys to the right sibling */
8438 x = mc->mc_ki[mc->mc_top] - split_indx;
8439 ksize = mc->mc_db->md_pad;
8440 split = LEAF2KEY(mp, split_indx, ksize);
8441 rsize = (nkeys - split_indx) * ksize;
8442 lsize = (nkeys - split_indx) * sizeof(indx_t);
8443 mp->mp_lower -= lsize;
8444 rp->mp_lower += lsize;
8445 mp->mp_upper += rsize - lsize;
8446 rp->mp_upper -= rsize - lsize;
8447 sepkey.mv_size = ksize;
8448 if (newindx == split_indx) {
8449 sepkey.mv_data = newkey->mv_data;
8451 sepkey.mv_data = split;
8454 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8455 memcpy(rp->mp_ptrs, split, rsize);
8456 sepkey.mv_data = rp->mp_ptrs;
8457 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8458 memcpy(ins, newkey->mv_data, ksize);
8459 mp->mp_lower += sizeof(indx_t);
8460 mp->mp_upper -= ksize - sizeof(indx_t);
8463 memcpy(rp->mp_ptrs, split, x * ksize);
8464 ins = LEAF2KEY(rp, x, ksize);
8465 memcpy(ins, newkey->mv_data, ksize);
8466 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8467 rp->mp_lower += sizeof(indx_t);
8468 rp->mp_upper -= ksize - sizeof(indx_t);
8469 mc->mc_ki[mc->mc_top] = x;
8472 int psize, nsize, k;
8473 /* Maximum free space in an empty page */
8474 pmax = env->me_psize - PAGEHDRSZ;
8476 nsize = mdb_leaf_size(env, newkey, newdata);
8478 nsize = mdb_branch_size(env, newkey);
8479 nsize = EVEN(nsize);
8481 /* grab a page to hold a temporary copy */
8482 copy = mdb_page_malloc(mc->mc_txn, 1);
8487 copy->mp_pgno = mp->mp_pgno;
8488 copy->mp_flags = mp->mp_flags;
8489 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8490 copy->mp_upper = env->me_psize - PAGEBASE;
8492 /* prepare to insert */
8493 for (i=0, j=0; i<nkeys; i++) {
8495 copy->mp_ptrs[j++] = 0;
8497 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8500 /* When items are relatively large the split point needs
8501 * to be checked, because being off-by-one will make the
8502 * difference between success or failure in mdb_node_add.
8504 * It's also relevant if a page happens to be laid out
8505 * such that one half of its nodes are all "small" and
8506 * the other half of its nodes are "large." If the new
8507 * item is also "large" and falls on the half with
8508 * "large" nodes, it also may not fit.
8510 * As a final tweak, if the new item goes on the last
8511 * spot on the page (and thus, onto the new page), bias
8512 * the split so the new page is emptier than the old page.
8513 * This yields better packing during sequential inserts.
8515 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8516 /* Find split point */
8518 if (newindx <= split_indx || newindx >= nkeys) {
8520 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8525 for (; i!=k; i+=j) {
8530 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8531 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8533 if (F_ISSET(node->mn_flags, F_BIGDATA))
8534 psize += sizeof(pgno_t);
8536 psize += NODEDSZ(node);
8538 psize = EVEN(psize);
8540 if (psize > pmax || i == k-j) {
8541 split_indx = i + (j<0);
8546 if (split_indx == newindx) {
8547 sepkey.mv_size = newkey->mv_size;
8548 sepkey.mv_data = newkey->mv_data;
8550 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8551 sepkey.mv_size = node->mn_ksize;
8552 sepkey.mv_data = NODEKEY(node);
8557 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8559 /* Copy separator key to the parent.
8561 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8562 int snum = mc->mc_snum;
8566 /* We want other splits to find mn when doing fixups */
8567 WITH_CURSOR_TRACKING(mn,
8568 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
8573 if (mc->mc_snum > snum) {
8576 /* Right page might now have changed parent.
8577 * Check if left page also changed parent.
8579 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8580 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8581 for (i=0; i<ptop; i++) {
8582 mc->mc_pg[i] = mn.mc_pg[i];
8583 mc->mc_ki[i] = mn.mc_ki[i];
8585 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8586 if (mn.mc_ki[ptop]) {
8587 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8589 /* find right page's left sibling */
8590 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8591 mdb_cursor_sibling(mc, 0);
8596 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8599 if (rc != MDB_SUCCESS) {
8602 if (nflags & MDB_APPEND) {
8603 mc->mc_pg[mc->mc_top] = rp;
8604 mc->mc_ki[mc->mc_top] = 0;
8605 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8608 for (i=0; i<mc->mc_top; i++)
8609 mc->mc_ki[i] = mn.mc_ki[i];
8610 } else if (!IS_LEAF2(mp)) {
8612 mc->mc_pg[mc->mc_top] = rp;
8617 rkey.mv_data = newkey->mv_data;
8618 rkey.mv_size = newkey->mv_size;
8624 /* Update index for the new key. */
8625 mc->mc_ki[mc->mc_top] = j;
8627 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8628 rkey.mv_data = NODEKEY(node);
8629 rkey.mv_size = node->mn_ksize;
8631 xdata.mv_data = NODEDATA(node);
8632 xdata.mv_size = NODEDSZ(node);
8635 pgno = NODEPGNO(node);
8636 flags = node->mn_flags;
8639 if (!IS_LEAF(mp) && j == 0) {
8640 /* First branch index doesn't need key data. */
8644 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8650 mc->mc_pg[mc->mc_top] = copy;
8655 } while (i != split_indx);
8657 nkeys = NUMKEYS(copy);
8658 for (i=0; i<nkeys; i++)
8659 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8660 mp->mp_lower = copy->mp_lower;
8661 mp->mp_upper = copy->mp_upper;
8662 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8663 env->me_psize - copy->mp_upper - PAGEBASE);
8665 /* reset back to original page */
8666 if (newindx < split_indx) {
8667 mc->mc_pg[mc->mc_top] = mp;
8669 mc->mc_pg[mc->mc_top] = rp;
8671 /* Make sure mc_ki is still valid.
8673 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8674 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8675 for (i=0; i<=ptop; i++) {
8676 mc->mc_pg[i] = mn.mc_pg[i];
8677 mc->mc_ki[i] = mn.mc_ki[i];
8681 if (nflags & MDB_RESERVE) {
8682 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8683 if (!(node->mn_flags & F_BIGDATA))
8684 newdata->mv_data = NODEDATA(node);
8687 if (newindx >= split_indx) {
8688 mc->mc_pg[mc->mc_top] = rp;
8690 /* Make sure mc_ki is still valid.
8692 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8693 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8694 for (i=0; i<=ptop; i++) {
8695 mc->mc_pg[i] = mn.mc_pg[i];
8696 mc->mc_ki[i] = mn.mc_ki[i];
8703 /* Adjust other cursors pointing to mp */
8704 MDB_cursor *m2, *m3;
8705 MDB_dbi dbi = mc->mc_dbi;
8706 nkeys = NUMKEYS(mp);
8708 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8709 if (mc->mc_flags & C_SUB)
8710 m3 = &m2->mc_xcursor->mx_cursor;
8715 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8719 /* sub cursors may be on different DB */
8720 if (m3->mc_pg[0] != mp)
8723 for (k=new_root; k>=0; k--) {
8724 m3->mc_ki[k+1] = m3->mc_ki[k];
8725 m3->mc_pg[k+1] = m3->mc_pg[k];
8727 if (m3->mc_ki[0] >= nkeys) {
8732 m3->mc_pg[0] = mc->mc_pg[0];
8736 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8737 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8738 m3->mc_ki[mc->mc_top]++;
8739 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8740 m3->mc_pg[mc->mc_top] = rp;
8741 m3->mc_ki[mc->mc_top] -= nkeys;
8742 for (i=0; i<mc->mc_top; i++) {
8743 m3->mc_ki[i] = mn.mc_ki[i];
8744 m3->mc_pg[i] = mn.mc_pg[i];
8747 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8748 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8751 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8753 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8754 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8755 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8759 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8762 if (copy) /* tmp page */
8763 mdb_page_free(env, copy);
8765 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8770 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8771 MDB_val *key, MDB_val *data, unsigned int flags)
8777 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8780 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8783 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8784 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8786 mdb_cursor_init(&mc, txn, dbi, &mx);
8787 mc.mc_next = txn->mt_cursors[dbi];
8788 txn->mt_cursors[dbi] = &mc;
8789 rc = mdb_cursor_put(&mc, key, data, flags);
8790 txn->mt_cursors[dbi] = mc.mc_next;
8795 #define MDB_WBUF (1024*1024)
8798 /** State needed for a compacting copy. */
8799 typedef struct mdb_copy {
8800 pthread_mutex_t mc_mutex;
8801 pthread_cond_t mc_cond;
8808 pgno_t mc_next_pgno;
8811 volatile int mc_new;
8816 /** Dedicated writer thread for compacting copy. */
8817 static THREAD_RET ESECT CALL_CONV
8818 mdb_env_copythr(void *arg)
8822 int toggle = 0, wsize, rc;
8825 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8828 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8831 pthread_mutex_lock(&my->mc_mutex);
8833 pthread_cond_signal(&my->mc_cond);
8836 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8837 if (my->mc_new < 0) {
8842 wsize = my->mc_wlen[toggle];
8843 ptr = my->mc_wbuf[toggle];
8846 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8850 } else if (len > 0) {
8864 /* If there's an overflow page tail, write it too */
8865 if (my->mc_olen[toggle]) {
8866 wsize = my->mc_olen[toggle];
8867 ptr = my->mc_over[toggle];
8868 my->mc_olen[toggle] = 0;
8871 my->mc_wlen[toggle] = 0;
8873 pthread_cond_signal(&my->mc_cond);
8875 pthread_cond_signal(&my->mc_cond);
8876 pthread_mutex_unlock(&my->mc_mutex);
8877 return (THREAD_RET)0;
8881 /** Tell the writer thread there's a buffer ready to write */
8883 mdb_env_cthr_toggle(mdb_copy *my, int st)
8885 int toggle = my->mc_toggle ^ 1;
8886 pthread_mutex_lock(&my->mc_mutex);
8887 if (my->mc_status) {
8888 pthread_mutex_unlock(&my->mc_mutex);
8889 return my->mc_status;
8891 while (my->mc_new == 1)
8892 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8894 my->mc_toggle = toggle;
8895 pthread_cond_signal(&my->mc_cond);
8896 pthread_mutex_unlock(&my->mc_mutex);
8900 /** Depth-first tree traversal for compacting copy. */
8902 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8904 MDB_cursor mc = {0};
8905 MDB_txn *txn = my->mc_txn;
8907 MDB_page *mo, *mp, *leaf;
8912 /* Empty DB, nothing to do */
8913 if (*pg == P_INVALID)
8919 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8922 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8926 /* Make cursor pages writable */
8927 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8931 for (i=0; i<mc.mc_top; i++) {
8932 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8933 mc.mc_pg[i] = (MDB_page *)ptr;
8934 ptr += my->mc_env->me_psize;
8937 /* This is writable space for a leaf page. Usually not needed. */
8938 leaf = (MDB_page *)ptr;
8940 toggle = my->mc_toggle;
8941 while (mc.mc_snum > 0) {
8943 mp = mc.mc_pg[mc.mc_top];
8947 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8948 for (i=0; i<n; i++) {
8949 ni = NODEPTR(mp, i);
8950 if (ni->mn_flags & F_BIGDATA) {
8954 /* Need writable leaf */
8956 mc.mc_pg[mc.mc_top] = leaf;
8957 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8959 ni = NODEPTR(mp, i);
8962 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8963 rc = mdb_page_get(txn, pg, &omp, NULL);
8966 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8967 rc = mdb_env_cthr_toggle(my, 1);
8970 toggle = my->mc_toggle;
8972 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8973 memcpy(mo, omp, my->mc_env->me_psize);
8974 mo->mp_pgno = my->mc_next_pgno;
8975 my->mc_next_pgno += omp->mp_pages;
8976 my->mc_wlen[toggle] += my->mc_env->me_psize;
8977 if (omp->mp_pages > 1) {
8978 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8979 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8980 rc = mdb_env_cthr_toggle(my, 1);
8983 toggle = my->mc_toggle;
8985 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8986 } else if (ni->mn_flags & F_SUBDATA) {
8989 /* Need writable leaf */
8991 mc.mc_pg[mc.mc_top] = leaf;
8992 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8994 ni = NODEPTR(mp, i);
8997 memcpy(&db, NODEDATA(ni), sizeof(db));
8998 my->mc_toggle = toggle;
8999 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9002 toggle = my->mc_toggle;
9003 memcpy(NODEDATA(ni), &db, sizeof(db));
9008 mc.mc_ki[mc.mc_top]++;
9009 if (mc.mc_ki[mc.mc_top] < n) {
9012 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9014 rc = mdb_page_get(txn, pg, &mp, NULL);
9019 mc.mc_ki[mc.mc_top] = 0;
9020 if (IS_BRANCH(mp)) {
9021 /* Whenever we advance to a sibling branch page,
9022 * we must proceed all the way down to its first leaf.
9024 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9027 mc.mc_pg[mc.mc_top] = mp;
9031 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9032 rc = mdb_env_cthr_toggle(my, 1);
9035 toggle = my->mc_toggle;
9037 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9038 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9039 mo->mp_pgno = my->mc_next_pgno++;
9040 my->mc_wlen[toggle] += my->mc_env->me_psize;
9042 /* Update parent if there is one */
9043 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9044 SETPGNO(ni, mo->mp_pgno);
9045 mdb_cursor_pop(&mc);
9047 /* Otherwise we're done */
9057 /** Copy environment with compaction. */
9059 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9064 MDB_txn *txn = NULL;
9069 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
9070 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
9071 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9072 if (my.mc_wbuf[0] == NULL)
9075 pthread_mutex_init(&my.mc_mutex, NULL);
9076 pthread_cond_init(&my.mc_cond, NULL);
9077 #ifdef HAVE_MEMALIGN
9078 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9079 if (my.mc_wbuf[0] == NULL)
9082 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
9087 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9088 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9093 my.mc_next_pgno = NUM_METAS;
9099 THREAD_CREATE(thr, mdb_env_copythr, &my);
9101 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9105 mp = (MDB_page *)my.mc_wbuf[0];
9106 memset(mp, 0, NUM_METAS * env->me_psize);
9108 mp->mp_flags = P_META;
9109 mm = (MDB_meta *)METADATA(mp);
9110 mdb_env_init_meta0(env, mm);
9111 mm->mm_address = env->me_metas[0]->mm_address;
9113 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9115 mp->mp_flags = P_META;
9116 *(MDB_meta *)METADATA(mp) = *mm;
9117 mm = (MDB_meta *)METADATA(mp);
9119 /* Count the number of free pages, subtract from lastpg to find
9120 * number of active pages
9123 MDB_ID freecount = 0;
9126 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9127 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9128 freecount += *(MDB_ID *)data.mv_data;
9129 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9130 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9131 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9133 /* Set metapage 1 */
9134 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
9135 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9136 if (mm->mm_last_pg > NUM_METAS-1) {
9137 mm->mm_dbs[MAIN_DBI].md_root = mm->mm_last_pg;
9140 mm->mm_dbs[MAIN_DBI].md_root = P_INVALID;
9143 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9145 pthread_mutex_lock(&my.mc_mutex);
9147 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9148 pthread_mutex_unlock(&my.mc_mutex);
9149 rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0);
9150 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
9151 rc = mdb_env_cthr_toggle(&my, 1);
9152 mdb_env_cthr_toggle(&my, -1);
9153 pthread_mutex_lock(&my.mc_mutex);
9155 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
9156 pthread_mutex_unlock(&my.mc_mutex);
9161 CloseHandle(my.mc_cond);
9162 CloseHandle(my.mc_mutex);
9163 _aligned_free(my.mc_wbuf[0]);
9165 pthread_cond_destroy(&my.mc_cond);
9166 pthread_mutex_destroy(&my.mc_mutex);
9167 free(my.mc_wbuf[0]);
9172 /** Copy environment as-is. */
9174 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9176 MDB_txn *txn = NULL;
9177 mdb_mutexref_t wmutex = NULL;
9183 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9187 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9190 /* Do the lock/unlock of the reader mutex before starting the
9191 * write txn. Otherwise other read txns could block writers.
9193 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9198 /* We must start the actual read txn after blocking writers */
9199 mdb_txn_end(txn, MDB_END_RESET_TMP);
9201 /* Temporarily block writers until we snapshot the meta pages */
9202 wmutex = env->me_wmutex;
9203 if (LOCK_MUTEX(rc, env, wmutex))
9206 rc = mdb_txn_renew0(txn);
9208 UNLOCK_MUTEX(wmutex);
9213 wsize = env->me_psize * NUM_METAS;
9217 DO_WRITE(rc, fd, ptr, w2, len);
9221 } else if (len > 0) {
9227 /* Non-blocking or async handles are not supported */
9233 UNLOCK_MUTEX(wmutex);
9238 w3 = txn->mt_next_pgno * env->me_psize;
9241 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9248 if (wsize > MAX_WRITE)
9252 DO_WRITE(rc, fd, ptr, w2, len);
9256 } else if (len > 0) {
9273 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9275 if (flags & MDB_CP_COMPACT)
9276 return mdb_env_copyfd1(env, fd);
9278 return mdb_env_copyfd0(env, fd);
9282 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9284 return mdb_env_copyfd2(env, fd, 0);
9288 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9292 HANDLE newfd = INVALID_HANDLE_VALUE;
9297 if (env->me_flags & MDB_NOSUBDIR) {
9298 lpath = (char *)path;
9301 len += sizeof(DATANAME);
9302 lpath = malloc(len);
9305 sprintf(lpath, "%s" DATANAME, path);
9308 /* The destination path must exist, but the destination file must not.
9309 * We don't want the OS to cache the writes, since the source data is
9310 * already in the OS cache.
9313 rc = utf8_to_utf16(lpath, -1, &wpath, NULL);
9316 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9317 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9320 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9322 if (newfd == INVALID_HANDLE_VALUE) {
9327 if (env->me_psize >= env->me_os_psize) {
9329 /* Set O_DIRECT if the file system supports it */
9330 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9331 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9333 #ifdef F_NOCACHE /* __APPLE__ */
9334 rc = fcntl(newfd, F_NOCACHE, 1);
9342 rc = mdb_env_copyfd2(env, newfd, flags);
9345 if (!(env->me_flags & MDB_NOSUBDIR))
9347 if (newfd != INVALID_HANDLE_VALUE)
9348 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9355 mdb_env_copy(MDB_env *env, const char *path)
9357 return mdb_env_copy2(env, path, 0);
9361 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9363 if (flag & ~CHANGEABLE)
9366 env->me_flags |= flag;
9368 env->me_flags &= ~flag;
9373 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9378 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9383 mdb_env_set_userctx(MDB_env *env, void *ctx)
9387 env->me_userctx = ctx;
9392 mdb_env_get_userctx(MDB_env *env)
9394 return env ? env->me_userctx : NULL;
9398 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9403 env->me_assert_func = func;
9409 mdb_env_get_path(MDB_env *env, const char **arg)
9414 *arg = env->me_path;
9419 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9428 /** Common code for #mdb_stat() and #mdb_env_stat().
9429 * @param[in] env the environment to operate in.
9430 * @param[in] db the #MDB_db record containing the stats to return.
9431 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9432 * @return 0, this function always succeeds.
9435 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9437 arg->ms_psize = env->me_psize;
9438 arg->ms_depth = db->md_depth;
9439 arg->ms_branch_pages = db->md_branch_pages;
9440 arg->ms_leaf_pages = db->md_leaf_pages;
9441 arg->ms_overflow_pages = db->md_overflow_pages;
9442 arg->ms_entries = db->md_entries;
9448 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9452 if (env == NULL || arg == NULL)
9455 meta = mdb_env_pick_meta(env);
9457 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9461 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9465 if (env == NULL || arg == NULL)
9468 meta = mdb_env_pick_meta(env);
9469 arg->me_mapaddr = meta->mm_address;
9470 arg->me_last_pgno = meta->mm_last_pg;
9471 arg->me_last_txnid = meta->mm_txnid;
9473 arg->me_mapsize = env->me_mapsize;
9474 arg->me_maxreaders = env->me_maxreaders;
9475 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9479 /** Set the default comparison functions for a database.
9480 * Called immediately after a database is opened to set the defaults.
9481 * The user can then override them with #mdb_set_compare() or
9482 * #mdb_set_dupsort().
9483 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9484 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9487 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9489 uint16_t f = txn->mt_dbs[dbi].md_flags;
9491 txn->mt_dbxs[dbi].md_cmp =
9492 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9493 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9495 txn->mt_dbxs[dbi].md_dcmp =
9496 !(f & MDB_DUPSORT) ? 0 :
9497 ((f & MDB_INTEGERDUP)
9498 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9499 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9502 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9508 int rc, dbflag, exact;
9509 unsigned int unused = 0, seq;
9513 if (flags & ~VALID_FLAGS)
9515 if (txn->mt_flags & MDB_TXN_BLOCKED)
9521 if (flags & PERSISTENT_FLAGS) {
9522 uint16_t f2 = flags & PERSISTENT_FLAGS;
9523 /* make sure flag changes get committed */
9524 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9525 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9526 txn->mt_flags |= MDB_TXN_DIRTY;
9529 mdb_default_cmp(txn, MAIN_DBI);
9533 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9534 mdb_default_cmp(txn, MAIN_DBI);
9537 /* Is the DB already open? */
9539 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9540 if (!txn->mt_dbxs[i].md_name.mv_size) {
9541 /* Remember this free slot */
9542 if (!unused) unused = i;
9545 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9546 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9552 /* If no free slot and max hit, fail */
9553 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9554 return MDB_DBS_FULL;
9556 /* Cannot mix named databases with some mainDB flags */
9557 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9558 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9560 /* Find the DB info */
9561 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9564 key.mv_data = (void *)name;
9565 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9566 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9567 if (rc == MDB_SUCCESS) {
9568 /* make sure this is actually a DB */
9569 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9570 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9571 return MDB_INCOMPATIBLE;
9572 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
9576 /* Done here so we cannot fail after creating a new DB */
9577 if ((namedup = strdup(name)) == NULL)
9581 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
9582 data.mv_size = sizeof(MDB_db);
9583 data.mv_data = &dummy;
9584 memset(&dummy, 0, sizeof(dummy));
9585 dummy.md_root = P_INVALID;
9586 dummy.md_flags = flags & PERSISTENT_FLAGS;
9587 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9594 /* Got info, register DBI in this txn */
9595 unsigned int slot = unused ? unused : txn->mt_numdbs;
9596 txn->mt_dbxs[slot].md_name.mv_data = namedup;
9597 txn->mt_dbxs[slot].md_name.mv_size = len;
9598 txn->mt_dbxs[slot].md_rel = NULL;
9599 txn->mt_dbflags[slot] = dbflag;
9600 /* txn-> and env-> are the same in read txns, use
9601 * tmp variable to avoid undefined assignment
9603 seq = ++txn->mt_env->me_dbiseqs[slot];
9604 txn->mt_dbiseqs[slot] = seq;
9606 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9608 mdb_default_cmp(txn, slot);
9618 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9620 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9623 if (txn->mt_flags & MDB_TXN_BLOCKED)
9626 if (txn->mt_dbflags[dbi] & DB_STALE) {
9629 /* Stale, must read the DB's root. cursor_init does it for us. */
9630 mdb_cursor_init(&mc, txn, dbi, &mx);
9632 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9635 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9638 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9640 ptr = env->me_dbxs[dbi].md_name.mv_data;
9641 /* If there was no name, this was already closed */
9643 env->me_dbxs[dbi].md_name.mv_data = NULL;
9644 env->me_dbxs[dbi].md_name.mv_size = 0;
9645 env->me_dbflags[dbi] = 0;
9646 env->me_dbiseqs[dbi]++;
9651 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9653 /* We could return the flags for the FREE_DBI too but what's the point? */
9654 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9656 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9660 /** Add all the DB's pages to the free list.
9661 * @param[in] mc Cursor on the DB to free.
9662 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9663 * @return 0 on success, non-zero on failure.
9666 mdb_drop0(MDB_cursor *mc, int subs)
9670 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9671 if (rc == MDB_SUCCESS) {
9672 MDB_txn *txn = mc->mc_txn;
9677 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9678 * This also avoids any P_LEAF2 pages, which have no nodes.
9679 * Also if the DB doesn't have sub-DBs and has no overflow
9680 * pages, omit scanning leaves.
9682 if ((mc->mc_flags & C_SUB) ||
9683 (!subs && !mc->mc_db->md_overflow_pages))
9686 mdb_cursor_copy(mc, &mx);
9687 while (mc->mc_snum > 0) {
9688 MDB_page *mp = mc->mc_pg[mc->mc_top];
9689 unsigned n = NUMKEYS(mp);
9691 for (i=0; i<n; i++) {
9692 ni = NODEPTR(mp, i);
9693 if (ni->mn_flags & F_BIGDATA) {
9696 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9697 rc = mdb_page_get(txn, pg, &omp, NULL);
9700 mdb_cassert(mc, IS_OVERFLOW(omp));
9701 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9705 mc->mc_db->md_overflow_pages -= omp->mp_pages;
9706 if (!mc->mc_db->md_overflow_pages && !subs)
9708 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9709 mdb_xcursor_init1(mc, ni);
9710 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9715 if (!subs && !mc->mc_db->md_overflow_pages)
9718 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9720 for (i=0; i<n; i++) {
9722 ni = NODEPTR(mp, i);
9725 mdb_midl_xappend(txn->mt_free_pgs, pg);
9730 mc->mc_ki[mc->mc_top] = i;
9731 rc = mdb_cursor_sibling(mc, 1);
9733 if (rc != MDB_NOTFOUND)
9735 /* no more siblings, go back to beginning
9736 * of previous level.
9741 for (i=1; i<mc->mc_snum; i++) {
9743 mc->mc_pg[i] = mx.mc_pg[i];
9748 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9751 txn->mt_flags |= MDB_TXN_ERROR;
9752 } else if (rc == MDB_NOTFOUND) {
9755 mc->mc_flags &= ~C_INITIALIZED;
9759 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9761 MDB_cursor *mc, *m2;
9764 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9767 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9770 if (TXN_DBI_CHANGED(txn, dbi))
9773 rc = mdb_cursor_open(txn, dbi, &mc);
9777 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9778 /* Invalidate the dropped DB's cursors */
9779 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9780 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9784 /* Can't delete the main DB */
9785 if (del && dbi >= CORE_DBS) {
9786 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9788 txn->mt_dbflags[dbi] = DB_STALE;
9789 mdb_dbi_close(txn->mt_env, dbi);
9791 txn->mt_flags |= MDB_TXN_ERROR;
9794 /* reset the DB record, mark it dirty */
9795 txn->mt_dbflags[dbi] |= DB_DIRTY;
9796 txn->mt_dbs[dbi].md_depth = 0;
9797 txn->mt_dbs[dbi].md_branch_pages = 0;
9798 txn->mt_dbs[dbi].md_leaf_pages = 0;
9799 txn->mt_dbs[dbi].md_overflow_pages = 0;
9800 txn->mt_dbs[dbi].md_entries = 0;
9801 txn->mt_dbs[dbi].md_root = P_INVALID;
9803 txn->mt_flags |= MDB_TXN_DIRTY;
9806 mdb_cursor_close(mc);
9810 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9812 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9815 txn->mt_dbxs[dbi].md_cmp = cmp;
9819 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9821 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9824 txn->mt_dbxs[dbi].md_dcmp = cmp;
9828 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9830 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9833 txn->mt_dbxs[dbi].md_rel = rel;
9837 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9839 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9842 txn->mt_dbxs[dbi].md_relctx = ctx;
9847 mdb_env_get_maxkeysize(MDB_env *env)
9849 return ENV_MAXKEY(env);
9853 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9855 unsigned int i, rdrs;
9858 int rc = 0, first = 1;
9862 if (!env->me_txns) {
9863 return func("(no reader locks)\n", ctx);
9865 rdrs = env->me_txns->mti_numreaders;
9866 mr = env->me_txns->mti_readers;
9867 for (i=0; i<rdrs; i++) {
9869 txnid_t txnid = mr[i].mr_txnid;
9870 sprintf(buf, txnid == (txnid_t)-1 ?
9871 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9872 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9875 rc = func(" pid thread txnid\n", ctx);
9879 rc = func(buf, ctx);
9885 rc = func("(no active readers)\n", ctx);
9890 /** Insert pid into list if not already present.
9891 * return -1 if already present.
9894 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9896 /* binary search of pid in list */
9898 unsigned cursor = 1;
9900 unsigned n = ids[0];
9903 unsigned pivot = n >> 1;
9904 cursor = base + pivot + 1;
9905 val = pid - ids[cursor];
9910 } else if ( val > 0 ) {
9915 /* found, so it's a duplicate */
9924 for (n = ids[0]; n > cursor; n--)
9931 mdb_reader_check(MDB_env *env, int *dead)
9937 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9940 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9942 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9944 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
9945 unsigned int i, j, rdrs;
9947 MDB_PID_T *pids, pid;
9948 int rc = MDB_SUCCESS, count = 0;
9950 rdrs = env->me_txns->mti_numreaders;
9951 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9955 mr = env->me_txns->mti_readers;
9956 for (i=0; i<rdrs; i++) {
9958 if (pid && pid != env->me_pid) {
9959 if (mdb_pid_insert(pids, pid) == 0) {
9960 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9961 /* Stale reader found */
9964 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9965 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9967 rdrs = 0; /* the above checked all readers */
9969 /* Recheck, a new process may have reused pid */
9970 if (mdb_reader_pid(env, Pidcheck, pid))
9975 if (mr[j].mr_pid == pid) {
9976 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9977 (unsigned) pid, mr[j].mr_txnid));
9982 UNLOCK_MUTEX(rmutex);
9993 #ifdef MDB_ROBUST_SUPPORTED
9994 /** Handle #LOCK_MUTEX0() failure.
9995 * Try to repair the lock file if the mutex owner died.
9996 * @param[in] env the environment handle
9997 * @param[in] mutex LOCK_MUTEX0() mutex
9998 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9999 * @return 0 on success with the mutex locked, or an error code on failure.
10002 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10007 if (rc == MDB_OWNERDEAD) {
10008 /* We own the mutex. Clean up after dead previous owner. */
10010 rlocked = (mutex == env->me_rmutex);
10012 /* Keep mti_txnid updated, otherwise next writer can
10013 * overwrite data which latest meta page refers to.
10015 meta = mdb_env_pick_meta(env);
10016 env->me_txns->mti_txnid = meta->mm_txnid;
10017 /* env is hosed if the dead thread was ours */
10019 env->me_flags |= MDB_FATAL_ERROR;
10020 env->me_txn = NULL;
10024 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10025 (rc ? "this process' env is hosed" : "recovering")));
10026 rc2 = mdb_reader_check0(env, rlocked, NULL);
10028 rc2 = mdb_mutex_consistent(mutex);
10029 if (rc || (rc = rc2)) {
10030 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10031 UNLOCK_MUTEX(mutex);
10037 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10042 #endif /* MDB_ROBUST_SUPPORTED */
10045 #if defined(_WIN32)
10046 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
10050 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
10051 if (need == 0xFFFD)
10055 result = malloc(sizeof(wchar_t) * need);
10058 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
10064 #endif /* defined(_WIN32) */
projects / openldap / blob