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 2011-2017 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.
38 #if defined(__WIN64__)
39 #define _FILE_OFFSET_BITS 64
44 #include <wchar.h> /* get wcscpy() */
46 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
47 * as int64 which is wrong. MSVC doesn't define it at all, so just
51 #define MDB_THR_T DWORD
52 #include <sys/types.h>
55 # include <sys/param.h>
57 # define LITTLE_ENDIAN 1234
58 # define BIG_ENDIAN 4321
59 # define BYTE_ORDER LITTLE_ENDIAN
61 # define SSIZE_MAX INT_MAX
65 #include <sys/types.h>
67 #define MDB_PID_T pid_t
68 #define MDB_THR_T pthread_t
69 #include <sys/param.h>
72 #ifdef HAVE_SYS_FILE_H
78 #if defined(__mips) && defined(__linux)
79 /* MIPS has cache coherency issues, requires explicit cache control */
80 #include <asm/cachectl.h>
81 extern int cacheflush(char *addr, int nbytes, int cache);
82 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
84 #define CACHEFLUSH(addr, bytes, cache)
87 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
88 /** fdatasync is broken on ext3/ext4fs on older kernels, see
89 * description in #mdb_env_open2 comments. You can safely
90 * define MDB_FDATASYNC_WORKS if this code will only be run
91 * on kernels 3.6 and newer.
93 #define BROKEN_FDATASYNC
107 typedef SSIZE_T ssize_t;
112 #if defined(__sun) || defined(ANDROID)
113 /* Most platforms have posix_memalign, older may only have memalign */
114 #define HAVE_MEMALIGN 1
118 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
119 #include <netinet/in.h>
120 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
123 #if defined(__APPLE__) || defined (BSD) || defined(__FreeBSD_kernel__)
124 # define MDB_USE_POSIX_SEM 1
125 # define MDB_FDATASYNC fsync
126 #elif defined(ANDROID)
127 # define MDB_FDATASYNC fsync
133 #ifdef MDB_USE_POSIX_SEM
134 # define MDB_USE_HASH 1
135 #include <semaphore.h>
137 #define MDB_USE_POSIX_MUTEX 1
141 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \
142 + defined(MDB_USE_POSIX_MUTEX) != 1
143 # error "Ambiguous shared-lock implementation"
147 #include <valgrind/memcheck.h>
148 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
149 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
150 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
151 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
152 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
154 #define VGMEMP_CREATE(h,r,z)
155 #define VGMEMP_ALLOC(h,a,s)
156 #define VGMEMP_FREE(h,a)
157 #define VGMEMP_DESTROY(h)
158 #define VGMEMP_DEFINED(a,s)
162 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
163 /* Solaris just defines one or the other */
164 # define LITTLE_ENDIAN 1234
165 # define BIG_ENDIAN 4321
166 # ifdef _LITTLE_ENDIAN
167 # define BYTE_ORDER LITTLE_ENDIAN
169 # define BYTE_ORDER BIG_ENDIAN
172 # define BYTE_ORDER __BYTE_ORDER
176 #ifndef LITTLE_ENDIAN
177 #define LITTLE_ENDIAN __LITTLE_ENDIAN
180 #define BIG_ENDIAN __BIG_ENDIAN
183 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
184 #define MISALIGNED_OK 1
190 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
191 # error "Unknown or unsupported endianness (BYTE_ORDER)"
192 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
193 # error "Two's complement, reasonably sized integer types, please"
197 /** Put infrequently used env functions in separate section */
199 # define ESECT __attribute__ ((section("__TEXT,text_env")))
201 # define ESECT __attribute__ ((section("text_env")))
208 #define CALL_CONV WINAPI
213 /** @defgroup internal LMDB Internals
216 /** @defgroup compat Compatibility Macros
217 * A bunch of macros to minimize the amount of platform-specific ifdefs
218 * needed throughout the rest of the code. When the features this library
219 * needs are similar enough to POSIX to be hidden in a one-or-two line
220 * replacement, this macro approach is used.
224 /** Features under development */
229 /** Wrapper around __func__, which is a C99 feature */
230 #if __STDC_VERSION__ >= 199901L
231 # define mdb_func_ __func__
232 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
233 # define mdb_func_ __FUNCTION__
235 /* If a debug message says <mdb_unknown>(), update the #if statements above */
236 # define mdb_func_ "<mdb_unknown>"
239 /* Internal error codes, not exposed outside liblmdb */
240 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
242 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
243 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
244 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
248 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
250 /** Some platforms define the EOWNERDEAD error code
251 * even though they don't support Robust Mutexes.
252 * Compile with -DMDB_USE_ROBUST=0, or use some other
253 * mechanism like -DMDB_USE_POSIX_SEM instead of
254 * -DMDB_USE_POSIX_MUTEX.
255 * (Posix semaphores are not robust.)
257 #ifndef MDB_USE_ROBUST
258 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
259 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
260 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
261 # define MDB_USE_ROBUST 0
263 # define MDB_USE_ROBUST 1
265 #endif /* !MDB_USE_ROBUST */
267 #if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST)
268 /* glibc < 2.12 only provided _np API */
269 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
270 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
271 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
272 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
273 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
275 #endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */
277 #if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST)
278 #define MDB_ROBUST_SUPPORTED 1
282 #define MDB_USE_HASH 1
283 #define MDB_PIDLOCK 0
284 #define THREAD_RET DWORD
285 #define pthread_t HANDLE
286 #define pthread_mutex_t HANDLE
287 #define pthread_cond_t HANDLE
288 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
289 #define pthread_key_t DWORD
290 #define pthread_self() GetCurrentThreadId()
291 #define pthread_key_create(x,y) \
292 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
293 #define pthread_key_delete(x) TlsFree(x)
294 #define pthread_getspecific(x) TlsGetValue(x)
295 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
296 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
297 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
298 #define pthread_cond_signal(x) SetEvent(*x)
299 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
300 #define THREAD_CREATE(thr,start,arg) \
301 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
302 #define THREAD_FINISH(thr) \
303 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
304 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
305 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
306 #define mdb_mutex_consistent(mutex) 0
307 #define getpid() GetCurrentProcessId()
308 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
309 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
310 #define ErrCode() GetLastError()
311 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
312 #define close(fd) (CloseHandle(fd) ? 0 : -1)
313 #define munmap(ptr,len) UnmapViewOfFile(ptr)
314 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
315 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
317 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
321 #define THREAD_RET void *
322 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
323 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
324 #define Z "z" /**< printf format modifier for size_t */
326 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
327 #define MDB_PIDLOCK 1
329 #ifdef MDB_USE_POSIX_SEM
331 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
332 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
333 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
336 mdb_sem_wait(sem_t *sem)
339 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
343 #else /* MDB_USE_POSIX_MUTEX: */
344 /** Shared mutex/semaphore as the original is stored.
346 * Not for copies. Instead it can be assigned to an #mdb_mutexref_t.
347 * When mdb_mutexref_t is a pointer and mdb_mutex_t is not, then it
348 * is array[size 1] so it can be assigned to the pointer.
350 typedef pthread_mutex_t mdb_mutex_t[1];
351 /** Reference to an #mdb_mutex_t */
352 typedef pthread_mutex_t *mdb_mutexref_t;
353 /** Lock the reader or writer mutex.
354 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
356 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
357 /** Unlock the reader or writer mutex.
359 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
360 /** Mark mutex-protected data as repaired, after death of previous owner.
362 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
363 #endif /* MDB_USE_POSIX_SEM */
365 /** Get the error code for the last failed system function.
367 #define ErrCode() errno
369 /** An abstraction for a file handle.
370 * On POSIX systems file handles are small integers. On Windows
371 * they're opaque pointers.
375 /** A value for an invalid file handle.
376 * Mainly used to initialize file variables and signify that they are
379 #define INVALID_HANDLE_VALUE (-1)
381 /** Get the size of a memory page for the system.
382 * This is the basic size that the platform's memory manager uses, and is
383 * fundamental to the use of memory-mapped files.
385 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
388 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
391 #define MNAME_LEN (sizeof(pthread_mutex_t))
396 #ifdef MDB_ROBUST_SUPPORTED
397 /** Lock mutex, handle any error, set rc = result.
398 * Return 0 on success, nonzero (not rc) on error.
400 #define LOCK_MUTEX(rc, env, mutex) \
401 (((rc) = LOCK_MUTEX0(mutex)) && \
402 ((rc) = mdb_mutex_failed(env, mutex, rc)))
403 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
405 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
406 #define mdb_mutex_failed(env, mutex, rc) (rc)
410 /** A flag for opening a file and requesting synchronous data writes.
411 * This is only used when writing a meta page. It's not strictly needed;
412 * we could just do a normal write and then immediately perform a flush.
413 * But if this flag is available it saves us an extra system call.
415 * @note If O_DSYNC is undefined but exists in /usr/include,
416 * preferably set some compiler flag to get the definition.
420 # define MDB_DSYNC O_DSYNC
422 # define MDB_DSYNC O_SYNC
427 /** Function for flushing the data of a file. Define this to fsync
428 * if fdatasync() is not supported.
430 #ifndef MDB_FDATASYNC
431 # define MDB_FDATASYNC fdatasync
435 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
446 /** A page number in the database.
447 * Note that 64 bit page numbers are overkill, since pages themselves
448 * already represent 12-13 bits of addressable memory, and the OS will
449 * always limit applications to a maximum of 63 bits of address space.
451 * @note In the #MDB_node structure, we only store 48 bits of this value,
452 * which thus limits us to only 60 bits of addressable data.
454 typedef MDB_ID pgno_t;
456 /** A transaction ID.
457 * See struct MDB_txn.mt_txnid for details.
459 typedef MDB_ID txnid_t;
461 /** @defgroup debug Debug Macros
465 /** Enable debug output. Needs variable argument macros (a C99 feature).
466 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
467 * read from and written to the database (used for free space management).
473 static int mdb_debug;
474 static txnid_t mdb_debug_start;
476 /** Print a debug message with printf formatting.
477 * Requires double parenthesis around 2 or more args.
479 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
480 # define DPRINTF0(fmt, ...) \
481 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
483 # define DPRINTF(args) ((void) 0)
485 /** Print a debug string.
486 * The string is printed literally, with no format processing.
488 #define DPUTS(arg) DPRINTF(("%s", arg))
489 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
491 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
494 /** @brief The maximum size of a database page.
496 * It is 32k or 64k, since value-PAGEBASE must fit in
497 * #MDB_page.%mp_upper.
499 * LMDB will use database pages < OS pages if needed.
500 * That causes more I/O in write transactions: The OS must
501 * know (read) the whole page before writing a partial page.
503 * Note that we don't currently support Huge pages. On Linux,
504 * regular data files cannot use Huge pages, and in general
505 * Huge pages aren't actually pageable. We rely on the OS
506 * demand-pager to read our data and page it out when memory
507 * pressure from other processes is high. So until OSs have
508 * actual paging support for Huge pages, they're not viable.
510 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
512 /** The minimum number of keys required in a database page.
513 * Setting this to a larger value will place a smaller bound on the
514 * maximum size of a data item. Data items larger than this size will
515 * be pushed into overflow pages instead of being stored directly in
516 * the B-tree node. This value used to default to 4. With a page size
517 * of 4096 bytes that meant that any item larger than 1024 bytes would
518 * go into an overflow page. That also meant that on average 2-3KB of
519 * each overflow page was wasted space. The value cannot be lower than
520 * 2 because then there would no longer be a tree structure. With this
521 * value, items larger than 2KB will go into overflow pages, and on
522 * average only 1KB will be wasted.
524 #define MDB_MINKEYS 2
526 /** A stamp that identifies a file as an LMDB file.
527 * There's nothing special about this value other than that it is easily
528 * recognizable, and it will reflect any byte order mismatches.
530 #define MDB_MAGIC 0xBEEFC0DE
532 /** The version number for a database's datafile format. */
533 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
534 /** The version number for a database's lockfile format. */
535 #define MDB_LOCK_VERSION 1
537 /** @brief The max size of a key we can write, or 0 for computed max.
539 * This macro should normally be left alone or set to 0.
540 * Note that a database with big keys or dupsort data cannot be
541 * reliably modified by a liblmdb which uses a smaller max.
542 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
544 * Other values are allowed, for backwards compat. However:
545 * A value bigger than the computed max can break if you do not
546 * know what you are doing, and liblmdb <= 0.9.10 can break when
547 * modifying a DB with keys/dupsort data bigger than its max.
549 * Data items in an #MDB_DUPSORT database are also limited to
550 * this size, since they're actually keys of a sub-DB. Keys and
551 * #MDB_DUPSORT data items must fit on a node in a regular page.
553 #ifndef MDB_MAXKEYSIZE
554 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
557 /** The maximum size of a key we can write to the environment. */
559 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
561 #define ENV_MAXKEY(env) ((env)->me_maxkey)
564 /** @brief The maximum size of a data item.
566 * We only store a 32 bit value for node sizes.
568 #define MAXDATASIZE 0xffffffffUL
571 /** Key size which fits in a #DKBUF.
574 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
577 * This is used for printing a hex dump of a key's contents.
579 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
580 /** Display a key in hex.
582 * Invoke a function to display a key in hex.
584 #define DKEY(x) mdb_dkey(x, kbuf)
590 /** An invalid page number.
591 * Mainly used to denote an empty tree.
593 #define P_INVALID (~(pgno_t)0)
595 /** Test if the flags \b f are set in a flag word \b w. */
596 #define F_ISSET(w, f) (((w) & (f)) == (f))
598 /** Round \b n up to an even number. */
599 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
601 /** Used for offsets within a single page.
602 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
605 typedef uint16_t indx_t;
607 /** Default size of memory map.
608 * This is certainly too small for any actual applications. Apps should always set
609 * the size explicitly using #mdb_env_set_mapsize().
611 #define DEFAULT_MAPSIZE 1048576
613 /** @defgroup readers Reader Lock Table
614 * Readers don't acquire any locks for their data access. Instead, they
615 * simply record their transaction ID in the reader table. The reader
616 * mutex is needed just to find an empty slot in the reader table. The
617 * slot's address is saved in thread-specific data so that subsequent read
618 * transactions started by the same thread need no further locking to proceed.
620 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
622 * No reader table is used if the database is on a read-only filesystem, or
623 * if #MDB_NOLOCK is set.
625 * Since the database uses multi-version concurrency control, readers don't
626 * actually need any locking. This table is used to keep track of which
627 * readers are using data from which old transactions, so that we'll know
628 * when a particular old transaction is no longer in use. Old transactions
629 * that have discarded any data pages can then have those pages reclaimed
630 * for use by a later write transaction.
632 * The lock table is constructed such that reader slots are aligned with the
633 * processor's cache line size. Any slot is only ever used by one thread.
634 * This alignment guarantees that there will be no contention or cache
635 * thrashing as threads update their own slot info, and also eliminates
636 * any need for locking when accessing a slot.
638 * A writer thread will scan every slot in the table to determine the oldest
639 * outstanding reader transaction. Any freed pages older than this will be
640 * reclaimed by the writer. The writer doesn't use any locks when scanning
641 * this table. This means that there's no guarantee that the writer will
642 * see the most up-to-date reader info, but that's not required for correct
643 * operation - all we need is to know the upper bound on the oldest reader,
644 * we don't care at all about the newest reader. So the only consequence of
645 * reading stale information here is that old pages might hang around a
646 * while longer before being reclaimed. That's actually good anyway, because
647 * the longer we delay reclaiming old pages, the more likely it is that a
648 * string of contiguous pages can be found after coalescing old pages from
649 * many old transactions together.
652 /** Number of slots in the reader table.
653 * This value was chosen somewhat arbitrarily. 126 readers plus a
654 * couple mutexes fit exactly into 8KB on my development machine.
655 * Applications should set the table size using #mdb_env_set_maxreaders().
657 #define DEFAULT_READERS 126
659 /** The size of a CPU cache line in bytes. We want our lock structures
660 * aligned to this size to avoid false cache line sharing in the
662 * This value works for most CPUs. For Itanium this should be 128.
668 /** The information we store in a single slot of the reader table.
669 * In addition to a transaction ID, we also record the process and
670 * thread ID that owns a slot, so that we can detect stale information,
671 * e.g. threads or processes that went away without cleaning up.
672 * @note We currently don't check for stale records. We simply re-init
673 * the table when we know that we're the only process opening the
676 typedef struct MDB_rxbody {
677 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
678 * Multiple readers that start at the same time will probably have the
679 * same ID here. Again, it's not important to exclude them from
680 * anything; all we need to know is which version of the DB they
681 * started from so we can avoid overwriting any data used in that
682 * particular version.
684 volatile txnid_t mrb_txnid;
685 /** The process ID of the process owning this reader txn. */
686 volatile MDB_PID_T mrb_pid;
687 /** The thread ID of the thread owning this txn. */
688 volatile MDB_THR_T mrb_tid;
691 /** The actual reader record, with cacheline padding. */
692 typedef struct MDB_reader {
695 /** shorthand for mrb_txnid */
696 #define mr_txnid mru.mrx.mrb_txnid
697 #define mr_pid mru.mrx.mrb_pid
698 #define mr_tid mru.mrx.mrb_tid
699 /** cache line alignment */
700 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
704 /** The header for the reader table.
705 * The table resides in a memory-mapped file. (This is a different file
706 * than is used for the main database.)
708 * For POSIX the actual mutexes reside in the shared memory of this
709 * mapped file. On Windows, mutexes are named objects allocated by the
710 * kernel; we store the mutex names in this mapped file so that other
711 * processes can grab them. This same approach is also used on
712 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
713 * process-shared POSIX mutexes. For these cases where a named object
714 * is used, the object name is derived from a 64 bit FNV hash of the
715 * environment pathname. As such, naming collisions are extremely
716 * unlikely. If a collision occurs, the results are unpredictable.
718 typedef struct MDB_txbody {
719 /** Stamp identifying this as an LMDB file. It must be set
722 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
724 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
725 char mtb_rmname[MNAME_LEN];
727 /** Mutex protecting access to this table.
728 * This is the reader table lock used with LOCK_MUTEX().
730 mdb_mutex_t mtb_rmutex;
732 /** The ID of the last transaction committed to the database.
733 * This is recorded here only for convenience; the value can always
734 * be determined by reading the main database meta pages.
736 volatile txnid_t mtb_txnid;
737 /** The number of slots that have been used in the reader table.
738 * This always records the maximum count, it is not decremented
739 * when readers release their slots.
741 volatile unsigned mtb_numreaders;
744 /** The actual reader table definition. */
745 typedef struct MDB_txninfo {
748 #define mti_magic mt1.mtb.mtb_magic
749 #define mti_format mt1.mtb.mtb_format
750 #define mti_rmutex mt1.mtb.mtb_rmutex
751 #define mti_rmname mt1.mtb.mtb_rmname
752 #define mti_txnid mt1.mtb.mtb_txnid
753 #define mti_numreaders mt1.mtb.mtb_numreaders
754 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
757 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
758 char mt2_wmname[MNAME_LEN];
759 #define mti_wmname mt2.mt2_wmname
761 mdb_mutex_t mt2_wmutex;
762 #define mti_wmutex mt2.mt2_wmutex
764 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
766 MDB_reader mti_readers[1];
769 /** Lockfile format signature: version, features and field layout */
770 #define MDB_LOCK_FORMAT \
772 ((MDB_LOCK_VERSION) \
773 /* Flags which describe functionality */ \
774 + (((MDB_PIDLOCK) != 0) << 16)))
777 /** Common header for all page types. The page type depends on #mp_flags.
779 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with
780 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages
781 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header.
783 * #P_OVERFLOW records occupy one or more contiguous pages where only the
784 * first has a page header. They hold the real data of #F_BIGDATA nodes.
786 * #P_SUBP sub-pages are small leaf "pages" with duplicate data.
787 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page.
788 * (Duplicate data can also go in sub-databases, which use normal pages.)
790 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot.
792 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once
793 * in the snapshot: Either used by a database or listed in a freeDB record.
795 typedef struct MDB_page {
796 #define mp_pgno mp_p.p_pgno
797 #define mp_next mp_p.p_next
799 pgno_t p_pgno; /**< page number */
800 struct MDB_page *p_next; /**< for in-memory list of freed pages */
802 uint16_t mp_pad; /**< key size if this is a LEAF2 page */
803 /** @defgroup mdb_page Page Flags
805 * Flags for the page headers.
808 #define P_BRANCH 0x01 /**< branch page */
809 #define P_LEAF 0x02 /**< leaf page */
810 #define P_OVERFLOW 0x04 /**< overflow page */
811 #define P_META 0x08 /**< meta page */
812 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
813 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
814 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
815 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
816 #define P_KEEP 0x8000 /**< leave this page alone during spill */
818 uint16_t mp_flags; /**< @ref mdb_page */
819 #define mp_lower mp_pb.pb.pb_lower
820 #define mp_upper mp_pb.pb.pb_upper
821 #define mp_pages mp_pb.pb_pages
824 indx_t pb_lower; /**< lower bound of free space */
825 indx_t pb_upper; /**< upper bound of free space */
827 uint32_t pb_pages; /**< number of overflow pages */
829 indx_t mp_ptrs[1]; /**< dynamic size */
832 /** Size of the page header, excluding dynamic data at the end */
833 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
835 /** Address of first usable data byte in a page, after the header */
836 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
838 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
839 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
841 /** Number of nodes on a page */
842 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
844 /** The amount of space remaining in the page */
845 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
847 /** The percentage of space used in the page, in tenths of a percent. */
848 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
849 ((env)->me_psize - PAGEHDRSZ))
850 /** The minimum page fill factor, in tenths of a percent.
851 * Pages emptier than this are candidates for merging.
853 #define FILL_THRESHOLD 250
855 /** Test if a page is a leaf page */
856 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
857 /** Test if a page is a LEAF2 page */
858 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
859 /** Test if a page is a branch page */
860 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
861 /** Test if a page is an overflow page */
862 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
863 /** Test if a page is a sub page */
864 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
866 /** The number of overflow pages needed to store the given size. */
867 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
869 /** Link in #MDB_txn.%mt_loose_pgs list.
870 * Kept outside the page header, which is needed when reusing the page.
872 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
874 /** Header for a single key/data pair within a page.
875 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
876 * We guarantee 2-byte alignment for 'MDB_node's.
878 * #mn_lo and #mn_hi are used for data size on leaf nodes, and for child
879 * pgno on branch nodes. On 64 bit platforms, #mn_flags is also used
880 * for pgno. (Branch nodes have no flags). Lo and hi are in host byte
881 * order in case some accesses can be optimized to 32-bit word access.
883 * Leaf node flags describe node contents. #F_BIGDATA says the node's
884 * data part is the page number of an overflow page with actual data.
885 * #F_DUPDATA and #F_SUBDATA can be combined giving duplicate data in
886 * a sub-page/sub-database, and named databases (just #F_SUBDATA).
888 typedef struct MDB_node {
889 /** part of data size or pgno
891 #if BYTE_ORDER == LITTLE_ENDIAN
892 unsigned short mn_lo, mn_hi;
894 unsigned short mn_hi, mn_lo;
897 /** @defgroup mdb_node Node Flags
899 * Flags for node headers.
902 #define F_BIGDATA 0x01 /**< data put on overflow page */
903 #define F_SUBDATA 0x02 /**< data is a sub-database */
904 #define F_DUPDATA 0x04 /**< data has duplicates */
906 /** valid flags for #mdb_node_add() */
907 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
910 unsigned short mn_flags; /**< @ref mdb_node */
911 unsigned short mn_ksize; /**< key size */
912 char mn_data[1]; /**< key and data are appended here */
915 /** Size of the node header, excluding dynamic data at the end */
916 #define NODESIZE offsetof(MDB_node, mn_data)
918 /** Bit position of top word in page number, for shifting mn_flags */
919 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
921 /** Size of a node in a branch page with a given key.
922 * This is just the node header plus the key, there is no data.
924 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
926 /** Size of a node in a leaf page with a given key and data.
927 * This is node header plus key plus data size.
929 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
931 /** Address of node \b i in page \b p */
932 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
934 /** Address of the key for the node */
935 #define NODEKEY(node) (void *)((node)->mn_data)
937 /** Address of the data for a node */
938 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
940 /** Get the page number pointed to by a branch node */
941 #define NODEPGNO(node) \
942 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
943 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
944 /** Set the page number in a branch node */
945 #define SETPGNO(node,pgno) do { \
946 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
947 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
949 /** Get the size of the data in a leaf node */
950 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
951 /** Set the size of the data for a leaf node */
952 #define SETDSZ(node,size) do { \
953 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
954 /** The size of a key in a node */
955 #define NODEKSZ(node) ((node)->mn_ksize)
957 /** Copy a page number from src to dst */
959 #define COPY_PGNO(dst,src) dst = src
961 #if SIZE_MAX > 4294967295UL
962 #define COPY_PGNO(dst,src) do { \
963 unsigned short *s, *d; \
964 s = (unsigned short *)&(src); \
965 d = (unsigned short *)&(dst); \
972 #define COPY_PGNO(dst,src) do { \
973 unsigned short *s, *d; \
974 s = (unsigned short *)&(src); \
975 d = (unsigned short *)&(dst); \
981 /** The address of a key in a LEAF2 page.
982 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
983 * There are no node headers, keys are stored contiguously.
985 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
987 /** Set the \b node's key into \b keyptr, if requested. */
988 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
989 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
991 /** Set the \b node's key into \b key. */
992 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
994 /** Information about a single database in the environment. */
995 typedef struct MDB_db {
996 uint32_t md_pad; /**< also ksize for LEAF2 pages */
997 uint16_t md_flags; /**< @ref mdb_dbi_open */
998 uint16_t md_depth; /**< depth of this tree */
999 pgno_t md_branch_pages; /**< number of internal pages */
1000 pgno_t md_leaf_pages; /**< number of leaf pages */
1001 pgno_t md_overflow_pages; /**< number of overflow pages */
1002 size_t md_entries; /**< number of data items */
1003 pgno_t md_root; /**< the root page of this tree */
1006 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1007 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1008 /** #mdb_dbi_open() flags */
1009 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1010 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1012 /** Handle for the DB used to track free pages. */
1014 /** Handle for the default DB. */
1016 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1019 /** Number of meta pages - also hardcoded elsewhere */
1022 /** Meta page content.
1023 * A meta page is the start point for accessing a database snapshot.
1024 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1026 typedef struct MDB_meta {
1027 /** Stamp identifying this as an LMDB file. It must be set
1030 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1031 uint32_t mm_version;
1032 void *mm_address; /**< address for fixed mapping */
1033 size_t mm_mapsize; /**< size of mmap region */
1034 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1035 /** The size of pages used in this DB */
1036 #define mm_psize mm_dbs[FREE_DBI].md_pad
1037 /** Any persistent environment flags. @ref mdb_env */
1038 #define mm_flags mm_dbs[FREE_DBI].md_flags
1039 /** Last used page in the datafile.
1040 * Actually the file may be shorter if the freeDB lists the final pages.
1043 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1046 /** Buffer for a stack-allocated meta page.
1047 * The members define size and alignment, and silence type
1048 * aliasing warnings. They are not used directly; that could
1049 * mean incorrectly using several union members in parallel.
1051 typedef union MDB_metabuf {
1054 char mm_pad[PAGEHDRSZ];
1059 /** Auxiliary DB info.
1060 * The information here is mostly static/read-only. There is
1061 * only a single copy of this record in the environment.
1063 typedef struct MDB_dbx {
1064 MDB_val md_name; /**< name of the database */
1065 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1066 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1067 MDB_rel_func *md_rel; /**< user relocate function */
1068 void *md_relctx; /**< user-provided context for md_rel */
1071 /** A database transaction.
1072 * Every operation requires a transaction handle.
1075 MDB_txn *mt_parent; /**< parent of a nested txn */
1076 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1078 pgno_t mt_next_pgno; /**< next unallocated page */
1079 /** The ID of this transaction. IDs are integers incrementing from 1.
1080 * Only committed write transactions increment the ID. If a transaction
1081 * aborts, the ID may be re-used by the next writer.
1084 MDB_env *mt_env; /**< the DB environment */
1085 /** The list of pages that became unused during this transaction.
1087 MDB_IDL mt_free_pgs;
1088 /** The list of loose pages that became unused and may be reused
1089 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1091 MDB_page *mt_loose_pgs;
1092 /** Number of loose pages (#mt_loose_pgs) */
1094 /** The sorted list of dirty pages we temporarily wrote to disk
1095 * because the dirty list was full. page numbers in here are
1096 * shifted left by 1, deleted slots have the LSB set.
1098 MDB_IDL mt_spill_pgs;
1100 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1101 MDB_ID2L dirty_list;
1102 /** For read txns: This thread/txn's reader table slot, or NULL. */
1105 /** Array of records for each DB known in the environment. */
1107 /** Array of MDB_db records for each known DB */
1109 /** Array of sequence numbers for each DB handle */
1110 unsigned int *mt_dbiseqs;
1111 /** @defgroup mt_dbflag Transaction DB Flags
1115 #define DB_DIRTY 0x01 /**< DB was written in this txn */
1116 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1117 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1118 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1119 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1120 #define DB_DUPDATA 0x20 /**< DB is #MDB_DUPSORT data */
1122 /** In write txns, array of cursors for each DB */
1123 MDB_cursor **mt_cursors;
1124 /** Array of flags for each DB */
1125 unsigned char *mt_dbflags;
1126 /** Number of DB records in use, or 0 when the txn is finished.
1127 * This number only ever increments until the txn finishes; we
1128 * don't decrement it when individual DB handles are closed.
1132 /** @defgroup mdb_txn Transaction Flags
1136 /** #mdb_txn_begin() flags */
1137 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1138 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1139 /* internal txn flags */
1140 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1141 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1142 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1143 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1144 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1145 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1146 /** most operations on the txn are currently illegal */
1147 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1149 unsigned int mt_flags; /**< @ref mdb_txn */
1150 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1151 * Includes ancestor txns' dirty pages not hidden by other txns'
1152 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1153 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1155 unsigned int mt_dirty_room;
1158 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1159 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1160 * raise this on a 64 bit machine.
1162 #define CURSOR_STACK 32
1166 /** Cursors are used for all DB operations.
1167 * A cursor holds a path of (page pointer, key index) from the DB
1168 * root to a position in the DB, plus other state. #MDB_DUPSORT
1169 * cursors include an xcursor to the current data item. Write txns
1170 * track their cursors and keep them up to date when data moves.
1171 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1172 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1175 /** Next cursor on this DB in this txn */
1176 MDB_cursor *mc_next;
1177 /** Backup of the original cursor if this cursor is a shadow */
1178 MDB_cursor *mc_backup;
1179 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1180 struct MDB_xcursor *mc_xcursor;
1181 /** The transaction that owns this cursor */
1183 /** The database handle this cursor operates on */
1185 /** The database record for this cursor */
1187 /** The database auxiliary record for this cursor */
1189 /** The @ref mt_dbflag for this database */
1190 unsigned char *mc_dbflag;
1191 unsigned short mc_snum; /**< number of pushed pages */
1192 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1193 /** @defgroup mdb_cursor Cursor Flags
1195 * Cursor state flags.
1198 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1199 #define C_EOF 0x02 /**< No more data */
1200 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1201 #define C_DEL 0x08 /**< last op was a cursor_del */
1202 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1204 unsigned int mc_flags; /**< @ref mdb_cursor */
1205 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1206 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1209 /** Context for sorted-dup records.
1210 * We could have gone to a fully recursive design, with arbitrarily
1211 * deep nesting of sub-databases. But for now we only handle these
1212 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1214 typedef struct MDB_xcursor {
1215 /** A sub-cursor for traversing the Dup DB */
1216 MDB_cursor mx_cursor;
1217 /** The database record for this Dup DB */
1219 /** The auxiliary DB record for this Dup DB */
1221 /** The @ref mt_dbflag for this Dup DB */
1222 unsigned char mx_dbflag;
1225 /** Check if there is an inited xcursor, so #XCURSOR_REFRESH() is proper */
1226 #define XCURSOR_INITED(mc) \
1227 ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
1229 /** Update sub-page pointer, if any, in \b mc->mc_xcursor. Needed
1230 * when the node which contains the sub-page may have moved. Called
1231 * with \b mp = mc->mc_pg[mc->mc_top], \b ki = mc->mc_ki[mc->mc_top].
1233 #define XCURSOR_REFRESH(mc, mp, ki) do { \
1234 MDB_page *xr_pg = (mp); \
1235 MDB_node *xr_node = NODEPTR(xr_pg, ki); \
1236 if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \
1237 (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \
1240 /** State of FreeDB old pages, stored in the MDB_env */
1241 typedef struct MDB_pgstate {
1242 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1243 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1246 /** The database environment. */
1248 HANDLE me_fd; /**< The main data file */
1249 HANDLE me_lfd; /**< The lock file */
1250 HANDLE me_mfd; /**< For writing and syncing the meta pages */
1251 /** Failed to update the meta page. Probably an I/O error. */
1252 #define MDB_FATAL_ERROR 0x80000000U
1253 /** Some fields are initialized. */
1254 #define MDB_ENV_ACTIVE 0x20000000U
1255 /** me_txkey is set */
1256 #define MDB_ENV_TXKEY 0x10000000U
1257 /** fdatasync is unreliable */
1258 #define MDB_FSYNCONLY 0x08000000U
1259 uint32_t me_flags; /**< @ref mdb_env */
1260 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1261 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1262 unsigned int me_maxreaders; /**< size of the reader table */
1263 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1264 volatile int me_close_readers;
1265 MDB_dbi me_numdbs; /**< number of DBs opened */
1266 MDB_dbi me_maxdbs; /**< size of the DB table */
1267 MDB_PID_T me_pid; /**< process ID of this env */
1268 char *me_path; /**< path to the DB files */
1269 char *me_map; /**< the memory map of the data file */
1270 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1271 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1272 void *me_pbuf; /**< scratch area for DUPSORT put() */
1273 MDB_txn *me_txn; /**< current write transaction */
1274 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1275 size_t me_mapsize; /**< size of the data memory map */
1276 off_t me_size; /**< current file size */
1277 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1278 MDB_dbx *me_dbxs; /**< array of static DB info */
1279 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1280 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1281 pthread_key_t me_txkey; /**< thread-key for readers */
1282 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1283 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1284 # define me_pglast me_pgstate.mf_pglast
1285 # define me_pghead me_pgstate.mf_pghead
1286 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1287 /** IDL of pages that became unused in a write txn */
1288 MDB_IDL me_free_pgs;
1289 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1290 MDB_ID2L me_dirty_list;
1291 /** Max number of freelist items that can fit in a single overflow page */
1293 /** Max size of a node on a page */
1294 unsigned int me_nodemax;
1295 #if !(MDB_MAXKEYSIZE)
1296 unsigned int me_maxkey; /**< max size of a key */
1298 int me_live_reader; /**< have liveness lock in reader table */
1300 int me_pidquery; /**< Used in OpenProcess */
1302 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1303 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1304 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1306 mdb_mutex_t me_rmutex;
1307 mdb_mutex_t me_wmutex;
1309 void *me_userctx; /**< User-settable context */
1310 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1313 /** Nested transaction */
1314 typedef struct MDB_ntxn {
1315 MDB_txn mnt_txn; /**< the transaction */
1316 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1319 /** max number of pages to commit in one writev() call */
1320 #define MDB_COMMIT_PAGES 64
1321 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1322 #undef MDB_COMMIT_PAGES
1323 #define MDB_COMMIT_PAGES IOV_MAX
1326 /** max bytes to write in one call */
1327 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1329 /** Check \b txn and \b dbi arguments to a function */
1330 #define TXN_DBI_EXIST(txn, dbi, validity) \
1331 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1333 /** Check for misused \b dbi handles */
1334 #define TXN_DBI_CHANGED(txn, dbi) \
1335 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1337 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1338 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1339 static int mdb_page_touch(MDB_cursor *mc);
1341 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1342 "reset-tmp", "fail-begin", "fail-beginchild"}
1344 /* mdb_txn_end operation number, for logging */
1345 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1346 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1348 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1349 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1350 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1351 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1352 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1354 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1355 static int mdb_page_search_root(MDB_cursor *mc,
1356 MDB_val *key, int modify);
1357 #define MDB_PS_MODIFY 1
1358 #define MDB_PS_ROOTONLY 2
1359 #define MDB_PS_FIRST 4
1360 #define MDB_PS_LAST 8
1361 static int mdb_page_search(MDB_cursor *mc,
1362 MDB_val *key, int flags);
1363 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1365 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1366 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1367 pgno_t newpgno, unsigned int nflags);
1369 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1370 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1371 static int mdb_env_write_meta(MDB_txn *txn);
1372 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1373 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1375 static void mdb_env_close0(MDB_env *env, int excl);
1377 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1378 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1379 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1380 static void mdb_node_del(MDB_cursor *mc, int ksize);
1381 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1382 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1383 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1384 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1385 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1387 static int mdb_rebalance(MDB_cursor *mc);
1388 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1390 static void mdb_cursor_pop(MDB_cursor *mc);
1391 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1393 static int mdb_cursor_del0(MDB_cursor *mc);
1394 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1395 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1396 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1397 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1398 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1400 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1401 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1403 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1404 static void mdb_xcursor_init0(MDB_cursor *mc);
1405 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1406 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1408 static int mdb_drop0(MDB_cursor *mc, int subs);
1409 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1410 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1413 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1416 /** Compare two items pointing at size_t's of unknown alignment. */
1417 #ifdef MISALIGNED_OK
1418 # define mdb_cmp_clong mdb_cmp_long
1420 # define mdb_cmp_clong mdb_cmp_cint
1424 static SECURITY_DESCRIPTOR mdb_null_sd;
1425 static SECURITY_ATTRIBUTES mdb_all_sa;
1426 static int mdb_sec_inited;
1429 static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra);
1432 /** Return the library version info. */
1434 mdb_version(int *major, int *minor, int *patch)
1436 if (major) *major = MDB_VERSION_MAJOR;
1437 if (minor) *minor = MDB_VERSION_MINOR;
1438 if (patch) *patch = MDB_VERSION_PATCH;
1439 return MDB_VERSION_STRING;
1442 /** Table of descriptions for LMDB @ref errors */
1443 static char *const mdb_errstr[] = {
1444 "MDB_KEYEXIST: Key/data pair already exists",
1445 "MDB_NOTFOUND: No matching key/data pair found",
1446 "MDB_PAGE_NOTFOUND: Requested page not found",
1447 "MDB_CORRUPTED: Located page was wrong type",
1448 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1449 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1450 "MDB_INVALID: File is not an LMDB file",
1451 "MDB_MAP_FULL: Environment mapsize limit reached",
1452 "MDB_DBS_FULL: Environment maxdbs limit reached",
1453 "MDB_READERS_FULL: Environment maxreaders limit reached",
1454 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1455 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1456 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1457 "MDB_PAGE_FULL: Internal error - page has no more space",
1458 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1459 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1460 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1461 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1462 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1463 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1467 mdb_strerror(int err)
1470 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1471 * This works as long as no function between the call to mdb_strerror
1472 * and the actual use of the message uses more than 4K of stack.
1474 #define MSGSIZE 1024
1475 #define PADSIZE 4096
1476 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1480 return ("Successful return: 0");
1482 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1483 i = err - MDB_KEYEXIST;
1484 return mdb_errstr[i];
1488 /* These are the C-runtime error codes we use. The comment indicates
1489 * their numeric value, and the Win32 error they would correspond to
1490 * if the error actually came from a Win32 API. A major mess, we should
1491 * have used LMDB-specific error codes for everything.
1494 case ENOENT: /* 2, FILE_NOT_FOUND */
1495 case EIO: /* 5, ACCESS_DENIED */
1496 case ENOMEM: /* 12, INVALID_ACCESS */
1497 case EACCES: /* 13, INVALID_DATA */
1498 case EBUSY: /* 16, CURRENT_DIRECTORY */
1499 case EINVAL: /* 22, BAD_COMMAND */
1500 case ENOSPC: /* 28, OUT_OF_PAPER */
1501 return strerror(err);
1506 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1507 FORMAT_MESSAGE_IGNORE_INSERTS,
1508 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1511 return strerror(err);
1515 /** assert(3) variant in cursor context */
1516 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1517 /** assert(3) variant in transaction context */
1518 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1519 /** assert(3) variant in environment context */
1520 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1523 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1524 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1527 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1528 const char *func, const char *file, int line)
1531 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1532 file, line, expr_txt, func);
1533 if (env->me_assert_func)
1534 env->me_assert_func(env, buf);
1535 fprintf(stderr, "%s\n", buf);
1539 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1543 /** Return the page number of \b mp which may be sub-page, for debug output */
1545 mdb_dbg_pgno(MDB_page *mp)
1548 COPY_PGNO(ret, mp->mp_pgno);
1552 /** Display a key in hexadecimal and return the address of the result.
1553 * @param[in] key the key to display
1554 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1555 * @return The key in hexadecimal form.
1558 mdb_dkey(MDB_val *key, char *buf)
1561 unsigned char *c = key->mv_data;
1567 if (key->mv_size > DKBUF_MAXKEYSIZE)
1568 return "MDB_MAXKEYSIZE";
1569 /* may want to make this a dynamic check: if the key is mostly
1570 * printable characters, print it as-is instead of converting to hex.
1574 for (i=0; i<key->mv_size; i++)
1575 ptr += sprintf(ptr, "%02x", *c++);
1577 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1583 mdb_leafnode_type(MDB_node *n)
1585 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1586 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1587 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1590 /** Display all the keys in the page. */
1592 mdb_page_list(MDB_page *mp)
1594 pgno_t pgno = mdb_dbg_pgno(mp);
1595 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1597 unsigned int i, nkeys, nsize, total = 0;
1601 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1602 case P_BRANCH: type = "Branch page"; break;
1603 case P_LEAF: type = "Leaf page"; break;
1604 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1605 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1606 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1608 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1609 pgno, mp->mp_pages, state);
1612 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1613 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1616 fprintf(stderr, "Bad page %"Z"u flags 0x%X\n", pgno, mp->mp_flags);
1620 nkeys = NUMKEYS(mp);
1621 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1623 for (i=0; i<nkeys; i++) {
1624 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1625 key.mv_size = nsize = mp->mp_pad;
1626 key.mv_data = LEAF2KEY(mp, i, nsize);
1628 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1631 node = NODEPTR(mp, i);
1632 key.mv_size = node->mn_ksize;
1633 key.mv_data = node->mn_data;
1634 nsize = NODESIZE + key.mv_size;
1635 if (IS_BRANCH(mp)) {
1636 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1640 if (F_ISSET(node->mn_flags, F_BIGDATA))
1641 nsize += sizeof(pgno_t);
1643 nsize += NODEDSZ(node);
1645 nsize += sizeof(indx_t);
1646 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1647 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1649 total = EVEN(total);
1651 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1652 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1656 mdb_cursor_chk(MDB_cursor *mc)
1662 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1663 for (i=0; i<mc->mc_top; i++) {
1665 node = NODEPTR(mp, mc->mc_ki[i]);
1666 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1669 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1671 if (XCURSOR_INITED(mc)) {
1672 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1673 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1674 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1682 /** Count all the pages in each DB and in the freelist
1683 * and make sure it matches the actual number of pages
1685 * All named DBs must be open for a correct count.
1687 static void mdb_audit(MDB_txn *txn)
1691 MDB_ID freecount, count;
1696 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1697 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1698 freecount += *(MDB_ID *)data.mv_data;
1699 mdb_tassert(txn, rc == MDB_NOTFOUND);
1702 for (i = 0; i<txn->mt_numdbs; i++) {
1704 if (!(txn->mt_dbflags[i] & DB_VALID))
1706 mdb_cursor_init(&mc, txn, i, &mx);
1707 if (txn->mt_dbs[i].md_root == P_INVALID)
1709 count += txn->mt_dbs[i].md_branch_pages +
1710 txn->mt_dbs[i].md_leaf_pages +
1711 txn->mt_dbs[i].md_overflow_pages;
1712 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1713 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1714 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1717 mp = mc.mc_pg[mc.mc_top];
1718 for (j=0; j<NUMKEYS(mp); j++) {
1719 MDB_node *leaf = NODEPTR(mp, j);
1720 if (leaf->mn_flags & F_SUBDATA) {
1722 memcpy(&db, NODEDATA(leaf), sizeof(db));
1723 count += db.md_branch_pages + db.md_leaf_pages +
1724 db.md_overflow_pages;
1728 mdb_tassert(txn, rc == MDB_NOTFOUND);
1731 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1732 fprintf(stderr, "audit: %"Z"u freecount: %"Z"u count: %"Z"u total: %"Z"u next_pgno: %"Z"u\n",
1733 txn->mt_txnid, freecount, count+NUM_METAS,
1734 freecount+count+NUM_METAS, txn->mt_next_pgno);
1740 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1742 return txn->mt_dbxs[dbi].md_cmp(a, b);
1746 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1748 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1749 #if UINT_MAX < SIZE_MAX
1750 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1751 dcmp = mdb_cmp_clong;
1756 /** Allocate memory for a page.
1757 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1758 * Set #MDB_TXN_ERROR on failure.
1761 mdb_page_malloc(MDB_txn *txn, unsigned num)
1763 MDB_env *env = txn->mt_env;
1764 MDB_page *ret = env->me_dpages;
1765 size_t psize = env->me_psize, sz = psize, off;
1766 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1767 * For a single page alloc, we init everything after the page header.
1768 * For multi-page, we init the final page; if the caller needed that
1769 * many pages they will be filling in at least up to the last page.
1773 VGMEMP_ALLOC(env, ret, sz);
1774 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1775 env->me_dpages = ret->mp_next;
1778 psize -= off = PAGEHDRSZ;
1783 if ((ret = malloc(sz)) != NULL) {
1784 VGMEMP_ALLOC(env, ret, sz);
1785 if (!(env->me_flags & MDB_NOMEMINIT)) {
1786 memset((char *)ret + off, 0, psize);
1790 txn->mt_flags |= MDB_TXN_ERROR;
1794 /** Free a single page.
1795 * Saves single pages to a list, for future reuse.
1796 * (This is not used for multi-page overflow pages.)
1799 mdb_page_free(MDB_env *env, MDB_page *mp)
1801 mp->mp_next = env->me_dpages;
1802 VGMEMP_FREE(env, mp);
1803 env->me_dpages = mp;
1806 /** Free a dirty page */
1808 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1810 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1811 mdb_page_free(env, dp);
1813 /* large pages just get freed directly */
1814 VGMEMP_FREE(env, dp);
1819 /** Return all dirty pages to dpage list */
1821 mdb_dlist_free(MDB_txn *txn)
1823 MDB_env *env = txn->mt_env;
1824 MDB_ID2L dl = txn->mt_u.dirty_list;
1825 unsigned i, n = dl[0].mid;
1827 for (i = 1; i <= n; i++) {
1828 mdb_dpage_free(env, dl[i].mptr);
1833 /** Loosen or free a single page.
1834 * Saves single pages to a list for future reuse
1835 * in this same txn. It has been pulled from the freeDB
1836 * and already resides on the dirty list, but has been
1837 * deleted. Use these pages first before pulling again
1840 * If the page wasn't dirtied in this txn, just add it
1841 * to this txn's free list.
1844 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1847 pgno_t pgno = mp->mp_pgno;
1848 MDB_txn *txn = mc->mc_txn;
1850 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1851 if (txn->mt_parent) {
1852 MDB_ID2 *dl = txn->mt_u.dirty_list;
1853 /* If txn has a parent, make sure the page is in our
1857 unsigned x = mdb_mid2l_search(dl, pgno);
1858 if (x <= dl[0].mid && dl[x].mid == pgno) {
1859 if (mp != dl[x].mptr) { /* bad cursor? */
1860 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1861 txn->mt_flags |= MDB_TXN_ERROR;
1862 return MDB_CORRUPTED;
1869 /* no parent txn, so it's just ours */
1874 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1876 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1877 txn->mt_loose_pgs = mp;
1878 txn->mt_loose_count++;
1879 mp->mp_flags |= P_LOOSE;
1881 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1889 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1890 * @param[in] mc A cursor handle for the current operation.
1891 * @param[in] pflags Flags of the pages to update:
1892 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1893 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1894 * @return 0 on success, non-zero on failure.
1897 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1899 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1900 MDB_txn *txn = mc->mc_txn;
1901 MDB_cursor *m3, *m0 = mc;
1906 int rc = MDB_SUCCESS, level;
1908 /* Mark pages seen by cursors */
1909 if (mc->mc_flags & C_UNTRACK)
1910 mc = NULL; /* will find mc in mt_cursors */
1911 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1912 for (; mc; mc=mc->mc_next) {
1913 if (!(mc->mc_flags & C_INITIALIZED))
1915 for (m3 = mc;; m3 = &mx->mx_cursor) {
1917 for (j=0; j<m3->mc_snum; j++) {
1919 if ((mp->mp_flags & Mask) == pflags)
1920 mp->mp_flags ^= P_KEEP;
1922 mx = m3->mc_xcursor;
1923 /* Proceed to mx if it is at a sub-database */
1924 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1926 if (! (mp && (mp->mp_flags & P_LEAF)))
1928 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1929 if (!(leaf->mn_flags & F_SUBDATA))
1938 /* Mark dirty root pages */
1939 for (i=0; i<txn->mt_numdbs; i++) {
1940 if (txn->mt_dbflags[i] & DB_DIRTY) {
1941 pgno_t pgno = txn->mt_dbs[i].md_root;
1942 if (pgno == P_INVALID)
1944 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
1946 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1947 dp->mp_flags ^= P_KEEP;
1955 static int mdb_page_flush(MDB_txn *txn, int keep);
1957 /** Spill pages from the dirty list back to disk.
1958 * This is intended to prevent running into #MDB_TXN_FULL situations,
1959 * but note that they may still occur in a few cases:
1960 * 1) our estimate of the txn size could be too small. Currently this
1961 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1962 * 2) child txns may run out of space if their parents dirtied a
1963 * lot of pages and never spilled them. TODO: we probably should do
1964 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1965 * the parent's dirty_room is below a given threshold.
1967 * Otherwise, if not using nested txns, it is expected that apps will
1968 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1969 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1970 * If the txn never references them again, they can be left alone.
1971 * If the txn only reads them, they can be used without any fuss.
1972 * If the txn writes them again, they can be dirtied immediately without
1973 * going thru all of the work of #mdb_page_touch(). Such references are
1974 * handled by #mdb_page_unspill().
1976 * Also note, we never spill DB root pages, nor pages of active cursors,
1977 * because we'll need these back again soon anyway. And in nested txns,
1978 * we can't spill a page in a child txn if it was already spilled in a
1979 * parent txn. That would alter the parent txns' data even though
1980 * the child hasn't committed yet, and we'd have no way to undo it if
1981 * the child aborted.
1983 * @param[in] m0 cursor A cursor handle identifying the transaction and
1984 * database for which we are checking space.
1985 * @param[in] key For a put operation, the key being stored.
1986 * @param[in] data For a put operation, the data being stored.
1987 * @return 0 on success, non-zero on failure.
1990 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1992 MDB_txn *txn = m0->mc_txn;
1994 MDB_ID2L dl = txn->mt_u.dirty_list;
1995 unsigned int i, j, need;
1998 if (m0->mc_flags & C_SUB)
2001 /* Estimate how much space this op will take */
2002 i = m0->mc_db->md_depth;
2003 /* Named DBs also dirty the main DB */
2004 if (m0->mc_dbi >= CORE_DBS)
2005 i += txn->mt_dbs[MAIN_DBI].md_depth;
2006 /* For puts, roughly factor in the key+data size */
2008 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2009 i += i; /* double it for good measure */
2012 if (txn->mt_dirty_room > i)
2015 if (!txn->mt_spill_pgs) {
2016 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2017 if (!txn->mt_spill_pgs)
2020 /* purge deleted slots */
2021 MDB_IDL sl = txn->mt_spill_pgs;
2022 unsigned int num = sl[0];
2024 for (i=1; i<=num; i++) {
2031 /* Preserve pages which may soon be dirtied again */
2032 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2035 /* Less aggressive spill - we originally spilled the entire dirty list,
2036 * with a few exceptions for cursor pages and DB root pages. But this
2037 * turns out to be a lot of wasted effort because in a large txn many
2038 * of those pages will need to be used again. So now we spill only 1/8th
2039 * of the dirty pages. Testing revealed this to be a good tradeoff,
2040 * better than 1/2, 1/4, or 1/10.
2042 if (need < MDB_IDL_UM_MAX / 8)
2043 need = MDB_IDL_UM_MAX / 8;
2045 /* Save the page IDs of all the pages we're flushing */
2046 /* flush from the tail forward, this saves a lot of shifting later on. */
2047 for (i=dl[0].mid; i && need; i--) {
2048 MDB_ID pn = dl[i].mid << 1;
2050 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2052 /* Can't spill twice, make sure it's not already in a parent's
2055 if (txn->mt_parent) {
2057 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2058 if (tx2->mt_spill_pgs) {
2059 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2060 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2061 dp->mp_flags |= P_KEEP;
2069 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2073 mdb_midl_sort(txn->mt_spill_pgs);
2075 /* Flush the spilled part of dirty list */
2076 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2079 /* Reset any dirty pages we kept that page_flush didn't see */
2080 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2083 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2087 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2089 mdb_find_oldest(MDB_txn *txn)
2092 txnid_t mr, oldest = txn->mt_txnid - 1;
2093 if (txn->mt_env->me_txns) {
2094 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2095 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2106 /** Add a page to the txn's dirty list */
2108 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2111 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2113 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2114 insert = mdb_mid2l_append;
2116 insert = mdb_mid2l_insert;
2118 mid.mid = mp->mp_pgno;
2120 rc = insert(txn->mt_u.dirty_list, &mid);
2121 mdb_tassert(txn, rc == 0);
2122 txn->mt_dirty_room--;
2125 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2126 * me_pghead and mt_next_pgno. Set #MDB_TXN_ERROR on failure.
2128 * If there are free pages available from older transactions, they
2129 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2130 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2131 * and move me_pglast to say which records were consumed. Only this
2132 * function can create me_pghead and move me_pglast/mt_next_pgno.
2133 * @param[in] mc cursor A cursor handle identifying the transaction and
2134 * database for which we are allocating.
2135 * @param[in] num the number of pages to allocate.
2136 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2137 * will always be satisfied by a single contiguous chunk of memory.
2138 * @return 0 on success, non-zero on failure.
2141 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2143 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2144 /* Get at most <Max_retries> more freeDB records once me_pghead
2145 * has enough pages. If not enough, use new pages from the map.
2146 * If <Paranoid> and mc is updating the freeDB, only get new
2147 * records if me_pghead is empty. Then the freelist cannot play
2148 * catch-up with itself by growing while trying to save it.
2150 enum { Paranoid = 1, Max_retries = 500 };
2152 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2154 int rc, retry = num * 60;
2155 MDB_txn *txn = mc->mc_txn;
2156 MDB_env *env = txn->mt_env;
2157 pgno_t pgno, *mop = env->me_pghead;
2158 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2160 txnid_t oldest = 0, last;
2165 /* If there are any loose pages, just use them */
2166 if (num == 1 && txn->mt_loose_pgs) {
2167 np = txn->mt_loose_pgs;
2168 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2169 txn->mt_loose_count--;
2170 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2178 /* If our dirty list is already full, we can't do anything */
2179 if (txn->mt_dirty_room == 0) {
2184 for (op = MDB_FIRST;; op = MDB_NEXT) {
2189 /* Seek a big enough contiguous page range. Prefer
2190 * pages at the tail, just truncating the list.
2196 if (mop[i-n2] == pgno+n2)
2203 if (op == MDB_FIRST) { /* 1st iteration */
2204 /* Prepare to fetch more and coalesce */
2205 last = env->me_pglast;
2206 oldest = env->me_pgoldest;
2207 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2210 key.mv_data = &last; /* will look up last+1 */
2211 key.mv_size = sizeof(last);
2213 if (Paranoid && mc->mc_dbi == FREE_DBI)
2216 if (Paranoid && retry < 0 && mop_len)
2220 /* Do not fetch more if the record will be too recent */
2221 if (oldest <= last) {
2223 oldest = mdb_find_oldest(txn);
2224 env->me_pgoldest = oldest;
2230 rc = mdb_cursor_get(&m2, &key, NULL, op);
2232 if (rc == MDB_NOTFOUND)
2236 last = *(txnid_t*)key.mv_data;
2237 if (oldest <= last) {
2239 oldest = mdb_find_oldest(txn);
2240 env->me_pgoldest = oldest;
2246 np = m2.mc_pg[m2.mc_top];
2247 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2248 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2251 idl = (MDB_ID *) data.mv_data;
2254 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2259 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2261 mop = env->me_pghead;
2263 env->me_pglast = last;
2265 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2266 last, txn->mt_dbs[FREE_DBI].md_root, i));
2268 DPRINTF(("IDL %"Z"u", idl[j]));
2270 /* Merge in descending sorted order */
2271 mdb_midl_xmerge(mop, idl);
2275 /* Use new pages from the map when nothing suitable in the freeDB */
2277 pgno = txn->mt_next_pgno;
2278 if (pgno + num >= env->me_maxpg) {
2279 DPUTS("DB size maxed out");
2285 if (env->me_flags & MDB_WRITEMAP) {
2286 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2288 if (!(np = mdb_page_malloc(txn, num))) {
2294 mop[0] = mop_len -= num;
2295 /* Move any stragglers down */
2296 for (j = i-num; j < mop_len; )
2297 mop[++j] = mop[++i];
2299 txn->mt_next_pgno = pgno + num;
2302 mdb_page_dirty(txn, np);
2308 txn->mt_flags |= MDB_TXN_ERROR;
2312 /** Copy the used portions of a non-overflow page.
2313 * @param[in] dst page to copy into
2314 * @param[in] src page to copy from
2315 * @param[in] psize size of a page
2318 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2320 enum { Align = sizeof(pgno_t) };
2321 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2323 /* If page isn't full, just copy the used portion. Adjust
2324 * alignment so memcpy may copy words instead of bytes.
2326 if ((unused &= -Align) && !IS_LEAF2(src)) {
2327 upper = (upper + PAGEBASE) & -Align;
2328 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2329 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2332 memcpy(dst, src, psize - unused);
2336 /** Pull a page off the txn's spill list, if present.
2337 * If a page being referenced was spilled to disk in this txn, bring
2338 * it back and make it dirty/writable again.
2339 * @param[in] txn the transaction handle.
2340 * @param[in] mp the page being referenced. It must not be dirty.
2341 * @param[out] ret the writable page, if any. ret is unchanged if
2342 * mp wasn't spilled.
2345 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2347 MDB_env *env = txn->mt_env;
2350 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2352 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2353 if (!tx2->mt_spill_pgs)
2355 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2356 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2359 if (txn->mt_dirty_room == 0)
2360 return MDB_TXN_FULL;
2361 if (IS_OVERFLOW(mp))
2365 if (env->me_flags & MDB_WRITEMAP) {
2368 np = mdb_page_malloc(txn, num);
2372 memcpy(np, mp, num * env->me_psize);
2374 mdb_page_copy(np, mp, env->me_psize);
2377 /* If in current txn, this page is no longer spilled.
2378 * If it happens to be the last page, truncate the spill list.
2379 * Otherwise mark it as deleted by setting the LSB.
2381 if (x == txn->mt_spill_pgs[0])
2382 txn->mt_spill_pgs[0]--;
2384 txn->mt_spill_pgs[x] |= 1;
2385 } /* otherwise, if belonging to a parent txn, the
2386 * page remains spilled until child commits
2389 mdb_page_dirty(txn, np);
2390 np->mp_flags |= P_DIRTY;
2398 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2399 * Set #MDB_TXN_ERROR on failure.
2400 * @param[in] mc cursor pointing to the page to be touched
2401 * @return 0 on success, non-zero on failure.
2404 mdb_page_touch(MDB_cursor *mc)
2406 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2407 MDB_txn *txn = mc->mc_txn;
2408 MDB_cursor *m2, *m3;
2412 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2413 if (txn->mt_flags & MDB_TXN_SPILLS) {
2415 rc = mdb_page_unspill(txn, mp, &np);
2421 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2422 (rc = mdb_page_alloc(mc, 1, &np)))
2425 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2426 mp->mp_pgno, pgno));
2427 mdb_cassert(mc, mp->mp_pgno != pgno);
2428 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2429 /* Update the parent page, if any, to point to the new page */
2431 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2432 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2433 SETPGNO(node, pgno);
2435 mc->mc_db->md_root = pgno;
2437 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2438 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2440 /* If txn has a parent, make sure the page is in our
2444 unsigned x = mdb_mid2l_search(dl, pgno);
2445 if (x <= dl[0].mid && dl[x].mid == pgno) {
2446 if (mp != dl[x].mptr) { /* bad cursor? */
2447 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2448 txn->mt_flags |= MDB_TXN_ERROR;
2449 return MDB_CORRUPTED;
2454 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2456 np = mdb_page_malloc(txn, 1);
2461 rc = mdb_mid2l_insert(dl, &mid);
2462 mdb_cassert(mc, rc == 0);
2467 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2469 np->mp_flags |= P_DIRTY;
2472 /* Adjust cursors pointing to mp */
2473 mc->mc_pg[mc->mc_top] = np;
2474 m2 = txn->mt_cursors[mc->mc_dbi];
2475 if (mc->mc_flags & C_SUB) {
2476 for (; m2; m2=m2->mc_next) {
2477 m3 = &m2->mc_xcursor->mx_cursor;
2478 if (m3->mc_snum < mc->mc_snum) continue;
2479 if (m3->mc_pg[mc->mc_top] == mp)
2480 m3->mc_pg[mc->mc_top] = np;
2483 for (; m2; m2=m2->mc_next) {
2484 if (m2->mc_snum < mc->mc_snum) continue;
2485 if (m2 == mc) continue;
2486 if (m2->mc_pg[mc->mc_top] == mp) {
2487 m2->mc_pg[mc->mc_top] = np;
2488 if (XCURSOR_INITED(m2) && IS_LEAF(np))
2489 XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]);
2496 txn->mt_flags |= MDB_TXN_ERROR;
2501 mdb_env_sync(MDB_env *env, int force)
2504 if (env->me_flags & MDB_RDONLY)
2506 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2507 if (env->me_flags & MDB_WRITEMAP) {
2508 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2509 ? MS_ASYNC : MS_SYNC;
2510 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2513 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2517 #ifdef BROKEN_FDATASYNC
2518 if (env->me_flags & MDB_FSYNCONLY) {
2519 if (fsync(env->me_fd))
2523 if (MDB_FDATASYNC(env->me_fd))
2530 /** Back up parent txn's cursors, then grab the originals for tracking */
2532 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2534 MDB_cursor *mc, *bk;
2539 for (i = src->mt_numdbs; --i >= 0; ) {
2540 if ((mc = src->mt_cursors[i]) != NULL) {
2541 size = sizeof(MDB_cursor);
2543 size += sizeof(MDB_xcursor);
2544 for (; mc; mc = bk->mc_next) {
2550 mc->mc_db = &dst->mt_dbs[i];
2551 /* Kill pointers into src to reduce abuse: The
2552 * user may not use mc until dst ends. But we need a valid
2553 * txn pointer here for cursor fixups to keep working.
2556 mc->mc_dbflag = &dst->mt_dbflags[i];
2557 if ((mx = mc->mc_xcursor) != NULL) {
2558 *(MDB_xcursor *)(bk+1) = *mx;
2559 mx->mx_cursor.mc_txn = dst;
2561 mc->mc_next = dst->mt_cursors[i];
2562 dst->mt_cursors[i] = mc;
2569 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2570 * @param[in] txn the transaction handle.
2571 * @param[in] merge true to keep changes to parent cursors, false to revert.
2572 * @return 0 on success, non-zero on failure.
2575 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2577 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2581 for (i = txn->mt_numdbs; --i >= 0; ) {
2582 for (mc = cursors[i]; mc; mc = next) {
2584 if ((bk = mc->mc_backup) != NULL) {
2586 /* Commit changes to parent txn */
2587 mc->mc_next = bk->mc_next;
2588 mc->mc_backup = bk->mc_backup;
2589 mc->mc_txn = bk->mc_txn;
2590 mc->mc_db = bk->mc_db;
2591 mc->mc_dbflag = bk->mc_dbflag;
2592 if ((mx = mc->mc_xcursor) != NULL)
2593 mx->mx_cursor.mc_txn = bk->mc_txn;
2595 /* Abort nested txn */
2597 if ((mx = mc->mc_xcursor) != NULL)
2598 *mx = *(MDB_xcursor *)(bk+1);
2602 /* Only malloced cursors are permanently tracked. */
2609 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2615 Pidset = F_SETLK, Pidcheck = F_GETLK
2619 /** Set or check a pid lock. Set returns 0 on success.
2620 * Check returns 0 if the process is certainly dead, nonzero if it may
2621 * be alive (the lock exists or an error happened so we do not know).
2623 * On Windows Pidset is a no-op, we merely check for the existence
2624 * of the process with the given pid. On POSIX we use a single byte
2625 * lock on the lockfile, set at an offset equal to the pid.
2628 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2630 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2633 if (op == Pidcheck) {
2634 h = OpenProcess(env->me_pidquery, FALSE, pid);
2635 /* No documented "no such process" code, but other program use this: */
2637 return ErrCode() != ERROR_INVALID_PARAMETER;
2638 /* A process exists until all handles to it close. Has it exited? */
2639 ret = WaitForSingleObject(h, 0) != 0;
2646 struct flock lock_info;
2647 memset(&lock_info, 0, sizeof(lock_info));
2648 lock_info.l_type = F_WRLCK;
2649 lock_info.l_whence = SEEK_SET;
2650 lock_info.l_start = pid;
2651 lock_info.l_len = 1;
2652 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2653 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2655 } else if ((rc = ErrCode()) == EINTR) {
2663 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2664 * @param[in] txn the transaction handle to initialize
2665 * @return 0 on success, non-zero on failure.
2668 mdb_txn_renew0(MDB_txn *txn)
2670 MDB_env *env = txn->mt_env;
2671 MDB_txninfo *ti = env->me_txns;
2673 unsigned int i, nr, flags = txn->mt_flags;
2675 int rc, new_notls = 0;
2677 if ((flags &= MDB_TXN_RDONLY) != 0) {
2679 meta = mdb_env_pick_meta(env);
2680 txn->mt_txnid = meta->mm_txnid;
2681 txn->mt_u.reader = NULL;
2683 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2684 pthread_getspecific(env->me_txkey);
2686 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2687 return MDB_BAD_RSLOT;
2689 MDB_PID_T pid = env->me_pid;
2690 MDB_THR_T tid = pthread_self();
2691 mdb_mutexref_t rmutex = env->me_rmutex;
2693 if (!env->me_live_reader) {
2694 rc = mdb_reader_pid(env, Pidset, pid);
2697 env->me_live_reader = 1;
2700 if (LOCK_MUTEX(rc, env, rmutex))
2702 nr = ti->mti_numreaders;
2703 for (i=0; i<nr; i++)
2704 if (ti->mti_readers[i].mr_pid == 0)
2706 if (i == env->me_maxreaders) {
2707 UNLOCK_MUTEX(rmutex);
2708 return MDB_READERS_FULL;
2710 r = &ti->mti_readers[i];
2711 /* Claim the reader slot, carefully since other code
2712 * uses the reader table un-mutexed: First reset the
2713 * slot, next publish it in mti_numreaders. After
2714 * that, it is safe for mdb_env_close() to touch it.
2715 * When it will be closed, we can finally claim it.
2718 r->mr_txnid = (txnid_t)-1;
2721 ti->mti_numreaders = ++nr;
2722 env->me_close_readers = nr;
2724 UNLOCK_MUTEX(rmutex);
2726 new_notls = (env->me_flags & MDB_NOTLS);
2727 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2732 do /* LY: Retry on a race, ITS#7970. */
2733 r->mr_txnid = ti->mti_txnid;
2734 while(r->mr_txnid != ti->mti_txnid);
2735 txn->mt_txnid = r->mr_txnid;
2736 txn->mt_u.reader = r;
2737 meta = env->me_metas[txn->mt_txnid & 1];
2741 /* Not yet touching txn == env->me_txn0, it may be active */
2743 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2745 txn->mt_txnid = ti->mti_txnid;
2746 meta = env->me_metas[txn->mt_txnid & 1];
2748 meta = mdb_env_pick_meta(env);
2749 txn->mt_txnid = meta->mm_txnid;
2753 if (txn->mt_txnid == mdb_debug_start)
2756 txn->mt_child = NULL;
2757 txn->mt_loose_pgs = NULL;
2758 txn->mt_loose_count = 0;
2759 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2760 txn->mt_u.dirty_list = env->me_dirty_list;
2761 txn->mt_u.dirty_list[0].mid = 0;
2762 txn->mt_free_pgs = env->me_free_pgs;
2763 txn->mt_free_pgs[0] = 0;
2764 txn->mt_spill_pgs = NULL;
2766 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2769 /* Copy the DB info and flags */
2770 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2772 /* Moved to here to avoid a data race in read TXNs */
2773 txn->mt_next_pgno = meta->mm_last_pg+1;
2775 txn->mt_flags = flags;
2778 txn->mt_numdbs = env->me_numdbs;
2779 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2780 x = env->me_dbflags[i];
2781 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2782 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2784 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2785 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2787 if (env->me_flags & MDB_FATAL_ERROR) {
2788 DPUTS("environment had fatal error, must shutdown!");
2790 } else if (env->me_maxpg < txn->mt_next_pgno) {
2791 rc = MDB_MAP_RESIZED;
2795 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
2800 mdb_txn_renew(MDB_txn *txn)
2804 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
2807 rc = mdb_txn_renew0(txn);
2808 if (rc == MDB_SUCCESS) {
2809 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2810 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2811 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2817 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2821 int rc, size, tsize;
2823 flags &= MDB_TXN_BEGIN_FLAGS;
2824 flags |= env->me_flags & MDB_WRITEMAP;
2826 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2830 /* Nested transactions: Max 1 child, write txns only, no writemap */
2831 flags |= parent->mt_flags;
2832 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
2833 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2835 /* Child txns save MDB_pgstate and use own copy of cursors */
2836 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
2837 size += tsize = sizeof(MDB_ntxn);
2838 } else if (flags & MDB_RDONLY) {
2839 size = env->me_maxdbs * (sizeof(MDB_db)+1);
2840 size += tsize = sizeof(MDB_txn);
2842 /* Reuse preallocated write txn. However, do not touch it until
2843 * mdb_txn_renew0() succeeds, since it currently may be active.
2848 if ((txn = calloc(1, size)) == NULL) {
2849 DPRINTF(("calloc: %s", strerror(errno)));
2852 txn->mt_dbxs = env->me_dbxs; /* static */
2853 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2854 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
2855 txn->mt_flags = flags;
2860 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2861 txn->mt_dbiseqs = parent->mt_dbiseqs;
2862 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2863 if (!txn->mt_u.dirty_list ||
2864 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2866 free(txn->mt_u.dirty_list);
2870 txn->mt_txnid = parent->mt_txnid;
2871 txn->mt_dirty_room = parent->mt_dirty_room;
2872 txn->mt_u.dirty_list[0].mid = 0;
2873 txn->mt_spill_pgs = NULL;
2874 txn->mt_next_pgno = parent->mt_next_pgno;
2875 parent->mt_flags |= MDB_TXN_HAS_CHILD;
2876 parent->mt_child = txn;
2877 txn->mt_parent = parent;
2878 txn->mt_numdbs = parent->mt_numdbs;
2879 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2880 /* Copy parent's mt_dbflags, but clear DB_NEW */
2881 for (i=0; i<txn->mt_numdbs; i++)
2882 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2884 ntxn = (MDB_ntxn *)txn;
2885 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2886 if (env->me_pghead) {
2887 size = MDB_IDL_SIZEOF(env->me_pghead);
2888 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2890 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2895 rc = mdb_cursor_shadow(parent, txn);
2897 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
2898 } else { /* MDB_RDONLY */
2899 txn->mt_dbiseqs = env->me_dbiseqs;
2901 rc = mdb_txn_renew0(txn);
2904 if (txn != env->me_txn0)
2907 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
2909 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2910 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2911 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2918 mdb_txn_env(MDB_txn *txn)
2920 if(!txn) return NULL;
2925 mdb_txn_id(MDB_txn *txn)
2928 return txn->mt_txnid;
2931 /** Export or close DBI handles opened in this txn. */
2933 mdb_dbis_update(MDB_txn *txn, int keep)
2936 MDB_dbi n = txn->mt_numdbs;
2937 MDB_env *env = txn->mt_env;
2938 unsigned char *tdbflags = txn->mt_dbflags;
2940 for (i = n; --i >= CORE_DBS;) {
2941 if (tdbflags[i] & DB_NEW) {
2943 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2945 char *ptr = env->me_dbxs[i].md_name.mv_data;
2947 env->me_dbxs[i].md_name.mv_data = NULL;
2948 env->me_dbxs[i].md_name.mv_size = 0;
2949 env->me_dbflags[i] = 0;
2950 env->me_dbiseqs[i]++;
2956 if (keep && env->me_numdbs < n)
2960 /** End a transaction, except successful commit of a nested transaction.
2961 * May be called twice for readonly txns: First reset it, then abort.
2962 * @param[in] txn the transaction handle to end
2963 * @param[in] mode why and how to end the transaction
2966 mdb_txn_end(MDB_txn *txn, unsigned mode)
2968 MDB_env *env = txn->mt_env;
2970 static const char *const names[] = MDB_END_NAMES;
2973 /* Export or close DBI handles opened in this txn */
2974 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
2976 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2977 names[mode & MDB_END_OPMASK],
2978 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2979 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2981 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2982 if (txn->mt_u.reader) {
2983 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2984 if (!(env->me_flags & MDB_NOTLS)) {
2985 txn->mt_u.reader = NULL; /* txn does not own reader */
2986 } else if (mode & MDB_END_SLOT) {
2987 txn->mt_u.reader->mr_pid = 0;
2988 txn->mt_u.reader = NULL;
2989 } /* else txn owns the slot until it does MDB_END_SLOT */
2991 txn->mt_numdbs = 0; /* prevent further DBI activity */
2992 txn->mt_flags |= MDB_TXN_FINISHED;
2994 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
2995 pgno_t *pghead = env->me_pghead;
2997 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
2998 mdb_cursors_close(txn, 0);
2999 if (!(env->me_flags & MDB_WRITEMAP)) {
3000 mdb_dlist_free(txn);
3004 txn->mt_flags = MDB_TXN_FINISHED;
3006 if (!txn->mt_parent) {
3007 mdb_midl_shrink(&txn->mt_free_pgs);
3008 env->me_free_pgs = txn->mt_free_pgs;
3010 env->me_pghead = NULL;
3014 mode = 0; /* txn == env->me_txn0, do not free() it */
3016 /* The writer mutex was locked in mdb_txn_begin. */
3018 UNLOCK_MUTEX(env->me_wmutex);
3020 txn->mt_parent->mt_child = NULL;
3021 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3022 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3023 mdb_midl_free(txn->mt_free_pgs);
3024 mdb_midl_free(txn->mt_spill_pgs);
3025 free(txn->mt_u.dirty_list);
3028 mdb_midl_free(pghead);
3031 if (mode & MDB_END_FREE)
3036 mdb_txn_reset(MDB_txn *txn)
3041 /* This call is only valid for read-only txns */
3042 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3045 mdb_txn_end(txn, MDB_END_RESET);
3049 mdb_txn_abort(MDB_txn *txn)
3055 mdb_txn_abort(txn->mt_child);
3057 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3060 /** Save the freelist as of this transaction to the freeDB.
3061 * This changes the freelist. Keep trying until it stabilizes.
3064 mdb_freelist_save(MDB_txn *txn)
3066 /* env->me_pghead[] can grow and shrink during this call.
3067 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3068 * Page numbers cannot disappear from txn->mt_free_pgs[].
3071 MDB_env *env = txn->mt_env;
3072 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3073 txnid_t pglast = 0, head_id = 0;
3074 pgno_t freecnt = 0, *free_pgs, *mop;
3075 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3077 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3079 if (env->me_pghead) {
3080 /* Make sure first page of freeDB is touched and on freelist */
3081 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3082 if (rc && rc != MDB_NOTFOUND)
3086 if (!env->me_pghead && txn->mt_loose_pgs) {
3087 /* Put loose page numbers in mt_free_pgs, since
3088 * we may be unable to return them to me_pghead.
3090 MDB_page *mp = txn->mt_loose_pgs;
3091 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3093 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3094 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3095 txn->mt_loose_pgs = NULL;
3096 txn->mt_loose_count = 0;
3099 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3100 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3101 ? SSIZE_MAX : maxfree_1pg;
3104 /* Come back here after each Put() in case freelist changed */
3109 /* If using records from freeDB which we have not yet
3110 * deleted, delete them and any we reserved for me_pghead.
3112 while (pglast < env->me_pglast) {
3113 rc = mdb_cursor_first(&mc, &key, NULL);
3116 pglast = head_id = *(txnid_t *)key.mv_data;
3117 total_room = head_room = 0;
3118 mdb_tassert(txn, pglast <= env->me_pglast);
3119 rc = mdb_cursor_del(&mc, 0);
3124 /* Save the IDL of pages freed by this txn, to a single record */
3125 if (freecnt < txn->mt_free_pgs[0]) {
3127 /* Make sure last page of freeDB is touched and on freelist */
3128 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3129 if (rc && rc != MDB_NOTFOUND)
3132 free_pgs = txn->mt_free_pgs;
3133 /* Write to last page of freeDB */
3134 key.mv_size = sizeof(txn->mt_txnid);
3135 key.mv_data = &txn->mt_txnid;
3137 freecnt = free_pgs[0];
3138 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3139 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3142 /* Retry if mt_free_pgs[] grew during the Put() */
3143 free_pgs = txn->mt_free_pgs;
3144 } while (freecnt < free_pgs[0]);
3145 mdb_midl_sort(free_pgs);
3146 memcpy(data.mv_data, free_pgs, data.mv_size);
3149 unsigned int i = free_pgs[0];
3150 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3151 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3153 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3159 mop = env->me_pghead;
3160 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3162 /* Reserve records for me_pghead[]. Split it if multi-page,
3163 * to avoid searching freeDB for a page range. Use keys in
3164 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3166 if (total_room >= mop_len) {
3167 if (total_room == mop_len || --more < 0)
3169 } else if (head_room >= maxfree_1pg && head_id > 1) {
3170 /* Keep current record (overflow page), add a new one */
3174 /* (Re)write {key = head_id, IDL length = head_room} */
3175 total_room -= head_room;
3176 head_room = mop_len - total_room;
3177 if (head_room > maxfree_1pg && head_id > 1) {
3178 /* Overflow multi-page for part of me_pghead */
3179 head_room /= head_id; /* amortize page sizes */
3180 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3181 } else if (head_room < 0) {
3182 /* Rare case, not bothering to delete this record */
3185 key.mv_size = sizeof(head_id);
3186 key.mv_data = &head_id;
3187 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3188 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3191 /* IDL is initially empty, zero out at least the length */
3192 pgs = (pgno_t *)data.mv_data;
3193 j = head_room > clean_limit ? head_room : 0;
3197 total_room += head_room;
3200 /* Return loose page numbers to me_pghead, though usually none are
3201 * left at this point. The pages themselves remain in dirty_list.
3203 if (txn->mt_loose_pgs) {
3204 MDB_page *mp = txn->mt_loose_pgs;
3205 unsigned count = txn->mt_loose_count;
3207 /* Room for loose pages + temp IDL with same */
3208 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3210 mop = env->me_pghead;
3211 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3212 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3213 loose[ ++count ] = mp->mp_pgno;
3215 mdb_midl_sort(loose);
3216 mdb_midl_xmerge(mop, loose);
3217 txn->mt_loose_pgs = NULL;
3218 txn->mt_loose_count = 0;
3222 /* Fill in the reserved me_pghead records */
3228 rc = mdb_cursor_first(&mc, &key, &data);
3229 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3230 txnid_t id = *(txnid_t *)key.mv_data;
3231 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3234 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3236 if (len > mop_len) {
3238 data.mv_size = (len + 1) * sizeof(MDB_ID);
3240 data.mv_data = mop -= len;
3243 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3245 if (rc || !(mop_len -= len))
3252 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3253 * @param[in] txn the transaction that's being committed
3254 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3255 * @return 0 on success, non-zero on failure.
3258 mdb_page_flush(MDB_txn *txn, int keep)
3260 MDB_env *env = txn->mt_env;
3261 MDB_ID2L dl = txn->mt_u.dirty_list;
3262 unsigned psize = env->me_psize, j;
3263 int i, pagecount = dl[0].mid, rc;
3264 size_t size = 0, pos = 0;
3266 MDB_page *dp = NULL;
3270 struct iovec iov[MDB_COMMIT_PAGES];
3271 ssize_t wpos = 0, wsize = 0, wres;
3272 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3278 if (env->me_flags & MDB_WRITEMAP) {
3279 /* Clear dirty flags */
3280 while (++i <= pagecount) {
3282 /* Don't flush this page yet */
3283 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3284 dp->mp_flags &= ~P_KEEP;
3288 dp->mp_flags &= ~P_DIRTY;
3293 /* Write the pages */
3295 if (++i <= pagecount) {
3297 /* Don't flush this page yet */
3298 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3299 dp->mp_flags &= ~P_KEEP;
3304 /* clear dirty flag */
3305 dp->mp_flags &= ~P_DIRTY;
3308 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3313 /* Windows actually supports scatter/gather I/O, but only on
3314 * unbuffered file handles. Since we're relying on the OS page
3315 * cache for all our data, that's self-defeating. So we just
3316 * write pages one at a time. We use the ov structure to set
3317 * the write offset, to at least save the overhead of a Seek
3320 DPRINTF(("committing page %"Z"u", pgno));
3321 memset(&ov, 0, sizeof(ov));
3322 ov.Offset = pos & 0xffffffff;
3323 ov.OffsetHigh = pos >> 16 >> 16;
3324 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3326 DPRINTF(("WriteFile: %d", rc));
3330 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3331 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3334 /* Write previous page(s) */
3335 #ifdef MDB_USE_PWRITEV
3336 wres = pwritev(env->me_fd, iov, n, wpos);
3339 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3342 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3346 DPRINTF(("lseek: %s", strerror(rc)));
3349 wres = writev(env->me_fd, iov, n);
3352 if (wres != wsize) {
3357 DPRINTF(("Write error: %s", strerror(rc)));
3359 rc = EIO; /* TODO: Use which error code? */
3360 DPUTS("short write, filesystem full?");
3371 DPRINTF(("committing page %"Z"u", pgno));
3372 next_pos = pos + size;
3373 iov[n].iov_len = size;
3374 iov[n].iov_base = (char *)dp;
3380 /* MIPS has cache coherency issues, this is a no-op everywhere else
3381 * Note: for any size >= on-chip cache size, entire on-chip cache is
3384 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3386 for (i = keep; ++i <= pagecount; ) {
3388 /* This is a page we skipped above */
3391 dl[j].mid = dp->mp_pgno;
3394 mdb_dpage_free(env, dp);
3399 txn->mt_dirty_room += i - j;
3405 mdb_txn_commit(MDB_txn *txn)
3408 unsigned int i, end_mode;
3414 /* mdb_txn_end() mode for a commit which writes nothing */
3415 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3417 if (txn->mt_child) {
3418 rc = mdb_txn_commit(txn->mt_child);
3425 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3429 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3430 DPUTS("txn has failed/finished, can't commit");
3432 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3437 if (txn->mt_parent) {
3438 MDB_txn *parent = txn->mt_parent;
3442 unsigned x, y, len, ps_len;
3444 /* Append our free list to parent's */
3445 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3448 mdb_midl_free(txn->mt_free_pgs);
3449 /* Failures after this must either undo the changes
3450 * to the parent or set MDB_TXN_ERROR in the parent.
3453 parent->mt_next_pgno = txn->mt_next_pgno;
3454 parent->mt_flags = txn->mt_flags;
3456 /* Merge our cursors into parent's and close them */
3457 mdb_cursors_close(txn, 1);
3459 /* Update parent's DB table. */
3460 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3461 parent->mt_numdbs = txn->mt_numdbs;
3462 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3463 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3464 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3465 /* preserve parent's DB_NEW status */
3466 x = parent->mt_dbflags[i] & DB_NEW;
3467 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3470 dst = parent->mt_u.dirty_list;
3471 src = txn->mt_u.dirty_list;
3472 /* Remove anything in our dirty list from parent's spill list */
3473 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3475 pspill[0] = (pgno_t)-1;
3476 /* Mark our dirty pages as deleted in parent spill list */
3477 for (i=0, len=src[0].mid; ++i <= len; ) {
3478 MDB_ID pn = src[i].mid << 1;
3479 while (pn > pspill[x])
3481 if (pn == pspill[x]) {
3486 /* Squash deleted pagenums if we deleted any */
3487 for (x=y; ++x <= ps_len; )
3488 if (!(pspill[x] & 1))
3489 pspill[++y] = pspill[x];
3493 /* Remove anything in our spill list from parent's dirty list */
3494 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3495 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3496 MDB_ID pn = txn->mt_spill_pgs[i];
3498 continue; /* deleted spillpg */
3500 y = mdb_mid2l_search(dst, pn);
3501 if (y <= dst[0].mid && dst[y].mid == pn) {
3503 while (y < dst[0].mid) {
3512 /* Find len = length of merging our dirty list with parent's */
3514 dst[0].mid = 0; /* simplify loops */
3515 if (parent->mt_parent) {
3516 len = x + src[0].mid;
3517 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3518 for (i = x; y && i; y--) {
3519 pgno_t yp = src[y].mid;
3520 while (yp < dst[i].mid)
3522 if (yp == dst[i].mid) {
3527 } else { /* Simplify the above for single-ancestor case */
3528 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3530 /* Merge our dirty list with parent's */
3532 for (i = len; y; dst[i--] = src[y--]) {
3533 pgno_t yp = src[y].mid;
3534 while (yp < dst[x].mid)
3535 dst[i--] = dst[x--];
3536 if (yp == dst[x].mid)
3537 free(dst[x--].mptr);
3539 mdb_tassert(txn, i == x);
3541 free(txn->mt_u.dirty_list);
3542 parent->mt_dirty_room = txn->mt_dirty_room;
3543 if (txn->mt_spill_pgs) {
3544 if (parent->mt_spill_pgs) {
3545 /* TODO: Prevent failure here, so parent does not fail */
3546 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3548 parent->mt_flags |= MDB_TXN_ERROR;
3549 mdb_midl_free(txn->mt_spill_pgs);
3550 mdb_midl_sort(parent->mt_spill_pgs);
3552 parent->mt_spill_pgs = txn->mt_spill_pgs;
3556 /* Append our loose page list to parent's */
3557 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3559 *lp = txn->mt_loose_pgs;
3560 parent->mt_loose_count += txn->mt_loose_count;
3562 parent->mt_child = NULL;
3563 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3568 if (txn != env->me_txn) {
3569 DPUTS("attempt to commit unknown transaction");
3574 mdb_cursors_close(txn, 0);
3576 if (!txn->mt_u.dirty_list[0].mid &&
3577 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3580 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3581 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3583 /* Update DB root pointers */
3584 if (txn->mt_numdbs > CORE_DBS) {
3588 data.mv_size = sizeof(MDB_db);
3590 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3591 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3592 if (txn->mt_dbflags[i] & DB_DIRTY) {
3593 if (TXN_DBI_CHANGED(txn, i)) {
3597 data.mv_data = &txn->mt_dbs[i];
3598 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3606 rc = mdb_freelist_save(txn);
3610 mdb_midl_free(env->me_pghead);
3611 env->me_pghead = NULL;
3612 mdb_midl_shrink(&txn->mt_free_pgs);
3618 if ((rc = mdb_page_flush(txn, 0)) ||
3619 (rc = mdb_env_sync(env, 0)) ||
3620 (rc = mdb_env_write_meta(txn)))
3622 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3625 mdb_txn_end(txn, end_mode);
3633 /** Read the environment parameters of a DB environment before
3634 * mapping it into memory.
3635 * @param[in] env the environment handle
3636 * @param[out] meta address of where to store the meta information
3637 * @return 0 on success, non-zero on failure.
3640 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3646 enum { Size = sizeof(pbuf) };
3648 /* We don't know the page size yet, so use a minimum value.
3649 * Read both meta pages so we can use the latest one.
3652 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3656 memset(&ov, 0, sizeof(ov));
3658 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3659 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3662 rc = pread(env->me_fd, &pbuf, Size, off);
3665 if (rc == 0 && off == 0)
3667 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3668 DPRINTF(("read: %s", mdb_strerror(rc)));
3672 p = (MDB_page *)&pbuf;
3674 if (!F_ISSET(p->mp_flags, P_META)) {
3675 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3680 if (m->mm_magic != MDB_MAGIC) {
3681 DPUTS("meta has invalid magic");
3685 if (m->mm_version != MDB_DATA_VERSION) {
3686 DPRINTF(("database is version %u, expected version %u",
3687 m->mm_version, MDB_DATA_VERSION));
3688 return MDB_VERSION_MISMATCH;
3691 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3697 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3699 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3701 meta->mm_magic = MDB_MAGIC;
3702 meta->mm_version = MDB_DATA_VERSION;
3703 meta->mm_mapsize = env->me_mapsize;
3704 meta->mm_psize = env->me_psize;
3705 meta->mm_last_pg = NUM_METAS-1;
3706 meta->mm_flags = env->me_flags & 0xffff;
3707 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3708 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3709 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3712 /** Write the environment parameters of a freshly created DB environment.
3713 * @param[in] env the environment handle
3714 * @param[in] meta the #MDB_meta to write
3715 * @return 0 on success, non-zero on failure.
3718 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3726 memset(&ov, 0, sizeof(ov));
3727 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3729 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3732 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3733 len = pwrite(fd, ptr, size, pos); \
3734 if (len == -1 && ErrCode() == EINTR) continue; \
3735 rc = (len >= 0); break; } while(1)
3738 DPUTS("writing new meta page");
3740 psize = env->me_psize;
3742 p = calloc(NUM_METAS, psize);
3747 p->mp_flags = P_META;
3748 *(MDB_meta *)METADATA(p) = *meta;
3750 q = (MDB_page *)((char *)p + psize);
3752 q->mp_flags = P_META;
3753 *(MDB_meta *)METADATA(q) = *meta;
3755 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
3758 else if ((unsigned) len == psize * NUM_METAS)
3766 /** Update the environment info to commit a transaction.
3767 * @param[in] txn the transaction that's being committed
3768 * @return 0 on success, non-zero on failure.
3771 mdb_env_write_meta(MDB_txn *txn)
3774 MDB_meta meta, metab, *mp;
3778 int rc, len, toggle;
3787 toggle = txn->mt_txnid & 1;
3788 DPRINTF(("writing meta page %d for root page %"Z"u",
3789 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3792 flags = env->me_flags;
3793 mp = env->me_metas[toggle];
3794 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3795 /* Persist any increases of mapsize config */
3796 if (mapsize < env->me_mapsize)
3797 mapsize = env->me_mapsize;
3799 if (flags & MDB_WRITEMAP) {
3800 mp->mm_mapsize = mapsize;
3801 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3802 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3803 mp->mm_last_pg = txn->mt_next_pgno - 1;
3804 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3805 !(defined(__i386__) || defined(__x86_64__))
3806 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3807 __sync_synchronize();
3809 mp->mm_txnid = txn->mt_txnid;
3810 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3811 unsigned meta_size = env->me_psize;
3812 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3813 ptr = (char *)mp - PAGEHDRSZ;
3814 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3815 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
3819 if (MDB_MSYNC(ptr, meta_size, rc)) {
3826 metab.mm_txnid = mp->mm_txnid;
3827 metab.mm_last_pg = mp->mm_last_pg;
3829 meta.mm_mapsize = mapsize;
3830 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
3831 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
3832 meta.mm_last_pg = txn->mt_next_pgno - 1;
3833 meta.mm_txnid = txn->mt_txnid;
3835 off = offsetof(MDB_meta, mm_mapsize);
3836 ptr = (char *)&meta + off;
3837 len = sizeof(MDB_meta) - off;
3838 off += (char *)mp - env->me_map;
3840 /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC.
3841 * (me_mfd goes to the same file as me_fd, but writing to it
3842 * also syncs to disk. Avoids a separate fdatasync() call.)
3844 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3847 memset(&ov, 0, sizeof(ov));
3849 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3854 rc = pwrite(mfd, ptr, len, off);
3857 rc = rc < 0 ? ErrCode() : EIO;
3862 DPUTS("write failed, disk error?");
3863 /* On a failure, the pagecache still contains the new data.
3864 * Write some old data back, to prevent it from being used.
3865 * Use the non-SYNC fd; we know it will fail anyway.
3867 meta.mm_last_pg = metab.mm_last_pg;
3868 meta.mm_txnid = metab.mm_txnid;
3870 memset(&ov, 0, sizeof(ov));
3872 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3874 r2 = pwrite(env->me_fd, ptr, len, off);
3875 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3878 env->me_flags |= MDB_FATAL_ERROR;
3881 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3882 CACHEFLUSH(env->me_map + off, len, DCACHE);
3884 /* Memory ordering issues are irrelevant; since the entire writer
3885 * is wrapped by wmutex, all of these changes will become visible
3886 * after the wmutex is unlocked. Since the DB is multi-version,
3887 * readers will get consistent data regardless of how fresh or
3888 * how stale their view of these values is.
3891 env->me_txns->mti_txnid = txn->mt_txnid;
3896 /** Check both meta pages to see which one is newer.
3897 * @param[in] env the environment handle
3898 * @return newest #MDB_meta.
3901 mdb_env_pick_meta(const MDB_env *env)
3903 MDB_meta *const *metas = env->me_metas;
3904 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
3908 mdb_env_create(MDB_env **env)
3912 e = calloc(1, sizeof(MDB_env));
3916 e->me_maxreaders = DEFAULT_READERS;
3917 e->me_maxdbs = e->me_numdbs = CORE_DBS;
3918 e->me_fd = INVALID_HANDLE_VALUE;
3919 e->me_lfd = INVALID_HANDLE_VALUE;
3920 e->me_mfd = INVALID_HANDLE_VALUE;
3921 #ifdef MDB_USE_POSIX_SEM
3922 e->me_rmutex = SEM_FAILED;
3923 e->me_wmutex = SEM_FAILED;
3925 e->me_pid = getpid();
3926 GET_PAGESIZE(e->me_os_psize);
3927 VGMEMP_CREATE(e,0,0);
3933 mdb_env_map(MDB_env *env, void *addr)
3936 unsigned int flags = env->me_flags;
3940 LONG sizelo, sizehi;
3943 if (flags & MDB_RDONLY) {
3944 /* Don't set explicit map size, use whatever exists */
3949 msize = env->me_mapsize;
3950 sizelo = msize & 0xffffffff;
3951 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3953 /* Windows won't create mappings for zero length files.
3954 * and won't map more than the file size.
3955 * Just set the maxsize right now.
3957 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3958 || !SetEndOfFile(env->me_fd)
3959 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3963 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3964 PAGE_READWRITE : PAGE_READONLY,
3965 sizehi, sizelo, NULL);
3968 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3969 FILE_MAP_WRITE : FILE_MAP_READ,
3971 rc = env->me_map ? 0 : ErrCode();
3976 int prot = PROT_READ;
3977 if (flags & MDB_WRITEMAP) {
3979 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3982 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3984 if (env->me_map == MAP_FAILED) {
3989 if (flags & MDB_NORDAHEAD) {
3990 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3992 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3994 #ifdef POSIX_MADV_RANDOM
3995 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3996 #endif /* POSIX_MADV_RANDOM */
3997 #endif /* MADV_RANDOM */
4001 /* Can happen because the address argument to mmap() is just a
4002 * hint. mmap() can pick another, e.g. if the range is in use.
4003 * The MAP_FIXED flag would prevent that, but then mmap could
4004 * instead unmap existing pages to make room for the new map.
4006 if (addr && env->me_map != addr)
4007 return EBUSY; /* TODO: Make a new MDB_* error code? */
4009 p = (MDB_page *)env->me_map;
4010 env->me_metas[0] = METADATA(p);
4011 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4017 mdb_env_set_mapsize(MDB_env *env, size_t size)
4019 /* If env is already open, caller is responsible for making
4020 * sure there are no active txns.
4028 meta = mdb_env_pick_meta(env);
4030 size = meta->mm_mapsize;
4032 /* Silently round up to minimum if the size is too small */
4033 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4037 munmap(env->me_map, env->me_mapsize);
4038 env->me_mapsize = size;
4039 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4040 rc = mdb_env_map(env, old);
4044 env->me_mapsize = size;
4046 env->me_maxpg = env->me_mapsize / env->me_psize;
4051 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4055 env->me_maxdbs = dbs + CORE_DBS;
4060 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4062 if (env->me_map || readers < 1)
4064 env->me_maxreaders = readers;
4069 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4071 if (!env || !readers)
4073 *readers = env->me_maxreaders;
4078 mdb_fsize(HANDLE fd, size_t *size)
4081 LARGE_INTEGER fsize;
4083 if (!GetFileSizeEx(fd, &fsize))
4086 *size = fsize.QuadPart;
4100 typedef wchar_t mdb_nchar_t;
4101 # define MDB_NAME(str) L##str
4102 # define mdb_name_cpy wcscpy
4104 /** Character type for file names: char on Unix, wchar_t on Windows */
4105 typedef char mdb_nchar_t;
4106 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */
4107 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */
4110 /** Filename - string of #mdb_nchar_t[] */
4111 typedef struct MDB_name {
4112 int mn_len; /**< Length */
4113 int mn_alloced; /**< True if #mn_val was malloced */
4114 mdb_nchar_t *mn_val; /**< Contents */
4117 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */
4118 static const mdb_nchar_t *const mdb_suffixes[2][2] = {
4119 { MDB_NAME("/data.mdb"), MDB_NAME("") },
4120 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") }
4123 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */
4125 /** Set up filename + scratch area for filename suffix, for opening files.
4126 * It should be freed with #mdb_fname_destroy().
4127 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16.
4129 * @param[in] path Pathname for #mdb_env_open().
4130 * @param[in] envflags Whether a subdir and/or lockfile will be used.
4131 * @param[out] fname Resulting filename, with room for a suffix if necessary.
4134 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname)
4136 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK);
4137 fname->mn_alloced = 0;
4139 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN);
4141 fname->mn_len = strlen(path);
4143 fname->mn_val = (char *) path;
4144 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) {
4145 fname->mn_alloced = 1;
4146 strcpy(fname->mn_val, path);
4154 /** Destroy \b fname from #mdb_fname_init() */
4155 #define mdb_fname_destroy(fname) \
4156 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0)
4158 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */
4159 # define MDB_CLOEXEC O_CLOEXEC
4161 # define MDB_CLOEXEC 0
4164 /** File type, access mode etc. for #mdb_fopen() */
4165 enum mdb_fopen_type {
4167 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS
4169 /* A comment in mdb_fopen() explains some O_* flag choices. */
4170 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */
4171 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */
4172 MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */
4173 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */
4174 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits
4175 * distinguish otherwise-equal MDB_O_* constants from each other.
4177 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY,
4178 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */
4182 /** Open an LMDB file.
4183 * @param[in] env The LMDB environment.
4184 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is
4185 * appended if necessary to create the filename, without changing mn_len.
4186 * @param[in] which Determines file type, access mode, etc.
4187 * @param[in] mode The Unix permissions for the file, if we create it.
4188 * @param[out] res Resulting file handle.
4189 * @return 0 on success, non-zero on failure.
4192 mdb_fopen(const MDB_env *env, MDB_name *fname,
4193 enum mdb_fopen_type which, mdb_mode_t mode,
4196 int rc = MDB_SUCCESS;
4199 DWORD acc, share, disp, attrs;
4204 if (fname->mn_alloced) /* modifiable copy */
4205 mdb_name_cpy(fname->mn_val + fname->mn_len,
4206 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]);
4208 /* The directory must already exist. Usually the file need not.
4209 * MDB_O_META requires the file because we already created it using
4210 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file.
4212 * With MDB_O_COPY we do not want the OS to cache the writes, since
4213 * the source data is already in the OS cache.
4215 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*())
4216 * to avoid the flock() issues noted under Caveats in lmdb.h.
4217 * Also set it for other filehandles which the user cannot get at
4218 * and close himself, which he may need after fork(). I.e. all but
4219 * me_fd, which programs do use via mdb_env_get_fd().
4223 acc = GENERIC_READ|GENERIC_WRITE;
4224 share = FILE_SHARE_READ|FILE_SHARE_WRITE;
4226 attrs = FILE_ATTRIBUTE_NORMAL;
4228 case MDB_O_RDONLY: /* read-only datafile */
4230 disp = OPEN_EXISTING;
4232 case MDB_O_META: /* for writing metapages */
4233 acc = GENERIC_WRITE;
4234 disp = OPEN_EXISTING;
4235 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH;
4237 case MDB_O_COPY: /* mdb_env_copy() & co */
4238 acc = GENERIC_WRITE;
4241 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH;
4243 default: break; /* silence gcc -Wswitch (not all enum values handled) */
4245 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL);
4247 fd = open(fname->mn_val, which & MDB_O_MASK, mode);
4250 if (fd == INVALID_HANDLE_VALUE)
4254 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) {
4255 /* Set CLOEXEC if we could not pass it to open() */
4256 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1)
4257 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
4259 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) {
4260 /* This may require buffer alignment. There is no portable
4261 * way to ask how much, so we require OS pagesize alignment.
4263 # ifdef F_NOCACHE /* __APPLE__ */
4264 (void) fcntl(fd, F_NOCACHE, 1);
4265 # elif defined O_DIRECT
4266 /* open(...O_DIRECT...) would break on filesystems without
4267 * O_DIRECT support (ITS#7682). Try to set it here instead.
4269 if ((flags = fcntl(fd, F_GETFL)) != -1)
4270 (void) fcntl(fd, F_SETFL, flags | O_DIRECT);
4274 #endif /* !_WIN32 */
4281 #ifdef BROKEN_FDATASYNC
4282 #include <sys/utsname.h>
4283 #include <sys/vfs.h>
4286 /** Further setup required for opening an LMDB environment
4289 mdb_env_open2(MDB_env *env)
4291 unsigned int flags = env->me_flags;
4292 int i, newenv = 0, rc;
4296 /* See if we should use QueryLimited */
4298 if ((rc & 0xff) > 5)
4299 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4301 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4304 #ifdef BROKEN_FDATASYNC
4305 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4306 * https://lkml.org/lkml/2012/9/3/83
4307 * Kernels after 3.6-rc6 are known good.
4308 * https://lkml.org/lkml/2012/9/10/556
4309 * See if the DB is on ext3/ext4, then check for new enough kernel
4310 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4315 fstatfs(env->me_fd, &st);
4316 while (st.f_type == 0xEF53) {
4320 if (uts.release[0] < '3') {
4321 if (!strncmp(uts.release, "2.6.32.", 7)) {
4322 i = atoi(uts.release+7);
4324 break; /* 2.6.32.60 and newer is OK */
4325 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4326 i = atoi(uts.release+7);
4328 break; /* 2.6.34.15 and newer is OK */
4330 } else if (uts.release[0] == '3') {
4331 i = atoi(uts.release+2);
4333 break; /* 3.6 and newer is OK */
4335 i = atoi(uts.release+4);
4337 break; /* 3.5.4 and newer is OK */
4338 } else if (i == 2) {
4339 i = atoi(uts.release+4);
4341 break; /* 3.2.30 and newer is OK */
4343 } else { /* 4.x and newer is OK */
4346 env->me_flags |= MDB_FSYNCONLY;
4352 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4355 DPUTS("new mdbenv");
4357 env->me_psize = env->me_os_psize;
4358 if (env->me_psize > MAX_PAGESIZE)
4359 env->me_psize = MAX_PAGESIZE;
4360 memset(&meta, 0, sizeof(meta));
4361 mdb_env_init_meta0(env, &meta);
4362 meta.mm_mapsize = DEFAULT_MAPSIZE;
4364 env->me_psize = meta.mm_psize;
4367 /* Was a mapsize configured? */
4368 if (!env->me_mapsize) {
4369 env->me_mapsize = meta.mm_mapsize;
4372 /* Make sure mapsize >= committed data size. Even when using
4373 * mm_mapsize, which could be broken in old files (ITS#7789).
4375 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4376 if (env->me_mapsize < minsize)
4377 env->me_mapsize = minsize;
4379 meta.mm_mapsize = env->me_mapsize;
4381 if (newenv && !(flags & MDB_FIXEDMAP)) {
4382 /* mdb_env_map() may grow the datafile. Write the metapages
4383 * first, so the file will be valid if initialization fails.
4384 * Except with FIXEDMAP, since we do not yet know mm_address.
4385 * We could fill in mm_address later, but then a different
4386 * program might end up doing that - one with a memory layout
4387 * and map address which does not suit the main program.
4389 rc = mdb_env_init_meta(env, &meta);
4395 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4400 if (flags & MDB_FIXEDMAP)
4401 meta.mm_address = env->me_map;
4402 i = mdb_env_init_meta(env, &meta);
4403 if (i != MDB_SUCCESS) {
4408 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4409 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4411 #if !(MDB_MAXKEYSIZE)
4412 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4414 env->me_maxpg = env->me_mapsize / env->me_psize;
4418 MDB_meta *meta = mdb_env_pick_meta(env);
4419 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4421 DPRINTF(("opened database version %u, pagesize %u",
4422 meta->mm_version, env->me_psize));
4423 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4424 DPRINTF(("depth: %u", db->md_depth));
4425 DPRINTF(("entries: %"Z"u", db->md_entries));
4426 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4427 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4428 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4429 DPRINTF(("root: %"Z"u", db->md_root));
4437 /** Release a reader thread's slot in the reader lock table.
4438 * This function is called automatically when a thread exits.
4439 * @param[in] ptr This points to the slot in the reader lock table.
4442 mdb_env_reader_dest(void *ptr)
4444 MDB_reader *reader = ptr;
4447 if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */
4449 /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */
4454 /** Junk for arranging thread-specific callbacks on Windows. This is
4455 * necessarily platform and compiler-specific. Windows supports up
4456 * to 1088 keys. Let's assume nobody opens more than 64 environments
4457 * in a single process, for now. They can override this if needed.
4459 #ifndef MAX_TLS_KEYS
4460 #define MAX_TLS_KEYS 64
4462 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4463 static int mdb_tls_nkeys;
4465 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4469 case DLL_PROCESS_ATTACH: break;
4470 case DLL_THREAD_ATTACH: break;
4471 case DLL_THREAD_DETACH:
4472 for (i=0; i<mdb_tls_nkeys; i++) {
4473 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4475 mdb_env_reader_dest(r);
4479 case DLL_PROCESS_DETACH: break;
4484 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4486 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4490 /* Force some symbol references.
4491 * _tls_used forces the linker to create the TLS directory if not already done
4492 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4494 #pragma comment(linker, "/INCLUDE:_tls_used")
4495 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4496 #pragma const_seg(".CRT$XLB")
4497 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4498 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4501 #pragma comment(linker, "/INCLUDE:__tls_used")
4502 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4503 #pragma data_seg(".CRT$XLB")
4504 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4506 #endif /* WIN 32/64 */
4507 #endif /* !__GNUC__ */
4510 /** Downgrade the exclusive lock on the region back to shared */
4512 mdb_env_share_locks(MDB_env *env, int *excl)
4515 MDB_meta *meta = mdb_env_pick_meta(env);
4517 env->me_txns->mti_txnid = meta->mm_txnid;
4522 /* First acquire a shared lock. The Unlock will
4523 * then release the existing exclusive lock.
4525 memset(&ov, 0, sizeof(ov));
4526 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4529 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4535 struct flock lock_info;
4536 /* The shared lock replaces the existing lock */
4537 memset((void *)&lock_info, 0, sizeof(lock_info));
4538 lock_info.l_type = F_RDLCK;
4539 lock_info.l_whence = SEEK_SET;
4540 lock_info.l_start = 0;
4541 lock_info.l_len = 1;
4542 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4543 (rc = ErrCode()) == EINTR) ;
4544 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4551 /** Try to get exclusive lock, otherwise shared.
4552 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4555 mdb_env_excl_lock(MDB_env *env, int *excl)
4559 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4563 memset(&ov, 0, sizeof(ov));
4564 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4571 struct flock lock_info;
4572 memset((void *)&lock_info, 0, sizeof(lock_info));
4573 lock_info.l_type = F_WRLCK;
4574 lock_info.l_whence = SEEK_SET;
4575 lock_info.l_start = 0;
4576 lock_info.l_len = 1;
4577 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4578 (rc = ErrCode()) == EINTR) ;
4582 # ifndef MDB_USE_POSIX_MUTEX
4583 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4586 lock_info.l_type = F_RDLCK;
4587 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4588 (rc = ErrCode()) == EINTR) ;
4598 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4600 * @(#) $Revision: 5.1 $
4601 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4602 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4604 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4608 * Please do not copyright this code. This code is in the public domain.
4610 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4611 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4612 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4613 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4614 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4615 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4616 * PERFORMANCE OF THIS SOFTWARE.
4619 * chongo <Landon Curt Noll> /\oo/\
4620 * http://www.isthe.com/chongo/
4622 * Share and Enjoy! :-)
4625 typedef unsigned long long mdb_hash_t;
4626 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4628 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4629 * @param[in] val value to hash
4630 * @param[in] hval initial value for hash
4631 * @return 64 bit hash
4633 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4634 * hval arg on the first call.
4637 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4639 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4640 unsigned char *end = s + val->mv_size;
4642 * FNV-1a hash each octet of the string
4645 /* xor the bottom with the current octet */
4646 hval ^= (mdb_hash_t)*s++;
4648 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4649 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4650 (hval << 7) + (hval << 8) + (hval << 40);
4652 /* return our new hash value */
4656 /** Hash the string and output the encoded hash.
4657 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4658 * very short name limits. We don't care about the encoding being reversible,
4659 * we just want to preserve as many bits of the input as possible in a
4660 * small printable string.
4661 * @param[in] str string to hash
4662 * @param[out] encbuf an array of 11 chars to hold the hash
4664 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4667 mdb_pack85(unsigned long l, char *out)
4671 for (i=0; i<5; i++) {
4672 *out++ = mdb_a85[l % 85];
4678 mdb_hash_enc(MDB_val *val, char *encbuf)
4680 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4682 mdb_pack85(h, encbuf);
4683 mdb_pack85(h>>32, encbuf+5);
4688 /** Open and/or initialize the lock region for the environment.
4689 * @param[in] env The LMDB environment.
4690 * @param[in] fname Filename + scratch area, from #mdb_fname_init().
4691 * @param[in] mode The Unix permissions for the file, if we create it.
4692 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4693 * @return 0 on success, non-zero on failure.
4696 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl)
4699 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4701 # define MDB_ERRCODE_ROFS EROFS
4706 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd);
4708 /* Omit lockfile if read-only env on read-only filesystem */
4709 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4715 if (!(env->me_flags & MDB_NOTLS)) {
4716 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4719 env->me_flags |= MDB_ENV_TXKEY;
4721 /* Windows TLS callbacks need help finding their TLS info. */
4722 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4726 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4730 /* Try to get exclusive lock. If we succeed, then
4731 * nobody is using the lock region and we should initialize it.
4733 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4736 size = GetFileSize(env->me_lfd, NULL);
4738 size = lseek(env->me_lfd, 0, SEEK_END);
4739 if (size == -1) goto fail_errno;
4741 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4742 if (size < rsize && *excl > 0) {
4744 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4745 || !SetEndOfFile(env->me_lfd))
4748 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4752 size = rsize - sizeof(MDB_txninfo);
4753 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4758 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4760 if (!mh) goto fail_errno;
4761 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4763 if (!env->me_txns) goto fail_errno;
4765 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4767 if (m == MAP_FAILED) goto fail_errno;
4773 BY_HANDLE_FILE_INFORMATION stbuf;
4782 if (!mdb_sec_inited) {
4783 InitializeSecurityDescriptor(&mdb_null_sd,
4784 SECURITY_DESCRIPTOR_REVISION);
4785 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4786 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4787 mdb_all_sa.bInheritHandle = FALSE;
4788 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4791 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4792 idbuf.volume = stbuf.dwVolumeSerialNumber;
4793 idbuf.nhigh = stbuf.nFileIndexHigh;
4794 idbuf.nlow = stbuf.nFileIndexLow;
4795 val.mv_data = &idbuf;
4796 val.mv_size = sizeof(idbuf);
4797 mdb_hash_enc(&val, encbuf);
4798 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4799 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4800 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4801 if (!env->me_rmutex) goto fail_errno;
4802 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4803 if (!env->me_wmutex) goto fail_errno;
4804 #elif defined(MDB_USE_POSIX_SEM)
4813 #if defined(__NetBSD__)
4814 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4816 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4817 idbuf.dev = stbuf.st_dev;
4818 idbuf.ino = stbuf.st_ino;
4819 val.mv_data = &idbuf;
4820 val.mv_size = sizeof(idbuf);
4821 mdb_hash_enc(&val, encbuf);
4822 #ifdef MDB_SHORT_SEMNAMES
4823 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4825 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4826 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4827 /* Clean up after a previous run, if needed: Try to
4828 * remove both semaphores before doing anything else.
4830 sem_unlink(env->me_txns->mti_rmname);
4831 sem_unlink(env->me_txns->mti_wmname);
4832 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4833 O_CREAT|O_EXCL, mode, 1);
4834 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4835 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4836 O_CREAT|O_EXCL, mode, 1);
4837 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4838 #else /* MDB_USE_POSIX_MUTEX: */
4839 pthread_mutexattr_t mattr;
4841 /* Solaris needs this before initing a robust mutex. Otherwise
4842 * it may skip the init and return EBUSY "seems someone already
4843 * inited" or EINVAL "it was inited differently".
4845 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
4846 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
4848 if ((rc = pthread_mutexattr_init(&mattr)))
4851 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
4852 #ifdef MDB_ROBUST_SUPPORTED
4853 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
4855 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
4856 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
4857 pthread_mutexattr_destroy(&mattr);
4860 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4862 env->me_txns->mti_magic = MDB_MAGIC;
4863 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4864 env->me_txns->mti_txnid = 0;
4865 env->me_txns->mti_numreaders = 0;
4868 if (env->me_txns->mti_magic != MDB_MAGIC) {
4869 DPUTS("lock region has invalid magic");
4873 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4874 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4875 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4876 rc = MDB_VERSION_MISMATCH;
4880 if (rc && rc != EACCES && rc != EAGAIN) {
4884 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4885 if (!env->me_rmutex) goto fail_errno;
4886 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4887 if (!env->me_wmutex) goto fail_errno;
4888 #elif defined(MDB_USE_POSIX_SEM)
4889 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4890 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4891 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4892 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4903 /** Only a subset of the @ref mdb_env flags can be changed
4904 * at runtime. Changing other flags requires closing the
4905 * environment and re-opening it with the new flags.
4907 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4908 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4909 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4911 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4912 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4916 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4921 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4924 flags |= env->me_flags;
4926 rc = mdb_fname_init(path, flags, &fname);
4930 if (flags & MDB_RDONLY) {
4931 /* silently ignore WRITEMAP when we're only getting read access */
4932 flags &= ~MDB_WRITEMAP;
4934 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4935 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4938 env->me_flags = flags |= MDB_ENV_ACTIVE;
4942 env->me_path = strdup(path);
4943 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4944 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4945 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4946 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4950 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
4952 /* For RDONLY, get lockfile after we know datafile exists */
4953 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4954 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
4959 rc = mdb_fopen(env, &fname,
4960 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR,
4965 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4966 rc = mdb_env_setup_locks(env, &fname, mode, &excl);
4971 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4972 if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) {
4973 /* Synchronous fd for meta writes. Needed even with
4974 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4976 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd);
4980 DPRINTF(("opened dbenv %p", (void *) env));
4982 rc = mdb_env_share_locks(env, &excl);
4986 if (!(flags & MDB_RDONLY)) {
4988 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4989 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4990 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4991 (txn = calloc(1, size)))
4993 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4994 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4995 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4996 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4998 txn->mt_dbxs = env->me_dbxs;
4999 txn->mt_flags = MDB_TXN_FINISHED;
5009 mdb_env_close0(env, excl);
5011 mdb_fname_destroy(fname);
5015 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5017 mdb_env_close0(MDB_env *env, int excl)
5021 if (!(env->me_flags & MDB_ENV_ACTIVE))
5024 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5026 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5027 free(env->me_dbxs[i].md_name.mv_data);
5032 free(env->me_dbiseqs);
5033 free(env->me_dbflags);
5035 free(env->me_dirty_list);
5037 mdb_midl_free(env->me_free_pgs);
5039 if (env->me_flags & MDB_ENV_TXKEY) {
5040 pthread_key_delete(env->me_txkey);
5042 /* Delete our key from the global list */
5043 for (i=0; i<mdb_tls_nkeys; i++)
5044 if (mdb_tls_keys[i] == env->me_txkey) {
5045 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5053 munmap(env->me_map, env->me_mapsize);
5055 if (env->me_mfd != INVALID_HANDLE_VALUE)
5056 (void) close(env->me_mfd);
5057 if (env->me_fd != INVALID_HANDLE_VALUE)
5058 (void) close(env->me_fd);
5060 MDB_PID_T pid = env->me_pid;
5061 /* Clearing readers is done in this function because
5062 * me_txkey with its destructor must be disabled first.
5064 * We skip the the reader mutex, so we touch only
5065 * data owned by this process (me_close_readers and
5066 * our readers), and clear each reader atomically.
5068 for (i = env->me_close_readers; --i >= 0; )
5069 if (env->me_txns->mti_readers[i].mr_pid == pid)
5070 env->me_txns->mti_readers[i].mr_pid = 0;
5072 if (env->me_rmutex) {
5073 CloseHandle(env->me_rmutex);
5074 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5076 /* Windows automatically destroys the mutexes when
5077 * the last handle closes.
5079 #elif defined(MDB_USE_POSIX_SEM)
5080 if (env->me_rmutex != SEM_FAILED) {
5081 sem_close(env->me_rmutex);
5082 if (env->me_wmutex != SEM_FAILED)
5083 sem_close(env->me_wmutex);
5084 /* If we have the filelock: If we are the
5085 * only remaining user, clean up semaphores.
5088 mdb_env_excl_lock(env, &excl);
5090 sem_unlink(env->me_txns->mti_rmname);
5091 sem_unlink(env->me_txns->mti_wmname);
5095 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5097 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5100 /* Unlock the lockfile. Windows would have unlocked it
5101 * after closing anyway, but not necessarily at once.
5103 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5106 (void) close(env->me_lfd);
5109 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5113 mdb_env_close(MDB_env *env)
5120 VGMEMP_DESTROY(env);
5121 while ((dp = env->me_dpages) != NULL) {
5122 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5123 env->me_dpages = dp->mp_next;
5127 mdb_env_close0(env, 0);
5131 /** Compare two items pointing at aligned size_t's */
5133 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5135 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
5136 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
5139 /** Compare two items pointing at aligned unsigned int's.
5141 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5142 * but #mdb_cmp_clong() is called instead if the data type is size_t.
5145 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5147 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5148 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5151 /** Compare two items pointing at unsigned ints of unknown alignment.
5152 * Nodes and keys are guaranteed to be 2-byte aligned.
5155 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5157 #if BYTE_ORDER == LITTLE_ENDIAN
5158 unsigned short *u, *c;
5161 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5162 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5165 } while(!x && u > (unsigned short *)a->mv_data);
5168 unsigned short *u, *c, *end;
5171 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5172 u = (unsigned short *)a->mv_data;
5173 c = (unsigned short *)b->mv_data;
5176 } while(!x && u < end);
5181 /** Compare two items lexically */
5183 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5190 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5196 diff = memcmp(a->mv_data, b->mv_data, len);
5197 return diff ? diff : len_diff<0 ? -1 : len_diff;
5200 /** Compare two items in reverse byte order */
5202 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5204 const unsigned char *p1, *p2, *p1_lim;
5208 p1_lim = (const unsigned char *)a->mv_data;
5209 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5210 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5212 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5218 while (p1 > p1_lim) {
5219 diff = *--p1 - *--p2;
5223 return len_diff<0 ? -1 : len_diff;
5226 /** Search for key within a page, using binary search.
5227 * Returns the smallest entry larger or equal to the key.
5228 * If exactp is non-null, stores whether the found entry was an exact match
5229 * in *exactp (1 or 0).
5230 * Updates the cursor index with the index of the found entry.
5231 * If no entry larger or equal to the key is found, returns NULL.
5234 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5236 unsigned int i = 0, nkeys;
5239 MDB_page *mp = mc->mc_pg[mc->mc_top];
5240 MDB_node *node = NULL;
5245 nkeys = NUMKEYS(mp);
5247 DPRINTF(("searching %u keys in %s %spage %"Z"u",
5248 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5251 low = IS_LEAF(mp) ? 0 : 1;
5253 cmp = mc->mc_dbx->md_cmp;
5255 /* Branch pages have no data, so if using integer keys,
5256 * alignment is guaranteed. Use faster mdb_cmp_int.
5258 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5259 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5266 nodekey.mv_size = mc->mc_db->md_pad;
5267 node = NODEPTR(mp, 0); /* fake */
5268 while (low <= high) {
5269 i = (low + high) >> 1;
5270 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5271 rc = cmp(key, &nodekey);
5272 DPRINTF(("found leaf index %u [%s], rc = %i",
5273 i, DKEY(&nodekey), rc));
5282 while (low <= high) {
5283 i = (low + high) >> 1;
5285 node = NODEPTR(mp, i);
5286 nodekey.mv_size = NODEKSZ(node);
5287 nodekey.mv_data = NODEKEY(node);
5289 rc = cmp(key, &nodekey);
5292 DPRINTF(("found leaf index %u [%s], rc = %i",
5293 i, DKEY(&nodekey), rc));
5295 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5296 i, DKEY(&nodekey), NODEPGNO(node), rc));
5307 if (rc > 0) { /* Found entry is less than the key. */
5308 i++; /* Skip to get the smallest entry larger than key. */
5310 node = NODEPTR(mp, i);
5313 *exactp = (rc == 0 && nkeys > 0);
5314 /* store the key index */
5315 mc->mc_ki[mc->mc_top] = i;
5317 /* There is no entry larger or equal to the key. */
5320 /* nodeptr is fake for LEAF2 */
5326 mdb_cursor_adjust(MDB_cursor *mc, func)
5330 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5331 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5338 /** Pop a page off the top of the cursor's stack. */
5340 mdb_cursor_pop(MDB_cursor *mc)
5343 DPRINTF(("popping page %"Z"u off db %d cursor %p",
5344 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5350 mc->mc_flags &= ~C_INITIALIZED;
5355 /** Push a page onto the top of the cursor's stack.
5356 * Set #MDB_TXN_ERROR on failure.
5359 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5361 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5362 DDBI(mc), (void *) mc));
5364 if (mc->mc_snum >= CURSOR_STACK) {
5365 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5366 return MDB_CURSOR_FULL;
5369 mc->mc_top = mc->mc_snum++;
5370 mc->mc_pg[mc->mc_top] = mp;
5371 mc->mc_ki[mc->mc_top] = 0;
5376 /** Find the address of the page corresponding to a given page number.
5377 * Set #MDB_TXN_ERROR on failure.
5378 * @param[in] mc the cursor accessing the page.
5379 * @param[in] pgno the page number for the page to retrieve.
5380 * @param[out] ret address of a pointer where the page's address will be stored.
5381 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5382 * @return 0 on success, non-zero on failure.
5385 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5387 MDB_txn *txn = mc->mc_txn;
5388 MDB_env *env = txn->mt_env;
5392 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5396 MDB_ID2L dl = tx2->mt_u.dirty_list;
5398 /* Spilled pages were dirtied in this txn and flushed
5399 * because the dirty list got full. Bring this page
5400 * back in from the map (but don't unspill it here,
5401 * leave that unless page_touch happens again).
5403 if (tx2->mt_spill_pgs) {
5404 MDB_ID pn = pgno << 1;
5405 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5406 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5407 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5412 unsigned x = mdb_mid2l_search(dl, pgno);
5413 if (x <= dl[0].mid && dl[x].mid == pgno) {
5419 } while ((tx2 = tx2->mt_parent) != NULL);
5422 if (pgno < txn->mt_next_pgno) {
5424 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5426 DPRINTF(("page %"Z"u not found", pgno));
5427 txn->mt_flags |= MDB_TXN_ERROR;
5428 return MDB_PAGE_NOTFOUND;
5438 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5439 * The cursor is at the root page, set up the rest of it.
5442 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5444 MDB_page *mp = mc->mc_pg[mc->mc_top];
5448 while (IS_BRANCH(mp)) {
5452 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5453 /* Don't assert on branch pages in the FreeDB. We can get here
5454 * while in the process of rebalancing a FreeDB branch page; we must
5455 * let that proceed. ITS#8336
5457 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
5458 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5460 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5462 if (flags & MDB_PS_LAST) {
5463 i = NUMKEYS(mp) - 1;
5464 /* if already init'd, see if we're already in right place */
5465 if (mc->mc_flags & C_INITIALIZED) {
5466 if (mc->mc_ki[mc->mc_top] == i) {
5467 mc->mc_top = mc->mc_snum++;
5468 mp = mc->mc_pg[mc->mc_top];
5475 node = mdb_node_search(mc, key, &exact);
5477 i = NUMKEYS(mp) - 1;
5479 i = mc->mc_ki[mc->mc_top];
5481 mdb_cassert(mc, i > 0);
5485 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5488 mdb_cassert(mc, i < NUMKEYS(mp));
5489 node = NODEPTR(mp, i);
5491 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5494 mc->mc_ki[mc->mc_top] = i;
5495 if ((rc = mdb_cursor_push(mc, mp)))
5499 if (flags & MDB_PS_MODIFY) {
5500 if ((rc = mdb_page_touch(mc)) != 0)
5502 mp = mc->mc_pg[mc->mc_top];
5507 DPRINTF(("internal error, index points to a %02X page!?",
5509 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5510 return MDB_CORRUPTED;
5513 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5514 key ? DKEY(key) : "null"));
5515 mc->mc_flags |= C_INITIALIZED;
5516 mc->mc_flags &= ~C_EOF;
5521 /** Search for the lowest key under the current branch page.
5522 * This just bypasses a NUMKEYS check in the current page
5523 * before calling mdb_page_search_root(), because the callers
5524 * are all in situations where the current page is known to
5528 mdb_page_search_lowest(MDB_cursor *mc)
5530 MDB_page *mp = mc->mc_pg[mc->mc_top];
5531 MDB_node *node = NODEPTR(mp, 0);
5534 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
5537 mc->mc_ki[mc->mc_top] = 0;
5538 if ((rc = mdb_cursor_push(mc, mp)))
5540 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5543 /** Search for the page a given key should be in.
5544 * Push it and its parent pages on the cursor stack.
5545 * @param[in,out] mc the cursor for this operation.
5546 * @param[in] key the key to search for, or NULL for first/last page.
5547 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5548 * are touched (updated with new page numbers).
5549 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5550 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5551 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5552 * @return 0 on success, non-zero on failure.
5555 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5560 /* Make sure the txn is still viable, then find the root from
5561 * the txn's db table and set it as the root of the cursor's stack.
5563 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
5564 DPUTS("transaction may not be used now");
5567 /* Make sure we're using an up-to-date root */
5568 if (*mc->mc_dbflag & DB_STALE) {
5570 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5572 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5573 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5580 MDB_node *leaf = mdb_node_search(&mc2,
5581 &mc->mc_dbx->md_name, &exact);
5583 return MDB_NOTFOUND;
5584 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5585 return MDB_INCOMPATIBLE; /* not a named DB */
5586 rc = mdb_node_read(&mc2, leaf, &data);
5589 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5591 /* The txn may not know this DBI, or another process may
5592 * have dropped and recreated the DB with other flags.
5594 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5595 return MDB_INCOMPATIBLE;
5596 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5598 *mc->mc_dbflag &= ~DB_STALE;
5600 root = mc->mc_db->md_root;
5602 if (root == P_INVALID) { /* Tree is empty. */
5603 DPUTS("tree is empty");
5604 return MDB_NOTFOUND;
5608 mdb_cassert(mc, root > 1);
5609 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5610 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
5616 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5617 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5619 if (flags & MDB_PS_MODIFY) {
5620 if ((rc = mdb_page_touch(mc)))
5624 if (flags & MDB_PS_ROOTONLY)
5627 return mdb_page_search_root(mc, key, flags);
5631 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5633 MDB_txn *txn = mc->mc_txn;
5634 pgno_t pg = mp->mp_pgno;
5635 unsigned x = 0, ovpages = mp->mp_pages;
5636 MDB_env *env = txn->mt_env;
5637 MDB_IDL sl = txn->mt_spill_pgs;
5638 MDB_ID pn = pg << 1;
5641 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5642 /* If the page is dirty or on the spill list we just acquired it,
5643 * so we should give it back to our current free list, if any.
5644 * Otherwise put it onto the list of pages we freed in this txn.
5646 * Won't create me_pghead: me_pglast must be inited along with it.
5647 * Unsupported in nested txns: They would need to hide the page
5648 * range in ancestor txns' dirty and spilled lists.
5650 if (env->me_pghead &&
5652 ((mp->mp_flags & P_DIRTY) ||
5653 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5657 MDB_ID2 *dl, ix, iy;
5658 rc = mdb_midl_need(&env->me_pghead, ovpages);
5661 if (!(mp->mp_flags & P_DIRTY)) {
5662 /* This page is no longer spilled */
5669 /* Remove from dirty list */
5670 dl = txn->mt_u.dirty_list;
5672 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5678 mdb_cassert(mc, x > 1);
5680 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5681 txn->mt_flags |= MDB_TXN_ERROR;
5682 return MDB_CORRUPTED;
5685 txn->mt_dirty_room++;
5686 if (!(env->me_flags & MDB_WRITEMAP))
5687 mdb_dpage_free(env, mp);
5689 /* Insert in me_pghead */
5690 mop = env->me_pghead;
5691 j = mop[0] + ovpages;
5692 for (i = mop[0]; i && mop[i] < pg; i--)
5698 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5702 mc->mc_db->md_overflow_pages -= ovpages;
5706 /** Return the data associated with a given node.
5707 * @param[in] mc The cursor for this operation.
5708 * @param[in] leaf The node being read.
5709 * @param[out] data Updated to point to the node's data.
5710 * @return 0 on success, non-zero on failure.
5713 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
5715 MDB_page *omp; /* overflow page */
5719 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5720 data->mv_size = NODEDSZ(leaf);
5721 data->mv_data = NODEDATA(leaf);
5725 /* Read overflow data.
5727 data->mv_size = NODEDSZ(leaf);
5728 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5729 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
5730 DPRINTF(("read overflow page %"Z"u failed", pgno));
5733 data->mv_data = METADATA(omp);
5739 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5740 MDB_val *key, MDB_val *data)
5747 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5749 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
5752 if (txn->mt_flags & MDB_TXN_BLOCKED)
5755 mdb_cursor_init(&mc, txn, dbi, &mx);
5756 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5759 /** Find a sibling for a page.
5760 * Replaces the page at the top of the cursor's stack with the
5761 * specified sibling, if one exists.
5762 * @param[in] mc The cursor for this operation.
5763 * @param[in] move_right Non-zero if the right sibling is requested,
5764 * otherwise the left sibling.
5765 * @return 0 on success, non-zero on failure.
5768 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5774 if (mc->mc_snum < 2) {
5775 return MDB_NOTFOUND; /* root has no siblings */
5779 DPRINTF(("parent page is page %"Z"u, index %u",
5780 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5782 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5783 : (mc->mc_ki[mc->mc_top] == 0)) {
5784 DPRINTF(("no more keys left, moving to %s sibling",
5785 move_right ? "right" : "left"));
5786 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5787 /* undo cursor_pop before returning */
5794 mc->mc_ki[mc->mc_top]++;
5796 mc->mc_ki[mc->mc_top]--;
5797 DPRINTF(("just moving to %s index key %u",
5798 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5800 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5802 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5803 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
5804 /* mc will be inconsistent if caller does mc_snum++ as above */
5805 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5809 mdb_cursor_push(mc, mp);
5811 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5816 /** Move the cursor to the next data item. */
5818 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5824 if ((mc->mc_flags & C_DEL && op == MDB_NEXT_DUP))
5825 return MDB_NOTFOUND;
5827 if (!(mc->mc_flags & C_INITIALIZED))
5828 return mdb_cursor_first(mc, key, data);
5830 mp = mc->mc_pg[mc->mc_top];
5832 if (mc->mc_flags & C_EOF) {
5833 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mp)-1)
5834 return MDB_NOTFOUND;
5835 mc->mc_flags ^= C_EOF;
5838 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5839 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5840 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5841 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5842 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5843 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5844 if (rc == MDB_SUCCESS)
5845 MDB_GET_KEY(leaf, key);
5850 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5851 if (op == MDB_NEXT_DUP)
5852 return MDB_NOTFOUND;
5856 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5857 mdb_dbg_pgno(mp), (void *) mc));
5858 if (mc->mc_flags & C_DEL) {
5859 mc->mc_flags ^= C_DEL;
5863 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5864 DPUTS("=====> move to next sibling page");
5865 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5866 mc->mc_flags |= C_EOF;
5869 mp = mc->mc_pg[mc->mc_top];
5870 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5872 mc->mc_ki[mc->mc_top]++;
5875 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5876 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5879 key->mv_size = mc->mc_db->md_pad;
5880 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5884 mdb_cassert(mc, IS_LEAF(mp));
5885 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5887 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5888 mdb_xcursor_init1(mc, leaf);
5891 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5894 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5895 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5896 if (rc != MDB_SUCCESS)
5901 MDB_GET_KEY(leaf, key);
5905 /** Move the cursor to the previous data item. */
5907 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5913 if (!(mc->mc_flags & C_INITIALIZED)) {
5914 rc = mdb_cursor_last(mc, key, data);
5917 mc->mc_ki[mc->mc_top]++;
5920 mp = mc->mc_pg[mc->mc_top];
5922 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5923 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5924 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5925 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5926 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5927 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5928 if (rc == MDB_SUCCESS) {
5929 MDB_GET_KEY(leaf, key);
5930 mc->mc_flags &= ~C_EOF;
5936 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5937 if (op == MDB_PREV_DUP)
5938 return MDB_NOTFOUND;
5942 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5943 mdb_dbg_pgno(mp), (void *) mc));
5945 mc->mc_flags &= ~(C_EOF|C_DEL);
5947 if (mc->mc_ki[mc->mc_top] == 0) {
5948 DPUTS("=====> move to prev sibling page");
5949 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5952 mp = mc->mc_pg[mc->mc_top];
5953 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5954 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5956 mc->mc_ki[mc->mc_top]--;
5958 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5959 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5962 key->mv_size = mc->mc_db->md_pad;
5963 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5967 mdb_cassert(mc, IS_LEAF(mp));
5968 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5970 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5971 mdb_xcursor_init1(mc, leaf);
5974 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
5977 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5978 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5979 if (rc != MDB_SUCCESS)
5984 MDB_GET_KEY(leaf, key);
5988 /** Set the cursor on a specific data item. */
5990 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5991 MDB_cursor_op op, int *exactp)
5995 MDB_node *leaf = NULL;
5998 if (key->mv_size == 0)
5999 return MDB_BAD_VALSIZE;
6002 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6004 /* See if we're already on the right page */
6005 if (mc->mc_flags & C_INITIALIZED) {
6008 mp = mc->mc_pg[mc->mc_top];
6010 mc->mc_ki[mc->mc_top] = 0;
6011 return MDB_NOTFOUND;
6013 if (mp->mp_flags & P_LEAF2) {
6014 nodekey.mv_size = mc->mc_db->md_pad;
6015 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
6017 leaf = NODEPTR(mp, 0);
6018 MDB_GET_KEY2(leaf, nodekey);
6020 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6022 /* Probably happens rarely, but first node on the page
6023 * was the one we wanted.
6025 mc->mc_ki[mc->mc_top] = 0;
6032 unsigned int nkeys = NUMKEYS(mp);
6034 if (mp->mp_flags & P_LEAF2) {
6035 nodekey.mv_data = LEAF2KEY(mp,
6036 nkeys-1, nodekey.mv_size);
6038 leaf = NODEPTR(mp, nkeys-1);
6039 MDB_GET_KEY2(leaf, nodekey);
6041 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6043 /* last node was the one we wanted */
6044 mc->mc_ki[mc->mc_top] = nkeys-1;
6050 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6051 /* This is definitely the right page, skip search_page */
6052 if (mp->mp_flags & P_LEAF2) {
6053 nodekey.mv_data = LEAF2KEY(mp,
6054 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6056 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6057 MDB_GET_KEY2(leaf, nodekey);
6059 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6061 /* current node was the one we wanted */
6068 mc->mc_flags &= ~C_EOF;
6072 /* If any parents have right-sibs, search.
6073 * Otherwise, there's nothing further.
6075 for (i=0; i<mc->mc_top; i++)
6077 NUMKEYS(mc->mc_pg[i])-1)
6079 if (i == mc->mc_top) {
6080 /* There are no other pages */
6081 mc->mc_ki[mc->mc_top] = nkeys;
6082 return MDB_NOTFOUND;
6086 /* There are no other pages */
6087 mc->mc_ki[mc->mc_top] = 0;
6088 if (op == MDB_SET_RANGE && !exactp) {
6092 return MDB_NOTFOUND;
6098 rc = mdb_page_search(mc, key, 0);
6099 if (rc != MDB_SUCCESS)
6102 mp = mc->mc_pg[mc->mc_top];
6103 mdb_cassert(mc, IS_LEAF(mp));
6106 leaf = mdb_node_search(mc, key, exactp);
6107 if (exactp != NULL && !*exactp) {
6108 /* MDB_SET specified and not an exact match. */
6109 return MDB_NOTFOUND;
6113 DPUTS("===> inexact leaf not found, goto sibling");
6114 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6115 mc->mc_flags |= C_EOF;
6116 return rc; /* no entries matched */
6118 mp = mc->mc_pg[mc->mc_top];
6119 mdb_cassert(mc, IS_LEAF(mp));
6120 leaf = NODEPTR(mp, 0);
6124 mc->mc_flags |= C_INITIALIZED;
6125 mc->mc_flags &= ~C_EOF;
6128 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6129 key->mv_size = mc->mc_db->md_pad;
6130 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6135 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6136 mdb_xcursor_init1(mc, leaf);
6139 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6140 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6141 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6144 if (op == MDB_GET_BOTH) {
6150 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6151 if (rc != MDB_SUCCESS)
6154 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6157 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6159 dcmp = mc->mc_dbx->md_dcmp;
6160 #if UINT_MAX < SIZE_MAX
6161 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6162 dcmp = mdb_cmp_clong;
6164 rc = dcmp(data, &olddata);
6166 if (op == MDB_GET_BOTH || rc > 0)
6167 return MDB_NOTFOUND;
6174 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6175 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6180 /* The key already matches in all other cases */
6181 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6182 MDB_GET_KEY(leaf, key);
6183 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6188 /** Move the cursor to the first item in the database. */
6190 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6196 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6198 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6199 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6200 if (rc != MDB_SUCCESS)
6203 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6205 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6206 mc->mc_flags |= C_INITIALIZED;
6207 mc->mc_flags &= ~C_EOF;
6209 mc->mc_ki[mc->mc_top] = 0;
6211 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6212 key->mv_size = mc->mc_db->md_pad;
6213 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6218 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6219 mdb_xcursor_init1(mc, leaf);
6220 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6224 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6228 MDB_GET_KEY(leaf, key);
6232 /** Move the cursor to the last item in the database. */
6234 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6240 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6242 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6243 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6244 if (rc != MDB_SUCCESS)
6247 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6249 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6250 mc->mc_flags |= C_INITIALIZED|C_EOF;
6251 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6253 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6254 key->mv_size = mc->mc_db->md_pad;
6255 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6260 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6261 mdb_xcursor_init1(mc, leaf);
6262 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6266 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6271 MDB_GET_KEY(leaf, key);
6276 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6281 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6286 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6290 case MDB_GET_CURRENT:
6291 if (!(mc->mc_flags & C_INITIALIZED)) {
6294 MDB_page *mp = mc->mc_pg[mc->mc_top];
6295 int nkeys = NUMKEYS(mp);
6296 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6297 mc->mc_ki[mc->mc_top] = nkeys;
6303 key->mv_size = mc->mc_db->md_pad;
6304 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6306 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6307 MDB_GET_KEY(leaf, key);
6309 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6310 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6312 rc = mdb_node_read(mc, leaf, data);
6319 case MDB_GET_BOTH_RANGE:
6324 if (mc->mc_xcursor == NULL) {
6325 rc = MDB_INCOMPATIBLE;
6335 rc = mdb_cursor_set(mc, key, data, op,
6336 op == MDB_SET_RANGE ? NULL : &exact);
6339 case MDB_GET_MULTIPLE:
6340 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6344 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6345 rc = MDB_INCOMPATIBLE;
6349 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6350 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6353 case MDB_NEXT_MULTIPLE:
6358 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6359 rc = MDB_INCOMPATIBLE;
6362 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6363 if (rc == MDB_SUCCESS) {
6364 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6367 mx = &mc->mc_xcursor->mx_cursor;
6368 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6370 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6371 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6377 case MDB_PREV_MULTIPLE:
6382 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6383 rc = MDB_INCOMPATIBLE;
6386 if (!(mc->mc_flags & C_INITIALIZED))
6387 rc = mdb_cursor_last(mc, key, data);
6390 if (rc == MDB_SUCCESS) {
6391 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
6392 if (mx->mc_flags & C_INITIALIZED) {
6393 rc = mdb_cursor_sibling(mx, 0);
6394 if (rc == MDB_SUCCESS)
6403 case MDB_NEXT_NODUP:
6404 rc = mdb_cursor_next(mc, key, data, op);
6408 case MDB_PREV_NODUP:
6409 rc = mdb_cursor_prev(mc, key, data, op);
6412 rc = mdb_cursor_first(mc, key, data);
6415 mfunc = mdb_cursor_first;
6417 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6421 if (mc->mc_xcursor == NULL) {
6422 rc = MDB_INCOMPATIBLE;
6426 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6427 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6428 MDB_GET_KEY(leaf, key);
6429 rc = mdb_node_read(mc, leaf, data);
6433 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6437 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6440 rc = mdb_cursor_last(mc, key, data);
6443 mfunc = mdb_cursor_last;
6446 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6451 if (mc->mc_flags & C_DEL)
6452 mc->mc_flags ^= C_DEL;
6457 /** Touch all the pages in the cursor stack. Set mc_top.
6458 * Makes sure all the pages are writable, before attempting a write operation.
6459 * @param[in] mc The cursor to operate on.
6462 mdb_cursor_touch(MDB_cursor *mc)
6464 int rc = MDB_SUCCESS;
6466 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) {
6467 /* Touch DB record of named DB */
6470 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6472 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6473 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6476 *mc->mc_dbflag |= DB_DIRTY;
6481 rc = mdb_page_touch(mc);
6482 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6483 mc->mc_top = mc->mc_snum-1;
6488 /** Do not spill pages to disk if txn is getting full, may fail instead */
6489 #define MDB_NOSPILL 0x8000
6492 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6496 MDB_node *leaf = NULL;
6497 MDB_page *fp, *mp, *sub_root = NULL;
6499 MDB_val xdata, *rdata, dkey, olddata;
6501 int do_sub = 0, insert_key, insert_data;
6502 unsigned int mcount = 0, dcount = 0, nospill;
6505 unsigned int nflags;
6508 if (mc == NULL || key == NULL)
6511 env = mc->mc_txn->mt_env;
6513 /* Check this first so counter will always be zero on any
6516 if (flags & MDB_MULTIPLE) {
6517 dcount = data[1].mv_size;
6518 data[1].mv_size = 0;
6519 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6520 return MDB_INCOMPATIBLE;
6523 nospill = flags & MDB_NOSPILL;
6524 flags &= ~MDB_NOSPILL;
6526 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
6527 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6529 if (key->mv_size-1 >= ENV_MAXKEY(env))
6530 return MDB_BAD_VALSIZE;
6532 #if SIZE_MAX > MAXDATASIZE
6533 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6534 return MDB_BAD_VALSIZE;
6536 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6537 return MDB_BAD_VALSIZE;
6540 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6541 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6545 if (flags == MDB_CURRENT) {
6546 if (!(mc->mc_flags & C_INITIALIZED))
6549 } else if (mc->mc_db->md_root == P_INVALID) {
6550 /* new database, cursor has nothing to point to */
6553 mc->mc_flags &= ~C_INITIALIZED;
6558 if (flags & MDB_APPEND) {
6560 rc = mdb_cursor_last(mc, &k2, &d2);
6562 rc = mc->mc_dbx->md_cmp(key, &k2);
6565 mc->mc_ki[mc->mc_top]++;
6567 /* new key is <= last key */
6572 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6574 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6575 DPRINTF(("duplicate key [%s]", DKEY(key)));
6577 return MDB_KEYEXIST;
6579 if (rc && rc != MDB_NOTFOUND)
6583 if (mc->mc_flags & C_DEL)
6584 mc->mc_flags ^= C_DEL;
6586 /* Cursor is positioned, check for room in the dirty list */
6588 if (flags & MDB_MULTIPLE) {
6590 xdata.mv_size = data->mv_size * dcount;
6594 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6598 if (rc == MDB_NO_ROOT) {
6600 /* new database, write a root leaf page */
6601 DPUTS("allocating new root leaf page");
6602 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6605 mdb_cursor_push(mc, np);
6606 mc->mc_db->md_root = np->mp_pgno;
6607 mc->mc_db->md_depth++;
6608 *mc->mc_dbflag |= DB_DIRTY;
6609 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6611 np->mp_flags |= P_LEAF2;
6612 mc->mc_flags |= C_INITIALIZED;
6614 /* make sure all cursor pages are writable */
6615 rc2 = mdb_cursor_touch(mc);
6620 insert_key = insert_data = rc;
6622 /* The key does not exist */
6623 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6624 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6625 LEAFSIZE(key, data) > env->me_nodemax)
6627 /* Too big for a node, insert in sub-DB. Set up an empty
6628 * "old sub-page" for prep_subDB to expand to a full page.
6630 fp_flags = P_LEAF|P_DIRTY;
6632 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6633 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6634 olddata.mv_size = PAGEHDRSZ;
6638 /* there's only a key anyway, so this is a no-op */
6639 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6641 unsigned int ksize = mc->mc_db->md_pad;
6642 if (key->mv_size != ksize)
6643 return MDB_BAD_VALSIZE;
6644 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6645 memcpy(ptr, key->mv_data, ksize);
6647 /* if overwriting slot 0 of leaf, need to
6648 * update branch key if there is a parent page
6650 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6651 unsigned short dtop = 1;
6653 /* slot 0 is always an empty key, find real slot */
6654 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6658 if (mc->mc_ki[mc->mc_top])
6659 rc2 = mdb_update_key(mc, key);
6670 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6671 olddata.mv_size = NODEDSZ(leaf);
6672 olddata.mv_data = NODEDATA(leaf);
6675 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6676 /* Prepare (sub-)page/sub-DB to accept the new item,
6677 * if needed. fp: old sub-page or a header faking
6678 * it. mp: new (sub-)page. offset: growth in page
6679 * size. xdata: node data with new page or DB.
6681 unsigned i, offset = 0;
6682 mp = fp = xdata.mv_data = env->me_pbuf;
6683 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6685 /* Was a single item before, must convert now */
6686 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6688 /* Just overwrite the current item */
6689 if (flags == MDB_CURRENT)
6691 dcmp = mc->mc_dbx->md_dcmp;
6692 #if UINT_MAX < SIZE_MAX
6693 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6694 dcmp = mdb_cmp_clong;
6696 /* does data match? */
6697 if (!dcmp(data, &olddata)) {
6698 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
6699 return MDB_KEYEXIST;
6704 /* Back up original data item */
6705 dkey.mv_size = olddata.mv_size;
6706 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6708 /* Make sub-page header for the dup items, with dummy body */
6709 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6710 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6711 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6712 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6713 fp->mp_flags |= P_LEAF2;
6714 fp->mp_pad = data->mv_size;
6715 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6717 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6718 (dkey.mv_size & 1) + (data->mv_size & 1);
6720 fp->mp_upper = xdata.mv_size - PAGEBASE;
6721 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6722 } else if (leaf->mn_flags & F_SUBDATA) {
6723 /* Data is on sub-DB, just store it */
6724 flags |= F_DUPDATA|F_SUBDATA;
6727 /* Data is on sub-page */
6728 fp = olddata.mv_data;
6731 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6732 offset = EVEN(NODESIZE + sizeof(indx_t) +
6736 offset = fp->mp_pad;
6737 if (SIZELEFT(fp) < offset) {
6738 offset *= 4; /* space for 4 more */
6741 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6743 fp->mp_flags |= P_DIRTY;
6744 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6745 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6749 xdata.mv_size = olddata.mv_size + offset;
6752 fp_flags = fp->mp_flags;
6753 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6754 /* Too big for a sub-page, convert to sub-DB */
6755 fp_flags &= ~P_SUBP;
6757 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6758 fp_flags |= P_LEAF2;
6759 dummy.md_pad = fp->mp_pad;
6760 dummy.md_flags = MDB_DUPFIXED;
6761 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6762 dummy.md_flags |= MDB_INTEGERKEY;
6768 dummy.md_branch_pages = 0;
6769 dummy.md_leaf_pages = 1;
6770 dummy.md_overflow_pages = 0;
6771 dummy.md_entries = NUMKEYS(fp);
6772 xdata.mv_size = sizeof(MDB_db);
6773 xdata.mv_data = &dummy;
6774 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6776 offset = env->me_psize - olddata.mv_size;
6777 flags |= F_DUPDATA|F_SUBDATA;
6778 dummy.md_root = mp->mp_pgno;
6782 mp->mp_flags = fp_flags | P_DIRTY;
6783 mp->mp_pad = fp->mp_pad;
6784 mp->mp_lower = fp->mp_lower;
6785 mp->mp_upper = fp->mp_upper + offset;
6786 if (fp_flags & P_LEAF2) {
6787 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6789 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6790 olddata.mv_size - fp->mp_upper - PAGEBASE);
6791 for (i=0; i<NUMKEYS(fp); i++)
6792 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6800 mdb_node_del(mc, 0);
6804 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6805 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6806 return MDB_INCOMPATIBLE;
6807 /* overflow page overwrites need special handling */
6808 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6811 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6813 memcpy(&pg, olddata.mv_data, sizeof(pg));
6814 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
6816 ovpages = omp->mp_pages;
6818 /* Is the ov page large enough? */
6819 if (ovpages >= dpages) {
6820 if (!(omp->mp_flags & P_DIRTY) &&
6821 (level || (env->me_flags & MDB_WRITEMAP)))
6823 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6826 level = 0; /* dirty in this txn or clean */
6829 if (omp->mp_flags & P_DIRTY) {
6830 /* yes, overwrite it. Note in this case we don't
6831 * bother to try shrinking the page if the new data
6832 * is smaller than the overflow threshold.
6835 /* It is writable only in a parent txn */
6836 size_t sz = (size_t) env->me_psize * ovpages, off;
6837 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6843 /* Note - this page is already counted in parent's dirty_room */
6844 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6845 mdb_cassert(mc, rc2 == 0);
6846 /* Currently we make the page look as with put() in the
6847 * parent txn, in case the user peeks at MDB_RESERVEd
6848 * or unused parts. Some users treat ovpages specially.
6850 if (!(flags & MDB_RESERVE)) {
6851 /* Skip the part where LMDB will put *data.
6852 * Copy end of page, adjusting alignment so
6853 * compiler may copy words instead of bytes.
6855 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6856 memcpy((size_t *)((char *)np + off),
6857 (size_t *)((char *)omp + off), sz - off);
6860 memcpy(np, omp, sz); /* Copy beginning of page */
6863 SETDSZ(leaf, data->mv_size);
6864 if (F_ISSET(flags, MDB_RESERVE))
6865 data->mv_data = METADATA(omp);
6867 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6871 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6873 } else if (data->mv_size == olddata.mv_size) {
6874 /* same size, just replace it. Note that we could
6875 * also reuse this node if the new data is smaller,
6876 * but instead we opt to shrink the node in that case.
6878 if (F_ISSET(flags, MDB_RESERVE))
6879 data->mv_data = olddata.mv_data;
6880 else if (!(mc->mc_flags & C_SUB))
6881 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6883 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6888 mdb_node_del(mc, 0);
6894 nflags = flags & NODE_ADD_FLAGS;
6895 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6896 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6897 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6898 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6900 nflags |= MDB_SPLIT_REPLACE;
6901 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6903 /* There is room already in this leaf page. */
6904 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6906 /* Adjust other cursors pointing to mp */
6907 MDB_cursor *m2, *m3;
6908 MDB_dbi dbi = mc->mc_dbi;
6909 unsigned i = mc->mc_top;
6910 MDB_page *mp = mc->mc_pg[i];
6912 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6913 if (mc->mc_flags & C_SUB)
6914 m3 = &m2->mc_xcursor->mx_cursor;
6917 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
6918 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
6921 if (XCURSOR_INITED(m3))
6922 XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]);
6927 if (rc == MDB_SUCCESS) {
6928 /* Now store the actual data in the child DB. Note that we're
6929 * storing the user data in the keys field, so there are strict
6930 * size limits on dupdata. The actual data fields of the child
6931 * DB are all zero size.
6934 int xflags, new_dupdata;
6939 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6940 if (flags & MDB_CURRENT) {
6941 xflags = MDB_CURRENT|MDB_NOSPILL;
6943 mdb_xcursor_init1(mc, leaf);
6944 xflags = (flags & MDB_NODUPDATA) ?
6945 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6948 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6949 new_dupdata = (int)dkey.mv_size;
6950 /* converted, write the original data first */
6952 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6955 /* we've done our job */
6958 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6959 /* Adjust other cursors pointing to mp */
6961 MDB_xcursor *mx = mc->mc_xcursor;
6962 unsigned i = mc->mc_top;
6963 MDB_page *mp = mc->mc_pg[i];
6964 int nkeys = NUMKEYS(mp);
6966 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6967 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6968 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6969 if (m2->mc_pg[i] == mp) {
6970 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6971 mdb_xcursor_init2(m2, mx, new_dupdata);
6972 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
6973 XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]);
6978 ecount = mc->mc_xcursor->mx_db.md_entries;
6979 if (flags & MDB_APPENDDUP)
6980 xflags |= MDB_APPEND;
6981 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6982 if (flags & F_SUBDATA) {
6983 void *db = NODEDATA(leaf);
6984 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6986 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6988 /* Increment count unless we just replaced an existing item. */
6990 mc->mc_db->md_entries++;
6992 /* Invalidate txn if we created an empty sub-DB */
6995 /* If we succeeded and the key didn't exist before,
6996 * make sure the cursor is marked valid.
6998 mc->mc_flags |= C_INITIALIZED;
7000 if (flags & MDB_MULTIPLE) {
7003 /* let caller know how many succeeded, if any */
7004 data[1].mv_size = mcount;
7005 if (mcount < dcount) {
7006 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
7007 insert_key = insert_data = 0;
7014 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7017 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7022 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7028 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7029 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7031 if (!(mc->mc_flags & C_INITIALIZED))
7034 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7035 return MDB_NOTFOUND;
7037 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7040 rc = mdb_cursor_touch(mc);
7044 mp = mc->mc_pg[mc->mc_top];
7047 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7049 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7050 if (flags & MDB_NODUPDATA) {
7051 /* mdb_cursor_del0() will subtract the final entry */
7052 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7053 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7055 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7056 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7058 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7061 /* If sub-DB still has entries, we're done */
7062 if (mc->mc_xcursor->mx_db.md_entries) {
7063 if (leaf->mn_flags & F_SUBDATA) {
7064 /* update subDB info */
7065 void *db = NODEDATA(leaf);
7066 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7069 /* shrink fake page */
7070 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7071 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7072 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7073 /* fix other sub-DB cursors pointed at fake pages on this page */
7074 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7075 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7076 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7077 if (m2->mc_pg[mc->mc_top] == mp) {
7078 MDB_node *n2 = leaf;
7079 if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) {
7080 n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7081 if (n2->mn_flags & F_SUBDATA) continue;
7083 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7087 mc->mc_db->md_entries--;
7090 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7092 /* otherwise fall thru and delete the sub-DB */
7095 if (leaf->mn_flags & F_SUBDATA) {
7096 /* add all the child DB's pages to the free list */
7097 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7102 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7103 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7104 rc = MDB_INCOMPATIBLE;
7108 /* add overflow pages to free list */
7109 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7113 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7114 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7115 (rc = mdb_ovpage_free(mc, omp)))
7120 return mdb_cursor_del0(mc);
7123 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7127 /** Allocate and initialize new pages for a database.
7128 * Set #MDB_TXN_ERROR on failure.
7129 * @param[in] mc a cursor on the database being added to.
7130 * @param[in] flags flags defining what type of page is being allocated.
7131 * @param[in] num the number of pages to allocate. This is usually 1,
7132 * unless allocating overflow pages for a large record.
7133 * @param[out] mp Address of a page, or NULL on failure.
7134 * @return 0 on success, non-zero on failure.
7137 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7142 if ((rc = mdb_page_alloc(mc, num, &np)))
7144 DPRINTF(("allocated new mpage %"Z"u, page size %u",
7145 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7146 np->mp_flags = flags | P_DIRTY;
7147 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7148 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7151 mc->mc_db->md_branch_pages++;
7152 else if (IS_LEAF(np))
7153 mc->mc_db->md_leaf_pages++;
7154 else if (IS_OVERFLOW(np)) {
7155 mc->mc_db->md_overflow_pages += num;
7163 /** Calculate the size of a leaf node.
7164 * The size depends on the environment's page size; if a data item
7165 * is too large it will be put onto an overflow page and the node
7166 * size will only include the key and not the data. Sizes are always
7167 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7168 * of the #MDB_node headers.
7169 * @param[in] env The environment handle.
7170 * @param[in] key The key for the node.
7171 * @param[in] data The data for the node.
7172 * @return The number of bytes needed to store the node.
7175 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7179 sz = LEAFSIZE(key, data);
7180 if (sz > env->me_nodemax) {
7181 /* put on overflow page */
7182 sz -= data->mv_size - sizeof(pgno_t);
7185 return EVEN(sz + sizeof(indx_t));
7188 /** Calculate the size of a branch node.
7189 * The size should depend on the environment's page size but since
7190 * we currently don't support spilling large keys onto overflow
7191 * pages, it's simply the size of the #MDB_node header plus the
7192 * size of the key. Sizes are always rounded up to an even number
7193 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7194 * @param[in] env The environment handle.
7195 * @param[in] key The key for the node.
7196 * @return The number of bytes needed to store the node.
7199 mdb_branch_size(MDB_env *env, MDB_val *key)
7204 if (sz > env->me_nodemax) {
7205 /* put on overflow page */
7206 /* not implemented */
7207 /* sz -= key->size - sizeof(pgno_t); */
7210 return sz + sizeof(indx_t);
7213 /** Add a node to the page pointed to by the cursor.
7214 * Set #MDB_TXN_ERROR on failure.
7215 * @param[in] mc The cursor for this operation.
7216 * @param[in] indx The index on the page where the new node should be added.
7217 * @param[in] key The key for the new node.
7218 * @param[in] data The data for the new node, if any.
7219 * @param[in] pgno The page number, if adding a branch node.
7220 * @param[in] flags Flags for the node.
7221 * @return 0 on success, non-zero on failure. Possible errors are:
7223 * <li>ENOMEM - failed to allocate overflow pages for the node.
7224 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7225 * should never happen since all callers already calculate the
7226 * page's free space before calling this function.
7230 mdb_node_add(MDB_cursor *mc, indx_t indx,
7231 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7234 size_t node_size = NODESIZE;
7238 MDB_page *mp = mc->mc_pg[mc->mc_top];
7239 MDB_page *ofp = NULL; /* overflow page */
7243 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7245 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
7246 IS_LEAF(mp) ? "leaf" : "branch",
7247 IS_SUBP(mp) ? "sub-" : "",
7248 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7249 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7252 /* Move higher keys up one slot. */
7253 int ksize = mc->mc_db->md_pad, dif;
7254 char *ptr = LEAF2KEY(mp, indx, ksize);
7255 dif = NUMKEYS(mp) - indx;
7257 memmove(ptr+ksize, ptr, dif*ksize);
7258 /* insert new key */
7259 memcpy(ptr, key->mv_data, ksize);
7261 /* Just using these for counting */
7262 mp->mp_lower += sizeof(indx_t);
7263 mp->mp_upper -= ksize - sizeof(indx_t);
7267 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7269 node_size += key->mv_size;
7271 mdb_cassert(mc, key && data);
7272 if (F_ISSET(flags, F_BIGDATA)) {
7273 /* Data already on overflow page. */
7274 node_size += sizeof(pgno_t);
7275 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7276 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7278 /* Put data on overflow page. */
7279 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7280 data->mv_size, node_size+data->mv_size));
7281 node_size = EVEN(node_size + sizeof(pgno_t));
7282 if ((ssize_t)node_size > room)
7284 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7286 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
7290 node_size += data->mv_size;
7293 node_size = EVEN(node_size);
7294 if ((ssize_t)node_size > room)
7298 /* Move higher pointers up one slot. */
7299 for (i = NUMKEYS(mp); i > indx; i--)
7300 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7302 /* Adjust free space offsets. */
7303 ofs = mp->mp_upper - node_size;
7304 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7305 mp->mp_ptrs[indx] = ofs;
7307 mp->mp_lower += sizeof(indx_t);
7309 /* Write the node data. */
7310 node = NODEPTR(mp, indx);
7311 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7312 node->mn_flags = flags;
7314 SETDSZ(node,data->mv_size);
7319 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7322 ndata = NODEDATA(node);
7324 if (F_ISSET(flags, F_BIGDATA))
7325 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7326 else if (F_ISSET(flags, MDB_RESERVE))
7327 data->mv_data = ndata;
7329 memcpy(ndata, data->mv_data, data->mv_size);
7331 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7332 ndata = METADATA(ofp);
7333 if (F_ISSET(flags, MDB_RESERVE))
7334 data->mv_data = ndata;
7336 memcpy(ndata, data->mv_data, data->mv_size);
7343 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7344 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7345 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7346 DPRINTF(("node size = %"Z"u", node_size));
7347 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7348 return MDB_PAGE_FULL;
7351 /** Delete the specified node from a page.
7352 * @param[in] mc Cursor pointing to the node to delete.
7353 * @param[in] ksize The size of a node. Only used if the page is
7354 * part of a #MDB_DUPFIXED database.
7357 mdb_node_del(MDB_cursor *mc, int ksize)
7359 MDB_page *mp = mc->mc_pg[mc->mc_top];
7360 indx_t indx = mc->mc_ki[mc->mc_top];
7362 indx_t i, j, numkeys, ptr;
7366 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7367 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7368 numkeys = NUMKEYS(mp);
7369 mdb_cassert(mc, indx < numkeys);
7372 int x = numkeys - 1 - indx;
7373 base = LEAF2KEY(mp, indx, ksize);
7375 memmove(base, base + ksize, x * ksize);
7376 mp->mp_lower -= sizeof(indx_t);
7377 mp->mp_upper += ksize - sizeof(indx_t);
7381 node = NODEPTR(mp, indx);
7382 sz = NODESIZE + node->mn_ksize;
7384 if (F_ISSET(node->mn_flags, F_BIGDATA))
7385 sz += sizeof(pgno_t);
7387 sz += NODEDSZ(node);
7391 ptr = mp->mp_ptrs[indx];
7392 for (i = j = 0; i < numkeys; i++) {
7394 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7395 if (mp->mp_ptrs[i] < ptr)
7396 mp->mp_ptrs[j] += sz;
7401 base = (char *)mp + mp->mp_upper + PAGEBASE;
7402 memmove(base + sz, base, ptr - mp->mp_upper);
7404 mp->mp_lower -= sizeof(indx_t);
7408 /** Compact the main page after deleting a node on a subpage.
7409 * @param[in] mp The main page to operate on.
7410 * @param[in] indx The index of the subpage on the main page.
7413 mdb_node_shrink(MDB_page *mp, indx_t indx)
7418 indx_t delta, nsize, len, ptr;
7421 node = NODEPTR(mp, indx);
7422 sp = (MDB_page *)NODEDATA(node);
7423 delta = SIZELEFT(sp);
7424 nsize = NODEDSZ(node) - delta;
7426 /* Prepare to shift upward, set len = length(subpage part to shift) */
7430 return; /* do not make the node uneven-sized */
7432 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
7433 for (i = NUMKEYS(sp); --i >= 0; )
7434 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7437 sp->mp_upper = sp->mp_lower;
7438 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
7439 SETDSZ(node, nsize);
7441 /* Shift <lower nodes...initial part of subpage> upward */
7442 base = (char *)mp + mp->mp_upper + PAGEBASE;
7443 memmove(base + delta, base, (char *)sp + len - base);
7445 ptr = mp->mp_ptrs[indx];
7446 for (i = NUMKEYS(mp); --i >= 0; ) {
7447 if (mp->mp_ptrs[i] <= ptr)
7448 mp->mp_ptrs[i] += delta;
7450 mp->mp_upper += delta;
7453 /** Initial setup of a sorted-dups cursor.
7454 * Sorted duplicates are implemented as a sub-database for the given key.
7455 * The duplicate data items are actually keys of the sub-database.
7456 * Operations on the duplicate data items are performed using a sub-cursor
7457 * initialized when the sub-database is first accessed. This function does
7458 * the preliminary setup of the sub-cursor, filling in the fields that
7459 * depend only on the parent DB.
7460 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7463 mdb_xcursor_init0(MDB_cursor *mc)
7465 MDB_xcursor *mx = mc->mc_xcursor;
7467 mx->mx_cursor.mc_xcursor = NULL;
7468 mx->mx_cursor.mc_txn = mc->mc_txn;
7469 mx->mx_cursor.mc_db = &mx->mx_db;
7470 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7471 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7472 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7473 mx->mx_cursor.mc_snum = 0;
7474 mx->mx_cursor.mc_top = 0;
7475 mx->mx_cursor.mc_flags = C_SUB;
7476 mx->mx_dbx.md_name.mv_size = 0;
7477 mx->mx_dbx.md_name.mv_data = NULL;
7478 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7479 mx->mx_dbx.md_dcmp = NULL;
7480 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7483 /** Final setup of a sorted-dups cursor.
7484 * Sets up the fields that depend on the data from the main cursor.
7485 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7486 * @param[in] node The data containing the #MDB_db record for the
7487 * sorted-dup database.
7490 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7492 MDB_xcursor *mx = mc->mc_xcursor;
7494 if (node->mn_flags & F_SUBDATA) {
7495 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7496 mx->mx_cursor.mc_pg[0] = 0;
7497 mx->mx_cursor.mc_snum = 0;
7498 mx->mx_cursor.mc_top = 0;
7499 mx->mx_cursor.mc_flags = C_SUB;
7501 MDB_page *fp = NODEDATA(node);
7502 mx->mx_db.md_pad = 0;
7503 mx->mx_db.md_flags = 0;
7504 mx->mx_db.md_depth = 1;
7505 mx->mx_db.md_branch_pages = 0;
7506 mx->mx_db.md_leaf_pages = 1;
7507 mx->mx_db.md_overflow_pages = 0;
7508 mx->mx_db.md_entries = NUMKEYS(fp);
7509 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7510 mx->mx_cursor.mc_snum = 1;
7511 mx->mx_cursor.mc_top = 0;
7512 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7513 mx->mx_cursor.mc_pg[0] = fp;
7514 mx->mx_cursor.mc_ki[0] = 0;
7515 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7516 mx->mx_db.md_flags = MDB_DUPFIXED;
7517 mx->mx_db.md_pad = fp->mp_pad;
7518 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7519 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7522 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7523 mx->mx_db.md_root));
7524 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7525 #if UINT_MAX < SIZE_MAX
7526 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7527 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7532 /** Fixup a sorted-dups cursor due to underlying update.
7533 * Sets up some fields that depend on the data from the main cursor.
7534 * Almost the same as init1, but skips initialization steps if the
7535 * xcursor had already been used.
7536 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7537 * @param[in] src_mx The xcursor of an up-to-date cursor.
7538 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7541 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7543 MDB_xcursor *mx = mc->mc_xcursor;
7546 mx->mx_cursor.mc_snum = 1;
7547 mx->mx_cursor.mc_top = 0;
7548 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7549 mx->mx_cursor.mc_ki[0] = 0;
7550 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA;
7551 #if UINT_MAX < SIZE_MAX
7552 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7554 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7557 mx->mx_db = src_mx->mx_db;
7558 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7559 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7560 mx->mx_db.md_root));
7563 /** Initialize a cursor for a given transaction and database. */
7565 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7568 mc->mc_backup = NULL;
7571 mc->mc_db = &txn->mt_dbs[dbi];
7572 mc->mc_dbx = &txn->mt_dbxs[dbi];
7573 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7579 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7580 mdb_tassert(txn, mx != NULL);
7581 mc->mc_xcursor = mx;
7582 mdb_xcursor_init0(mc);
7584 mc->mc_xcursor = NULL;
7586 if (*mc->mc_dbflag & DB_STALE) {
7587 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7592 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7595 size_t size = sizeof(MDB_cursor);
7597 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
7600 if (txn->mt_flags & MDB_TXN_BLOCKED)
7603 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7606 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7607 size += sizeof(MDB_xcursor);
7609 if ((mc = malloc(size)) != NULL) {
7610 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7611 if (txn->mt_cursors) {
7612 mc->mc_next = txn->mt_cursors[dbi];
7613 txn->mt_cursors[dbi] = mc;
7614 mc->mc_flags |= C_UNTRACK;
7626 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7628 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
7631 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7634 if (txn->mt_flags & MDB_TXN_BLOCKED)
7637 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7641 /* Return the count of duplicate data items for the current key */
7643 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7647 if (mc == NULL || countp == NULL)
7650 if (mc->mc_xcursor == NULL)
7651 return MDB_INCOMPATIBLE;
7653 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
7656 if (!(mc->mc_flags & C_INITIALIZED))
7660 return MDB_NOTFOUND;
7662 if (mc->mc_flags & C_EOF) {
7663 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7664 return MDB_NOTFOUND;
7665 mc->mc_flags ^= C_EOF;
7668 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7669 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7672 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7675 *countp = mc->mc_xcursor->mx_db.md_entries;
7681 mdb_cursor_close(MDB_cursor *mc)
7683 if (mc && !mc->mc_backup) {
7684 /* remove from txn, if tracked */
7685 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7686 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7687 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7689 *prev = mc->mc_next;
7696 mdb_cursor_txn(MDB_cursor *mc)
7698 if (!mc) return NULL;
7703 mdb_cursor_dbi(MDB_cursor *mc)
7708 /** Replace the key for a branch node with a new key.
7709 * Set #MDB_TXN_ERROR on failure.
7710 * @param[in] mc Cursor pointing to the node to operate on.
7711 * @param[in] key The new key to use.
7712 * @return 0 on success, non-zero on failure.
7715 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7721 int delta, ksize, oksize;
7722 indx_t ptr, i, numkeys, indx;
7725 indx = mc->mc_ki[mc->mc_top];
7726 mp = mc->mc_pg[mc->mc_top];
7727 node = NODEPTR(mp, indx);
7728 ptr = mp->mp_ptrs[indx];
7732 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7733 k2.mv_data = NODEKEY(node);
7734 k2.mv_size = node->mn_ksize;
7735 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7737 mdb_dkey(&k2, kbuf2),
7743 /* Sizes must be 2-byte aligned. */
7744 ksize = EVEN(key->mv_size);
7745 oksize = EVEN(node->mn_ksize);
7746 delta = ksize - oksize;
7748 /* Shift node contents if EVEN(key length) changed. */
7750 if (delta > 0 && SIZELEFT(mp) < delta) {
7752 /* not enough space left, do a delete and split */
7753 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7754 pgno = NODEPGNO(node);
7755 mdb_node_del(mc, 0);
7756 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7759 numkeys = NUMKEYS(mp);
7760 for (i = 0; i < numkeys; i++) {
7761 if (mp->mp_ptrs[i] <= ptr)
7762 mp->mp_ptrs[i] -= delta;
7765 base = (char *)mp + mp->mp_upper + PAGEBASE;
7766 len = ptr - mp->mp_upper + NODESIZE;
7767 memmove(base - delta, base, len);
7768 mp->mp_upper -= delta;
7770 node = NODEPTR(mp, indx);
7773 /* But even if no shift was needed, update ksize */
7774 if (node->mn_ksize != key->mv_size)
7775 node->mn_ksize = key->mv_size;
7778 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7784 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7786 /** Perform \b act while tracking temporary cursor \b mn */
7787 #define WITH_CURSOR_TRACKING(mn, act) do { \
7788 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
7789 if ((mn).mc_flags & C_SUB) { \
7790 dummy.mc_flags = C_INITIALIZED; \
7791 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
7796 tracked->mc_next = *tp; \
7799 *tp = tracked->mc_next; \
7802 /** Move a node from csrc to cdst.
7805 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
7812 unsigned short flags;
7816 /* Mark src and dst as dirty. */
7817 if ((rc = mdb_page_touch(csrc)) ||
7818 (rc = mdb_page_touch(cdst)))
7821 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7822 key.mv_size = csrc->mc_db->md_pad;
7823 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7825 data.mv_data = NULL;
7829 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7830 mdb_cassert(csrc, !((size_t)srcnode & 1));
7831 srcpg = NODEPGNO(srcnode);
7832 flags = srcnode->mn_flags;
7833 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7834 unsigned int snum = csrc->mc_snum;
7836 /* must find the lowest key below src */
7837 rc = mdb_page_search_lowest(csrc);
7840 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7841 key.mv_size = csrc->mc_db->md_pad;
7842 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7844 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7845 key.mv_size = NODEKSZ(s2);
7846 key.mv_data = NODEKEY(s2);
7848 csrc->mc_snum = snum--;
7849 csrc->mc_top = snum;
7851 key.mv_size = NODEKSZ(srcnode);
7852 key.mv_data = NODEKEY(srcnode);
7854 data.mv_size = NODEDSZ(srcnode);
7855 data.mv_data = NODEDATA(srcnode);
7857 mn.mc_xcursor = NULL;
7858 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7859 unsigned int snum = cdst->mc_snum;
7862 /* must find the lowest key below dst */
7863 mdb_cursor_copy(cdst, &mn);
7864 rc = mdb_page_search_lowest(&mn);
7867 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7868 bkey.mv_size = mn.mc_db->md_pad;
7869 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7871 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7872 bkey.mv_size = NODEKSZ(s2);
7873 bkey.mv_data = NODEKEY(s2);
7875 mn.mc_snum = snum--;
7878 rc = mdb_update_key(&mn, &bkey);
7883 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7884 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7885 csrc->mc_ki[csrc->mc_top],
7887 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7888 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7890 /* Add the node to the destination page.
7892 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7893 if (rc != MDB_SUCCESS)
7896 /* Delete the node from the source page.
7898 mdb_node_del(csrc, key.mv_size);
7901 /* Adjust other cursors pointing to mp */
7902 MDB_cursor *m2, *m3;
7903 MDB_dbi dbi = csrc->mc_dbi;
7904 MDB_page *mpd, *mps;
7906 mps = csrc->mc_pg[csrc->mc_top];
7907 /* If we're adding on the left, bump others up */
7909 mpd = cdst->mc_pg[csrc->mc_top];
7910 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7911 if (csrc->mc_flags & C_SUB)
7912 m3 = &m2->mc_xcursor->mx_cursor;
7915 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7918 m3->mc_pg[csrc->mc_top] == mpd &&
7919 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
7920 m3->mc_ki[csrc->mc_top]++;
7923 m3->mc_pg[csrc->mc_top] == mps &&
7924 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
7925 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7926 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7927 m3->mc_ki[csrc->mc_top-1]++;
7929 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7930 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7933 /* Adding on the right, bump others down */
7935 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7936 if (csrc->mc_flags & C_SUB)
7937 m3 = &m2->mc_xcursor->mx_cursor;
7940 if (m3 == csrc) continue;
7941 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
7943 if (m3->mc_pg[csrc->mc_top] == mps) {
7944 if (!m3->mc_ki[csrc->mc_top]) {
7945 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7946 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7947 m3->mc_ki[csrc->mc_top-1]--;
7949 m3->mc_ki[csrc->mc_top]--;
7951 if (XCURSOR_INITED(m3) && IS_LEAF(mps))
7952 XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
7958 /* Update the parent separators.
7960 if (csrc->mc_ki[csrc->mc_top] == 0) {
7961 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7962 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7963 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7965 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7966 key.mv_size = NODEKSZ(srcnode);
7967 key.mv_data = NODEKEY(srcnode);
7969 DPRINTF(("update separator for source page %"Z"u to [%s]",
7970 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7971 mdb_cursor_copy(csrc, &mn);
7974 /* We want mdb_rebalance to find mn when doing fixups */
7975 WITH_CURSOR_TRACKING(mn,
7976 rc = mdb_update_key(&mn, &key));
7980 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7982 indx_t ix = csrc->mc_ki[csrc->mc_top];
7983 nullkey.mv_size = 0;
7984 csrc->mc_ki[csrc->mc_top] = 0;
7985 rc = mdb_update_key(csrc, &nullkey);
7986 csrc->mc_ki[csrc->mc_top] = ix;
7987 mdb_cassert(csrc, rc == MDB_SUCCESS);
7991 if (cdst->mc_ki[cdst->mc_top] == 0) {
7992 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7993 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7994 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7996 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7997 key.mv_size = NODEKSZ(srcnode);
7998 key.mv_data = NODEKEY(srcnode);
8000 DPRINTF(("update separator for destination page %"Z"u to [%s]",
8001 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
8002 mdb_cursor_copy(cdst, &mn);
8005 /* We want mdb_rebalance to find mn when doing fixups */
8006 WITH_CURSOR_TRACKING(mn,
8007 rc = mdb_update_key(&mn, &key));
8011 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
8013 indx_t ix = cdst->mc_ki[cdst->mc_top];
8014 nullkey.mv_size = 0;
8015 cdst->mc_ki[cdst->mc_top] = 0;
8016 rc = mdb_update_key(cdst, &nullkey);
8017 cdst->mc_ki[cdst->mc_top] = ix;
8018 mdb_cassert(cdst, rc == MDB_SUCCESS);
8025 /** Merge one page into another.
8026 * The nodes from the page pointed to by \b csrc will
8027 * be copied to the page pointed to by \b cdst and then
8028 * the \b csrc page will be freed.
8029 * @param[in] csrc Cursor pointing to the source page.
8030 * @param[in] cdst Cursor pointing to the destination page.
8031 * @return 0 on success, non-zero on failure.
8034 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8036 MDB_page *psrc, *pdst;
8043 psrc = csrc->mc_pg[csrc->mc_top];
8044 pdst = cdst->mc_pg[cdst->mc_top];
8046 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
8048 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8049 mdb_cassert(csrc, cdst->mc_snum > 1);
8051 /* Mark dst as dirty. */
8052 if ((rc = mdb_page_touch(cdst)))
8055 /* get dst page again now that we've touched it. */
8056 pdst = cdst->mc_pg[cdst->mc_top];
8058 /* Move all nodes from src to dst.
8060 j = nkeys = NUMKEYS(pdst);
8061 if (IS_LEAF2(psrc)) {
8062 key.mv_size = csrc->mc_db->md_pad;
8063 key.mv_data = METADATA(psrc);
8064 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8065 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8066 if (rc != MDB_SUCCESS)
8068 key.mv_data = (char *)key.mv_data + key.mv_size;
8071 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8072 srcnode = NODEPTR(psrc, i);
8073 if (i == 0 && IS_BRANCH(psrc)) {
8076 mdb_cursor_copy(csrc, &mn);
8077 mn.mc_xcursor = NULL;
8078 /* must find the lowest key below src */
8079 rc = mdb_page_search_lowest(&mn);
8082 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8083 key.mv_size = mn.mc_db->md_pad;
8084 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8086 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8087 key.mv_size = NODEKSZ(s2);
8088 key.mv_data = NODEKEY(s2);
8091 key.mv_size = srcnode->mn_ksize;
8092 key.mv_data = NODEKEY(srcnode);
8095 data.mv_size = NODEDSZ(srcnode);
8096 data.mv_data = NODEDATA(srcnode);
8097 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8098 if (rc != MDB_SUCCESS)
8103 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
8104 pdst->mp_pgno, NUMKEYS(pdst),
8105 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8107 /* Unlink the src page from parent and add to free list.
8110 mdb_node_del(csrc, 0);
8111 if (csrc->mc_ki[csrc->mc_top] == 0) {
8113 rc = mdb_update_key(csrc, &key);
8121 psrc = csrc->mc_pg[csrc->mc_top];
8122 /* If not operating on FreeDB, allow this page to be reused
8123 * in this txn. Otherwise just add to free list.
8125 rc = mdb_page_loose(csrc, psrc);
8129 csrc->mc_db->md_leaf_pages--;
8131 csrc->mc_db->md_branch_pages--;
8133 /* Adjust other cursors pointing to mp */
8134 MDB_cursor *m2, *m3;
8135 MDB_dbi dbi = csrc->mc_dbi;
8136 unsigned int top = csrc->mc_top;
8138 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8139 if (csrc->mc_flags & C_SUB)
8140 m3 = &m2->mc_xcursor->mx_cursor;
8143 if (m3 == csrc) continue;
8144 if (m3->mc_snum < csrc->mc_snum) continue;
8145 if (m3->mc_pg[top] == psrc) {
8146 m3->mc_pg[top] = pdst;
8147 m3->mc_ki[top] += nkeys;
8148 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8149 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8150 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8153 if (XCURSOR_INITED(m3) && IS_LEAF(psrc))
8154 XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]);
8158 unsigned int snum = cdst->mc_snum;
8159 uint16_t depth = cdst->mc_db->md_depth;
8160 mdb_cursor_pop(cdst);
8161 rc = mdb_rebalance(cdst);
8162 /* Did the tree height change? */
8163 if (depth != cdst->mc_db->md_depth)
8164 snum += cdst->mc_db->md_depth - depth;
8165 cdst->mc_snum = snum;
8166 cdst->mc_top = snum-1;
8171 /** Copy the contents of a cursor.
8172 * @param[in] csrc The cursor to copy from.
8173 * @param[out] cdst The cursor to copy to.
8176 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8180 cdst->mc_txn = csrc->mc_txn;
8181 cdst->mc_dbi = csrc->mc_dbi;
8182 cdst->mc_db = csrc->mc_db;
8183 cdst->mc_dbx = csrc->mc_dbx;
8184 cdst->mc_snum = csrc->mc_snum;
8185 cdst->mc_top = csrc->mc_top;
8186 cdst->mc_flags = csrc->mc_flags;
8188 for (i=0; i<csrc->mc_snum; i++) {
8189 cdst->mc_pg[i] = csrc->mc_pg[i];
8190 cdst->mc_ki[i] = csrc->mc_ki[i];
8194 /** Rebalance the tree after a delete operation.
8195 * @param[in] mc Cursor pointing to the page where rebalancing
8197 * @return 0 on success, non-zero on failure.
8200 mdb_rebalance(MDB_cursor *mc)
8204 unsigned int ptop, minkeys, thresh;
8208 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8213 thresh = FILL_THRESHOLD;
8215 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
8216 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8217 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8218 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8220 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8221 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8222 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
8223 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8227 if (mc->mc_snum < 2) {
8228 MDB_page *mp = mc->mc_pg[0];
8230 DPUTS("Can't rebalance a subpage, ignoring");
8233 if (NUMKEYS(mp) == 0) {
8234 DPUTS("tree is completely empty");
8235 mc->mc_db->md_root = P_INVALID;
8236 mc->mc_db->md_depth = 0;
8237 mc->mc_db->md_leaf_pages = 0;
8238 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8241 /* Adjust cursors pointing to mp */
8244 mc->mc_flags &= ~C_INITIALIZED;
8246 MDB_cursor *m2, *m3;
8247 MDB_dbi dbi = mc->mc_dbi;
8249 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8250 if (mc->mc_flags & C_SUB)
8251 m3 = &m2->mc_xcursor->mx_cursor;
8254 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8256 if (m3->mc_pg[0] == mp) {
8259 m3->mc_flags &= ~C_INITIALIZED;
8263 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8265 DPUTS("collapsing root page!");
8266 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8269 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8270 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8273 mc->mc_db->md_depth--;
8274 mc->mc_db->md_branch_pages--;
8275 mc->mc_ki[0] = mc->mc_ki[1];
8276 for (i = 1; i<mc->mc_db->md_depth; i++) {
8277 mc->mc_pg[i] = mc->mc_pg[i+1];
8278 mc->mc_ki[i] = mc->mc_ki[i+1];
8281 /* Adjust other cursors pointing to mp */
8282 MDB_cursor *m2, *m3;
8283 MDB_dbi dbi = mc->mc_dbi;
8285 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8286 if (mc->mc_flags & C_SUB)
8287 m3 = &m2->mc_xcursor->mx_cursor;
8290 if (m3 == mc) continue;
8291 if (!(m3->mc_flags & C_INITIALIZED))
8293 if (m3->mc_pg[0] == mp) {
8294 for (i=0; i<mc->mc_db->md_depth; i++) {
8295 m3->mc_pg[i] = m3->mc_pg[i+1];
8296 m3->mc_ki[i] = m3->mc_ki[i+1];
8304 DPUTS("root page doesn't need rebalancing");
8308 /* The parent (branch page) must have at least 2 pointers,
8309 * otherwise the tree is invalid.
8311 ptop = mc->mc_top-1;
8312 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8314 /* Leaf page fill factor is below the threshold.
8315 * Try to move keys from left or right neighbor, or
8316 * merge with a neighbor page.
8321 mdb_cursor_copy(mc, &mn);
8322 mn.mc_xcursor = NULL;
8324 oldki = mc->mc_ki[mc->mc_top];
8325 if (mc->mc_ki[ptop] == 0) {
8326 /* We're the leftmost leaf in our parent.
8328 DPUTS("reading right neighbor");
8330 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8331 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8334 mn.mc_ki[mn.mc_top] = 0;
8335 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8338 /* There is at least one neighbor to the left.
8340 DPUTS("reading left neighbor");
8342 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8343 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8346 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8347 mc->mc_ki[mc->mc_top] = 0;
8351 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
8352 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8353 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
8355 /* If the neighbor page is above threshold and has enough keys,
8356 * move one key from it. Otherwise we should try to merge them.
8357 * (A branch page must never have less than 2 keys.)
8359 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
8360 rc = mdb_node_move(&mn, mc, fromleft);
8362 /* if we inserted on left, bump position up */
8367 rc = mdb_page_merge(&mn, mc);
8369 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
8370 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
8371 /* We want mdb_rebalance to find mn when doing fixups */
8372 WITH_CURSOR_TRACKING(mn,
8373 rc = mdb_page_merge(mc, &mn));
8374 mdb_cursor_copy(&mn, mc);
8376 mc->mc_flags &= ~C_EOF;
8378 mc->mc_ki[mc->mc_top] = oldki;
8382 /** Complete a delete operation started by #mdb_cursor_del(). */
8384 mdb_cursor_del0(MDB_cursor *mc)
8390 MDB_cursor *m2, *m3;
8391 MDB_dbi dbi = mc->mc_dbi;
8393 ki = mc->mc_ki[mc->mc_top];
8394 mp = mc->mc_pg[mc->mc_top];
8395 mdb_node_del(mc, mc->mc_db->md_pad);
8396 mc->mc_db->md_entries--;
8398 /* Adjust other cursors pointing to mp */
8399 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8400 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8401 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8403 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8405 if (m3->mc_pg[mc->mc_top] == mp) {
8406 if (m3->mc_ki[mc->mc_top] == ki) {
8407 m3->mc_flags |= C_DEL;
8408 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8409 /* Sub-cursor referred into dataset which is gone */
8410 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
8413 } else if (m3->mc_ki[mc->mc_top] > ki) {
8414 m3->mc_ki[mc->mc_top]--;
8416 if (XCURSOR_INITED(m3))
8417 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8421 rc = mdb_rebalance(mc);
8423 if (rc == MDB_SUCCESS) {
8424 /* DB is totally empty now, just bail out.
8425 * Other cursors adjustments were already done
8426 * by mdb_rebalance and aren't needed here.
8431 mp = mc->mc_pg[mc->mc_top];
8432 nkeys = NUMKEYS(mp);
8434 /* Adjust other cursors pointing to mp */
8435 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8436 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8437 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8439 if (m3->mc_snum < mc->mc_snum)
8441 if (m3->mc_pg[mc->mc_top] == mp) {
8442 /* if m3 points past last node in page, find next sibling */
8443 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
8444 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8445 rc = mdb_cursor_sibling(m3, 1);
8446 if (rc == MDB_NOTFOUND) {
8447 m3->mc_flags |= C_EOF;
8452 if (mc->mc_db->md_flags & MDB_DUPSORT) {
8453 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
8454 /* If this node has dupdata, it may need to be reinited
8455 * because its data has moved.
8456 * If the xcursor was not initd it must be reinited.
8457 * Else if node points to a subDB, nothing is needed.
8458 * Else (xcursor was initd, not a subDB) needs mc_pg[0] reset.
8460 if (node->mn_flags & F_DUPDATA) {
8461 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
8462 if (!(node->mn_flags & F_SUBDATA))
8463 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8465 mdb_xcursor_init1(m3, node);
8471 mc->mc_flags |= C_DEL;
8475 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8480 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8481 MDB_val *key, MDB_val *data)
8483 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8486 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8487 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8489 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8490 /* must ignore any data */
8494 return mdb_del0(txn, dbi, key, data, 0);
8498 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8499 MDB_val *key, MDB_val *data, unsigned flags)
8504 MDB_val rdata, *xdata;
8508 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8510 mdb_cursor_init(&mc, txn, dbi, &mx);
8519 flags |= MDB_NODUPDATA;
8521 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8523 /* let mdb_page_split know about this cursor if needed:
8524 * delete will trigger a rebalance; if it needs to move
8525 * a node from one page to another, it will have to
8526 * update the parent's separator key(s). If the new sepkey
8527 * is larger than the current one, the parent page may
8528 * run out of space, triggering a split. We need this
8529 * cursor to be consistent until the end of the rebalance.
8531 mc.mc_flags |= C_UNTRACK;
8532 mc.mc_next = txn->mt_cursors[dbi];
8533 txn->mt_cursors[dbi] = &mc;
8534 rc = mdb_cursor_del(&mc, flags);
8535 txn->mt_cursors[dbi] = mc.mc_next;
8540 /** Split a page and insert a new node.
8541 * Set #MDB_TXN_ERROR on failure.
8542 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8543 * The cursor will be updated to point to the actual page and index where
8544 * the node got inserted after the split.
8545 * @param[in] newkey The key for the newly inserted node.
8546 * @param[in] newdata The data for the newly inserted node.
8547 * @param[in] newpgno The page number, if the new node is a branch node.
8548 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8549 * @return 0 on success, non-zero on failure.
8552 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8553 unsigned int nflags)
8556 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8559 int i, j, split_indx, nkeys, pmax;
8560 MDB_env *env = mc->mc_txn->mt_env;
8562 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8563 MDB_page *copy = NULL;
8564 MDB_page *mp, *rp, *pp;
8569 mp = mc->mc_pg[mc->mc_top];
8570 newindx = mc->mc_ki[mc->mc_top];
8571 nkeys = NUMKEYS(mp);
8573 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8574 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8575 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8577 /* Create a right sibling. */
8578 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8580 rp->mp_pad = mp->mp_pad;
8581 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8583 /* Usually when splitting the root page, the cursor
8584 * height is 1. But when called from mdb_update_key,
8585 * the cursor height may be greater because it walks
8586 * up the stack while finding the branch slot to update.
8588 if (mc->mc_top < 1) {
8589 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8591 /* shift current top to make room for new parent */
8592 for (i=mc->mc_snum; i>0; i--) {
8593 mc->mc_pg[i] = mc->mc_pg[i-1];
8594 mc->mc_ki[i] = mc->mc_ki[i-1];
8598 mc->mc_db->md_root = pp->mp_pgno;
8599 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8600 new_root = mc->mc_db->md_depth++;
8602 /* Add left (implicit) pointer. */
8603 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8604 /* undo the pre-push */
8605 mc->mc_pg[0] = mc->mc_pg[1];
8606 mc->mc_ki[0] = mc->mc_ki[1];
8607 mc->mc_db->md_root = mp->mp_pgno;
8608 mc->mc_db->md_depth--;
8615 ptop = mc->mc_top-1;
8616 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8619 mdb_cursor_copy(mc, &mn);
8620 mn.mc_xcursor = NULL;
8621 mn.mc_pg[mn.mc_top] = rp;
8622 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8624 if (nflags & MDB_APPEND) {
8625 mn.mc_ki[mn.mc_top] = 0;
8627 split_indx = newindx;
8631 split_indx = (nkeys+1) / 2;
8636 unsigned int lsize, rsize, ksize;
8637 /* Move half of the keys to the right sibling */
8638 x = mc->mc_ki[mc->mc_top] - split_indx;
8639 ksize = mc->mc_db->md_pad;
8640 split = LEAF2KEY(mp, split_indx, ksize);
8641 rsize = (nkeys - split_indx) * ksize;
8642 lsize = (nkeys - split_indx) * sizeof(indx_t);
8643 mp->mp_lower -= lsize;
8644 rp->mp_lower += lsize;
8645 mp->mp_upper += rsize - lsize;
8646 rp->mp_upper -= rsize - lsize;
8647 sepkey.mv_size = ksize;
8648 if (newindx == split_indx) {
8649 sepkey.mv_data = newkey->mv_data;
8651 sepkey.mv_data = split;
8654 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8655 memcpy(rp->mp_ptrs, split, rsize);
8656 sepkey.mv_data = rp->mp_ptrs;
8657 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8658 memcpy(ins, newkey->mv_data, ksize);
8659 mp->mp_lower += sizeof(indx_t);
8660 mp->mp_upper -= ksize - sizeof(indx_t);
8663 memcpy(rp->mp_ptrs, split, x * ksize);
8664 ins = LEAF2KEY(rp, x, ksize);
8665 memcpy(ins, newkey->mv_data, ksize);
8666 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8667 rp->mp_lower += sizeof(indx_t);
8668 rp->mp_upper -= ksize - sizeof(indx_t);
8669 mc->mc_ki[mc->mc_top] = x;
8672 int psize, nsize, k;
8673 /* Maximum free space in an empty page */
8674 pmax = env->me_psize - PAGEHDRSZ;
8676 nsize = mdb_leaf_size(env, newkey, newdata);
8678 nsize = mdb_branch_size(env, newkey);
8679 nsize = EVEN(nsize);
8681 /* grab a page to hold a temporary copy */
8682 copy = mdb_page_malloc(mc->mc_txn, 1);
8687 copy->mp_pgno = mp->mp_pgno;
8688 copy->mp_flags = mp->mp_flags;
8689 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8690 copy->mp_upper = env->me_psize - PAGEBASE;
8692 /* prepare to insert */
8693 for (i=0, j=0; i<nkeys; i++) {
8695 copy->mp_ptrs[j++] = 0;
8697 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8700 /* When items are relatively large the split point needs
8701 * to be checked, because being off-by-one will make the
8702 * difference between success or failure in mdb_node_add.
8704 * It's also relevant if a page happens to be laid out
8705 * such that one half of its nodes are all "small" and
8706 * the other half of its nodes are "large." If the new
8707 * item is also "large" and falls on the half with
8708 * "large" nodes, it also may not fit.
8710 * As a final tweak, if the new item goes on the last
8711 * spot on the page (and thus, onto the new page), bias
8712 * the split so the new page is emptier than the old page.
8713 * This yields better packing during sequential inserts.
8715 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8716 /* Find split point */
8718 if (newindx <= split_indx || newindx >= nkeys) {
8720 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8725 for (; i!=k; i+=j) {
8730 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8731 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8733 if (F_ISSET(node->mn_flags, F_BIGDATA))
8734 psize += sizeof(pgno_t);
8736 psize += NODEDSZ(node);
8738 psize = EVEN(psize);
8740 if (psize > pmax || i == k-j) {
8741 split_indx = i + (j<0);
8746 if (split_indx == newindx) {
8747 sepkey.mv_size = newkey->mv_size;
8748 sepkey.mv_data = newkey->mv_data;
8750 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8751 sepkey.mv_size = node->mn_ksize;
8752 sepkey.mv_data = NODEKEY(node);
8757 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8759 /* Copy separator key to the parent.
8761 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8762 int snum = mc->mc_snum;
8766 /* We want other splits to find mn when doing fixups */
8767 WITH_CURSOR_TRACKING(mn,
8768 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
8773 if (mc->mc_snum > snum) {
8776 /* Right page might now have changed parent.
8777 * Check if left page also changed parent.
8779 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8780 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8781 for (i=0; i<ptop; i++) {
8782 mc->mc_pg[i] = mn.mc_pg[i];
8783 mc->mc_ki[i] = mn.mc_ki[i];
8785 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8786 if (mn.mc_ki[ptop]) {
8787 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8789 /* find right page's left sibling */
8790 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8791 mdb_cursor_sibling(mc, 0);
8796 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8799 if (rc != MDB_SUCCESS) {
8802 if (nflags & MDB_APPEND) {
8803 mc->mc_pg[mc->mc_top] = rp;
8804 mc->mc_ki[mc->mc_top] = 0;
8805 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8808 for (i=0; i<mc->mc_top; i++)
8809 mc->mc_ki[i] = mn.mc_ki[i];
8810 } else if (!IS_LEAF2(mp)) {
8812 mc->mc_pg[mc->mc_top] = rp;
8817 rkey.mv_data = newkey->mv_data;
8818 rkey.mv_size = newkey->mv_size;
8824 /* Update index for the new key. */
8825 mc->mc_ki[mc->mc_top] = j;
8827 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8828 rkey.mv_data = NODEKEY(node);
8829 rkey.mv_size = node->mn_ksize;
8831 xdata.mv_data = NODEDATA(node);
8832 xdata.mv_size = NODEDSZ(node);
8835 pgno = NODEPGNO(node);
8836 flags = node->mn_flags;
8839 if (!IS_LEAF(mp) && j == 0) {
8840 /* First branch index doesn't need key data. */
8844 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8850 mc->mc_pg[mc->mc_top] = copy;
8855 } while (i != split_indx);
8857 nkeys = NUMKEYS(copy);
8858 for (i=0; i<nkeys; i++)
8859 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8860 mp->mp_lower = copy->mp_lower;
8861 mp->mp_upper = copy->mp_upper;
8862 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8863 env->me_psize - copy->mp_upper - PAGEBASE);
8865 /* reset back to original page */
8866 if (newindx < split_indx) {
8867 mc->mc_pg[mc->mc_top] = mp;
8869 mc->mc_pg[mc->mc_top] = rp;
8871 /* Make sure mc_ki is still valid.
8873 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8874 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8875 for (i=0; i<=ptop; i++) {
8876 mc->mc_pg[i] = mn.mc_pg[i];
8877 mc->mc_ki[i] = mn.mc_ki[i];
8881 if (nflags & MDB_RESERVE) {
8882 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8883 if (!(node->mn_flags & F_BIGDATA))
8884 newdata->mv_data = NODEDATA(node);
8887 if (newindx >= split_indx) {
8888 mc->mc_pg[mc->mc_top] = rp;
8890 /* Make sure mc_ki is still valid.
8892 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8893 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8894 for (i=0; i<=ptop; i++) {
8895 mc->mc_pg[i] = mn.mc_pg[i];
8896 mc->mc_ki[i] = mn.mc_ki[i];
8903 /* Adjust other cursors pointing to mp */
8904 MDB_cursor *m2, *m3;
8905 MDB_dbi dbi = mc->mc_dbi;
8906 nkeys = NUMKEYS(mp);
8908 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8909 if (mc->mc_flags & C_SUB)
8910 m3 = &m2->mc_xcursor->mx_cursor;
8915 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8919 /* sub cursors may be on different DB */
8920 if (m3->mc_pg[0] != mp)
8923 for (k=new_root; k>=0; k--) {
8924 m3->mc_ki[k+1] = m3->mc_ki[k];
8925 m3->mc_pg[k+1] = m3->mc_pg[k];
8927 if (m3->mc_ki[0] >= nkeys) {
8932 m3->mc_pg[0] = mc->mc_pg[0];
8936 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8937 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8938 m3->mc_ki[mc->mc_top]++;
8939 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8940 m3->mc_pg[mc->mc_top] = rp;
8941 m3->mc_ki[mc->mc_top] -= nkeys;
8942 for (i=0; i<mc->mc_top; i++) {
8943 m3->mc_ki[i] = mn.mc_ki[i];
8944 m3->mc_pg[i] = mn.mc_pg[i];
8947 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8948 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8951 if (XCURSOR_INITED(m3) && IS_LEAF(mp))
8952 XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
8955 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8958 if (copy) /* tmp page */
8959 mdb_page_free(env, copy);
8961 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8966 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8967 MDB_val *key, MDB_val *data, unsigned int flags)
8973 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
8976 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
8979 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
8980 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8982 mdb_cursor_init(&mc, txn, dbi, &mx);
8983 mc.mc_next = txn->mt_cursors[dbi];
8984 txn->mt_cursors[dbi] = &mc;
8985 rc = mdb_cursor_put(&mc, key, data, flags);
8986 txn->mt_cursors[dbi] = mc.mc_next;
8991 #define MDB_WBUF (1024*1024)
8993 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
8995 /** State needed for a double-buffering compacting copy. */
8996 typedef struct mdb_copy {
8999 pthread_mutex_t mc_mutex;
9000 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
9005 pgno_t mc_next_pgno;
9007 int mc_toggle; /**< Buffer number in provider */
9008 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
9009 /** Error code. Never cleared if set. Both threads can set nonzero
9010 * to fail the copy. Not mutex-protected, LMDB expects atomic int.
9012 volatile int mc_error;
9015 /** Dedicated writer thread for compacting copy. */
9016 static THREAD_RET ESECT CALL_CONV
9017 mdb_env_copythr(void *arg)
9021 int toggle = 0, wsize, rc;
9024 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9027 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9031 sigaddset(&set, SIGPIPE);
9032 if ((rc = pthread_sigmask(SIG_BLOCK, &set, NULL)) != 0)
9037 pthread_mutex_lock(&my->mc_mutex);
9040 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9041 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9043 wsize = my->mc_wlen[toggle];
9044 ptr = my->mc_wbuf[toggle];
9047 while (wsize > 0 && !my->mc_error) {
9048 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9051 #if defined(SIGPIPE) && !defined(_WIN32)
9053 /* Collect the pending SIGPIPE, otherwise at least OS X
9054 * gives it to the process on thread-exit (ITS#8504).
9057 sigwait(&set, &tmp);
9061 } else if (len > 0) {
9074 /* If there's an overflow page tail, write it too */
9075 if (my->mc_olen[toggle]) {
9076 wsize = my->mc_olen[toggle];
9077 ptr = my->mc_over[toggle];
9078 my->mc_olen[toggle] = 0;
9081 my->mc_wlen[toggle] = 0;
9083 /* Return the empty buffer to provider */
9085 pthread_cond_signal(&my->mc_cond);
9087 pthread_mutex_unlock(&my->mc_mutex);
9088 return (THREAD_RET)0;
9092 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9094 * @param[in] my control structure.
9095 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9098 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9100 pthread_mutex_lock(&my->mc_mutex);
9101 my->mc_new += adjust;
9102 pthread_cond_signal(&my->mc_cond);
9103 while (my->mc_new & 2) /* both buffers in use */
9104 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9105 pthread_mutex_unlock(&my->mc_mutex);
9107 my->mc_toggle ^= (adjust & 1);
9108 /* Both threads reset mc_wlen, to be safe from threading errors */
9109 my->mc_wlen[my->mc_toggle] = 0;
9110 return my->mc_error;
9113 /** Depth-first tree traversal for compacting copy.
9114 * @param[in] my control structure.
9115 * @param[in,out] pg database root.
9116 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB.
9119 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9121 MDB_cursor mc = {0};
9123 MDB_page *mo, *mp, *leaf;
9128 /* Empty DB, nothing to do */
9129 if (*pg == P_INVALID)
9133 mc.mc_txn = my->mc_txn;
9135 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9138 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9142 /* Make cursor pages writable */
9143 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9147 for (i=0; i<mc.mc_top; i++) {
9148 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9149 mc.mc_pg[i] = (MDB_page *)ptr;
9150 ptr += my->mc_env->me_psize;
9153 /* This is writable space for a leaf page. Usually not needed. */
9154 leaf = (MDB_page *)ptr;
9156 toggle = my->mc_toggle;
9157 while (mc.mc_snum > 0) {
9159 mp = mc.mc_pg[mc.mc_top];
9163 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9164 for (i=0; i<n; i++) {
9165 ni = NODEPTR(mp, i);
9166 if (ni->mn_flags & F_BIGDATA) {
9170 /* Need writable leaf */
9172 mc.mc_pg[mc.mc_top] = leaf;
9173 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9175 ni = NODEPTR(mp, i);
9178 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9179 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9180 rc = mdb_page_get(&mc, pg, &omp, NULL);
9183 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9184 rc = mdb_env_cthr_toggle(my, 1);
9187 toggle = my->mc_toggle;
9189 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9190 memcpy(mo, omp, my->mc_env->me_psize);
9191 mo->mp_pgno = my->mc_next_pgno;
9192 my->mc_next_pgno += omp->mp_pages;
9193 my->mc_wlen[toggle] += my->mc_env->me_psize;
9194 if (omp->mp_pages > 1) {
9195 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9196 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9197 rc = mdb_env_cthr_toggle(my, 1);
9200 toggle = my->mc_toggle;
9202 } else if (ni->mn_flags & F_SUBDATA) {
9205 /* Need writable leaf */
9207 mc.mc_pg[mc.mc_top] = leaf;
9208 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9210 ni = NODEPTR(mp, i);
9213 memcpy(&db, NODEDATA(ni), sizeof(db));
9214 my->mc_toggle = toggle;
9215 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9218 toggle = my->mc_toggle;
9219 memcpy(NODEDATA(ni), &db, sizeof(db));
9224 mc.mc_ki[mc.mc_top]++;
9225 if (mc.mc_ki[mc.mc_top] < n) {
9228 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9230 rc = mdb_page_get(&mc, pg, &mp, NULL);
9235 mc.mc_ki[mc.mc_top] = 0;
9236 if (IS_BRANCH(mp)) {
9237 /* Whenever we advance to a sibling branch page,
9238 * we must proceed all the way down to its first leaf.
9240 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9243 mc.mc_pg[mc.mc_top] = mp;
9247 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9248 rc = mdb_env_cthr_toggle(my, 1);
9251 toggle = my->mc_toggle;
9253 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9254 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9255 mo->mp_pgno = my->mc_next_pgno++;
9256 my->mc_wlen[toggle] += my->mc_env->me_psize;
9258 /* Update parent if there is one */
9259 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9260 SETPGNO(ni, mo->mp_pgno);
9261 mdb_cursor_pop(&mc);
9263 /* Otherwise we're done */
9273 /** Copy environment with compaction. */
9275 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9280 MDB_txn *txn = NULL;
9282 pgno_t root, new_root;
9283 int rc = MDB_SUCCESS;
9286 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
9287 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
9291 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9292 if (my.mc_wbuf[0] == NULL) {
9293 /* _aligned_malloc() sets errno, but we use Windows error codes */
9294 rc = ERROR_NOT_ENOUGH_MEMORY;
9298 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
9300 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
9302 #ifdef HAVE_MEMALIGN
9303 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9304 if (my.mc_wbuf[0] == NULL) {
9311 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
9317 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9318 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9319 my.mc_next_pgno = NUM_METAS;
9322 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
9326 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9330 mp = (MDB_page *)my.mc_wbuf[0];
9331 memset(mp, 0, NUM_METAS * env->me_psize);
9333 mp->mp_flags = P_META;
9334 mm = (MDB_meta *)METADATA(mp);
9335 mdb_env_init_meta0(env, mm);
9336 mm->mm_address = env->me_metas[0]->mm_address;
9338 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9340 mp->mp_flags = P_META;
9341 *(MDB_meta *)METADATA(mp) = *mm;
9342 mm = (MDB_meta *)METADATA(mp);
9344 /* Set metapage 1 with current main DB */
9345 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
9346 if (root != P_INVALID) {
9347 /* Count free pages + freeDB pages. Subtract from last_pg
9348 * to find the new last_pg, which also becomes the new root.
9350 MDB_ID freecount = 0;
9353 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9354 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9355 freecount += *(MDB_ID *)data.mv_data;
9356 if (rc != MDB_NOTFOUND)
9358 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9359 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9360 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9362 new_root = txn->mt_next_pgno - 1 - freecount;
9363 mm->mm_last_pg = new_root;
9364 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9365 mm->mm_dbs[MAIN_DBI].md_root = new_root;
9367 /* When the DB is empty, handle it specially to
9368 * fix any breakage like page leaks from ITS#8174.
9370 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
9372 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
9373 mm->mm_txnid = 1; /* use metapage 1 */
9376 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9378 rc = mdb_env_cwalk(&my, &root, 0);
9379 if (rc == MDB_SUCCESS && root != new_root) {
9380 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
9386 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
9387 rc = THREAD_FINISH(thr);
9392 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
9393 if (my.mc_cond) CloseHandle(my.mc_cond);
9394 if (my.mc_mutex) CloseHandle(my.mc_mutex);
9396 free(my.mc_wbuf[0]);
9397 pthread_cond_destroy(&my.mc_cond);
9399 pthread_mutex_destroy(&my.mc_mutex);
9401 return rc ? rc : my.mc_error;
9404 /** Copy environment as-is. */
9406 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
9408 MDB_txn *txn = NULL;
9409 mdb_mutexref_t wmutex = NULL;
9415 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9419 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9422 /* Do the lock/unlock of the reader mutex before starting the
9423 * write txn. Otherwise other read txns could block writers.
9425 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9430 /* We must start the actual read txn after blocking writers */
9431 mdb_txn_end(txn, MDB_END_RESET_TMP);
9433 /* Temporarily block writers until we snapshot the meta pages */
9434 wmutex = env->me_wmutex;
9435 if (LOCK_MUTEX(rc, env, wmutex))
9438 rc = mdb_txn_renew0(txn);
9440 UNLOCK_MUTEX(wmutex);
9445 wsize = env->me_psize * NUM_METAS;
9449 DO_WRITE(rc, fd, ptr, w2, len);
9453 } else if (len > 0) {
9459 /* Non-blocking or async handles are not supported */
9465 UNLOCK_MUTEX(wmutex);
9470 w3 = txn->mt_next_pgno * env->me_psize;
9473 if ((rc = mdb_fsize(env->me_fd, &fsize)))
9480 if (wsize > MAX_WRITE)
9484 DO_WRITE(rc, fd, ptr, w2, len);
9488 } else if (len > 0) {
9505 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9507 if (flags & MDB_CP_COMPACT)
9508 return mdb_env_copyfd1(env, fd);
9510 return mdb_env_copyfd0(env, fd);
9514 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9516 return mdb_env_copyfd2(env, fd, 0);
9520 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9524 HANDLE newfd = INVALID_HANDLE_VALUE;
9526 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname);
9527 if (rc == MDB_SUCCESS) {
9528 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd);
9529 mdb_fname_destroy(fname);
9531 if (rc == MDB_SUCCESS) {
9532 rc = mdb_env_copyfd2(env, newfd, flags);
9533 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9540 mdb_env_copy(MDB_env *env, const char *path)
9542 return mdb_env_copy2(env, path, 0);
9546 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9548 if (flag & ~CHANGEABLE)
9551 env->me_flags |= flag;
9553 env->me_flags &= ~flag;
9558 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9563 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
9568 mdb_env_set_userctx(MDB_env *env, void *ctx)
9572 env->me_userctx = ctx;
9577 mdb_env_get_userctx(MDB_env *env)
9579 return env ? env->me_userctx : NULL;
9583 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9588 env->me_assert_func = func;
9594 mdb_env_get_path(MDB_env *env, const char **arg)
9599 *arg = env->me_path;
9604 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9613 /** Common code for #mdb_stat() and #mdb_env_stat().
9614 * @param[in] env the environment to operate in.
9615 * @param[in] db the #MDB_db record containing the stats to return.
9616 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9617 * @return 0, this function always succeeds.
9620 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9622 arg->ms_psize = env->me_psize;
9623 arg->ms_depth = db->md_depth;
9624 arg->ms_branch_pages = db->md_branch_pages;
9625 arg->ms_leaf_pages = db->md_leaf_pages;
9626 arg->ms_overflow_pages = db->md_overflow_pages;
9627 arg->ms_entries = db->md_entries;
9633 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9637 if (env == NULL || arg == NULL)
9640 meta = mdb_env_pick_meta(env);
9642 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
9646 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9650 if (env == NULL || arg == NULL)
9653 meta = mdb_env_pick_meta(env);
9654 arg->me_mapaddr = meta->mm_address;
9655 arg->me_last_pgno = meta->mm_last_pg;
9656 arg->me_last_txnid = meta->mm_txnid;
9658 arg->me_mapsize = env->me_mapsize;
9659 arg->me_maxreaders = env->me_maxreaders;
9660 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9664 /** Set the default comparison functions for a database.
9665 * Called immediately after a database is opened to set the defaults.
9666 * The user can then override them with #mdb_set_compare() or
9667 * #mdb_set_dupsort().
9668 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9669 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9672 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9674 uint16_t f = txn->mt_dbs[dbi].md_flags;
9676 txn->mt_dbxs[dbi].md_cmp =
9677 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9678 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9680 txn->mt_dbxs[dbi].md_dcmp =
9681 !(f & MDB_DUPSORT) ? 0 :
9682 ((f & MDB_INTEGERDUP)
9683 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9684 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9687 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9693 int rc, dbflag, exact;
9694 unsigned int unused = 0, seq;
9698 if (flags & ~VALID_FLAGS)
9700 if (txn->mt_flags & MDB_TXN_BLOCKED)
9706 if (flags & PERSISTENT_FLAGS) {
9707 uint16_t f2 = flags & PERSISTENT_FLAGS;
9708 /* make sure flag changes get committed */
9709 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9710 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9711 txn->mt_flags |= MDB_TXN_DIRTY;
9714 mdb_default_cmp(txn, MAIN_DBI);
9718 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9719 mdb_default_cmp(txn, MAIN_DBI);
9722 /* Is the DB already open? */
9724 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
9725 if (!txn->mt_dbxs[i].md_name.mv_size) {
9726 /* Remember this free slot */
9727 if (!unused) unused = i;
9730 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9731 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9737 /* If no free slot and max hit, fail */
9738 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9739 return MDB_DBS_FULL;
9741 /* Cannot mix named databases with some mainDB flags */
9742 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9743 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9745 /* Find the DB info */
9746 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
9749 key.mv_data = (void *)name;
9750 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9751 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9752 if (rc == MDB_SUCCESS) {
9753 /* make sure this is actually a DB */
9754 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9755 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9756 return MDB_INCOMPATIBLE;
9757 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
9761 /* Done here so we cannot fail after creating a new DB */
9762 if ((namedup = strdup(name)) == NULL)
9766 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
9767 data.mv_size = sizeof(MDB_db);
9768 data.mv_data = &dummy;
9769 memset(&dummy, 0, sizeof(dummy));
9770 dummy.md_root = P_INVALID;
9771 dummy.md_flags = flags & PERSISTENT_FLAGS;
9772 WITH_CURSOR_TRACKING(mc,
9773 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA));
9780 /* Got info, register DBI in this txn */
9781 unsigned int slot = unused ? unused : txn->mt_numdbs;
9782 txn->mt_dbxs[slot].md_name.mv_data = namedup;
9783 txn->mt_dbxs[slot].md_name.mv_size = len;
9784 txn->mt_dbxs[slot].md_rel = NULL;
9785 txn->mt_dbflags[slot] = dbflag;
9786 /* txn-> and env-> are the same in read txns, use
9787 * tmp variable to avoid undefined assignment
9789 seq = ++txn->mt_env->me_dbiseqs[slot];
9790 txn->mt_dbiseqs[slot] = seq;
9792 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9794 mdb_default_cmp(txn, slot);
9804 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9806 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
9809 if (txn->mt_flags & MDB_TXN_BLOCKED)
9812 if (txn->mt_dbflags[dbi] & DB_STALE) {
9815 /* Stale, must read the DB's root. cursor_init does it for us. */
9816 mdb_cursor_init(&mc, txn, dbi, &mx);
9818 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9821 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9824 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
9826 ptr = env->me_dbxs[dbi].md_name.mv_data;
9827 /* If there was no name, this was already closed */
9829 env->me_dbxs[dbi].md_name.mv_data = NULL;
9830 env->me_dbxs[dbi].md_name.mv_size = 0;
9831 env->me_dbflags[dbi] = 0;
9832 env->me_dbiseqs[dbi]++;
9837 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9839 /* We could return the flags for the FREE_DBI too but what's the point? */
9840 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9842 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9846 /** Add all the DB's pages to the free list.
9847 * @param[in] mc Cursor on the DB to free.
9848 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9849 * @return 0 on success, non-zero on failure.
9852 mdb_drop0(MDB_cursor *mc, int subs)
9856 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9857 if (rc == MDB_SUCCESS) {
9858 MDB_txn *txn = mc->mc_txn;
9863 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
9864 * This also avoids any P_LEAF2 pages, which have no nodes.
9865 * Also if the DB doesn't have sub-DBs and has no overflow
9866 * pages, omit scanning leaves.
9868 if ((mc->mc_flags & C_SUB) ||
9869 (!subs && !mc->mc_db->md_overflow_pages))
9872 mdb_cursor_copy(mc, &mx);
9873 while (mc->mc_snum > 0) {
9874 MDB_page *mp = mc->mc_pg[mc->mc_top];
9875 unsigned n = NUMKEYS(mp);
9877 for (i=0; i<n; i++) {
9878 ni = NODEPTR(mp, i);
9879 if (ni->mn_flags & F_BIGDATA) {
9882 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9883 rc = mdb_page_get(mc, pg, &omp, NULL);
9886 mdb_cassert(mc, IS_OVERFLOW(omp));
9887 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9891 mc->mc_db->md_overflow_pages -= omp->mp_pages;
9892 if (!mc->mc_db->md_overflow_pages && !subs)
9894 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9895 mdb_xcursor_init1(mc, ni);
9896 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9901 if (!subs && !mc->mc_db->md_overflow_pages)
9904 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9906 for (i=0; i<n; i++) {
9908 ni = NODEPTR(mp, i);
9911 mdb_midl_xappend(txn->mt_free_pgs, pg);
9916 mc->mc_ki[mc->mc_top] = i;
9917 rc = mdb_cursor_sibling(mc, 1);
9919 if (rc != MDB_NOTFOUND)
9921 /* no more siblings, go back to beginning
9922 * of previous level.
9927 for (i=1; i<mc->mc_snum; i++) {
9929 mc->mc_pg[i] = mx.mc_pg[i];
9934 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9937 txn->mt_flags |= MDB_TXN_ERROR;
9938 } else if (rc == MDB_NOTFOUND) {
9941 mc->mc_flags &= ~C_INITIALIZED;
9945 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9947 MDB_cursor *mc, *m2;
9950 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9953 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9956 if (TXN_DBI_CHANGED(txn, dbi))
9959 rc = mdb_cursor_open(txn, dbi, &mc);
9963 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9964 /* Invalidate the dropped DB's cursors */
9965 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9966 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9970 /* Can't delete the main DB */
9971 if (del && dbi >= CORE_DBS) {
9972 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9974 txn->mt_dbflags[dbi] = DB_STALE;
9975 mdb_dbi_close(txn->mt_env, dbi);
9977 txn->mt_flags |= MDB_TXN_ERROR;
9980 /* reset the DB record, mark it dirty */
9981 txn->mt_dbflags[dbi] |= DB_DIRTY;
9982 txn->mt_dbs[dbi].md_depth = 0;
9983 txn->mt_dbs[dbi].md_branch_pages = 0;
9984 txn->mt_dbs[dbi].md_leaf_pages = 0;
9985 txn->mt_dbs[dbi].md_overflow_pages = 0;
9986 txn->mt_dbs[dbi].md_entries = 0;
9987 txn->mt_dbs[dbi].md_root = P_INVALID;
9989 txn->mt_flags |= MDB_TXN_DIRTY;
9992 mdb_cursor_close(mc);
9996 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9998 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10001 txn->mt_dbxs[dbi].md_cmp = cmp;
10002 return MDB_SUCCESS;
10005 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10007 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10010 txn->mt_dbxs[dbi].md_dcmp = cmp;
10011 return MDB_SUCCESS;
10014 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
10016 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10019 txn->mt_dbxs[dbi].md_rel = rel;
10020 return MDB_SUCCESS;
10023 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
10025 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10028 txn->mt_dbxs[dbi].md_relctx = ctx;
10029 return MDB_SUCCESS;
10033 mdb_env_get_maxkeysize(MDB_env *env)
10035 return ENV_MAXKEY(env);
10039 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10041 unsigned int i, rdrs;
10044 int rc = 0, first = 1;
10048 if (!env->me_txns) {
10049 return func("(no reader locks)\n", ctx);
10051 rdrs = env->me_txns->mti_numreaders;
10052 mr = env->me_txns->mti_readers;
10053 for (i=0; i<rdrs; i++) {
10054 if (mr[i].mr_pid) {
10055 txnid_t txnid = mr[i].mr_txnid;
10056 sprintf(buf, txnid == (txnid_t)-1 ?
10057 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
10058 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10061 rc = func(" pid thread txnid\n", ctx);
10065 rc = func(buf, ctx);
10071 rc = func("(no active readers)\n", ctx);
10076 /** Insert pid into list if not already present.
10077 * return -1 if already present.
10080 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10082 /* binary search of pid in list */
10084 unsigned cursor = 1;
10086 unsigned n = ids[0];
10089 unsigned pivot = n >> 1;
10090 cursor = base + pivot + 1;
10091 val = pid - ids[cursor];
10096 } else if ( val > 0 ) {
10101 /* found, so it's a duplicate */
10110 for (n = ids[0]; n > cursor; n--)
10117 mdb_reader_check(MDB_env *env, int *dead)
10123 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10126 /** As #mdb_reader_check(). \b rlocked is set if caller locked #me_rmutex. */
10128 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10130 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10131 unsigned int i, j, rdrs;
10133 MDB_PID_T *pids, pid;
10134 int rc = MDB_SUCCESS, count = 0;
10136 rdrs = env->me_txns->mti_numreaders;
10137 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10141 mr = env->me_txns->mti_readers;
10142 for (i=0; i<rdrs; i++) {
10143 pid = mr[i].mr_pid;
10144 if (pid && pid != env->me_pid) {
10145 if (mdb_pid_insert(pids, pid) == 0) {
10146 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10147 /* Stale reader found */
10150 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10151 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10153 rdrs = 0; /* the above checked all readers */
10155 /* Recheck, a new process may have reused pid */
10156 if (mdb_reader_pid(env, Pidcheck, pid))
10160 for (; j<rdrs; j++)
10161 if (mr[j].mr_pid == pid) {
10162 DPRINTF(("clear stale reader pid %u txn %"Z"d",
10163 (unsigned) pid, mr[j].mr_txnid));
10168 UNLOCK_MUTEX(rmutex);
10179 #ifdef MDB_ROBUST_SUPPORTED
10180 /** Handle #LOCK_MUTEX0() failure.
10181 * Try to repair the lock file if the mutex owner died.
10182 * @param[in] env the environment handle
10183 * @param[in] mutex LOCK_MUTEX0() mutex
10184 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10185 * @return 0 on success with the mutex locked, or an error code on failure.
10188 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10193 if (rc == MDB_OWNERDEAD) {
10194 /* We own the mutex. Clean up after dead previous owner. */
10196 rlocked = (mutex == env->me_rmutex);
10198 /* Keep mti_txnid updated, otherwise next writer can
10199 * overwrite data which latest meta page refers to.
10201 meta = mdb_env_pick_meta(env);
10202 env->me_txns->mti_txnid = meta->mm_txnid;
10203 /* env is hosed if the dead thread was ours */
10205 env->me_flags |= MDB_FATAL_ERROR;
10206 env->me_txn = NULL;
10210 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10211 (rc ? "this process' env is hosed" : "recovering")));
10212 rc2 = mdb_reader_check0(env, rlocked, NULL);
10214 rc2 = mdb_mutex_consistent(mutex);
10215 if (rc || (rc = rc2)) {
10216 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10217 UNLOCK_MUTEX(mutex);
10223 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10228 #endif /* MDB_ROBUST_SUPPORTED */
10230 #if defined(_WIN32)
10231 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */
10233 utf8_to_utf16(const char *src, MDB_name *dst, int xtra)
10236 wchar_t *result = NULL;
10237 for (;;) { /* malloc result, then fill it in */
10238 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need);
10245 result = malloc(sizeof(wchar_t) * (need + xtra));
10250 dst->mn_alloced = 1;
10251 dst->mn_len = need - 1;
10252 dst->mn_val = result;
10253 return MDB_SUCCESS;
10256 #endif /* defined(_WIN32) */