2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2014 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments. You can safely
85 * define MDB_FDATASYNC_WORKS if this code will only be run
86 * on kernels 3.6 and newer.
88 #define BROKEN_FDATASYNC
101 #if defined(__sun) || defined(ANDROID)
102 /* Most platforms have posix_memalign, older may only have memalign */
103 #define HAVE_MEMALIGN 1
107 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
108 #include <netinet/in.h>
109 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
112 #if defined(__APPLE__) || defined (BSD)
113 # define MDB_USE_SYSV_SEM 1
114 # define MDB_FDATASYNC fsync
115 #elif defined(ANDROID)
116 # define MDB_FDATASYNC fsync
121 #ifdef MDB_USE_SYSV_SEM
124 #ifdef _SEM_SEMUN_UNDEFINED
127 struct semid_ds *buf;
128 unsigned short *array;
130 #endif /* _SEM_SEMUN_UNDEFINED */
131 #endif /* MDB_USE_SYSV_SEM */
135 #include <valgrind/memcheck.h>
136 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
137 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
138 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
139 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
140 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
142 #define VGMEMP_CREATE(h,r,z)
143 #define VGMEMP_ALLOC(h,a,s)
144 #define VGMEMP_FREE(h,a)
145 #define VGMEMP_DESTROY(h)
146 #define VGMEMP_DEFINED(a,s)
150 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
151 /* Solaris just defines one or the other */
152 # define LITTLE_ENDIAN 1234
153 # define BIG_ENDIAN 4321
154 # ifdef _LITTLE_ENDIAN
155 # define BYTE_ORDER LITTLE_ENDIAN
157 # define BYTE_ORDER BIG_ENDIAN
160 # define BYTE_ORDER __BYTE_ORDER
164 #ifndef LITTLE_ENDIAN
165 #define LITTLE_ENDIAN __LITTLE_ENDIAN
168 #define BIG_ENDIAN __BIG_ENDIAN
171 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
172 #define MISALIGNED_OK 1
178 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
179 # error "Unknown or unsupported endianness (BYTE_ORDER)"
180 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
181 # error "Two's complement, reasonably sized integer types, please"
185 /** Put infrequently used env functions in separate section */
187 # define ESECT __attribute__ ((section("__TEXT,text_env")))
189 # define ESECT __attribute__ ((section("text_env")))
195 /** @defgroup internal LMDB Internals
198 /** @defgroup compat Compatibility Macros
199 * A bunch of macros to minimize the amount of platform-specific ifdefs
200 * needed throughout the rest of the code. When the features this library
201 * needs are similar enough to POSIX to be hidden in a one-or-two line
202 * replacement, this macro approach is used.
206 /** Features under development */
211 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM) || defined(EOWNERDEAD)
212 #define MDB_ROBUST_SUPPORTED 1
215 /** Wrapper around __func__, which is a C99 feature */
216 #if __STDC_VERSION__ >= 199901L
217 # define mdb_func_ __func__
218 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
219 # define mdb_func_ __FUNCTION__
221 /* If a debug message says <mdb_unknown>(), update the #if statements above */
222 # define mdb_func_ "<mdb_unknown>"
225 /* Internal error codes, not exposed outside liblmdb */
226 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
228 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
229 #elif defined MDB_USE_SYSV_SEM
230 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
232 #define MDB_OWNERDEAD EOWNERDEAD
236 #define MDB_USE_HASH 1
237 #define MDB_PIDLOCK 0
238 #define THREAD_RET DWORD
239 #define pthread_t HANDLE
240 #define pthread_mutex_t HANDLE
241 #define pthread_cond_t HANDLE
242 typedef HANDLE mdb_mutex_t;
243 #define pthread_key_t DWORD
244 #define pthread_self() GetCurrentThreadId()
245 #define pthread_key_create(x,y) \
246 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
247 #define pthread_key_delete(x) TlsFree(x)
248 #define pthread_getspecific(x) TlsGetValue(x)
249 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
250 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
251 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
252 #define pthread_cond_signal(x) SetEvent(*x)
253 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
254 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
255 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
256 #define MDB_MUTEX(env, rw) ((env)->me_##rw##mutex)
257 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
258 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
259 #define mdb_mutex_consistent(mutex) 0
260 #define getpid() GetCurrentProcessId()
261 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
262 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
263 #define ErrCode() GetLastError()
264 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
265 #define close(fd) (CloseHandle(fd) ? 0 : -1)
266 #define munmap(ptr,len) UnmapViewOfFile(ptr)
267 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
268 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
270 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
274 #define THREAD_RET void *
275 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
276 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
277 #define Z "z" /**< printf format modifier for size_t */
279 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
280 #define MDB_PIDLOCK 1
282 #ifdef MDB_USE_SYSV_SEM
284 typedef struct mdb_mutex {
290 #define MDB_MUTEX(env, rw) (&(env)->me_##rw##mutex)
291 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
292 #define UNLOCK_MUTEX(mutex) do { \
293 struct sembuf sb = { 0, 1, SEM_UNDO }; \
294 sb.sem_num = (mutex)->semnum; \
295 *(mutex)->locked = 0; \
296 semop((mutex)->semid, &sb, 1); \
300 mdb_sem_wait(mdb_mutex_t *sem)
302 int rc, *locked = sem->locked;
303 struct sembuf sb = { 0, -1, SEM_UNDO };
304 sb.sem_num = sem->semnum;
306 if (!semop(sem->semid, &sb, 1)) {
307 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
311 } while ((rc = errno) == EINTR);
315 #define mdb_mutex_consistent(mutex) 0
318 /** Pointer/HANDLE type of shared mutex/semaphore.
320 typedef pthread_mutex_t mdb_mutex_t;
321 /** Mutex for the reader table (rw = r) or write transaction (rw = w).
323 #define MDB_MUTEX(env, rw) (&(env)->me_txns->mti_##rw##mutex)
324 /** Lock the reader or writer mutex.
325 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
327 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
328 /** Unlock the reader or writer mutex.
330 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
331 /** Mark mutex-protected data as repaired, after death of previous owner.
333 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
334 #endif /* MDB_USE_SYSV_SEM */
336 /** Get the error code for the last failed system function.
338 #define ErrCode() errno
340 /** An abstraction for a file handle.
341 * On POSIX systems file handles are small integers. On Windows
342 * they're opaque pointers.
346 /** A value for an invalid file handle.
347 * Mainly used to initialize file variables and signify that they are
350 #define INVALID_HANDLE_VALUE (-1)
352 /** Get the size of a memory page for the system.
353 * This is the basic size that the platform's memory manager uses, and is
354 * fundamental to the use of memory-mapped files.
356 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
361 #elif defined(MDB_USE_SYSV_SEM)
362 #define MNAME_LEN (sizeof(int))
364 #define MNAME_LEN (sizeof(pthread_mutex_t))
367 #ifdef MDB_USE_SYSV_SEM
368 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
370 #define SYSV_SEM_FLAG 0
375 #ifdef MDB_ROBUST_SUPPORTED
376 /** Lock mutex, handle any error, set rc = result.
377 * Return 0 on success, nonzero (not rc) on error.
379 #define LOCK_MUTEX(rc, env, mutex) \
380 (((rc) = LOCK_MUTEX0(mutex)) && \
381 ((rc) = mdb_mutex_failed(env, mutex, rc)))
382 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc);
384 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
385 #define mdb_mutex_failed(env, mutex, rc) (rc)
389 /** A flag for opening a file and requesting synchronous data writes.
390 * This is only used when writing a meta page. It's not strictly needed;
391 * we could just do a normal write and then immediately perform a flush.
392 * But if this flag is available it saves us an extra system call.
394 * @note If O_DSYNC is undefined but exists in /usr/include,
395 * preferably set some compiler flag to get the definition.
396 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
399 # define MDB_DSYNC O_DSYNC
403 /** Function for flushing the data of a file. Define this to fsync
404 * if fdatasync() is not supported.
406 #ifndef MDB_FDATASYNC
407 # define MDB_FDATASYNC fdatasync
411 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
422 /** A page number in the database.
423 * Note that 64 bit page numbers are overkill, since pages themselves
424 * already represent 12-13 bits of addressable memory, and the OS will
425 * always limit applications to a maximum of 63 bits of address space.
427 * @note In the #MDB_node structure, we only store 48 bits of this value,
428 * which thus limits us to only 60 bits of addressable data.
430 typedef MDB_ID pgno_t;
432 /** A transaction ID.
433 * See struct MDB_txn.mt_txnid for details.
435 typedef MDB_ID txnid_t;
437 /** @defgroup debug Debug Macros
441 /** Enable debug output. Needs variable argument macros (a C99 feature).
442 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
443 * read from and written to the database (used for free space management).
449 static int mdb_debug;
450 static txnid_t mdb_debug_start;
452 /** Print a debug message with printf formatting.
453 * Requires double parenthesis around 2 or more args.
455 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
456 # define DPRINTF0(fmt, ...) \
457 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
459 # define DPRINTF(args) ((void) 0)
461 /** Print a debug string.
462 * The string is printed literally, with no format processing.
464 #define DPUTS(arg) DPRINTF(("%s", arg))
465 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
467 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
470 /** @brief The maximum size of a database page.
472 * It is 32k or 64k, since value-PAGEBASE must fit in
473 * #MDB_page.%mp_upper.
475 * LMDB will use database pages < OS pages if needed.
476 * That causes more I/O in write transactions: The OS must
477 * know (read) the whole page before writing a partial page.
479 * Note that we don't currently support Huge pages. On Linux,
480 * regular data files cannot use Huge pages, and in general
481 * Huge pages aren't actually pageable. We rely on the OS
482 * demand-pager to read our data and page it out when memory
483 * pressure from other processes is high. So until OSs have
484 * actual paging support for Huge pages, they're not viable.
486 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
488 /** The minimum number of keys required in a database page.
489 * Setting this to a larger value will place a smaller bound on the
490 * maximum size of a data item. Data items larger than this size will
491 * be pushed into overflow pages instead of being stored directly in
492 * the B-tree node. This value used to default to 4. With a page size
493 * of 4096 bytes that meant that any item larger than 1024 bytes would
494 * go into an overflow page. That also meant that on average 2-3KB of
495 * each overflow page was wasted space. The value cannot be lower than
496 * 2 because then there would no longer be a tree structure. With this
497 * value, items larger than 2KB will go into overflow pages, and on
498 * average only 1KB will be wasted.
500 #define MDB_MINKEYS 2
502 /** A stamp that identifies a file as an LMDB file.
503 * There's nothing special about this value other than that it is easily
504 * recognizable, and it will reflect any byte order mismatches.
506 #define MDB_MAGIC 0xBEEFC0DE
508 /** The version number for a database's datafile format. */
509 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
510 /** The version number for a database's lockfile format. */
511 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
513 /** @brief The max size of a key we can write, or 0 for dynamic max.
515 * Define this as 0 to compute the max from the page size. 511
516 * is default for backwards compat: liblmdb <= 0.9.10 can break
517 * when modifying a DB with keys/dupsort data bigger than its max.
518 * #MDB_DEVEL sets the default to 0.
520 * Data items in an #MDB_DUPSORT database are also limited to
521 * this size, since they're actually keys of a sub-DB. Keys and
522 * #MDB_DUPSORT data items must fit on a node in a regular page.
524 #ifndef MDB_MAXKEYSIZE
525 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
528 /** The maximum size of a key we can write to the environment. */
530 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
532 #define ENV_MAXKEY(env) ((env)->me_maxkey)
535 /** @brief The maximum size of a data item.
537 * We only store a 32 bit value for node sizes.
539 #define MAXDATASIZE 0xffffffffUL
542 /** Key size which fits in a #DKBUF.
545 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
548 * This is used for printing a hex dump of a key's contents.
550 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
551 /** Display a key in hex.
553 * Invoke a function to display a key in hex.
555 #define DKEY(x) mdb_dkey(x, kbuf)
561 /** An invalid page number.
562 * Mainly used to denote an empty tree.
564 #define P_INVALID (~(pgno_t)0)
566 /** Test if the flags \b f are set in a flag word \b w. */
567 #define F_ISSET(w, f) (((w) & (f)) == (f))
569 /** Round \b n up to an even number. */
570 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
572 /** Used for offsets within a single page.
573 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
576 typedef uint16_t indx_t;
578 /** Default size of memory map.
579 * This is certainly too small for any actual applications. Apps should always set
580 * the size explicitly using #mdb_env_set_mapsize().
582 #define DEFAULT_MAPSIZE 1048576
584 /** @defgroup readers Reader Lock Table
585 * Readers don't acquire any locks for their data access. Instead, they
586 * simply record their transaction ID in the reader table. The reader
587 * mutex is needed just to find an empty slot in the reader table. The
588 * slot's address is saved in thread-specific data so that subsequent read
589 * transactions started by the same thread need no further locking to proceed.
591 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
593 * No reader table is used if the database is on a read-only filesystem, or
594 * if #MDB_NOLOCK is set.
596 * Since the database uses multi-version concurrency control, readers don't
597 * actually need any locking. This table is used to keep track of which
598 * readers are using data from which old transactions, so that we'll know
599 * when a particular old transaction is no longer in use. Old transactions
600 * that have discarded any data pages can then have those pages reclaimed
601 * for use by a later write transaction.
603 * The lock table is constructed such that reader slots are aligned with the
604 * processor's cache line size. Any slot is only ever used by one thread.
605 * This alignment guarantees that there will be no contention or cache
606 * thrashing as threads update their own slot info, and also eliminates
607 * any need for locking when accessing a slot.
609 * A writer thread will scan every slot in the table to determine the oldest
610 * outstanding reader transaction. Any freed pages older than this will be
611 * reclaimed by the writer. The writer doesn't use any locks when scanning
612 * this table. This means that there's no guarantee that the writer will
613 * see the most up-to-date reader info, but that's not required for correct
614 * operation - all we need is to know the upper bound on the oldest reader,
615 * we don't care at all about the newest reader. So the only consequence of
616 * reading stale information here is that old pages might hang around a
617 * while longer before being reclaimed. That's actually good anyway, because
618 * the longer we delay reclaiming old pages, the more likely it is that a
619 * string of contiguous pages can be found after coalescing old pages from
620 * many old transactions together.
623 /** Number of slots in the reader table.
624 * This value was chosen somewhat arbitrarily. 126 readers plus a
625 * couple mutexes fit exactly into 8KB on my development machine.
626 * Applications should set the table size using #mdb_env_set_maxreaders().
628 #define DEFAULT_READERS 126
630 /** The size of a CPU cache line in bytes. We want our lock structures
631 * aligned to this size to avoid false cache line sharing in the
633 * This value works for most CPUs. For Itanium this should be 128.
639 /** The information we store in a single slot of the reader table.
640 * In addition to a transaction ID, we also record the process and
641 * thread ID that owns a slot, so that we can detect stale information,
642 * e.g. threads or processes that went away without cleaning up.
643 * @note We currently don't check for stale records. We simply re-init
644 * the table when we know that we're the only process opening the
647 typedef struct MDB_rxbody {
648 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
649 * Multiple readers that start at the same time will probably have the
650 * same ID here. Again, it's not important to exclude them from
651 * anything; all we need to know is which version of the DB they
652 * started from so we can avoid overwriting any data used in that
653 * particular version.
655 volatile txnid_t mrb_txnid;
656 /** The process ID of the process owning this reader txn. */
657 volatile MDB_PID_T mrb_pid;
658 /** The thread ID of the thread owning this txn. */
659 volatile MDB_THR_T mrb_tid;
662 /** The actual reader record, with cacheline padding. */
663 typedef struct MDB_reader {
666 /** shorthand for mrb_txnid */
667 #define mr_txnid mru.mrx.mrb_txnid
668 #define mr_pid mru.mrx.mrb_pid
669 #define mr_tid mru.mrx.mrb_tid
670 /** cache line alignment */
671 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
675 /** The header for the reader table.
676 * The table resides in a memory-mapped file. (This is a different file
677 * than is used for the main database.)
679 * For POSIX the actual mutexes reside in the shared memory of this
680 * mapped file. On Windows, mutexes are named objects allocated by the
681 * kernel; we store the mutex names in this mapped file so that other
682 * processes can grab them. This same approach is also used on
683 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
684 * process-shared POSIX mutexes. For these cases where a named object
685 * is used, the object name is derived from a 64 bit FNV hash of the
686 * environment pathname. As such, naming collisions are extremely
687 * unlikely. If a collision occurs, the results are unpredictable.
689 typedef struct MDB_txbody {
690 /** Stamp identifying this as an LMDB file. It must be set
693 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
696 char mtb_rmname[MNAME_LEN];
697 #elif defined(MDB_USE_SYSV_SEM)
701 /** Mutex protecting access to this table.
702 * This is the #MDB_MUTEX(env,r) reader table lock.
704 pthread_mutex_t mtb_rmutex;
706 /** The ID of the last transaction committed to the database.
707 * This is recorded here only for convenience; the value can always
708 * be determined by reading the main database meta pages.
710 volatile txnid_t mtb_txnid;
711 /** The number of slots that have been used in the reader table.
712 * This always records the maximum count, it is not decremented
713 * when readers release their slots.
715 volatile unsigned mtb_numreaders;
718 /** The actual reader table definition. */
719 typedef struct MDB_txninfo {
722 #define mti_magic mt1.mtb.mtb_magic
723 #define mti_format mt1.mtb.mtb_format
724 #define mti_rmutex mt1.mtb.mtb_rmutex
725 #define mti_rmname mt1.mtb.mtb_rmname
726 #define mti_txnid mt1.mtb.mtb_txnid
727 #define mti_numreaders mt1.mtb.mtb_numreaders
728 #ifdef MDB_USE_SYSV_SEM
729 #define mti_semid mt1.mtb.mtb_semid
730 #define mti_rlocked mt1.mtb.mtb_rlocked
732 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
736 char mt2_wmname[MNAME_LEN];
737 #define mti_wmname mt2.mt2_wmname
738 #elif defined MDB_USE_SYSV_SEM
740 #define mti_wlocked mt2.mt2_wlocked
742 pthread_mutex_t mt2_wmutex;
743 #define mti_wmutex mt2.mt2_wmutex
745 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
747 MDB_reader mti_readers[1];
750 /** Lockfile format signature: version, features and field layout */
751 #define MDB_LOCK_FORMAT \
753 ((MDB_LOCK_VERSION) \
754 /* Flags which describe functionality */ \
755 + (SYSV_SEM_FLAG << 18) \
756 + (((MDB_PIDLOCK) != 0) << 16)))
759 /** Common header for all page types.
760 * Overflow records occupy a number of contiguous pages with no
761 * headers on any page after the first.
763 typedef struct MDB_page {
764 #define mp_pgno mp_p.p_pgno
765 #define mp_next mp_p.p_next
767 pgno_t p_pgno; /**< page number */
768 struct MDB_page *p_next; /**< for in-memory list of freed pages */
771 /** @defgroup mdb_page Page Flags
773 * Flags for the page headers.
776 #define P_BRANCH 0x01 /**< branch page */
777 #define P_LEAF 0x02 /**< leaf page */
778 #define P_OVERFLOW 0x04 /**< overflow page */
779 #define P_META 0x08 /**< meta page */
780 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
781 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
782 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
783 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
784 #define P_KEEP 0x8000 /**< leave this page alone during spill */
786 uint16_t mp_flags; /**< @ref mdb_page */
787 #define mp_lower mp_pb.pb.pb_lower
788 #define mp_upper mp_pb.pb.pb_upper
789 #define mp_pages mp_pb.pb_pages
792 indx_t pb_lower; /**< lower bound of free space */
793 indx_t pb_upper; /**< upper bound of free space */
795 uint32_t pb_pages; /**< number of overflow pages */
797 indx_t mp_ptrs[1]; /**< dynamic size */
800 /** Size of the page header, excluding dynamic data at the end */
801 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
803 /** Address of first usable data byte in a page, after the header */
804 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
806 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
807 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
809 /** Number of nodes on a page */
810 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
812 /** The amount of space remaining in the page */
813 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
815 /** The percentage of space used in the page, in tenths of a percent. */
816 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
817 ((env)->me_psize - PAGEHDRSZ))
818 /** The minimum page fill factor, in tenths of a percent.
819 * Pages emptier than this are candidates for merging.
821 #define FILL_THRESHOLD 250
823 /** Test if a page is a leaf page */
824 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
825 /** Test if a page is a LEAF2 page */
826 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
827 /** Test if a page is a branch page */
828 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
829 /** Test if a page is an overflow page */
830 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
831 /** Test if a page is a sub page */
832 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
834 /** The number of overflow pages needed to store the given size. */
835 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
837 /** Link in #MDB_txn.%mt_loose_pgs list */
838 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
840 /** Header for a single key/data pair within a page.
841 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
842 * We guarantee 2-byte alignment for 'MDB_node's.
844 typedef struct MDB_node {
845 /** lo and hi are used for data size on leaf nodes and for
846 * child pgno on branch nodes. On 64 bit platforms, flags
847 * is also used for pgno. (Branch nodes have no flags).
848 * They are in host byte order in case that lets some
849 * accesses be optimized into a 32-bit word access.
851 #if BYTE_ORDER == LITTLE_ENDIAN
852 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
854 unsigned short mn_hi, mn_lo;
856 /** @defgroup mdb_node Node Flags
858 * Flags for node headers.
861 #define F_BIGDATA 0x01 /**< data put on overflow page */
862 #define F_SUBDATA 0x02 /**< data is a sub-database */
863 #define F_DUPDATA 0x04 /**< data has duplicates */
865 /** valid flags for #mdb_node_add() */
866 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
869 unsigned short mn_flags; /**< @ref mdb_node */
870 unsigned short mn_ksize; /**< key size */
871 char mn_data[1]; /**< key and data are appended here */
874 /** Size of the node header, excluding dynamic data at the end */
875 #define NODESIZE offsetof(MDB_node, mn_data)
877 /** Bit position of top word in page number, for shifting mn_flags */
878 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
880 /** Size of a node in a branch page with a given key.
881 * This is just the node header plus the key, there is no data.
883 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
885 /** Size of a node in a leaf page with a given key and data.
886 * This is node header plus key plus data size.
888 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
890 /** Address of node \b i in page \b p */
891 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
893 /** Address of the key for the node */
894 #define NODEKEY(node) (void *)((node)->mn_data)
896 /** Address of the data for a node */
897 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
899 /** Get the page number pointed to by a branch node */
900 #define NODEPGNO(node) \
901 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
902 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
903 /** Set the page number in a branch node */
904 #define SETPGNO(node,pgno) do { \
905 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
906 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
908 /** Get the size of the data in a leaf node */
909 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
910 /** Set the size of the data for a leaf node */
911 #define SETDSZ(node,size) do { \
912 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
913 /** The size of a key in a node */
914 #define NODEKSZ(node) ((node)->mn_ksize)
916 /** Copy a page number from src to dst */
918 #define COPY_PGNO(dst,src) dst = src
920 #if SIZE_MAX > 4294967295UL
921 #define COPY_PGNO(dst,src) do { \
922 unsigned short *s, *d; \
923 s = (unsigned short *)&(src); \
924 d = (unsigned short *)&(dst); \
931 #define COPY_PGNO(dst,src) do { \
932 unsigned short *s, *d; \
933 s = (unsigned short *)&(src); \
934 d = (unsigned short *)&(dst); \
940 /** The address of a key in a LEAF2 page.
941 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
942 * There are no node headers, keys are stored contiguously.
944 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
946 /** Set the \b node's key into \b keyptr, if requested. */
947 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
948 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
950 /** Set the \b node's key into \b key. */
951 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
953 /** Information about a single database in the environment. */
954 typedef struct MDB_db {
955 uint32_t md_pad; /**< also ksize for LEAF2 pages */
956 uint16_t md_flags; /**< @ref mdb_dbi_open */
957 uint16_t md_depth; /**< depth of this tree */
958 pgno_t md_branch_pages; /**< number of internal pages */
959 pgno_t md_leaf_pages; /**< number of leaf pages */
960 pgno_t md_overflow_pages; /**< number of overflow pages */
961 size_t md_entries; /**< number of data items */
962 pgno_t md_root; /**< the root page of this tree */
965 /** mdb_dbi_open flags */
966 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
967 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
968 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
969 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
971 /** Handle for the DB used to track free pages. */
973 /** Handle for the default DB. */
976 /** Meta page content.
977 * A meta page is the start point for accessing a database snapshot.
978 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
980 typedef struct MDB_meta {
981 /** Stamp identifying this as an LMDB file. It must be set
984 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
986 void *mm_address; /**< address for fixed mapping */
987 size_t mm_mapsize; /**< size of mmap region */
988 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
989 /** The size of pages used in this DB */
990 #define mm_psize mm_dbs[0].md_pad
991 /** Any persistent environment flags. @ref mdb_env */
992 #define mm_flags mm_dbs[0].md_flags
993 pgno_t mm_last_pg; /**< last used page in file */
994 volatile txnid_t mm_txnid; /**< txnid that committed this page */
997 /** Buffer for a stack-allocated meta page.
998 * The members define size and alignment, and silence type
999 * aliasing warnings. They are not used directly; that could
1000 * mean incorrectly using several union members in parallel.
1002 typedef union MDB_metabuf {
1005 char mm_pad[PAGEHDRSZ];
1010 /** Auxiliary DB info.
1011 * The information here is mostly static/read-only. There is
1012 * only a single copy of this record in the environment.
1014 typedef struct MDB_dbx {
1015 MDB_val md_name; /**< name of the database */
1016 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1017 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1018 MDB_rel_func *md_rel; /**< user relocate function */
1019 void *md_relctx; /**< user-provided context for md_rel */
1022 /** A database transaction.
1023 * Every operation requires a transaction handle.
1026 MDB_txn *mt_parent; /**< parent of a nested txn */
1027 MDB_txn *mt_child; /**< nested txn under this txn */
1028 pgno_t mt_next_pgno; /**< next unallocated page */
1029 /** The ID of this transaction. IDs are integers incrementing from 1.
1030 * Only committed write transactions increment the ID. If a transaction
1031 * aborts, the ID may be re-used by the next writer.
1034 MDB_env *mt_env; /**< the DB environment */
1035 /** The list of pages that became unused during this transaction.
1037 MDB_IDL mt_free_pgs;
1038 /** The list of loose pages that became unused and may be reused
1039 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1041 MDB_page *mt_loose_pgs;
1042 /* #Number of loose pages (#mt_loose_pgs) */
1044 /** The sorted list of dirty pages we temporarily wrote to disk
1045 * because the dirty list was full. page numbers in here are
1046 * shifted left by 1, deleted slots have the LSB set.
1048 MDB_IDL mt_spill_pgs;
1050 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1051 MDB_ID2L dirty_list;
1052 /** For read txns: This thread/txn's reader table slot, or NULL. */
1055 /** Array of records for each DB known in the environment. */
1057 /** Array of MDB_db records for each known DB */
1059 /** Array of sequence numbers for each DB handle */
1060 unsigned int *mt_dbiseqs;
1061 /** @defgroup mt_dbflag Transaction DB Flags
1065 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1066 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1067 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1068 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1070 /** In write txns, array of cursors for each DB */
1071 MDB_cursor **mt_cursors;
1072 /** Array of flags for each DB */
1073 unsigned char *mt_dbflags;
1074 /** Number of DB records in use. This number only ever increments;
1075 * we don't decrement it when individual DB handles are closed.
1079 /** @defgroup mdb_txn Transaction Flags
1083 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1084 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1085 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1086 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1087 #define MDB_TXN_NOSYNC 0x10 /**< don't sync this txn on commit */
1088 #define MDB_TXN_NOMETASYNC 0x20 /**< don't sync meta for this txn on commit */
1090 unsigned int mt_flags; /**< @ref mdb_txn */
1091 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1092 * Includes ancestor txns' dirty pages not hidden by other txns'
1093 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1094 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1096 unsigned int mt_dirty_room;
1099 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1100 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1101 * raise this on a 64 bit machine.
1103 #define CURSOR_STACK 32
1107 /** Cursors are used for all DB operations.
1108 * A cursor holds a path of (page pointer, key index) from the DB
1109 * root to a position in the DB, plus other state. #MDB_DUPSORT
1110 * cursors include an xcursor to the current data item. Write txns
1111 * track their cursors and keep them up to date when data moves.
1112 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1113 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1116 /** Next cursor on this DB in this txn */
1117 MDB_cursor *mc_next;
1118 /** Backup of the original cursor if this cursor is a shadow */
1119 MDB_cursor *mc_backup;
1120 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1121 struct MDB_xcursor *mc_xcursor;
1122 /** The transaction that owns this cursor */
1124 /** The database handle this cursor operates on */
1126 /** The database record for this cursor */
1128 /** The database auxiliary record for this cursor */
1130 /** The @ref mt_dbflag for this database */
1131 unsigned char *mc_dbflag;
1132 unsigned short mc_snum; /**< number of pushed pages */
1133 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1134 /** @defgroup mdb_cursor Cursor Flags
1136 * Cursor state flags.
1139 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1140 #define C_EOF 0x02 /**< No more data */
1141 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1142 #define C_DEL 0x08 /**< last op was a cursor_del */
1143 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1144 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1146 unsigned int mc_flags; /**< @ref mdb_cursor */
1147 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1148 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1151 /** Context for sorted-dup records.
1152 * We could have gone to a fully recursive design, with arbitrarily
1153 * deep nesting of sub-databases. But for now we only handle these
1154 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1156 typedef struct MDB_xcursor {
1157 /** A sub-cursor for traversing the Dup DB */
1158 MDB_cursor mx_cursor;
1159 /** The database record for this Dup DB */
1161 /** The auxiliary DB record for this Dup DB */
1163 /** The @ref mt_dbflag for this Dup DB */
1164 unsigned char mx_dbflag;
1167 /** State of FreeDB old pages, stored in the MDB_env */
1168 typedef struct MDB_pgstate {
1169 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1170 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1173 /** The database environment. */
1175 HANDLE me_fd; /**< The main data file */
1176 HANDLE me_lfd; /**< The lock file */
1177 HANDLE me_mfd; /**< just for writing the meta pages */
1178 /** Failed to update the meta page. Probably an I/O error. */
1179 #define MDB_FATAL_ERROR 0x80000000U
1180 /** Some fields are initialized. */
1181 #define MDB_ENV_ACTIVE 0x20000000U
1182 /** me_txkey is set */
1183 #define MDB_ENV_TXKEY 0x10000000U
1184 /** fdatasync is unreliable */
1185 #define MDB_FSYNCONLY 0x08000000U
1186 uint32_t me_flags; /**< @ref mdb_env */
1187 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1188 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1189 unsigned int me_maxreaders; /**< size of the reader table */
1190 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1191 volatile int me_close_readers;
1192 MDB_dbi me_numdbs; /**< number of DBs opened */
1193 MDB_dbi me_maxdbs; /**< size of the DB table */
1194 MDB_PID_T me_pid; /**< process ID of this env */
1195 char *me_path; /**< path to the DB files */
1196 char *me_map; /**< the memory map of the data file */
1197 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1198 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1199 void *me_pbuf; /**< scratch area for DUPSORT put() */
1200 MDB_txn *me_txn; /**< current write transaction */
1201 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1202 size_t me_mapsize; /**< size of the data memory map */
1203 off_t me_size; /**< current file size */
1204 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1205 MDB_dbx *me_dbxs; /**< array of static DB info */
1206 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1207 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1208 pthread_key_t me_txkey; /**< thread-key for readers */
1209 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1210 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1211 # define me_pglast me_pgstate.mf_pglast
1212 # define me_pghead me_pgstate.mf_pghead
1213 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1214 /** IDL of pages that became unused in a write txn */
1215 MDB_IDL me_free_pgs;
1216 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1217 MDB_ID2L me_dirty_list;
1218 /** Max number of freelist items that can fit in a single overflow page */
1220 /** Max size of a node on a page */
1221 unsigned int me_nodemax;
1222 #if !(MDB_MAXKEYSIZE)
1223 unsigned int me_maxkey; /**< max size of a key */
1225 int me_live_reader; /**< have liveness lock in reader table */
1227 int me_pidquery; /**< Used in OpenProcess */
1229 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM)
1230 /* Windows mutexes/SysV semaphores do not reside in shared mem */
1231 mdb_mutex_t me_rmutex;
1232 mdb_mutex_t me_wmutex;
1234 void *me_userctx; /**< User-settable context */
1235 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1238 /** Nested transaction */
1239 typedef struct MDB_ntxn {
1240 MDB_txn mnt_txn; /**< the transaction */
1241 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1244 /** max number of pages to commit in one writev() call */
1245 #define MDB_COMMIT_PAGES 64
1246 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1247 #undef MDB_COMMIT_PAGES
1248 #define MDB_COMMIT_PAGES IOV_MAX
1251 /** max bytes to write in one call */
1252 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1254 /** Check \b txn and \b dbi arguments to a function */
1255 #define TXN_DBI_EXIST(txn, dbi) \
1256 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1258 /** Check for misused \b dbi handles */
1259 #define TXN_DBI_CHANGED(txn, dbi) \
1260 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1262 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1263 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1264 static int mdb_page_touch(MDB_cursor *mc);
1266 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1267 static int mdb_page_search_root(MDB_cursor *mc,
1268 MDB_val *key, int modify);
1269 #define MDB_PS_MODIFY 1
1270 #define MDB_PS_ROOTONLY 2
1271 #define MDB_PS_FIRST 4
1272 #define MDB_PS_LAST 8
1273 static int mdb_page_search(MDB_cursor *mc,
1274 MDB_val *key, int flags);
1275 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1277 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1278 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1279 pgno_t newpgno, unsigned int nflags);
1281 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1282 static int mdb_env_pick_meta(const MDB_env *env);
1283 static int mdb_env_write_meta(MDB_txn *txn);
1284 #if !(defined(_WIN32) || defined(MDB_USE_SYSV_SEM)) /* Drop unused excl arg */
1285 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1287 static void mdb_env_close0(MDB_env *env, int excl);
1289 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1290 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1291 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1292 static void mdb_node_del(MDB_cursor *mc, int ksize);
1293 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1294 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1295 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1296 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1297 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1299 static int mdb_rebalance(MDB_cursor *mc);
1300 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1302 static void mdb_cursor_pop(MDB_cursor *mc);
1303 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1305 static int mdb_cursor_del0(MDB_cursor *mc);
1306 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1307 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1308 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1309 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1310 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1312 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1313 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1315 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1316 static void mdb_xcursor_init0(MDB_cursor *mc);
1317 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1319 static int mdb_drop0(MDB_cursor *mc, int subs);
1320 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1321 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1324 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1328 static SECURITY_DESCRIPTOR mdb_null_sd;
1329 static SECURITY_ATTRIBUTES mdb_all_sa;
1330 static int mdb_sec_inited;
1333 /** Return the library version info. */
1335 mdb_version(int *major, int *minor, int *patch)
1337 if (major) *major = MDB_VERSION_MAJOR;
1338 if (minor) *minor = MDB_VERSION_MINOR;
1339 if (patch) *patch = MDB_VERSION_PATCH;
1340 return MDB_VERSION_STRING;
1343 /** Table of descriptions for LMDB @ref errors */
1344 static char *const mdb_errstr[] = {
1345 "MDB_KEYEXIST: Key/data pair already exists",
1346 "MDB_NOTFOUND: No matching key/data pair found",
1347 "MDB_PAGE_NOTFOUND: Requested page not found",
1348 "MDB_CORRUPTED: Located page was wrong type",
1349 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1350 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1351 "MDB_INVALID: File is not an LMDB file",
1352 "MDB_MAP_FULL: Environment mapsize limit reached",
1353 "MDB_DBS_FULL: Environment maxdbs limit reached",
1354 "MDB_READERS_FULL: Environment maxreaders limit reached",
1355 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1356 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1357 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1358 "MDB_PAGE_FULL: Internal error - page has no more space",
1359 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1360 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1361 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1362 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1363 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1364 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1368 mdb_strerror(int err)
1371 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1372 * This works as long as no function between the call to mdb_strerror
1373 * and the actual use of the message uses more than 4K of stack.
1376 char buf[1024], *ptr = buf;
1380 return ("Successful return: 0");
1382 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1383 i = err - MDB_KEYEXIST;
1384 return mdb_errstr[i];
1388 /* These are the C-runtime error codes we use. The comment indicates
1389 * their numeric value, and the Win32 error they would correspond to
1390 * if the error actually came from a Win32 API. A major mess, we should
1391 * have used LMDB-specific error codes for everything.
1394 case ENOENT: /* 2, FILE_NOT_FOUND */
1395 case EIO: /* 5, ACCESS_DENIED */
1396 case ENOMEM: /* 12, INVALID_ACCESS */
1397 case EACCES: /* 13, INVALID_DATA */
1398 case EBUSY: /* 16, CURRENT_DIRECTORY */
1399 case EINVAL: /* 22, BAD_COMMAND */
1400 case ENOSPC: /* 28, OUT_OF_PAPER */
1401 return strerror(err);
1406 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1407 FORMAT_MESSAGE_IGNORE_INSERTS,
1408 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1411 return strerror(err);
1415 /** assert(3) variant in cursor context */
1416 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1417 /** assert(3) variant in transaction context */
1418 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1419 /** assert(3) variant in environment context */
1420 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1423 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1424 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1427 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1428 const char *func, const char *file, int line)
1431 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1432 file, line, expr_txt, func);
1433 if (env->me_assert_func)
1434 env->me_assert_func(env, buf);
1435 fprintf(stderr, "%s\n", buf);
1439 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1443 /** Return the page number of \b mp which may be sub-page, for debug output */
1445 mdb_dbg_pgno(MDB_page *mp)
1448 COPY_PGNO(ret, mp->mp_pgno);
1452 /** Display a key in hexadecimal and return the address of the result.
1453 * @param[in] key the key to display
1454 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1455 * @return The key in hexadecimal form.
1458 mdb_dkey(MDB_val *key, char *buf)
1461 unsigned char *c = key->mv_data;
1467 if (key->mv_size > DKBUF_MAXKEYSIZE)
1468 return "MDB_MAXKEYSIZE";
1469 /* may want to make this a dynamic check: if the key is mostly
1470 * printable characters, print it as-is instead of converting to hex.
1474 for (i=0; i<key->mv_size; i++)
1475 ptr += sprintf(ptr, "%02x", *c++);
1477 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1483 mdb_leafnode_type(MDB_node *n)
1485 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1486 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1487 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1490 /** Display all the keys in the page. */
1492 mdb_page_list(MDB_page *mp)
1494 pgno_t pgno = mdb_dbg_pgno(mp);
1495 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1497 unsigned int i, nkeys, nsize, total = 0;
1501 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1502 case P_BRANCH: type = "Branch page"; break;
1503 case P_LEAF: type = "Leaf page"; break;
1504 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1505 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1506 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1508 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1509 pgno, mp->mp_pages, state);
1512 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1513 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1516 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1520 nkeys = NUMKEYS(mp);
1521 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1523 for (i=0; i<nkeys; i++) {
1524 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1525 key.mv_size = nsize = mp->mp_pad;
1526 key.mv_data = LEAF2KEY(mp, i, nsize);
1528 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1531 node = NODEPTR(mp, i);
1532 key.mv_size = node->mn_ksize;
1533 key.mv_data = node->mn_data;
1534 nsize = NODESIZE + key.mv_size;
1535 if (IS_BRANCH(mp)) {
1536 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1540 if (F_ISSET(node->mn_flags, F_BIGDATA))
1541 nsize += sizeof(pgno_t);
1543 nsize += NODEDSZ(node);
1545 nsize += sizeof(indx_t);
1546 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1547 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1549 total = EVEN(total);
1551 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1552 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1556 mdb_cursor_chk(MDB_cursor *mc)
1562 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1563 for (i=0; i<mc->mc_top; i++) {
1565 node = NODEPTR(mp, mc->mc_ki[i]);
1566 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1569 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1575 /** Count all the pages in each DB and in the freelist
1576 * and make sure it matches the actual number of pages
1578 * All named DBs must be open for a correct count.
1580 static void mdb_audit(MDB_txn *txn)
1584 MDB_ID freecount, count;
1589 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1590 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1591 freecount += *(MDB_ID *)data.mv_data;
1592 mdb_tassert(txn, rc == MDB_NOTFOUND);
1595 for (i = 0; i<txn->mt_numdbs; i++) {
1597 if (!(txn->mt_dbflags[i] & DB_VALID))
1599 mdb_cursor_init(&mc, txn, i, &mx);
1600 if (txn->mt_dbs[i].md_root == P_INVALID)
1602 count += txn->mt_dbs[i].md_branch_pages +
1603 txn->mt_dbs[i].md_leaf_pages +
1604 txn->mt_dbs[i].md_overflow_pages;
1605 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1606 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1607 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1610 mp = mc.mc_pg[mc.mc_top];
1611 for (j=0; j<NUMKEYS(mp); j++) {
1612 MDB_node *leaf = NODEPTR(mp, j);
1613 if (leaf->mn_flags & F_SUBDATA) {
1615 memcpy(&db, NODEDATA(leaf), sizeof(db));
1616 count += db.md_branch_pages + db.md_leaf_pages +
1617 db.md_overflow_pages;
1621 mdb_tassert(txn, rc == MDB_NOTFOUND);
1624 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1625 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1626 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1632 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1634 return txn->mt_dbxs[dbi].md_cmp(a, b);
1638 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1640 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1643 /** Allocate memory for a page.
1644 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1647 mdb_page_malloc(MDB_txn *txn, unsigned num)
1649 MDB_env *env = txn->mt_env;
1650 MDB_page *ret = env->me_dpages;
1651 size_t psize = env->me_psize, sz = psize, off;
1652 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1653 * For a single page alloc, we init everything after the page header.
1654 * For multi-page, we init the final page; if the caller needed that
1655 * many pages they will be filling in at least up to the last page.
1659 VGMEMP_ALLOC(env, ret, sz);
1660 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1661 env->me_dpages = ret->mp_next;
1664 psize -= off = PAGEHDRSZ;
1669 if ((ret = malloc(sz)) != NULL) {
1670 VGMEMP_ALLOC(env, ret, sz);
1671 if (!(env->me_flags & MDB_NOMEMINIT)) {
1672 memset((char *)ret + off, 0, psize);
1676 txn->mt_flags |= MDB_TXN_ERROR;
1680 /** Free a single page.
1681 * Saves single pages to a list, for future reuse.
1682 * (This is not used for multi-page overflow pages.)
1685 mdb_page_free(MDB_env *env, MDB_page *mp)
1687 mp->mp_next = env->me_dpages;
1688 VGMEMP_FREE(env, mp);
1689 env->me_dpages = mp;
1692 /** Free a dirty page */
1694 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1696 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1697 mdb_page_free(env, dp);
1699 /* large pages just get freed directly */
1700 VGMEMP_FREE(env, dp);
1705 /** Return all dirty pages to dpage list */
1707 mdb_dlist_free(MDB_txn *txn)
1709 MDB_env *env = txn->mt_env;
1710 MDB_ID2L dl = txn->mt_u.dirty_list;
1711 unsigned i, n = dl[0].mid;
1713 for (i = 1; i <= n; i++) {
1714 mdb_dpage_free(env, dl[i].mptr);
1719 /** Loosen or free a single page.
1720 * Saves single pages to a list for future reuse
1721 * in this same txn. It has been pulled from the freeDB
1722 * and already resides on the dirty list, but has been
1723 * deleted. Use these pages first before pulling again
1726 * If the page wasn't dirtied in this txn, just add it
1727 * to this txn's free list.
1730 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1733 pgno_t pgno = mp->mp_pgno;
1734 MDB_txn *txn = mc->mc_txn;
1736 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1737 if (txn->mt_parent) {
1738 MDB_ID2 *dl = txn->mt_u.dirty_list;
1739 /* If txn has a parent, make sure the page is in our
1743 unsigned x = mdb_mid2l_search(dl, pgno);
1744 if (x <= dl[0].mid && dl[x].mid == pgno) {
1745 if (mp != dl[x].mptr) { /* bad cursor? */
1746 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1747 txn->mt_flags |= MDB_TXN_ERROR;
1748 return MDB_CORRUPTED;
1755 /* no parent txn, so it's just ours */
1760 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1762 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1763 txn->mt_loose_pgs = mp;
1764 txn->mt_loose_count++;
1765 mp->mp_flags |= P_LOOSE;
1767 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1775 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1776 * @param[in] mc A cursor handle for the current operation.
1777 * @param[in] pflags Flags of the pages to update:
1778 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1779 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1780 * @return 0 on success, non-zero on failure.
1783 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1785 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1786 MDB_txn *txn = mc->mc_txn;
1792 int rc = MDB_SUCCESS, level;
1794 /* Mark pages seen by cursors */
1795 if (mc->mc_flags & C_UNTRACK)
1796 mc = NULL; /* will find mc in mt_cursors */
1797 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1798 for (; mc; mc=mc->mc_next) {
1799 if (!(mc->mc_flags & C_INITIALIZED))
1801 for (m3 = mc;; m3 = &mx->mx_cursor) {
1803 for (j=0; j<m3->mc_snum; j++) {
1805 if ((mp->mp_flags & Mask) == pflags)
1806 mp->mp_flags ^= P_KEEP;
1808 mx = m3->mc_xcursor;
1809 /* Proceed to mx if it is at a sub-database */
1810 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1812 if (! (mp && (mp->mp_flags & P_LEAF)))
1814 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1815 if (!(leaf->mn_flags & F_SUBDATA))
1824 /* Mark dirty root pages */
1825 for (i=0; i<txn->mt_numdbs; i++) {
1826 if (txn->mt_dbflags[i] & DB_DIRTY) {
1827 pgno_t pgno = txn->mt_dbs[i].md_root;
1828 if (pgno == P_INVALID)
1830 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1832 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1833 dp->mp_flags ^= P_KEEP;
1841 static int mdb_page_flush(MDB_txn *txn, int keep);
1843 /** Spill pages from the dirty list back to disk.
1844 * This is intended to prevent running into #MDB_TXN_FULL situations,
1845 * but note that they may still occur in a few cases:
1846 * 1) our estimate of the txn size could be too small. Currently this
1847 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1848 * 2) child txns may run out of space if their parents dirtied a
1849 * lot of pages and never spilled them. TODO: we probably should do
1850 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1851 * the parent's dirty_room is below a given threshold.
1853 * Otherwise, if not using nested txns, it is expected that apps will
1854 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1855 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1856 * If the txn never references them again, they can be left alone.
1857 * If the txn only reads them, they can be used without any fuss.
1858 * If the txn writes them again, they can be dirtied immediately without
1859 * going thru all of the work of #mdb_page_touch(). Such references are
1860 * handled by #mdb_page_unspill().
1862 * Also note, we never spill DB root pages, nor pages of active cursors,
1863 * because we'll need these back again soon anyway. And in nested txns,
1864 * we can't spill a page in a child txn if it was already spilled in a
1865 * parent txn. That would alter the parent txns' data even though
1866 * the child hasn't committed yet, and we'd have no way to undo it if
1867 * the child aborted.
1869 * @param[in] m0 cursor A cursor handle identifying the transaction and
1870 * database for which we are checking space.
1871 * @param[in] key For a put operation, the key being stored.
1872 * @param[in] data For a put operation, the data being stored.
1873 * @return 0 on success, non-zero on failure.
1876 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1878 MDB_txn *txn = m0->mc_txn;
1880 MDB_ID2L dl = txn->mt_u.dirty_list;
1881 unsigned int i, j, need;
1884 if (m0->mc_flags & C_SUB)
1887 /* Estimate how much space this op will take */
1888 i = m0->mc_db->md_depth;
1889 /* Named DBs also dirty the main DB */
1890 if (m0->mc_dbi > MAIN_DBI)
1891 i += txn->mt_dbs[MAIN_DBI].md_depth;
1892 /* For puts, roughly factor in the key+data size */
1894 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1895 i += i; /* double it for good measure */
1898 if (txn->mt_dirty_room > i)
1901 if (!txn->mt_spill_pgs) {
1902 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1903 if (!txn->mt_spill_pgs)
1906 /* purge deleted slots */
1907 MDB_IDL sl = txn->mt_spill_pgs;
1908 unsigned int num = sl[0];
1910 for (i=1; i<=num; i++) {
1917 /* Preserve pages which may soon be dirtied again */
1918 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1921 /* Less aggressive spill - we originally spilled the entire dirty list,
1922 * with a few exceptions for cursor pages and DB root pages. But this
1923 * turns out to be a lot of wasted effort because in a large txn many
1924 * of those pages will need to be used again. So now we spill only 1/8th
1925 * of the dirty pages. Testing revealed this to be a good tradeoff,
1926 * better than 1/2, 1/4, or 1/10.
1928 if (need < MDB_IDL_UM_MAX / 8)
1929 need = MDB_IDL_UM_MAX / 8;
1931 /* Save the page IDs of all the pages we're flushing */
1932 /* flush from the tail forward, this saves a lot of shifting later on. */
1933 for (i=dl[0].mid; i && need; i--) {
1934 MDB_ID pn = dl[i].mid << 1;
1936 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1938 /* Can't spill twice, make sure it's not already in a parent's
1941 if (txn->mt_parent) {
1943 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1944 if (tx2->mt_spill_pgs) {
1945 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1946 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1947 dp->mp_flags |= P_KEEP;
1955 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1959 mdb_midl_sort(txn->mt_spill_pgs);
1961 /* Flush the spilled part of dirty list */
1962 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1965 /* Reset any dirty pages we kept that page_flush didn't see */
1966 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1969 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1973 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1975 mdb_find_oldest(MDB_txn *txn)
1978 txnid_t mr, oldest = txn->mt_txnid - 1;
1979 if (txn->mt_env->me_txns) {
1980 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1981 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1992 /** Add a page to the txn's dirty list */
1994 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1997 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1999 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
2000 insert = mdb_mid2l_append;
2002 insert = mdb_mid2l_insert;
2004 mid.mid = mp->mp_pgno;
2006 rc = insert(txn->mt_u.dirty_list, &mid);
2007 mdb_tassert(txn, rc == 0);
2008 txn->mt_dirty_room--;
2011 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2012 * me_pghead and mt_next_pgno.
2014 * If there are free pages available from older transactions, they
2015 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2016 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2017 * and move me_pglast to say which records were consumed. Only this
2018 * function can create me_pghead and move me_pglast/mt_next_pgno.
2019 * @param[in] mc cursor A cursor handle identifying the transaction and
2020 * database for which we are allocating.
2021 * @param[in] num the number of pages to allocate.
2022 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2023 * will always be satisfied by a single contiguous chunk of memory.
2024 * @return 0 on success, non-zero on failure.
2027 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2029 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2030 /* Get at most <Max_retries> more freeDB records once me_pghead
2031 * has enough pages. If not enough, use new pages from the map.
2032 * If <Paranoid> and mc is updating the freeDB, only get new
2033 * records if me_pghead is empty. Then the freelist cannot play
2034 * catch-up with itself by growing while trying to save it.
2036 enum { Paranoid = 1, Max_retries = 500 };
2038 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2040 int rc, retry = num * 60;
2041 MDB_txn *txn = mc->mc_txn;
2042 MDB_env *env = txn->mt_env;
2043 pgno_t pgno, *mop = env->me_pghead;
2044 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2046 txnid_t oldest = 0, last;
2051 /* If there are any loose pages, just use them */
2052 if (num == 1 && txn->mt_loose_pgs) {
2053 np = txn->mt_loose_pgs;
2054 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2055 txn->mt_loose_count--;
2056 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2064 /* If our dirty list is already full, we can't do anything */
2065 if (txn->mt_dirty_room == 0) {
2070 for (op = MDB_FIRST;; op = MDB_NEXT) {
2075 /* Seek a big enough contiguous page range. Prefer
2076 * pages at the tail, just truncating the list.
2082 if (mop[i-n2] == pgno+n2)
2089 if (op == MDB_FIRST) { /* 1st iteration */
2090 /* Prepare to fetch more and coalesce */
2091 last = env->me_pglast;
2092 oldest = env->me_pgoldest;
2093 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2096 key.mv_data = &last; /* will look up last+1 */
2097 key.mv_size = sizeof(last);
2099 if (Paranoid && mc->mc_dbi == FREE_DBI)
2102 if (Paranoid && retry < 0 && mop_len)
2106 /* Do not fetch more if the record will be too recent */
2107 if (oldest <= last) {
2109 oldest = mdb_find_oldest(txn);
2110 env->me_pgoldest = oldest;
2116 rc = mdb_cursor_get(&m2, &key, NULL, op);
2118 if (rc == MDB_NOTFOUND)
2122 last = *(txnid_t*)key.mv_data;
2123 if (oldest <= last) {
2125 oldest = mdb_find_oldest(txn);
2126 env->me_pgoldest = oldest;
2132 np = m2.mc_pg[m2.mc_top];
2133 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2134 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2137 idl = (MDB_ID *) data.mv_data;
2140 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2145 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2147 mop = env->me_pghead;
2149 env->me_pglast = last;
2151 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2152 last, txn->mt_dbs[FREE_DBI].md_root, i));
2154 DPRINTF(("IDL %"Z"u", idl[j]));
2156 /* Merge in descending sorted order */
2157 mdb_midl_xmerge(mop, idl);
2161 /* Use new pages from the map when nothing suitable in the freeDB */
2163 pgno = txn->mt_next_pgno;
2164 if (pgno + num >= env->me_maxpg) {
2165 DPUTS("DB size maxed out");
2171 if (env->me_flags & MDB_WRITEMAP) {
2172 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2174 if (!(np = mdb_page_malloc(txn, num))) {
2180 mop[0] = mop_len -= num;
2181 /* Move any stragglers down */
2182 for (j = i-num; j < mop_len; )
2183 mop[++j] = mop[++i];
2185 txn->mt_next_pgno = pgno + num;
2188 mdb_page_dirty(txn, np);
2194 txn->mt_flags |= MDB_TXN_ERROR;
2198 /** Copy the used portions of a non-overflow page.
2199 * @param[in] dst page to copy into
2200 * @param[in] src page to copy from
2201 * @param[in] psize size of a page
2204 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2206 enum { Align = sizeof(pgno_t) };
2207 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2209 /* If page isn't full, just copy the used portion. Adjust
2210 * alignment so memcpy may copy words instead of bytes.
2212 if ((unused &= -Align) && !IS_LEAF2(src)) {
2213 upper = (upper + PAGEBASE) & -Align;
2214 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2215 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2218 memcpy(dst, src, psize - unused);
2222 /** Pull a page off the txn's spill list, if present.
2223 * If a page being referenced was spilled to disk in this txn, bring
2224 * it back and make it dirty/writable again.
2225 * @param[in] txn the transaction handle.
2226 * @param[in] mp the page being referenced. It must not be dirty.
2227 * @param[out] ret the writable page, if any. ret is unchanged if
2228 * mp wasn't spilled.
2231 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2233 MDB_env *env = txn->mt_env;
2236 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2238 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2239 if (!tx2->mt_spill_pgs)
2241 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2242 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2245 if (txn->mt_dirty_room == 0)
2246 return MDB_TXN_FULL;
2247 if (IS_OVERFLOW(mp))
2251 if (env->me_flags & MDB_WRITEMAP) {
2254 np = mdb_page_malloc(txn, num);
2258 memcpy(np, mp, num * env->me_psize);
2260 mdb_page_copy(np, mp, env->me_psize);
2263 /* If in current txn, this page is no longer spilled.
2264 * If it happens to be the last page, truncate the spill list.
2265 * Otherwise mark it as deleted by setting the LSB.
2267 if (x == txn->mt_spill_pgs[0])
2268 txn->mt_spill_pgs[0]--;
2270 txn->mt_spill_pgs[x] |= 1;
2271 } /* otherwise, if belonging to a parent txn, the
2272 * page remains spilled until child commits
2275 mdb_page_dirty(txn, np);
2276 np->mp_flags |= P_DIRTY;
2284 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2285 * @param[in] mc cursor pointing to the page to be touched
2286 * @return 0 on success, non-zero on failure.
2289 mdb_page_touch(MDB_cursor *mc)
2291 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2292 MDB_txn *txn = mc->mc_txn;
2293 MDB_cursor *m2, *m3;
2297 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2298 if (txn->mt_flags & MDB_TXN_SPILLS) {
2300 rc = mdb_page_unspill(txn, mp, &np);
2306 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2307 (rc = mdb_page_alloc(mc, 1, &np)))
2310 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2311 mp->mp_pgno, pgno));
2312 mdb_cassert(mc, mp->mp_pgno != pgno);
2313 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2314 /* Update the parent page, if any, to point to the new page */
2316 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2317 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2318 SETPGNO(node, pgno);
2320 mc->mc_db->md_root = pgno;
2322 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2323 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2325 /* If txn has a parent, make sure the page is in our
2329 unsigned x = mdb_mid2l_search(dl, pgno);
2330 if (x <= dl[0].mid && dl[x].mid == pgno) {
2331 if (mp != dl[x].mptr) { /* bad cursor? */
2332 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2333 txn->mt_flags |= MDB_TXN_ERROR;
2334 return MDB_CORRUPTED;
2339 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2341 np = mdb_page_malloc(txn, 1);
2346 rc = mdb_mid2l_insert(dl, &mid);
2347 mdb_cassert(mc, rc == 0);
2352 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2354 np->mp_flags |= P_DIRTY;
2357 /* Adjust cursors pointing to mp */
2358 mc->mc_pg[mc->mc_top] = np;
2359 m2 = txn->mt_cursors[mc->mc_dbi];
2360 if (mc->mc_flags & C_SUB) {
2361 for (; m2; m2=m2->mc_next) {
2362 m3 = &m2->mc_xcursor->mx_cursor;
2363 if (m3->mc_snum < mc->mc_snum) continue;
2364 if (m3->mc_pg[mc->mc_top] == mp)
2365 m3->mc_pg[mc->mc_top] = np;
2368 for (; m2; m2=m2->mc_next) {
2369 if (m2->mc_snum < mc->mc_snum) continue;
2370 if (m2->mc_pg[mc->mc_top] == mp) {
2371 m2->mc_pg[mc->mc_top] = np;
2372 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2374 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2376 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2377 if (!(leaf->mn_flags & F_SUBDATA))
2378 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2386 txn->mt_flags |= MDB_TXN_ERROR;
2391 mdb_env_sync(MDB_env *env, int force)
2394 if (env->me_flags & MDB_RDONLY)
2396 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2397 if (env->me_flags & MDB_WRITEMAP) {
2398 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2399 ? MS_ASYNC : MS_SYNC;
2400 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2403 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2407 #ifdef BROKEN_FDATASYNC
2408 if (env->me_flags & MDB_FSYNCONLY) {
2409 if (fsync(env->me_fd))
2413 if (MDB_FDATASYNC(env->me_fd))
2420 /** Back up parent txn's cursors, then grab the originals for tracking */
2422 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2424 MDB_cursor *mc, *bk;
2429 for (i = src->mt_numdbs; --i >= 0; ) {
2430 if ((mc = src->mt_cursors[i]) != NULL) {
2431 size = sizeof(MDB_cursor);
2433 size += sizeof(MDB_xcursor);
2434 for (; mc; mc = bk->mc_next) {
2440 mc->mc_db = &dst->mt_dbs[i];
2441 /* Kill pointers into src - and dst to reduce abuse: The
2442 * user may not use mc until dst ends. Otherwise we'd...
2444 mc->mc_txn = NULL; /* ...set this to dst */
2445 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2446 if ((mx = mc->mc_xcursor) != NULL) {
2447 *(MDB_xcursor *)(bk+1) = *mx;
2448 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2450 mc->mc_next = dst->mt_cursors[i];
2451 dst->mt_cursors[i] = mc;
2458 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2459 * @param[in] txn the transaction handle.
2460 * @param[in] merge true to keep changes to parent cursors, false to revert.
2461 * @return 0 on success, non-zero on failure.
2464 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2466 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2470 for (i = txn->mt_numdbs; --i >= 0; ) {
2471 for (mc = cursors[i]; mc; mc = next) {
2473 if ((bk = mc->mc_backup) != NULL) {
2475 /* Commit changes to parent txn */
2476 mc->mc_next = bk->mc_next;
2477 mc->mc_backup = bk->mc_backup;
2478 mc->mc_txn = bk->mc_txn;
2479 mc->mc_db = bk->mc_db;
2480 mc->mc_dbflag = bk->mc_dbflag;
2481 if ((mx = mc->mc_xcursor) != NULL)
2482 mx->mx_cursor.mc_txn = bk->mc_txn;
2484 /* Abort nested txn */
2486 if ((mx = mc->mc_xcursor) != NULL)
2487 *mx = *(MDB_xcursor *)(bk+1);
2491 /* Only malloced cursors are permanently tracked. */
2499 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2502 mdb_txn_reset0(MDB_txn *txn, const char *act);
2504 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2510 Pidset = F_SETLK, Pidcheck = F_GETLK
2514 /** Set or check a pid lock. Set returns 0 on success.
2515 * Check returns 0 if the process is certainly dead, nonzero if it may
2516 * be alive (the lock exists or an error happened so we do not know).
2518 * On Windows Pidset is a no-op, we merely check for the existence
2519 * of the process with the given pid. On POSIX we use a single byte
2520 * lock on the lockfile, set at an offset equal to the pid.
2523 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2525 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2528 if (op == Pidcheck) {
2529 h = OpenProcess(env->me_pidquery, FALSE, pid);
2530 /* No documented "no such process" code, but other program use this: */
2532 return ErrCode() != ERROR_INVALID_PARAMETER;
2533 /* A process exists until all handles to it close. Has it exited? */
2534 ret = WaitForSingleObject(h, 0) != 0;
2541 struct flock lock_info;
2542 memset(&lock_info, 0, sizeof(lock_info));
2543 lock_info.l_type = F_WRLCK;
2544 lock_info.l_whence = SEEK_SET;
2545 lock_info.l_start = pid;
2546 lock_info.l_len = 1;
2547 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2548 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2550 } else if ((rc = ErrCode()) == EINTR) {
2558 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2559 * @param[in] txn the transaction handle to initialize
2560 * @return 0 on success, non-zero on failure.
2563 mdb_txn_renew0(MDB_txn *txn)
2565 MDB_env *env = txn->mt_env;
2566 MDB_txninfo *ti = env->me_txns;
2570 int rc, new_notls = 0;
2572 if (txn->mt_flags & MDB_TXN_RDONLY) {
2574 txn->mt_numdbs = env->me_numdbs;
2575 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2577 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2578 txn->mt_txnid = meta->mm_txnid;
2579 txn->mt_u.reader = NULL;
2581 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2582 pthread_getspecific(env->me_txkey);
2584 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2585 return MDB_BAD_RSLOT;
2587 MDB_PID_T pid = env->me_pid;
2588 MDB_THR_T tid = pthread_self();
2589 mdb_mutex_t *rmutex = MDB_MUTEX(env, r);
2591 if (!env->me_live_reader) {
2592 rc = mdb_reader_pid(env, Pidset, pid);
2595 env->me_live_reader = 1;
2598 if (LOCK_MUTEX(rc, env, rmutex))
2600 nr = ti->mti_numreaders;
2601 for (i=0; i<nr; i++)
2602 if (ti->mti_readers[i].mr_pid == 0)
2604 if (i == env->me_maxreaders) {
2605 UNLOCK_MUTEX(rmutex);
2606 return MDB_READERS_FULL;
2608 r = &ti->mti_readers[i];
2609 /* Claim the reader slot, carefully since other code
2610 * uses the reader table un-mutexed: First reset the
2611 * slot, next publish it in mti_numreaders. After
2612 * that, it is safe for mdb_env_close() to touch it.
2613 * When it will be closed, we can finally claim it.
2616 r->mr_txnid = (txnid_t)-1;
2619 ti->mti_numreaders = ++nr;
2620 env->me_close_readers = nr;
2622 UNLOCK_MUTEX(rmutex);
2624 new_notls = (env->me_flags & MDB_NOTLS);
2625 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2630 do /* LY: Retry on a race, ITS#7970. */
2631 r->mr_txnid = ti->mti_txnid;
2632 while(r->mr_txnid != ti->mti_txnid);
2633 txn->mt_txnid = r->mr_txnid;
2634 txn->mt_u.reader = r;
2635 meta = env->me_metas[txn->mt_txnid & 1];
2639 if (LOCK_MUTEX(rc, env, MDB_MUTEX(env, w)))
2641 txn->mt_txnid = ti->mti_txnid;
2642 meta = env->me_metas[txn->mt_txnid & 1];
2644 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2645 txn->mt_txnid = meta->mm_txnid;
2648 txn->mt_numdbs = env->me_numdbs;
2651 if (txn->mt_txnid == mdb_debug_start)
2655 txn->mt_child = NULL;
2656 txn->mt_loose_pgs = NULL;
2657 txn->mt_loose_count = 0;
2658 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2659 txn->mt_u.dirty_list = env->me_dirty_list;
2660 txn->mt_u.dirty_list[0].mid = 0;
2661 txn->mt_free_pgs = env->me_free_pgs;
2662 txn->mt_free_pgs[0] = 0;
2663 txn->mt_spill_pgs = NULL;
2665 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2668 /* Copy the DB info and flags */
2669 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2671 /* Moved to here to avoid a data race in read TXNs */
2672 txn->mt_next_pgno = meta->mm_last_pg+1;
2674 for (i=2; i<txn->mt_numdbs; i++) {
2675 x = env->me_dbflags[i];
2676 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2677 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2679 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2681 if (env->me_maxpg < txn->mt_next_pgno) {
2682 mdb_txn_reset0(txn, "renew0-mapfail");
2684 txn->mt_u.reader->mr_pid = 0;
2685 txn->mt_u.reader = NULL;
2687 return MDB_MAP_RESIZED;
2694 mdb_txn_renew(MDB_txn *txn)
2698 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2701 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2702 DPUTS("environment had fatal error, must shutdown!");
2706 rc = mdb_txn_renew0(txn);
2707 if (rc == MDB_SUCCESS) {
2708 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2709 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2710 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2716 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2720 int rc, size, tsize = sizeof(MDB_txn);
2722 if (env->me_flags & MDB_FATAL_ERROR) {
2723 DPUTS("environment had fatal error, must shutdown!");
2726 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2729 /* Nested transactions: Max 1 child, write txns only, no writemap */
2730 if (parent->mt_child ||
2731 (flags & MDB_RDONLY) ||
2732 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2733 (env->me_flags & MDB_WRITEMAP))
2735 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2737 tsize = sizeof(MDB_ntxn);
2740 if (!(flags & MDB_RDONLY)) {
2742 txn = env->me_txn0; /* just reuse preallocated write txn */
2745 /* child txns use own copy of cursors */
2746 size += env->me_maxdbs * sizeof(MDB_cursor *);
2748 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2750 if ((txn = calloc(1, size)) == NULL) {
2751 DPRINTF(("calloc: %s", strerror(errno)));
2754 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2755 if (flags & MDB_RDONLY) {
2756 txn->mt_flags |= MDB_TXN_RDONLY;
2757 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2758 txn->mt_dbiseqs = env->me_dbiseqs;
2760 if (flags & MDB_NOSYNC)
2761 txn->mt_flags |= MDB_TXN_NOSYNC;
2762 if (flags & MDB_NOMETASYNC)
2763 txn->mt_flags |= MDB_TXN_NOMETASYNC;
2764 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2766 txn->mt_dbiseqs = parent->mt_dbiseqs;
2767 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2769 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2770 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2778 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2779 if (!txn->mt_u.dirty_list ||
2780 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2782 free(txn->mt_u.dirty_list);
2786 txn->mt_txnid = parent->mt_txnid;
2787 txn->mt_dirty_room = parent->mt_dirty_room;
2788 txn->mt_u.dirty_list[0].mid = 0;
2789 txn->mt_spill_pgs = NULL;
2790 txn->mt_next_pgno = parent->mt_next_pgno;
2791 parent->mt_child = txn;
2792 txn->mt_parent = parent;
2793 txn->mt_numdbs = parent->mt_numdbs;
2794 txn->mt_flags = parent->mt_flags;
2795 txn->mt_dbxs = parent->mt_dbxs;
2796 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2797 /* Copy parent's mt_dbflags, but clear DB_NEW */
2798 for (i=0; i<txn->mt_numdbs; i++)
2799 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2801 ntxn = (MDB_ntxn *)txn;
2802 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2803 if (env->me_pghead) {
2804 size = MDB_IDL_SIZEOF(env->me_pghead);
2805 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2807 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2812 rc = mdb_cursor_shadow(parent, txn);
2814 mdb_txn_reset0(txn, "beginchild-fail");
2816 rc = mdb_txn_renew0(txn);
2819 if (txn != env->me_txn0)
2823 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2824 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2825 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2832 mdb_txn_env(MDB_txn *txn)
2834 if(!txn) return NULL;
2839 mdb_txn_id(MDB_txn *txn)
2842 return txn->mt_txnid;
2845 /** Export or close DBI handles opened in this txn. */
2847 mdb_dbis_update(MDB_txn *txn, int keep)
2850 MDB_dbi n = txn->mt_numdbs;
2851 MDB_env *env = txn->mt_env;
2852 unsigned char *tdbflags = txn->mt_dbflags;
2854 for (i = n; --i >= 2;) {
2855 if (tdbflags[i] & DB_NEW) {
2857 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2859 char *ptr = env->me_dbxs[i].md_name.mv_data;
2861 env->me_dbxs[i].md_name.mv_data = NULL;
2862 env->me_dbxs[i].md_name.mv_size = 0;
2863 env->me_dbflags[i] = 0;
2864 env->me_dbiseqs[i]++;
2870 if (keep && env->me_numdbs < n)
2874 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2875 * May be called twice for readonly txns: First reset it, then abort.
2876 * @param[in] txn the transaction handle to reset
2877 * @param[in] act why the transaction is being reset
2880 mdb_txn_reset0(MDB_txn *txn, const char *act)
2882 MDB_env *env = txn->mt_env;
2884 /* Close any DBI handles opened in this txn */
2885 mdb_dbis_update(txn, 0);
2887 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2888 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2889 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2891 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2892 if (txn->mt_u.reader) {
2893 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2894 if (!(env->me_flags & MDB_NOTLS))
2895 txn->mt_u.reader = NULL; /* txn does not own reader */
2897 txn->mt_numdbs = 0; /* close nothing if called again */
2898 txn->mt_dbxs = NULL; /* mark txn as reset */
2900 pgno_t *pghead = env->me_pghead;
2902 mdb_cursors_close(txn, 0);
2903 if (!(env->me_flags & MDB_WRITEMAP)) {
2904 mdb_dlist_free(txn);
2907 if (!txn->mt_parent) {
2908 if (mdb_midl_shrink(&txn->mt_free_pgs))
2909 env->me_free_pgs = txn->mt_free_pgs;
2911 env->me_pghead = NULL;
2915 /* The writer mutex was locked in mdb_txn_begin. */
2917 UNLOCK_MUTEX(MDB_MUTEX(env, w));
2919 txn->mt_parent->mt_child = NULL;
2920 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2921 mdb_midl_free(txn->mt_free_pgs);
2922 mdb_midl_free(txn->mt_spill_pgs);
2923 free(txn->mt_u.dirty_list);
2926 mdb_midl_free(pghead);
2931 mdb_txn_reset(MDB_txn *txn)
2936 /* This call is only valid for read-only txns */
2937 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2940 mdb_txn_reset0(txn, "reset");
2944 mdb_txn_abort(MDB_txn *txn)
2950 mdb_txn_abort(txn->mt_child);
2952 mdb_txn_reset0(txn, "abort");
2953 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2954 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2955 txn->mt_u.reader->mr_pid = 0;
2957 if (txn != txn->mt_env->me_txn0)
2961 /** Save the freelist as of this transaction to the freeDB.
2962 * This changes the freelist. Keep trying until it stabilizes.
2965 mdb_freelist_save(MDB_txn *txn)
2967 /* env->me_pghead[] can grow and shrink during this call.
2968 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2969 * Page numbers cannot disappear from txn->mt_free_pgs[].
2972 MDB_env *env = txn->mt_env;
2973 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2974 txnid_t pglast = 0, head_id = 0;
2975 pgno_t freecnt = 0, *free_pgs, *mop;
2976 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2978 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2980 if (env->me_pghead) {
2981 /* Make sure first page of freeDB is touched and on freelist */
2982 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2983 if (rc && rc != MDB_NOTFOUND)
2987 if (!env->me_pghead && txn->mt_loose_pgs) {
2988 /* Put loose page numbers in mt_free_pgs, since
2989 * we may be unable to return them to me_pghead.
2991 MDB_page *mp = txn->mt_loose_pgs;
2992 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2994 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2995 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2996 txn->mt_loose_pgs = NULL;
2997 txn->mt_loose_count = 0;
3000 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3001 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3002 ? SSIZE_MAX : maxfree_1pg;
3005 /* Come back here after each Put() in case freelist changed */
3010 /* If using records from freeDB which we have not yet
3011 * deleted, delete them and any we reserved for me_pghead.
3013 while (pglast < env->me_pglast) {
3014 rc = mdb_cursor_first(&mc, &key, NULL);
3017 pglast = head_id = *(txnid_t *)key.mv_data;
3018 total_room = head_room = 0;
3019 mdb_tassert(txn, pglast <= env->me_pglast);
3020 rc = mdb_cursor_del(&mc, 0);
3025 /* Save the IDL of pages freed by this txn, to a single record */
3026 if (freecnt < txn->mt_free_pgs[0]) {
3028 /* Make sure last page of freeDB is touched and on freelist */
3029 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3030 if (rc && rc != MDB_NOTFOUND)
3033 free_pgs = txn->mt_free_pgs;
3034 /* Write to last page of freeDB */
3035 key.mv_size = sizeof(txn->mt_txnid);
3036 key.mv_data = &txn->mt_txnid;
3038 freecnt = free_pgs[0];
3039 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3040 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3043 /* Retry if mt_free_pgs[] grew during the Put() */
3044 free_pgs = txn->mt_free_pgs;
3045 } while (freecnt < free_pgs[0]);
3046 mdb_midl_sort(free_pgs);
3047 memcpy(data.mv_data, free_pgs, data.mv_size);
3050 unsigned int i = free_pgs[0];
3051 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3052 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3054 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3060 mop = env->me_pghead;
3061 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3063 /* Reserve records for me_pghead[]. Split it if multi-page,
3064 * to avoid searching freeDB for a page range. Use keys in
3065 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3067 if (total_room >= mop_len) {
3068 if (total_room == mop_len || --more < 0)
3070 } else if (head_room >= maxfree_1pg && head_id > 1) {
3071 /* Keep current record (overflow page), add a new one */
3075 /* (Re)write {key = head_id, IDL length = head_room} */
3076 total_room -= head_room;
3077 head_room = mop_len - total_room;
3078 if (head_room > maxfree_1pg && head_id > 1) {
3079 /* Overflow multi-page for part of me_pghead */
3080 head_room /= head_id; /* amortize page sizes */
3081 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3082 } else if (head_room < 0) {
3083 /* Rare case, not bothering to delete this record */
3086 key.mv_size = sizeof(head_id);
3087 key.mv_data = &head_id;
3088 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3089 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3092 /* IDL is initially empty, zero out at least the length */
3093 pgs = (pgno_t *)data.mv_data;
3094 j = head_room > clean_limit ? head_room : 0;
3098 total_room += head_room;
3101 /* Return loose page numbers to me_pghead, though usually none are
3102 * left at this point. The pages themselves remain in dirty_list.
3104 if (txn->mt_loose_pgs) {
3105 MDB_page *mp = txn->mt_loose_pgs;
3106 unsigned count = txn->mt_loose_count;
3108 /* Room for loose pages + temp IDL with same */
3109 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3111 mop = env->me_pghead;
3112 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3113 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3114 loose[ ++count ] = mp->mp_pgno;
3116 mdb_midl_sort(loose);
3117 mdb_midl_xmerge(mop, loose);
3118 txn->mt_loose_pgs = NULL;
3119 txn->mt_loose_count = 0;
3123 /* Fill in the reserved me_pghead records */
3129 rc = mdb_cursor_first(&mc, &key, &data);
3130 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3131 txnid_t id = *(txnid_t *)key.mv_data;
3132 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3135 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3137 if (len > mop_len) {
3139 data.mv_size = (len + 1) * sizeof(MDB_ID);
3141 data.mv_data = mop -= len;
3144 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3146 if (rc || !(mop_len -= len))
3153 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3154 * @param[in] txn the transaction that's being committed
3155 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3156 * @return 0 on success, non-zero on failure.
3159 mdb_page_flush(MDB_txn *txn, int keep)
3161 MDB_env *env = txn->mt_env;
3162 MDB_ID2L dl = txn->mt_u.dirty_list;
3163 unsigned psize = env->me_psize, j;
3164 int i, pagecount = dl[0].mid, rc;
3165 size_t size = 0, pos = 0;
3167 MDB_page *dp = NULL;
3171 struct iovec iov[MDB_COMMIT_PAGES];
3172 ssize_t wpos = 0, wsize = 0, wres;
3173 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3179 if (env->me_flags & MDB_WRITEMAP) {
3180 /* Clear dirty flags */
3181 while (++i <= pagecount) {
3183 /* Don't flush this page yet */
3184 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3185 dp->mp_flags &= ~P_KEEP;
3189 dp->mp_flags &= ~P_DIRTY;
3194 /* Write the pages */
3196 if (++i <= pagecount) {
3198 /* Don't flush this page yet */
3199 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3200 dp->mp_flags &= ~P_KEEP;
3205 /* clear dirty flag */
3206 dp->mp_flags &= ~P_DIRTY;
3209 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3214 /* Windows actually supports scatter/gather I/O, but only on
3215 * unbuffered file handles. Since we're relying on the OS page
3216 * cache for all our data, that's self-defeating. So we just
3217 * write pages one at a time. We use the ov structure to set
3218 * the write offset, to at least save the overhead of a Seek
3221 DPRINTF(("committing page %"Z"u", pgno));
3222 memset(&ov, 0, sizeof(ov));
3223 ov.Offset = pos & 0xffffffff;
3224 ov.OffsetHigh = pos >> 16 >> 16;
3225 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3227 DPRINTF(("WriteFile: %d", rc));
3231 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3232 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3234 /* Write previous page(s) */
3235 #ifdef MDB_USE_PWRITEV
3236 wres = pwritev(env->me_fd, iov, n, wpos);
3239 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3241 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3243 DPRINTF(("lseek: %s", strerror(rc)));
3246 wres = writev(env->me_fd, iov, n);
3249 if (wres != wsize) {
3252 DPRINTF(("Write error: %s", strerror(rc)));
3254 rc = EIO; /* TODO: Use which error code? */
3255 DPUTS("short write, filesystem full?");
3266 DPRINTF(("committing page %"Z"u", pgno));
3267 next_pos = pos + size;
3268 iov[n].iov_len = size;
3269 iov[n].iov_base = (char *)dp;
3275 /* MIPS has cache coherency issues, this is a no-op everywhere else
3276 * Note: for any size >= on-chip cache size, entire on-chip cache is
3279 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3281 for (i = keep; ++i <= pagecount; ) {
3283 /* This is a page we skipped above */
3286 dl[j].mid = dp->mp_pgno;
3289 mdb_dpage_free(env, dp);
3294 txn->mt_dirty_room += i - j;
3300 mdb_txn_commit(MDB_txn *txn)
3306 if (txn == NULL || txn->mt_env == NULL)
3309 if (txn->mt_child) {
3310 rc = mdb_txn_commit(txn->mt_child);
3311 txn->mt_child = NULL;
3318 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3319 mdb_dbis_update(txn, 1);
3320 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3325 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3326 DPUTS("error flag is set, can't commit");
3328 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3333 if (txn->mt_parent) {
3334 MDB_txn *parent = txn->mt_parent;
3338 unsigned x, y, len, ps_len;
3340 /* Append our free list to parent's */
3341 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3344 mdb_midl_free(txn->mt_free_pgs);
3345 /* Failures after this must either undo the changes
3346 * to the parent or set MDB_TXN_ERROR in the parent.
3349 parent->mt_next_pgno = txn->mt_next_pgno;
3350 parent->mt_flags = txn->mt_flags;
3352 /* Merge our cursors into parent's and close them */
3353 mdb_cursors_close(txn, 1);
3355 /* Update parent's DB table. */
3356 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3357 parent->mt_numdbs = txn->mt_numdbs;
3358 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3359 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3360 for (i=2; i<txn->mt_numdbs; i++) {
3361 /* preserve parent's DB_NEW status */
3362 x = parent->mt_dbflags[i] & DB_NEW;
3363 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3366 dst = parent->mt_u.dirty_list;
3367 src = txn->mt_u.dirty_list;
3368 /* Remove anything in our dirty list from parent's spill list */
3369 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3371 pspill[0] = (pgno_t)-1;
3372 /* Mark our dirty pages as deleted in parent spill list */
3373 for (i=0, len=src[0].mid; ++i <= len; ) {
3374 MDB_ID pn = src[i].mid << 1;
3375 while (pn > pspill[x])
3377 if (pn == pspill[x]) {
3382 /* Squash deleted pagenums if we deleted any */
3383 for (x=y; ++x <= ps_len; )
3384 if (!(pspill[x] & 1))
3385 pspill[++y] = pspill[x];
3389 /* Find len = length of merging our dirty list with parent's */
3391 dst[0].mid = 0; /* simplify loops */
3392 if (parent->mt_parent) {
3393 len = x + src[0].mid;
3394 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3395 for (i = x; y && i; y--) {
3396 pgno_t yp = src[y].mid;
3397 while (yp < dst[i].mid)
3399 if (yp == dst[i].mid) {
3404 } else { /* Simplify the above for single-ancestor case */
3405 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3407 /* Merge our dirty list with parent's */
3409 for (i = len; y; dst[i--] = src[y--]) {
3410 pgno_t yp = src[y].mid;
3411 while (yp < dst[x].mid)
3412 dst[i--] = dst[x--];
3413 if (yp == dst[x].mid)
3414 free(dst[x--].mptr);
3416 mdb_tassert(txn, i == x);
3418 free(txn->mt_u.dirty_list);
3419 parent->mt_dirty_room = txn->mt_dirty_room;
3420 if (txn->mt_spill_pgs) {
3421 if (parent->mt_spill_pgs) {
3422 /* TODO: Prevent failure here, so parent does not fail */
3423 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3425 parent->mt_flags |= MDB_TXN_ERROR;
3426 mdb_midl_free(txn->mt_spill_pgs);
3427 mdb_midl_sort(parent->mt_spill_pgs);
3429 parent->mt_spill_pgs = txn->mt_spill_pgs;
3433 /* Append our loose page list to parent's */
3434 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3436 *lp = txn->mt_loose_pgs;
3437 parent->mt_loose_count += txn->mt_loose_count;
3439 parent->mt_child = NULL;
3440 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3445 if (txn != env->me_txn) {
3446 DPUTS("attempt to commit unknown transaction");
3451 mdb_cursors_close(txn, 0);
3453 if (!txn->mt_u.dirty_list[0].mid &&
3454 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3457 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3458 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3460 /* Update DB root pointers */
3461 if (txn->mt_numdbs > 2) {
3465 data.mv_size = sizeof(MDB_db);
3467 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3468 for (i = 2; i < txn->mt_numdbs; i++) {
3469 if (txn->mt_dbflags[i] & DB_DIRTY) {
3470 if (TXN_DBI_CHANGED(txn, i)) {
3474 data.mv_data = &txn->mt_dbs[i];
3475 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3482 rc = mdb_freelist_save(txn);
3486 mdb_midl_free(env->me_pghead);
3487 env->me_pghead = NULL;
3488 if (mdb_midl_shrink(&txn->mt_free_pgs))
3489 env->me_free_pgs = txn->mt_free_pgs;
3495 if ((rc = mdb_page_flush(txn, 0)))
3497 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3498 (rc = mdb_env_sync(env, 0)))
3500 if ((rc = mdb_env_write_meta(txn)))
3503 /* Free P_LOOSE pages left behind in dirty_list */
3504 if (!(env->me_flags & MDB_WRITEMAP))
3505 mdb_dlist_free(txn);
3510 mdb_dbis_update(txn, 1);
3513 UNLOCK_MUTEX(MDB_MUTEX(env, w));
3514 if (txn != env->me_txn0)
3524 /** Read the environment parameters of a DB environment before
3525 * mapping it into memory.
3526 * @param[in] env the environment handle
3527 * @param[out] meta address of where to store the meta information
3528 * @return 0 on success, non-zero on failure.
3531 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3537 enum { Size = sizeof(pbuf) };
3539 /* We don't know the page size yet, so use a minimum value.
3540 * Read both meta pages so we can use the latest one.
3543 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3547 memset(&ov, 0, sizeof(ov));
3549 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3550 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3553 rc = pread(env->me_fd, &pbuf, Size, off);
3556 if (rc == 0 && off == 0)
3558 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3559 DPRINTF(("read: %s", mdb_strerror(rc)));
3563 p = (MDB_page *)&pbuf;
3565 if (!F_ISSET(p->mp_flags, P_META)) {
3566 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3571 if (m->mm_magic != MDB_MAGIC) {
3572 DPUTS("meta has invalid magic");
3576 if (m->mm_version != MDB_DATA_VERSION) {
3577 DPRINTF(("database is version %u, expected version %u",
3578 m->mm_version, MDB_DATA_VERSION));
3579 return MDB_VERSION_MISMATCH;
3582 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3588 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3590 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3592 meta->mm_magic = MDB_MAGIC;
3593 meta->mm_version = MDB_DATA_VERSION;
3594 meta->mm_mapsize = env->me_mapsize;
3595 meta->mm_psize = env->me_psize;
3596 meta->mm_last_pg = 1;
3597 meta->mm_flags = env->me_flags & 0xffff;
3598 meta->mm_flags |= MDB_INTEGERKEY;
3599 meta->mm_dbs[0].md_root = P_INVALID;
3600 meta->mm_dbs[1].md_root = P_INVALID;
3603 /** Write the environment parameters of a freshly created DB environment.
3604 * @param[in] env the environment handle
3605 * @param[in] meta the #MDB_meta to write
3606 * @return 0 on success, non-zero on failure.
3609 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3617 memset(&ov, 0, sizeof(ov));
3618 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3620 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3623 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3624 len = pwrite(fd, ptr, size, pos); \
3625 rc = (len >= 0); } while(0)
3628 DPUTS("writing new meta page");
3630 psize = env->me_psize;
3632 p = calloc(2, psize);
3634 p->mp_flags = P_META;
3635 *(MDB_meta *)METADATA(p) = *meta;
3637 q = (MDB_page *)((char *)p + psize);
3639 q->mp_flags = P_META;
3640 *(MDB_meta *)METADATA(q) = *meta;
3642 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3645 else if ((unsigned) len == psize * 2)
3653 /** Update the environment info to commit a transaction.
3654 * @param[in] txn the transaction that's being committed
3655 * @return 0 on success, non-zero on failure.
3658 mdb_env_write_meta(MDB_txn *txn)
3661 MDB_meta meta, metab, *mp;
3664 int rc, len, toggle;
3673 toggle = txn->mt_txnid & 1;
3674 DPRINTF(("writing meta page %d for root page %"Z"u",
3675 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3678 mp = env->me_metas[toggle];
3679 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3680 /* Persist any increases of mapsize config */
3681 if (mapsize < env->me_mapsize)
3682 mapsize = env->me_mapsize;
3684 if (env->me_flags & MDB_WRITEMAP) {
3685 mp->mm_mapsize = mapsize;
3686 mp->mm_dbs[0] = txn->mt_dbs[0];
3687 mp->mm_dbs[1] = txn->mt_dbs[1];
3688 mp->mm_last_pg = txn->mt_next_pgno - 1;
3689 #if !(defined(_MSC_VER) || defined(__i386__) || defined(__x86_64__))
3690 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3691 __sync_synchronize();
3693 mp->mm_txnid = txn->mt_txnid;
3694 if (txn->mt_flags & (MDB_TXN_NOSYNC|MDB_TXN_NOMETASYNC))
3696 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3697 unsigned meta_size = env->me_psize;
3698 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3701 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3702 if (meta_size < env->me_os_psize)
3703 meta_size += meta_size;
3708 if (MDB_MSYNC(ptr, meta_size, rc)) {
3715 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3716 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3718 meta.mm_mapsize = mapsize;
3719 meta.mm_dbs[0] = txn->mt_dbs[0];
3720 meta.mm_dbs[1] = txn->mt_dbs[1];
3721 meta.mm_last_pg = txn->mt_next_pgno - 1;
3722 meta.mm_txnid = txn->mt_txnid;
3724 off = offsetof(MDB_meta, mm_mapsize);
3725 ptr = (char *)&meta + off;
3726 len = sizeof(MDB_meta) - off;
3728 off += env->me_psize;
3731 /* Write to the SYNC fd */
3732 mfd = ((env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ||
3733 (txn->mt_flags & (MDB_TXN_NOSYNC|MDB_TXN_NOMETASYNC))) ?
3734 env->me_fd : env->me_mfd;
3737 memset(&ov, 0, sizeof(ov));
3739 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3743 rc = pwrite(mfd, ptr, len, off);
3746 rc = rc < 0 ? ErrCode() : EIO;
3747 DPUTS("write failed, disk error?");
3748 /* On a failure, the pagecache still contains the new data.
3749 * Write some old data back, to prevent it from being used.
3750 * Use the non-SYNC fd; we know it will fail anyway.
3752 meta.mm_last_pg = metab.mm_last_pg;
3753 meta.mm_txnid = metab.mm_txnid;
3755 memset(&ov, 0, sizeof(ov));
3757 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3759 r2 = pwrite(env->me_fd, ptr, len, off);
3760 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3763 env->me_flags |= MDB_FATAL_ERROR;
3766 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3767 CACHEFLUSH(env->me_map + off, len, DCACHE);
3769 /* Memory ordering issues are irrelevant; since the entire writer
3770 * is wrapped by wmutex, all of these changes will become visible
3771 * after the wmutex is unlocked. Since the DB is multi-version,
3772 * readers will get consistent data regardless of how fresh or
3773 * how stale their view of these values is.
3776 env->me_txns->mti_txnid = txn->mt_txnid;
3781 /** Check both meta pages to see which one is newer.
3782 * @param[in] env the environment handle
3783 * @return meta toggle (0 or 1).
3786 mdb_env_pick_meta(const MDB_env *env)
3788 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3792 mdb_env_create(MDB_env **env)
3796 e = calloc(1, sizeof(MDB_env));
3800 e->me_maxreaders = DEFAULT_READERS;
3801 e->me_maxdbs = e->me_numdbs = 2;
3802 e->me_fd = INVALID_HANDLE_VALUE;
3803 e->me_lfd = INVALID_HANDLE_VALUE;
3804 e->me_mfd = INVALID_HANDLE_VALUE;
3805 #ifdef MDB_USE_SYSV_SEM
3806 e->me_rmutex.semid = -1;
3807 e->me_wmutex.semid = -1;
3809 e->me_pid = getpid();
3810 GET_PAGESIZE(e->me_os_psize);
3811 VGMEMP_CREATE(e,0,0);
3817 mdb_env_map(MDB_env *env, void *addr)
3820 unsigned int flags = env->me_flags;
3824 LONG sizelo, sizehi;
3827 if (flags & MDB_RDONLY) {
3828 /* Don't set explicit map size, use whatever exists */
3833 msize = env->me_mapsize;
3834 sizelo = msize & 0xffffffff;
3835 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3837 /* Windows won't create mappings for zero length files.
3838 * and won't map more than the file size.
3839 * Just set the maxsize right now.
3841 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3842 || !SetEndOfFile(env->me_fd)
3843 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3847 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3848 PAGE_READWRITE : PAGE_READONLY,
3849 sizehi, sizelo, NULL);
3852 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3853 FILE_MAP_WRITE : FILE_MAP_READ,
3855 rc = env->me_map ? 0 : ErrCode();
3860 int prot = PROT_READ;
3861 if (flags & MDB_WRITEMAP) {
3863 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3866 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3868 if (env->me_map == MAP_FAILED) {
3873 if (flags & MDB_NORDAHEAD) {
3874 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3876 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3878 #ifdef POSIX_MADV_RANDOM
3879 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3880 #endif /* POSIX_MADV_RANDOM */
3881 #endif /* MADV_RANDOM */
3885 /* Can happen because the address argument to mmap() is just a
3886 * hint. mmap() can pick another, e.g. if the range is in use.
3887 * The MAP_FIXED flag would prevent that, but then mmap could
3888 * instead unmap existing pages to make room for the new map.
3890 if (addr && env->me_map != addr)
3891 return EBUSY; /* TODO: Make a new MDB_* error code? */
3893 p = (MDB_page *)env->me_map;
3894 env->me_metas[0] = METADATA(p);
3895 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3901 mdb_env_set_mapsize(MDB_env *env, size_t size)
3903 /* If env is already open, caller is responsible for making
3904 * sure there are no active txns.
3912 meta = env->me_metas[mdb_env_pick_meta(env)];
3914 size = meta->mm_mapsize;
3916 /* Silently round up to minimum if the size is too small */
3917 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3921 munmap(env->me_map, env->me_mapsize);
3922 env->me_mapsize = size;
3923 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3924 rc = mdb_env_map(env, old);
3928 env->me_mapsize = size;
3930 env->me_maxpg = env->me_mapsize / env->me_psize;
3935 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3939 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3944 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3946 if (env->me_map || readers < 1)
3948 env->me_maxreaders = readers;
3953 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3955 if (!env || !readers)
3957 *readers = env->me_maxreaders;
3962 mdb_fsize(HANDLE fd, size_t *size)
3965 LARGE_INTEGER fsize;
3967 if (!GetFileSizeEx(fd, &fsize))
3970 *size = fsize.QuadPart;
3982 #ifdef BROKEN_FDATASYNC
3983 #include <sys/utsname.h>
3984 #include <sys/vfs.h>
3987 /** Further setup required for opening an LMDB environment
3990 mdb_env_open2(MDB_env *env)
3992 unsigned int flags = env->me_flags;
3993 int i, newenv = 0, rc;
3997 /* See if we should use QueryLimited */
3999 if ((rc & 0xff) > 5)
4000 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4002 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4005 #ifdef BROKEN_FDATASYNC
4006 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4007 * https://lkml.org/lkml/2012/9/3/83
4008 * Kernels after 3.6-rc6 are known good.
4009 * https://lkml.org/lkml/2012/9/10/556
4010 * See if the DB is on ext3/ext4, then check for new enough kernel
4011 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4016 fstatfs(env->me_fd, &st);
4017 while (st.f_type == 0xEF53) {
4021 if (uts.release[0] < '3') {
4022 if (!strncmp(uts.release, "2.6.32.", 7)) {
4023 i = atoi(uts.release+7);
4025 break; /* 2.6.32.60 and newer is OK */
4026 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4027 i = atoi(uts.release+7);
4029 break; /* 2.6.34.15 and newer is OK */
4031 } else if (uts.release[0] == '3') {
4032 i = atoi(uts.release+2);
4034 break; /* 3.6 and newer is OK */
4036 i = atoi(uts.release+4);
4038 break; /* 3.5.4 and newer is OK */
4039 } else if (i == 2) {
4040 i = atoi(uts.release+4);
4042 break; /* 3.2.30 and newer is OK */
4044 } else { /* 4.x and newer is OK */
4047 env->me_flags |= MDB_FSYNCONLY;
4053 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4056 DPUTS("new mdbenv");
4058 env->me_psize = env->me_os_psize;
4059 if (env->me_psize > MAX_PAGESIZE)
4060 env->me_psize = MAX_PAGESIZE;
4061 memset(&meta, 0, sizeof(meta));
4062 mdb_env_init_meta0(env, &meta);
4063 meta.mm_mapsize = DEFAULT_MAPSIZE;
4065 env->me_psize = meta.mm_psize;
4068 /* Was a mapsize configured? */
4069 if (!env->me_mapsize) {
4070 env->me_mapsize = meta.mm_mapsize;
4073 /* Make sure mapsize >= committed data size. Even when using
4074 * mm_mapsize, which could be broken in old files (ITS#7789).
4076 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4077 if (env->me_mapsize < minsize)
4078 env->me_mapsize = minsize;
4080 meta.mm_mapsize = env->me_mapsize;
4082 if (newenv && !(flags & MDB_FIXEDMAP)) {
4083 /* mdb_env_map() may grow the datafile. Write the metapages
4084 * first, so the file will be valid if initialization fails.
4085 * Except with FIXEDMAP, since we do not yet know mm_address.
4086 * We could fill in mm_address later, but then a different
4087 * program might end up doing that - one with a memory layout
4088 * and map address which does not suit the main program.
4090 rc = mdb_env_init_meta(env, &meta);
4096 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4101 if (flags & MDB_FIXEDMAP)
4102 meta.mm_address = env->me_map;
4103 i = mdb_env_init_meta(env, &meta);
4104 if (i != MDB_SUCCESS) {
4109 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4110 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4112 #if !(MDB_MAXKEYSIZE)
4113 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4115 env->me_maxpg = env->me_mapsize / env->me_psize;
4119 int toggle = mdb_env_pick_meta(env);
4120 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4122 DPRINTF(("opened database version %u, pagesize %u",
4123 env->me_metas[0]->mm_version, env->me_psize));
4124 DPRINTF(("using meta page %d", toggle));
4125 DPRINTF(("depth: %u", db->md_depth));
4126 DPRINTF(("entries: %"Z"u", db->md_entries));
4127 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4128 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4129 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4130 DPRINTF(("root: %"Z"u", db->md_root));
4138 /** Release a reader thread's slot in the reader lock table.
4139 * This function is called automatically when a thread exits.
4140 * @param[in] ptr This points to the slot in the reader lock table.
4143 mdb_env_reader_dest(void *ptr)
4145 MDB_reader *reader = ptr;
4151 /** Junk for arranging thread-specific callbacks on Windows. This is
4152 * necessarily platform and compiler-specific. Windows supports up
4153 * to 1088 keys. Let's assume nobody opens more than 64 environments
4154 * in a single process, for now. They can override this if needed.
4156 #ifndef MAX_TLS_KEYS
4157 #define MAX_TLS_KEYS 64
4159 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4160 static int mdb_tls_nkeys;
4162 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4166 case DLL_PROCESS_ATTACH: break;
4167 case DLL_THREAD_ATTACH: break;
4168 case DLL_THREAD_DETACH:
4169 for (i=0; i<mdb_tls_nkeys; i++) {
4170 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4172 mdb_env_reader_dest(r);
4176 case DLL_PROCESS_DETACH: break;
4181 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4183 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4187 /* Force some symbol references.
4188 * _tls_used forces the linker to create the TLS directory if not already done
4189 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4191 #pragma comment(linker, "/INCLUDE:_tls_used")
4192 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4193 #pragma const_seg(".CRT$XLB")
4194 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4195 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4198 #pragma comment(linker, "/INCLUDE:__tls_used")
4199 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4200 #pragma data_seg(".CRT$XLB")
4201 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4203 #endif /* WIN 32/64 */
4204 #endif /* !__GNUC__ */
4207 /** Downgrade the exclusive lock on the region back to shared */
4209 mdb_env_share_locks(MDB_env *env, int *excl)
4211 int rc = 0, toggle = mdb_env_pick_meta(env);
4213 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4218 /* First acquire a shared lock. The Unlock will
4219 * then release the existing exclusive lock.
4221 memset(&ov, 0, sizeof(ov));
4222 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4225 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4231 struct flock lock_info;
4232 /* The shared lock replaces the existing lock */
4233 memset((void *)&lock_info, 0, sizeof(lock_info));
4234 lock_info.l_type = F_RDLCK;
4235 lock_info.l_whence = SEEK_SET;
4236 lock_info.l_start = 0;
4237 lock_info.l_len = 1;
4238 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4239 (rc = ErrCode()) == EINTR) ;
4240 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4247 /** Try to get exclusive lock, otherwise shared.
4248 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4251 mdb_env_excl_lock(MDB_env *env, int *excl)
4255 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4259 memset(&ov, 0, sizeof(ov));
4260 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4267 struct flock lock_info;
4268 memset((void *)&lock_info, 0, sizeof(lock_info));
4269 lock_info.l_type = F_WRLCK;
4270 lock_info.l_whence = SEEK_SET;
4271 lock_info.l_start = 0;
4272 lock_info.l_len = 1;
4273 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4274 (rc = ErrCode()) == EINTR) ;
4278 # ifdef MDB_USE_SYSV_SEM
4279 if (*excl < 0) /* always true when !MDB_USE_SYSV_SEM */
4282 lock_info.l_type = F_RDLCK;
4283 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4284 (rc = ErrCode()) == EINTR) ;
4294 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4296 * @(#) $Revision: 5.1 $
4297 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4298 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4300 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4304 * Please do not copyright this code. This code is in the public domain.
4306 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4307 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4308 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4309 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4310 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4311 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4312 * PERFORMANCE OF THIS SOFTWARE.
4315 * chongo <Landon Curt Noll> /\oo/\
4316 * http://www.isthe.com/chongo/
4318 * Share and Enjoy! :-)
4321 typedef unsigned long long mdb_hash_t;
4322 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4324 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4325 * @param[in] val value to hash
4326 * @param[in] hval initial value for hash
4327 * @return 64 bit hash
4329 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4330 * hval arg on the first call.
4333 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4335 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4336 unsigned char *end = s + val->mv_size;
4338 * FNV-1a hash each octet of the string
4341 /* xor the bottom with the current octet */
4342 hval ^= (mdb_hash_t)*s++;
4344 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4345 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4346 (hval << 7) + (hval << 8) + (hval << 40);
4348 /* return our new hash value */
4352 /** Hash the string and output the encoded hash.
4353 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4354 * very short name limits. We don't care about the encoding being reversible,
4355 * we just want to preserve as many bits of the input as possible in a
4356 * small printable string.
4357 * @param[in] str string to hash
4358 * @param[out] encbuf an array of 11 chars to hold the hash
4360 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4363 mdb_pack85(unsigned long l, char *out)
4367 for (i=0; i<5; i++) {
4368 *out++ = mdb_a85[l % 85];
4374 mdb_hash_enc(MDB_val *val, char *encbuf)
4376 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4378 mdb_pack85(h, encbuf);
4379 mdb_pack85(h>>32, encbuf+5);
4384 /** Open and/or initialize the lock region for the environment.
4385 * @param[in] env The LMDB environment.
4386 * @param[in] lpath The pathname of the file used for the lock region.
4387 * @param[in] mode The Unix permissions for the file, if we create it.
4388 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4389 * @return 0 on success, non-zero on failure.
4392 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4395 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4397 # define MDB_ERRCODE_ROFS EROFS
4398 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4399 # define MDB_CLOEXEC O_CLOEXEC
4402 # define MDB_CLOEXEC 0
4405 #ifdef MDB_USE_SYSV_SEM
4413 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4414 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4415 FILE_ATTRIBUTE_NORMAL, NULL);
4417 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4419 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4421 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4426 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4427 /* Lose record locks when exec*() */
4428 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4429 fcntl(env->me_lfd, F_SETFD, fdflags);
4432 if (!(env->me_flags & MDB_NOTLS)) {
4433 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4436 env->me_flags |= MDB_ENV_TXKEY;
4438 /* Windows TLS callbacks need help finding their TLS info. */
4439 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4443 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4447 /* Try to get exclusive lock. If we succeed, then
4448 * nobody is using the lock region and we should initialize it.
4450 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4453 size = GetFileSize(env->me_lfd, NULL);
4455 size = lseek(env->me_lfd, 0, SEEK_END);
4456 if (size == -1) goto fail_errno;
4458 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4459 if (size < rsize && *excl > 0) {
4461 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4462 || !SetEndOfFile(env->me_lfd))
4465 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4469 size = rsize - sizeof(MDB_txninfo);
4470 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4475 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4477 if (!mh) goto fail_errno;
4478 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4480 if (!env->me_txns) goto fail_errno;
4482 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4484 if (m == MAP_FAILED) goto fail_errno;
4490 BY_HANDLE_FILE_INFORMATION stbuf;
4499 if (!mdb_sec_inited) {
4500 InitializeSecurityDescriptor(&mdb_null_sd,
4501 SECURITY_DESCRIPTOR_REVISION);
4502 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4503 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4504 mdb_all_sa.bInheritHandle = FALSE;
4505 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4508 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4509 idbuf.volume = stbuf.dwVolumeSerialNumber;
4510 idbuf.nhigh = stbuf.nFileIndexHigh;
4511 idbuf.nlow = stbuf.nFileIndexLow;
4512 val.mv_data = &idbuf;
4513 val.mv_size = sizeof(idbuf);
4514 mdb_hash_enc(&val, encbuf);
4515 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4516 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4517 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4518 if (!env->me_rmutex) goto fail_errno;
4519 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4520 if (!env->me_wmutex) goto fail_errno;
4521 #elif defined(MDB_USE_SYSV_SEM)
4522 unsigned short vals[2] = {1, 1};
4523 semid = semget(IPC_PRIVATE, 2, mode);
4527 if (semctl(semid, 0, SETALL, semu) < 0)
4529 env->me_txns->mti_semid = semid;
4530 #else /* MDB_USE_SYSV_SEM */
4531 pthread_mutexattr_t mattr;
4533 if ((rc = pthread_mutexattr_init(&mattr))
4534 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4535 #ifdef MDB_ROBUST_SUPPORTED
4536 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4538 || (rc = pthread_mutex_init(&env->me_txns->mti_rmutex, &mattr))
4539 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4541 pthread_mutexattr_destroy(&mattr);
4542 #endif /* _WIN32 || MDB_USE_SYSV_SEM */
4544 env->me_txns->mti_magic = MDB_MAGIC;
4545 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4546 env->me_txns->mti_txnid = 0;
4547 env->me_txns->mti_numreaders = 0;
4550 #ifdef MDB_USE_SYSV_SEM
4551 struct semid_ds buf;
4553 if (env->me_txns->mti_magic != MDB_MAGIC) {
4554 DPUTS("lock region has invalid magic");
4558 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4559 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4560 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4561 rc = MDB_VERSION_MISMATCH;
4565 if (rc && rc != EACCES && rc != EAGAIN) {
4569 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4570 if (!env->me_rmutex) goto fail_errno;
4571 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4572 if (!env->me_wmutex) goto fail_errno;
4573 #elif defined(MDB_USE_SYSV_SEM)
4574 semid = env->me_txns->mti_semid;
4576 /* check for read access */
4577 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4579 /* check for write access */
4580 if (semctl(semid, 0, IPC_SET, semu) < 0)
4584 #ifdef MDB_USE_SYSV_SEM
4585 env->me_rmutex.semid = semid;
4586 env->me_wmutex.semid = semid;
4587 env->me_rmutex.semnum = 0;
4588 env->me_wmutex.semnum = 1;
4589 env->me_rmutex.locked = &env->me_txns->mti_rlocked;
4590 env->me_wmutex.locked = &env->me_txns->mti_wlocked;
4601 /** The name of the lock file in the DB environment */
4602 #define LOCKNAME "/lock.mdb"
4603 /** The name of the data file in the DB environment */
4604 #define DATANAME "/data.mdb"
4605 /** The suffix of the lock file when no subdir is used */
4606 #define LOCKSUFF "-lock"
4607 /** Only a subset of the @ref mdb_env flags can be changed
4608 * at runtime. Changing other flags requires closing the
4609 * environment and re-opening it with the new flags.
4611 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4612 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4613 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4615 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4616 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4620 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4622 int oflags, rc, len, excl = -1;
4623 char *lpath, *dpath;
4625 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4629 if (flags & MDB_NOSUBDIR) {
4630 rc = len + sizeof(LOCKSUFF) + len + 1;
4632 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4637 if (flags & MDB_NOSUBDIR) {
4638 dpath = lpath + len + sizeof(LOCKSUFF);
4639 sprintf(lpath, "%s" LOCKSUFF, path);
4640 strcpy(dpath, path);
4642 dpath = lpath + len + sizeof(LOCKNAME);
4643 sprintf(lpath, "%s" LOCKNAME, path);
4644 sprintf(dpath, "%s" DATANAME, path);
4648 flags |= env->me_flags;
4649 if (flags & MDB_RDONLY) {
4650 /* silently ignore WRITEMAP when we're only getting read access */
4651 flags &= ~MDB_WRITEMAP;
4653 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4654 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4657 env->me_flags = flags |= MDB_ENV_ACTIVE;
4661 env->me_path = strdup(path);
4662 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4663 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4664 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4665 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4670 /* For RDONLY, get lockfile after we know datafile exists */
4671 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4672 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4678 if (F_ISSET(flags, MDB_RDONLY)) {
4679 oflags = GENERIC_READ;
4680 len = OPEN_EXISTING;
4682 oflags = GENERIC_READ|GENERIC_WRITE;
4685 mode = FILE_ATTRIBUTE_NORMAL;
4686 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4687 NULL, len, mode, NULL);
4689 if (F_ISSET(flags, MDB_RDONLY))
4692 oflags = O_RDWR | O_CREAT;
4694 env->me_fd = open(dpath, oflags, mode);
4696 if (env->me_fd == INVALID_HANDLE_VALUE) {
4701 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4702 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4707 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4708 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4709 env->me_mfd = env->me_fd;
4711 /* Synchronous fd for meta writes. Needed even with
4712 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4715 len = OPEN_EXISTING;
4716 env->me_mfd = CreateFile(dpath, oflags,
4717 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4718 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4721 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4723 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4728 DPRINTF(("opened dbenv %p", (void *) env));
4730 rc = mdb_env_share_locks(env, &excl);
4734 if (!((flags & MDB_RDONLY) ||
4735 (env->me_pbuf = calloc(1, env->me_psize))))
4737 if (!(flags & MDB_RDONLY)) {
4739 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4740 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4741 txn = calloc(1, size);
4743 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4744 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4745 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4746 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4748 txn->mt_dbxs = env->me_dbxs;
4758 mdb_env_close0(env, excl);
4764 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4766 mdb_env_close0(MDB_env *env, int excl)
4770 if (!(env->me_flags & MDB_ENV_ACTIVE))
4773 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4775 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4776 free(env->me_dbxs[i].md_name.mv_data);
4781 free(env->me_dbiseqs);
4782 free(env->me_dbflags);
4784 free(env->me_dirty_list);
4786 mdb_midl_free(env->me_free_pgs);
4788 if (env->me_flags & MDB_ENV_TXKEY) {
4789 pthread_key_delete(env->me_txkey);
4791 /* Delete our key from the global list */
4792 for (i=0; i<mdb_tls_nkeys; i++)
4793 if (mdb_tls_keys[i] == env->me_txkey) {
4794 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4802 munmap(env->me_map, env->me_mapsize);
4804 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4805 (void) close(env->me_mfd);
4806 if (env->me_fd != INVALID_HANDLE_VALUE)
4807 (void) close(env->me_fd);
4809 MDB_PID_T pid = env->me_pid;
4810 /* Clearing readers is done in this function because
4811 * me_txkey with its destructor must be disabled first.
4813 * We skip the the reader mutex, so we touch only
4814 * data owned by this process (me_close_readers and
4815 * our readers), and clear each reader atomically.
4817 for (i = env->me_close_readers; --i >= 0; )
4818 if (env->me_txns->mti_readers[i].mr_pid == pid)
4819 env->me_txns->mti_readers[i].mr_pid = 0;
4821 if (env->me_rmutex) {
4822 CloseHandle(env->me_rmutex);
4823 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4825 /* Windows automatically destroys the mutexes when
4826 * the last handle closes.
4828 #elif defined(MDB_USE_SYSV_SEM)
4829 if (env->me_rmutex.semid != -1) {
4830 /* If we have the filelock: If we are the
4831 * only remaining user, clean up semaphores.
4834 mdb_env_excl_lock(env, &excl);
4836 semctl(env->me_rmutex.semid, 0, IPC_RMID);
4839 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4841 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4844 /* Unlock the lockfile. Windows would have unlocked it
4845 * after closing anyway, but not necessarily at once.
4847 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4850 (void) close(env->me_lfd);
4853 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4857 mdb_env_close(MDB_env *env)
4864 VGMEMP_DESTROY(env);
4865 while ((dp = env->me_dpages) != NULL) {
4866 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4867 env->me_dpages = dp->mp_next;
4871 mdb_env_close0(env, 0);
4875 /** Compare two items pointing at aligned size_t's */
4877 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4879 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4880 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4883 /** Compare two items pointing at aligned unsigned int's */
4885 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4887 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4888 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4891 /** Compare two items pointing at unsigned ints of unknown alignment.
4892 * Nodes and keys are guaranteed to be 2-byte aligned.
4895 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4897 #if BYTE_ORDER == LITTLE_ENDIAN
4898 unsigned short *u, *c;
4901 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4902 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4905 } while(!x && u > (unsigned short *)a->mv_data);
4908 unsigned short *u, *c, *end;
4911 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4912 u = (unsigned short *)a->mv_data;
4913 c = (unsigned short *)b->mv_data;
4916 } while(!x && u < end);
4921 /** Compare two items pointing at size_t's of unknown alignment. */
4922 #ifdef MISALIGNED_OK
4923 # define mdb_cmp_clong mdb_cmp_long
4925 # define mdb_cmp_clong mdb_cmp_cint
4928 /** Compare two items lexically */
4930 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4937 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4943 diff = memcmp(a->mv_data, b->mv_data, len);
4944 return diff ? diff : len_diff<0 ? -1 : len_diff;
4947 /** Compare two items in reverse byte order */
4949 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4951 const unsigned char *p1, *p2, *p1_lim;
4955 p1_lim = (const unsigned char *)a->mv_data;
4956 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4957 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4959 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4965 while (p1 > p1_lim) {
4966 diff = *--p1 - *--p2;
4970 return len_diff<0 ? -1 : len_diff;
4973 /** Search for key within a page, using binary search.
4974 * Returns the smallest entry larger or equal to the key.
4975 * If exactp is non-null, stores whether the found entry was an exact match
4976 * in *exactp (1 or 0).
4977 * Updates the cursor index with the index of the found entry.
4978 * If no entry larger or equal to the key is found, returns NULL.
4981 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4983 unsigned int i = 0, nkeys;
4986 MDB_page *mp = mc->mc_pg[mc->mc_top];
4987 MDB_node *node = NULL;
4992 nkeys = NUMKEYS(mp);
4994 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4995 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4998 low = IS_LEAF(mp) ? 0 : 1;
5000 cmp = mc->mc_dbx->md_cmp;
5002 /* Branch pages have no data, so if using integer keys,
5003 * alignment is guaranteed. Use faster mdb_cmp_int.
5005 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5006 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5013 nodekey.mv_size = mc->mc_db->md_pad;
5014 node = NODEPTR(mp, 0); /* fake */
5015 while (low <= high) {
5016 i = (low + high) >> 1;
5017 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5018 rc = cmp(key, &nodekey);
5019 DPRINTF(("found leaf index %u [%s], rc = %i",
5020 i, DKEY(&nodekey), rc));
5029 while (low <= high) {
5030 i = (low + high) >> 1;
5032 node = NODEPTR(mp, i);
5033 nodekey.mv_size = NODEKSZ(node);
5034 nodekey.mv_data = NODEKEY(node);
5036 rc = cmp(key, &nodekey);
5039 DPRINTF(("found leaf index %u [%s], rc = %i",
5040 i, DKEY(&nodekey), rc));
5042 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5043 i, DKEY(&nodekey), NODEPGNO(node), rc));
5054 if (rc > 0) { /* Found entry is less than the key. */
5055 i++; /* Skip to get the smallest entry larger than key. */
5057 node = NODEPTR(mp, i);
5060 *exactp = (rc == 0 && nkeys > 0);
5061 /* store the key index */
5062 mc->mc_ki[mc->mc_top] = i;
5064 /* There is no entry larger or equal to the key. */
5067 /* nodeptr is fake for LEAF2 */
5073 mdb_cursor_adjust(MDB_cursor *mc, func)
5077 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5078 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5085 /** Pop a page off the top of the cursor's stack. */
5087 mdb_cursor_pop(MDB_cursor *mc)
5091 MDB_page *top = mc->mc_pg[mc->mc_top];
5097 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5098 DDBI(mc), (void *) mc));
5102 /** Push a page onto the top of the cursor's stack. */
5104 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5106 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5107 DDBI(mc), (void *) mc));
5109 if (mc->mc_snum >= CURSOR_STACK) {
5110 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5111 return MDB_CURSOR_FULL;
5114 mc->mc_top = mc->mc_snum++;
5115 mc->mc_pg[mc->mc_top] = mp;
5116 mc->mc_ki[mc->mc_top] = 0;
5121 /** Find the address of the page corresponding to a given page number.
5122 * @param[in] txn the transaction for this access.
5123 * @param[in] pgno the page number for the page to retrieve.
5124 * @param[out] ret address of a pointer where the page's address will be stored.
5125 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5126 * @return 0 on success, non-zero on failure.
5129 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5131 MDB_env *env = txn->mt_env;
5135 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
5139 MDB_ID2L dl = tx2->mt_u.dirty_list;
5141 /* Spilled pages were dirtied in this txn and flushed
5142 * because the dirty list got full. Bring this page
5143 * back in from the map (but don't unspill it here,
5144 * leave that unless page_touch happens again).
5146 if (tx2->mt_spill_pgs) {
5147 MDB_ID pn = pgno << 1;
5148 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5149 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5150 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5155 unsigned x = mdb_mid2l_search(dl, pgno);
5156 if (x <= dl[0].mid && dl[x].mid == pgno) {
5162 } while ((tx2 = tx2->mt_parent) != NULL);
5165 if (pgno < txn->mt_next_pgno) {
5167 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5169 DPRINTF(("page %"Z"u not found", pgno));
5170 txn->mt_flags |= MDB_TXN_ERROR;
5171 return MDB_PAGE_NOTFOUND;
5181 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5182 * The cursor is at the root page, set up the rest of it.
5185 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5187 MDB_page *mp = mc->mc_pg[mc->mc_top];
5191 while (IS_BRANCH(mp)) {
5195 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5196 mdb_cassert(mc, NUMKEYS(mp) > 1);
5197 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5199 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5201 if (flags & MDB_PS_LAST)
5202 i = NUMKEYS(mp) - 1;
5205 node = mdb_node_search(mc, key, &exact);
5207 i = NUMKEYS(mp) - 1;
5209 i = mc->mc_ki[mc->mc_top];
5211 mdb_cassert(mc, i > 0);
5215 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5218 mdb_cassert(mc, i < NUMKEYS(mp));
5219 node = NODEPTR(mp, i);
5221 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5224 mc->mc_ki[mc->mc_top] = i;
5225 if ((rc = mdb_cursor_push(mc, mp)))
5228 if (flags & MDB_PS_MODIFY) {
5229 if ((rc = mdb_page_touch(mc)) != 0)
5231 mp = mc->mc_pg[mc->mc_top];
5236 DPRINTF(("internal error, index points to a %02X page!?",
5238 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5239 return MDB_CORRUPTED;
5242 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5243 key ? DKEY(key) : "null"));
5244 mc->mc_flags |= C_INITIALIZED;
5245 mc->mc_flags &= ~C_EOF;
5250 /** Search for the lowest key under the current branch page.
5251 * This just bypasses a NUMKEYS check in the current page
5252 * before calling mdb_page_search_root(), because the callers
5253 * are all in situations where the current page is known to
5257 mdb_page_search_lowest(MDB_cursor *mc)
5259 MDB_page *mp = mc->mc_pg[mc->mc_top];
5260 MDB_node *node = NODEPTR(mp, 0);
5263 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5266 mc->mc_ki[mc->mc_top] = 0;
5267 if ((rc = mdb_cursor_push(mc, mp)))
5269 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5272 /** Search for the page a given key should be in.
5273 * Push it and its parent pages on the cursor stack.
5274 * @param[in,out] mc the cursor for this operation.
5275 * @param[in] key the key to search for, or NULL for first/last page.
5276 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5277 * are touched (updated with new page numbers).
5278 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5279 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5280 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5281 * @return 0 on success, non-zero on failure.
5284 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5289 /* Make sure the txn is still viable, then find the root from
5290 * the txn's db table and set it as the root of the cursor's stack.
5292 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5293 DPUTS("transaction has failed, must abort");
5296 /* Make sure we're using an up-to-date root */
5297 if (*mc->mc_dbflag & DB_STALE) {
5299 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5301 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5302 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5309 MDB_node *leaf = mdb_node_search(&mc2,
5310 &mc->mc_dbx->md_name, &exact);
5312 return MDB_NOTFOUND;
5313 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5316 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5318 /* The txn may not know this DBI, or another process may
5319 * have dropped and recreated the DB with other flags.
5321 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5322 return MDB_INCOMPATIBLE;
5323 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5325 *mc->mc_dbflag &= ~DB_STALE;
5327 root = mc->mc_db->md_root;
5329 if (root == P_INVALID) { /* Tree is empty. */
5330 DPUTS("tree is empty");
5331 return MDB_NOTFOUND;
5335 mdb_cassert(mc, root > 1);
5336 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5337 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5343 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5344 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5346 if (flags & MDB_PS_MODIFY) {
5347 if ((rc = mdb_page_touch(mc)))
5351 if (flags & MDB_PS_ROOTONLY)
5354 return mdb_page_search_root(mc, key, flags);
5358 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5360 MDB_txn *txn = mc->mc_txn;
5361 pgno_t pg = mp->mp_pgno;
5362 unsigned x = 0, ovpages = mp->mp_pages;
5363 MDB_env *env = txn->mt_env;
5364 MDB_IDL sl = txn->mt_spill_pgs;
5365 MDB_ID pn = pg << 1;
5368 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5369 /* If the page is dirty or on the spill list we just acquired it,
5370 * so we should give it back to our current free list, if any.
5371 * Otherwise put it onto the list of pages we freed in this txn.
5373 * Won't create me_pghead: me_pglast must be inited along with it.
5374 * Unsupported in nested txns: They would need to hide the page
5375 * range in ancestor txns' dirty and spilled lists.
5377 if (env->me_pghead &&
5379 ((mp->mp_flags & P_DIRTY) ||
5380 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5384 MDB_ID2 *dl, ix, iy;
5385 rc = mdb_midl_need(&env->me_pghead, ovpages);
5388 if (!(mp->mp_flags & P_DIRTY)) {
5389 /* This page is no longer spilled */
5396 /* Remove from dirty list */
5397 dl = txn->mt_u.dirty_list;
5399 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5405 mdb_cassert(mc, x > 1);
5407 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5408 txn->mt_flags |= MDB_TXN_ERROR;
5409 return MDB_CORRUPTED;
5412 if (!(env->me_flags & MDB_WRITEMAP))
5413 mdb_dpage_free(env, mp);
5415 /* Insert in me_pghead */
5416 mop = env->me_pghead;
5417 j = mop[0] + ovpages;
5418 for (i = mop[0]; i && mop[i] < pg; i--)
5424 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5428 mc->mc_db->md_overflow_pages -= ovpages;
5432 /** Return the data associated with a given node.
5433 * @param[in] txn The transaction for this operation.
5434 * @param[in] leaf The node being read.
5435 * @param[out] data Updated to point to the node's data.
5436 * @return 0 on success, non-zero on failure.
5439 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5441 MDB_page *omp; /* overflow page */
5445 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5446 data->mv_size = NODEDSZ(leaf);
5447 data->mv_data = NODEDATA(leaf);
5451 /* Read overflow data.
5453 data->mv_size = NODEDSZ(leaf);
5454 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5455 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5456 DPRINTF(("read overflow page %"Z"u failed", pgno));
5459 data->mv_data = METADATA(omp);
5465 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5466 MDB_val *key, MDB_val *data)
5473 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5475 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5478 if (txn->mt_flags & MDB_TXN_ERROR)
5481 mdb_cursor_init(&mc, txn, dbi, &mx);
5482 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5485 /** Find a sibling for a page.
5486 * Replaces the page at the top of the cursor's stack with the
5487 * specified sibling, if one exists.
5488 * @param[in] mc The cursor for this operation.
5489 * @param[in] move_right Non-zero if the right sibling is requested,
5490 * otherwise the left sibling.
5491 * @return 0 on success, non-zero on failure.
5494 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5500 if (mc->mc_snum < 2) {
5501 return MDB_NOTFOUND; /* root has no siblings */
5505 DPRINTF(("parent page is page %"Z"u, index %u",
5506 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5508 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5509 : (mc->mc_ki[mc->mc_top] == 0)) {
5510 DPRINTF(("no more keys left, moving to %s sibling",
5511 move_right ? "right" : "left"));
5512 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5513 /* undo cursor_pop before returning */
5520 mc->mc_ki[mc->mc_top]++;
5522 mc->mc_ki[mc->mc_top]--;
5523 DPRINTF(("just moving to %s index key %u",
5524 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5526 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5528 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5529 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5530 /* mc will be inconsistent if caller does mc_snum++ as above */
5531 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5535 mdb_cursor_push(mc, mp);
5537 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5542 /** Move the cursor to the next data item. */
5544 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5550 if (mc->mc_flags & C_EOF) {
5551 return MDB_NOTFOUND;
5554 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5556 mp = mc->mc_pg[mc->mc_top];
5558 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5559 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5560 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5561 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5562 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5563 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5564 if (rc == MDB_SUCCESS)
5565 MDB_GET_KEY(leaf, key);
5570 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5571 if (op == MDB_NEXT_DUP)
5572 return MDB_NOTFOUND;
5576 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5577 mdb_dbg_pgno(mp), (void *) mc));
5578 if (mc->mc_flags & C_DEL)
5581 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5582 DPUTS("=====> move to next sibling page");
5583 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5584 mc->mc_flags |= C_EOF;
5587 mp = mc->mc_pg[mc->mc_top];
5588 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5590 mc->mc_ki[mc->mc_top]++;
5593 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5594 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5597 key->mv_size = mc->mc_db->md_pad;
5598 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5602 mdb_cassert(mc, IS_LEAF(mp));
5603 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5605 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5606 mdb_xcursor_init1(mc, leaf);
5609 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5612 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5613 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5614 if (rc != MDB_SUCCESS)
5619 MDB_GET_KEY(leaf, key);
5623 /** Move the cursor to the previous data item. */
5625 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5631 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5633 mp = mc->mc_pg[mc->mc_top];
5635 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5636 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5637 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5638 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5639 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5640 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5641 if (rc == MDB_SUCCESS) {
5642 MDB_GET_KEY(leaf, key);
5643 mc->mc_flags &= ~C_EOF;
5649 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5650 if (op == MDB_PREV_DUP)
5651 return MDB_NOTFOUND;
5655 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5656 mdb_dbg_pgno(mp), (void *) mc));
5658 if (mc->mc_ki[mc->mc_top] == 0) {
5659 DPUTS("=====> move to prev sibling page");
5660 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5663 mp = mc->mc_pg[mc->mc_top];
5664 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5665 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5667 mc->mc_ki[mc->mc_top]--;
5669 mc->mc_flags &= ~C_EOF;
5671 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5672 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5675 key->mv_size = mc->mc_db->md_pad;
5676 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5680 mdb_cassert(mc, IS_LEAF(mp));
5681 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5683 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5684 mdb_xcursor_init1(mc, leaf);
5687 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5690 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5691 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5692 if (rc != MDB_SUCCESS)
5697 MDB_GET_KEY(leaf, key);
5701 /** Set the cursor on a specific data item. */
5703 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5704 MDB_cursor_op op, int *exactp)
5708 MDB_node *leaf = NULL;
5711 if (key->mv_size == 0)
5712 return MDB_BAD_VALSIZE;
5715 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5717 /* See if we're already on the right page */
5718 if (mc->mc_flags & C_INITIALIZED) {
5721 mp = mc->mc_pg[mc->mc_top];
5723 mc->mc_ki[mc->mc_top] = 0;
5724 return MDB_NOTFOUND;
5726 if (mp->mp_flags & P_LEAF2) {
5727 nodekey.mv_size = mc->mc_db->md_pad;
5728 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5730 leaf = NODEPTR(mp, 0);
5731 MDB_GET_KEY2(leaf, nodekey);
5733 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5735 /* Probably happens rarely, but first node on the page
5736 * was the one we wanted.
5738 mc->mc_ki[mc->mc_top] = 0;
5745 unsigned int nkeys = NUMKEYS(mp);
5747 if (mp->mp_flags & P_LEAF2) {
5748 nodekey.mv_data = LEAF2KEY(mp,
5749 nkeys-1, nodekey.mv_size);
5751 leaf = NODEPTR(mp, nkeys-1);
5752 MDB_GET_KEY2(leaf, nodekey);
5754 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5756 /* last node was the one we wanted */
5757 mc->mc_ki[mc->mc_top] = nkeys-1;
5763 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5764 /* This is definitely the right page, skip search_page */
5765 if (mp->mp_flags & P_LEAF2) {
5766 nodekey.mv_data = LEAF2KEY(mp,
5767 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5769 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5770 MDB_GET_KEY2(leaf, nodekey);
5772 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5774 /* current node was the one we wanted */
5784 /* If any parents have right-sibs, search.
5785 * Otherwise, there's nothing further.
5787 for (i=0; i<mc->mc_top; i++)
5789 NUMKEYS(mc->mc_pg[i])-1)
5791 if (i == mc->mc_top) {
5792 /* There are no other pages */
5793 mc->mc_ki[mc->mc_top] = nkeys;
5794 return MDB_NOTFOUND;
5798 /* There are no other pages */
5799 mc->mc_ki[mc->mc_top] = 0;
5800 if (op == MDB_SET_RANGE && !exactp) {
5804 return MDB_NOTFOUND;
5808 rc = mdb_page_search(mc, key, 0);
5809 if (rc != MDB_SUCCESS)
5812 mp = mc->mc_pg[mc->mc_top];
5813 mdb_cassert(mc, IS_LEAF(mp));
5816 leaf = mdb_node_search(mc, key, exactp);
5817 if (exactp != NULL && !*exactp) {
5818 /* MDB_SET specified and not an exact match. */
5819 return MDB_NOTFOUND;
5823 DPUTS("===> inexact leaf not found, goto sibling");
5824 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5825 return rc; /* no entries matched */
5826 mp = mc->mc_pg[mc->mc_top];
5827 mdb_cassert(mc, IS_LEAF(mp));
5828 leaf = NODEPTR(mp, 0);
5832 mc->mc_flags |= C_INITIALIZED;
5833 mc->mc_flags &= ~C_EOF;
5836 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5837 key->mv_size = mc->mc_db->md_pad;
5838 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5843 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5844 mdb_xcursor_init1(mc, leaf);
5847 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5848 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5849 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5852 if (op == MDB_GET_BOTH) {
5858 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5859 if (rc != MDB_SUCCESS)
5862 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5864 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5866 rc = mc->mc_dbx->md_dcmp(data, &d2);
5868 if (op == MDB_GET_BOTH || rc > 0)
5869 return MDB_NOTFOUND;
5876 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5877 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5882 /* The key already matches in all other cases */
5883 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5884 MDB_GET_KEY(leaf, key);
5885 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5890 /** Move the cursor to the first item in the database. */
5892 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5898 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5900 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5901 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5902 if (rc != MDB_SUCCESS)
5905 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5907 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5908 mc->mc_flags |= C_INITIALIZED;
5909 mc->mc_flags &= ~C_EOF;
5911 mc->mc_ki[mc->mc_top] = 0;
5913 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5914 key->mv_size = mc->mc_db->md_pad;
5915 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5920 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5921 mdb_xcursor_init1(mc, leaf);
5922 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5926 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5930 MDB_GET_KEY(leaf, key);
5934 /** Move the cursor to the last item in the database. */
5936 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5942 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5944 if (!(mc->mc_flags & C_EOF)) {
5946 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5947 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5948 if (rc != MDB_SUCCESS)
5951 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5954 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5955 mc->mc_flags |= C_INITIALIZED|C_EOF;
5956 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5958 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5959 key->mv_size = mc->mc_db->md_pad;
5960 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5965 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5966 mdb_xcursor_init1(mc, leaf);
5967 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5971 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5976 MDB_GET_KEY(leaf, key);
5981 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5986 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5991 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5995 case MDB_GET_CURRENT:
5996 if (!(mc->mc_flags & C_INITIALIZED)) {
5999 MDB_page *mp = mc->mc_pg[mc->mc_top];
6000 int nkeys = NUMKEYS(mp);
6001 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6002 mc->mc_ki[mc->mc_top] = nkeys;
6008 key->mv_size = mc->mc_db->md_pad;
6009 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6011 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6012 MDB_GET_KEY(leaf, key);
6014 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6015 if (mc->mc_flags & C_DEL)
6016 mdb_xcursor_init1(mc, leaf);
6017 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6019 rc = mdb_node_read(mc->mc_txn, leaf, data);
6026 case MDB_GET_BOTH_RANGE:
6031 if (mc->mc_xcursor == NULL) {
6032 rc = MDB_INCOMPATIBLE;
6042 rc = mdb_cursor_set(mc, key, data, op,
6043 op == MDB_SET_RANGE ? NULL : &exact);
6046 case MDB_GET_MULTIPLE:
6047 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6051 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6052 rc = MDB_INCOMPATIBLE;
6056 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6057 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6060 case MDB_NEXT_MULTIPLE:
6065 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6066 rc = MDB_INCOMPATIBLE;
6069 if (!(mc->mc_flags & C_INITIALIZED))
6070 rc = mdb_cursor_first(mc, key, data);
6072 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6073 if (rc == MDB_SUCCESS) {
6074 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6077 mx = &mc->mc_xcursor->mx_cursor;
6078 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6080 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6081 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6089 case MDB_NEXT_NODUP:
6090 if (!(mc->mc_flags & C_INITIALIZED))
6091 rc = mdb_cursor_first(mc, key, data);
6093 rc = mdb_cursor_next(mc, key, data, op);
6097 case MDB_PREV_NODUP:
6098 if (!(mc->mc_flags & C_INITIALIZED)) {
6099 rc = mdb_cursor_last(mc, key, data);
6102 mc->mc_flags |= C_INITIALIZED;
6103 mc->mc_ki[mc->mc_top]++;
6105 rc = mdb_cursor_prev(mc, key, data, op);
6108 rc = mdb_cursor_first(mc, key, data);
6111 mfunc = mdb_cursor_first;
6113 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6117 if (mc->mc_xcursor == NULL) {
6118 rc = MDB_INCOMPATIBLE;
6122 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6123 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6124 MDB_GET_KEY(leaf, key);
6125 rc = mdb_node_read(mc->mc_txn, leaf, data);
6129 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6133 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6136 rc = mdb_cursor_last(mc, key, data);
6139 mfunc = mdb_cursor_last;
6142 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6147 if (mc->mc_flags & C_DEL)
6148 mc->mc_flags ^= C_DEL;
6153 /** Touch all the pages in the cursor stack. Set mc_top.
6154 * Makes sure all the pages are writable, before attempting a write operation.
6155 * @param[in] mc The cursor to operate on.
6158 mdb_cursor_touch(MDB_cursor *mc)
6160 int rc = MDB_SUCCESS;
6162 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6165 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6167 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6168 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6171 *mc->mc_dbflag |= DB_DIRTY;
6176 rc = mdb_page_touch(mc);
6177 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6178 mc->mc_top = mc->mc_snum-1;
6183 /** Do not spill pages to disk if txn is getting full, may fail instead */
6184 #define MDB_NOSPILL 0x8000
6187 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6191 MDB_node *leaf = NULL;
6194 MDB_val xdata, *rdata, dkey, olddata;
6196 int do_sub = 0, insert_key, insert_data;
6197 unsigned int mcount = 0, dcount = 0, nospill;
6200 unsigned int nflags;
6203 if (mc == NULL || key == NULL)
6206 env = mc->mc_txn->mt_env;
6208 /* Check this first so counter will always be zero on any
6211 if (flags & MDB_MULTIPLE) {
6212 dcount = data[1].mv_size;
6213 data[1].mv_size = 0;
6214 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6215 return MDB_INCOMPATIBLE;
6218 nospill = flags & MDB_NOSPILL;
6219 flags &= ~MDB_NOSPILL;
6221 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6222 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6224 if (key->mv_size-1 >= ENV_MAXKEY(env))
6225 return MDB_BAD_VALSIZE;
6227 #if SIZE_MAX > MAXDATASIZE
6228 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6229 return MDB_BAD_VALSIZE;
6231 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6232 return MDB_BAD_VALSIZE;
6235 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6236 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6240 if (flags == MDB_CURRENT) {
6241 if (!(mc->mc_flags & C_INITIALIZED))
6244 } else if (mc->mc_db->md_root == P_INVALID) {
6245 /* new database, cursor has nothing to point to */
6248 mc->mc_flags &= ~C_INITIALIZED;
6253 if (flags & MDB_APPEND) {
6255 rc = mdb_cursor_last(mc, &k2, &d2);
6257 rc = mc->mc_dbx->md_cmp(key, &k2);
6260 mc->mc_ki[mc->mc_top]++;
6262 /* new key is <= last key */
6267 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6269 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6270 DPRINTF(("duplicate key [%s]", DKEY(key)));
6272 return MDB_KEYEXIST;
6274 if (rc && rc != MDB_NOTFOUND)
6278 if (mc->mc_flags & C_DEL)
6279 mc->mc_flags ^= C_DEL;
6281 /* Cursor is positioned, check for room in the dirty list */
6283 if (flags & MDB_MULTIPLE) {
6285 xdata.mv_size = data->mv_size * dcount;
6289 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6293 if (rc == MDB_NO_ROOT) {
6295 /* new database, write a root leaf page */
6296 DPUTS("allocating new root leaf page");
6297 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6300 mdb_cursor_push(mc, np);
6301 mc->mc_db->md_root = np->mp_pgno;
6302 mc->mc_db->md_depth++;
6303 *mc->mc_dbflag |= DB_DIRTY;
6304 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6306 np->mp_flags |= P_LEAF2;
6307 mc->mc_flags |= C_INITIALIZED;
6309 /* make sure all cursor pages are writable */
6310 rc2 = mdb_cursor_touch(mc);
6315 insert_key = insert_data = rc;
6317 /* The key does not exist */
6318 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6319 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6320 LEAFSIZE(key, data) > env->me_nodemax)
6322 /* Too big for a node, insert in sub-DB. Set up an empty
6323 * "old sub-page" for prep_subDB to expand to a full page.
6325 fp_flags = P_LEAF|P_DIRTY;
6327 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6328 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6329 olddata.mv_size = PAGEHDRSZ;
6333 /* there's only a key anyway, so this is a no-op */
6334 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6336 unsigned int ksize = mc->mc_db->md_pad;
6337 if (key->mv_size != ksize)
6338 return MDB_BAD_VALSIZE;
6339 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6340 memcpy(ptr, key->mv_data, ksize);
6342 /* if overwriting slot 0 of leaf, need to
6343 * update branch key if there is a parent page
6345 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6346 unsigned short top = mc->mc_top;
6348 /* slot 0 is always an empty key, find real slot */
6349 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6351 if (mc->mc_ki[mc->mc_top])
6352 rc2 = mdb_update_key(mc, key);
6363 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6364 olddata.mv_size = NODEDSZ(leaf);
6365 olddata.mv_data = NODEDATA(leaf);
6368 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6369 /* Prepare (sub-)page/sub-DB to accept the new item,
6370 * if needed. fp: old sub-page or a header faking
6371 * it. mp: new (sub-)page. offset: growth in page
6372 * size. xdata: node data with new page or DB.
6374 unsigned i, offset = 0;
6375 mp = fp = xdata.mv_data = env->me_pbuf;
6376 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6378 /* Was a single item before, must convert now */
6379 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6380 /* Just overwrite the current item */
6381 if (flags == MDB_CURRENT)
6384 #if UINT_MAX < SIZE_MAX
6385 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6386 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6388 /* does data match? */
6389 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6390 if (flags & MDB_NODUPDATA)
6391 return MDB_KEYEXIST;
6396 /* Back up original data item */
6397 dkey.mv_size = olddata.mv_size;
6398 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6400 /* Make sub-page header for the dup items, with dummy body */
6401 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6402 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6403 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6404 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6405 fp->mp_flags |= P_LEAF2;
6406 fp->mp_pad = data->mv_size;
6407 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6409 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6410 (dkey.mv_size & 1) + (data->mv_size & 1);
6412 fp->mp_upper = xdata.mv_size - PAGEBASE;
6413 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6414 } else if (leaf->mn_flags & F_SUBDATA) {
6415 /* Data is on sub-DB, just store it */
6416 flags |= F_DUPDATA|F_SUBDATA;
6419 /* Data is on sub-page */
6420 fp = olddata.mv_data;
6423 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6424 offset = EVEN(NODESIZE + sizeof(indx_t) +
6428 offset = fp->mp_pad;
6429 if (SIZELEFT(fp) < offset) {
6430 offset *= 4; /* space for 4 more */
6433 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6435 fp->mp_flags |= P_DIRTY;
6436 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6437 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6441 xdata.mv_size = olddata.mv_size + offset;
6444 fp_flags = fp->mp_flags;
6445 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6446 /* Too big for a sub-page, convert to sub-DB */
6447 fp_flags &= ~P_SUBP;
6449 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6450 fp_flags |= P_LEAF2;
6451 dummy.md_pad = fp->mp_pad;
6452 dummy.md_flags = MDB_DUPFIXED;
6453 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6454 dummy.md_flags |= MDB_INTEGERKEY;
6460 dummy.md_branch_pages = 0;
6461 dummy.md_leaf_pages = 1;
6462 dummy.md_overflow_pages = 0;
6463 dummy.md_entries = NUMKEYS(fp);
6464 xdata.mv_size = sizeof(MDB_db);
6465 xdata.mv_data = &dummy;
6466 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6468 offset = env->me_psize - olddata.mv_size;
6469 flags |= F_DUPDATA|F_SUBDATA;
6470 dummy.md_root = mp->mp_pgno;
6473 mp->mp_flags = fp_flags | P_DIRTY;
6474 mp->mp_pad = fp->mp_pad;
6475 mp->mp_lower = fp->mp_lower;
6476 mp->mp_upper = fp->mp_upper + offset;
6477 if (fp_flags & P_LEAF2) {
6478 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6480 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6481 olddata.mv_size - fp->mp_upper - PAGEBASE);
6482 for (i=0; i<NUMKEYS(fp); i++)
6483 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6491 mdb_node_del(mc, 0);
6495 /* overflow page overwrites need special handling */
6496 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6499 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6501 memcpy(&pg, olddata.mv_data, sizeof(pg));
6502 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6504 ovpages = omp->mp_pages;
6506 /* Is the ov page large enough? */
6507 if (ovpages >= dpages) {
6508 if (!(omp->mp_flags & P_DIRTY) &&
6509 (level || (env->me_flags & MDB_WRITEMAP)))
6511 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6514 level = 0; /* dirty in this txn or clean */
6517 if (omp->mp_flags & P_DIRTY) {
6518 /* yes, overwrite it. Note in this case we don't
6519 * bother to try shrinking the page if the new data
6520 * is smaller than the overflow threshold.
6523 /* It is writable only in a parent txn */
6524 size_t sz = (size_t) env->me_psize * ovpages, off;
6525 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6531 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6532 mdb_cassert(mc, rc2 == 0);
6533 if (!(flags & MDB_RESERVE)) {
6534 /* Copy end of page, adjusting alignment so
6535 * compiler may copy words instead of bytes.
6537 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6538 memcpy((size_t *)((char *)np + off),
6539 (size_t *)((char *)omp + off), sz - off);
6542 memcpy(np, omp, sz); /* Copy beginning of page */
6545 SETDSZ(leaf, data->mv_size);
6546 if (F_ISSET(flags, MDB_RESERVE))
6547 data->mv_data = METADATA(omp);
6549 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6553 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6555 } else if (data->mv_size == olddata.mv_size) {
6556 /* same size, just replace it. Note that we could
6557 * also reuse this node if the new data is smaller,
6558 * but instead we opt to shrink the node in that case.
6560 if (F_ISSET(flags, MDB_RESERVE))
6561 data->mv_data = olddata.mv_data;
6562 else if (!(mc->mc_flags & C_SUB))
6563 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6565 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6570 mdb_node_del(mc, 0);
6576 nflags = flags & NODE_ADD_FLAGS;
6577 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6578 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6579 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6580 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6582 nflags |= MDB_SPLIT_REPLACE;
6583 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6585 /* There is room already in this leaf page. */
6586 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6587 if (rc == 0 && insert_key) {
6588 /* Adjust other cursors pointing to mp */
6589 MDB_cursor *m2, *m3;
6590 MDB_dbi dbi = mc->mc_dbi;
6591 unsigned i = mc->mc_top;
6592 MDB_page *mp = mc->mc_pg[i];
6594 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6595 if (mc->mc_flags & C_SUB)
6596 m3 = &m2->mc_xcursor->mx_cursor;
6599 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6600 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6607 if (rc == MDB_SUCCESS) {
6608 /* Now store the actual data in the child DB. Note that we're
6609 * storing the user data in the keys field, so there are strict
6610 * size limits on dupdata. The actual data fields of the child
6611 * DB are all zero size.
6619 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6620 if (flags & MDB_CURRENT) {
6621 xflags = MDB_CURRENT|MDB_NOSPILL;
6623 mdb_xcursor_init1(mc, leaf);
6624 xflags = (flags & MDB_NODUPDATA) ?
6625 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6627 /* converted, write the original data first */
6629 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6633 /* Adjust other cursors pointing to mp */
6635 unsigned i = mc->mc_top;
6636 MDB_page *mp = mc->mc_pg[i];
6638 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6639 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6640 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6641 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6642 mdb_xcursor_init1(m2, leaf);
6646 /* we've done our job */
6649 ecount = mc->mc_xcursor->mx_db.md_entries;
6650 if (flags & MDB_APPENDDUP)
6651 xflags |= MDB_APPEND;
6652 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6653 if (flags & F_SUBDATA) {
6654 void *db = NODEDATA(leaf);
6655 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6657 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6659 /* Increment count unless we just replaced an existing item. */
6661 mc->mc_db->md_entries++;
6663 /* Invalidate txn if we created an empty sub-DB */
6666 /* If we succeeded and the key didn't exist before,
6667 * make sure the cursor is marked valid.
6669 mc->mc_flags |= C_INITIALIZED;
6671 if (flags & MDB_MULTIPLE) {
6674 /* let caller know how many succeeded, if any */
6675 data[1].mv_size = mcount;
6676 if (mcount < dcount) {
6677 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6678 insert_key = insert_data = 0;
6685 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6688 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6693 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6699 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6700 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6702 if (!(mc->mc_flags & C_INITIALIZED))
6705 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6706 return MDB_NOTFOUND;
6708 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6711 rc = mdb_cursor_touch(mc);
6715 mp = mc->mc_pg[mc->mc_top];
6718 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6720 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6721 if (flags & MDB_NODUPDATA) {
6722 /* mdb_cursor_del0() will subtract the final entry */
6723 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6725 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6726 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6728 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6731 /* If sub-DB still has entries, we're done */
6732 if (mc->mc_xcursor->mx_db.md_entries) {
6733 if (leaf->mn_flags & F_SUBDATA) {
6734 /* update subDB info */
6735 void *db = NODEDATA(leaf);
6736 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6739 /* shrink fake page */
6740 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6741 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6742 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6743 /* fix other sub-DB cursors pointed at this fake page */
6744 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6745 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6746 if (m2->mc_pg[mc->mc_top] == mp &&
6747 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6748 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6751 mc->mc_db->md_entries--;
6752 mc->mc_flags |= C_DEL;
6755 /* otherwise fall thru and delete the sub-DB */
6758 if (leaf->mn_flags & F_SUBDATA) {
6759 /* add all the child DB's pages to the free list */
6760 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6766 /* add overflow pages to free list */
6767 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6771 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6772 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6773 (rc = mdb_ovpage_free(mc, omp)))
6778 return mdb_cursor_del0(mc);
6781 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6785 /** Allocate and initialize new pages for a database.
6786 * @param[in] mc a cursor on the database being added to.
6787 * @param[in] flags flags defining what type of page is being allocated.
6788 * @param[in] num the number of pages to allocate. This is usually 1,
6789 * unless allocating overflow pages for a large record.
6790 * @param[out] mp Address of a page, or NULL on failure.
6791 * @return 0 on success, non-zero on failure.
6794 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6799 if ((rc = mdb_page_alloc(mc, num, &np)))
6801 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6802 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6803 np->mp_flags = flags | P_DIRTY;
6804 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6805 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6808 mc->mc_db->md_branch_pages++;
6809 else if (IS_LEAF(np))
6810 mc->mc_db->md_leaf_pages++;
6811 else if (IS_OVERFLOW(np)) {
6812 mc->mc_db->md_overflow_pages += num;
6820 /** Calculate the size of a leaf node.
6821 * The size depends on the environment's page size; if a data item
6822 * is too large it will be put onto an overflow page and the node
6823 * size will only include the key and not the data. Sizes are always
6824 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6825 * of the #MDB_node headers.
6826 * @param[in] env The environment handle.
6827 * @param[in] key The key for the node.
6828 * @param[in] data The data for the node.
6829 * @return The number of bytes needed to store the node.
6832 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6836 sz = LEAFSIZE(key, data);
6837 if (sz > env->me_nodemax) {
6838 /* put on overflow page */
6839 sz -= data->mv_size - sizeof(pgno_t);
6842 return EVEN(sz + sizeof(indx_t));
6845 /** Calculate the size of a branch node.
6846 * The size should depend on the environment's page size but since
6847 * we currently don't support spilling large keys onto overflow
6848 * pages, it's simply the size of the #MDB_node header plus the
6849 * size of the key. Sizes are always rounded up to an even number
6850 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6851 * @param[in] env The environment handle.
6852 * @param[in] key The key for the node.
6853 * @return The number of bytes needed to store the node.
6856 mdb_branch_size(MDB_env *env, MDB_val *key)
6861 if (sz > env->me_nodemax) {
6862 /* put on overflow page */
6863 /* not implemented */
6864 /* sz -= key->size - sizeof(pgno_t); */
6867 return sz + sizeof(indx_t);
6870 /** Add a node to the page pointed to by the cursor.
6871 * @param[in] mc The cursor for this operation.
6872 * @param[in] indx The index on the page where the new node should be added.
6873 * @param[in] key The key for the new node.
6874 * @param[in] data The data for the new node, if any.
6875 * @param[in] pgno The page number, if adding a branch node.
6876 * @param[in] flags Flags for the node.
6877 * @return 0 on success, non-zero on failure. Possible errors are:
6879 * <li>ENOMEM - failed to allocate overflow pages for the node.
6880 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6881 * should never happen since all callers already calculate the
6882 * page's free space before calling this function.
6886 mdb_node_add(MDB_cursor *mc, indx_t indx,
6887 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6890 size_t node_size = NODESIZE;
6894 MDB_page *mp = mc->mc_pg[mc->mc_top];
6895 MDB_page *ofp = NULL; /* overflow page */
6898 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6900 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6901 IS_LEAF(mp) ? "leaf" : "branch",
6902 IS_SUBP(mp) ? "sub-" : "",
6903 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6904 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6907 /* Move higher keys up one slot. */
6908 int ksize = mc->mc_db->md_pad, dif;
6909 char *ptr = LEAF2KEY(mp, indx, ksize);
6910 dif = NUMKEYS(mp) - indx;
6912 memmove(ptr+ksize, ptr, dif*ksize);
6913 /* insert new key */
6914 memcpy(ptr, key->mv_data, ksize);
6916 /* Just using these for counting */
6917 mp->mp_lower += sizeof(indx_t);
6918 mp->mp_upper -= ksize - sizeof(indx_t);
6922 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6924 node_size += key->mv_size;
6926 mdb_cassert(mc, data);
6927 if (F_ISSET(flags, F_BIGDATA)) {
6928 /* Data already on overflow page. */
6929 node_size += sizeof(pgno_t);
6930 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6931 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6933 /* Put data on overflow page. */
6934 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6935 data->mv_size, node_size+data->mv_size));
6936 node_size = EVEN(node_size + sizeof(pgno_t));
6937 if ((ssize_t)node_size > room)
6939 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6941 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6945 node_size += data->mv_size;
6948 node_size = EVEN(node_size);
6949 if ((ssize_t)node_size > room)
6953 /* Move higher pointers up one slot. */
6954 for (i = NUMKEYS(mp); i > indx; i--)
6955 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6957 /* Adjust free space offsets. */
6958 ofs = mp->mp_upper - node_size;
6959 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6960 mp->mp_ptrs[indx] = ofs;
6962 mp->mp_lower += sizeof(indx_t);
6964 /* Write the node data. */
6965 node = NODEPTR(mp, indx);
6966 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6967 node->mn_flags = flags;
6969 SETDSZ(node,data->mv_size);
6974 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6977 mdb_cassert(mc, key);
6979 if (F_ISSET(flags, F_BIGDATA))
6980 memcpy(node->mn_data + key->mv_size, data->mv_data,
6982 else if (F_ISSET(flags, MDB_RESERVE))
6983 data->mv_data = node->mn_data + key->mv_size;
6985 memcpy(node->mn_data + key->mv_size, data->mv_data,
6988 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6990 if (F_ISSET(flags, MDB_RESERVE))
6991 data->mv_data = METADATA(ofp);
6993 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
7000 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7001 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7002 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7003 DPRINTF(("node size = %"Z"u", node_size));
7004 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7005 return MDB_PAGE_FULL;
7008 /** Delete the specified node from a page.
7009 * @param[in] mc Cursor pointing to the node to delete.
7010 * @param[in] ksize The size of a node. Only used if the page is
7011 * part of a #MDB_DUPFIXED database.
7014 mdb_node_del(MDB_cursor *mc, int ksize)
7016 MDB_page *mp = mc->mc_pg[mc->mc_top];
7017 indx_t indx = mc->mc_ki[mc->mc_top];
7019 indx_t i, j, numkeys, ptr;
7023 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7024 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7025 numkeys = NUMKEYS(mp);
7026 mdb_cassert(mc, indx < numkeys);
7029 int x = numkeys - 1 - indx;
7030 base = LEAF2KEY(mp, indx, ksize);
7032 memmove(base, base + ksize, x * ksize);
7033 mp->mp_lower -= sizeof(indx_t);
7034 mp->mp_upper += ksize - sizeof(indx_t);
7038 node = NODEPTR(mp, indx);
7039 sz = NODESIZE + node->mn_ksize;
7041 if (F_ISSET(node->mn_flags, F_BIGDATA))
7042 sz += sizeof(pgno_t);
7044 sz += NODEDSZ(node);
7048 ptr = mp->mp_ptrs[indx];
7049 for (i = j = 0; i < numkeys; i++) {
7051 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7052 if (mp->mp_ptrs[i] < ptr)
7053 mp->mp_ptrs[j] += sz;
7058 base = (char *)mp + mp->mp_upper + PAGEBASE;
7059 memmove(base + sz, base, ptr - mp->mp_upper);
7061 mp->mp_lower -= sizeof(indx_t);
7065 /** Compact the main page after deleting a node on a subpage.
7066 * @param[in] mp The main page to operate on.
7067 * @param[in] indx The index of the subpage on the main page.
7070 mdb_node_shrink(MDB_page *mp, indx_t indx)
7076 indx_t i, numkeys, ptr;
7078 node = NODEPTR(mp, indx);
7079 sp = (MDB_page *)NODEDATA(node);
7080 delta = SIZELEFT(sp);
7081 xp = (MDB_page *)((char *)sp + delta);
7083 /* shift subpage upward */
7085 nsize = NUMKEYS(sp) * sp->mp_pad;
7087 return; /* do not make the node uneven-sized */
7088 memmove(METADATA(xp), METADATA(sp), nsize);
7091 numkeys = NUMKEYS(sp);
7092 for (i=numkeys-1; i>=0; i--)
7093 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7095 xp->mp_upper = sp->mp_lower;
7096 xp->mp_lower = sp->mp_lower;
7097 xp->mp_flags = sp->mp_flags;
7098 xp->mp_pad = sp->mp_pad;
7099 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7101 nsize = NODEDSZ(node) - delta;
7102 SETDSZ(node, nsize);
7104 /* shift lower nodes upward */
7105 ptr = mp->mp_ptrs[indx];
7106 numkeys = NUMKEYS(mp);
7107 for (i = 0; i < numkeys; i++) {
7108 if (mp->mp_ptrs[i] <= ptr)
7109 mp->mp_ptrs[i] += delta;
7112 base = (char *)mp + mp->mp_upper + PAGEBASE;
7113 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7114 mp->mp_upper += delta;
7117 /** Initial setup of a sorted-dups cursor.
7118 * Sorted duplicates are implemented as a sub-database for the given key.
7119 * The duplicate data items are actually keys of the sub-database.
7120 * Operations on the duplicate data items are performed using a sub-cursor
7121 * initialized when the sub-database is first accessed. This function does
7122 * the preliminary setup of the sub-cursor, filling in the fields that
7123 * depend only on the parent DB.
7124 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7127 mdb_xcursor_init0(MDB_cursor *mc)
7129 MDB_xcursor *mx = mc->mc_xcursor;
7131 mx->mx_cursor.mc_xcursor = NULL;
7132 mx->mx_cursor.mc_txn = mc->mc_txn;
7133 mx->mx_cursor.mc_db = &mx->mx_db;
7134 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7135 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7136 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7137 mx->mx_cursor.mc_snum = 0;
7138 mx->mx_cursor.mc_top = 0;
7139 mx->mx_cursor.mc_flags = C_SUB;
7140 mx->mx_dbx.md_name.mv_size = 0;
7141 mx->mx_dbx.md_name.mv_data = NULL;
7142 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7143 mx->mx_dbx.md_dcmp = NULL;
7144 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7147 /** Final setup of a sorted-dups cursor.
7148 * Sets up the fields that depend on the data from the main cursor.
7149 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7150 * @param[in] node The data containing the #MDB_db record for the
7151 * sorted-dup database.
7154 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7156 MDB_xcursor *mx = mc->mc_xcursor;
7158 if (node->mn_flags & F_SUBDATA) {
7159 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7160 mx->mx_cursor.mc_pg[0] = 0;
7161 mx->mx_cursor.mc_snum = 0;
7162 mx->mx_cursor.mc_top = 0;
7163 mx->mx_cursor.mc_flags = C_SUB;
7165 MDB_page *fp = NODEDATA(node);
7166 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7167 mx->mx_db.md_flags = 0;
7168 mx->mx_db.md_depth = 1;
7169 mx->mx_db.md_branch_pages = 0;
7170 mx->mx_db.md_leaf_pages = 1;
7171 mx->mx_db.md_overflow_pages = 0;
7172 mx->mx_db.md_entries = NUMKEYS(fp);
7173 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7174 mx->mx_cursor.mc_snum = 1;
7175 mx->mx_cursor.mc_top = 0;
7176 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7177 mx->mx_cursor.mc_pg[0] = fp;
7178 mx->mx_cursor.mc_ki[0] = 0;
7179 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7180 mx->mx_db.md_flags = MDB_DUPFIXED;
7181 mx->mx_db.md_pad = fp->mp_pad;
7182 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7183 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7186 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7187 mx->mx_db.md_root));
7188 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7189 #if UINT_MAX < SIZE_MAX
7190 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7191 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7195 /** Initialize a cursor for a given transaction and database. */
7197 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7200 mc->mc_backup = NULL;
7203 mc->mc_db = &txn->mt_dbs[dbi];
7204 mc->mc_dbx = &txn->mt_dbxs[dbi];
7205 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7210 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7211 mdb_tassert(txn, mx != NULL);
7212 mc->mc_xcursor = mx;
7213 mdb_xcursor_init0(mc);
7215 mc->mc_xcursor = NULL;
7217 if (*mc->mc_dbflag & DB_STALE) {
7218 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7223 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7226 size_t size = sizeof(MDB_cursor);
7228 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7231 if (txn->mt_flags & MDB_TXN_ERROR)
7234 /* Allow read access to the freelist */
7235 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7238 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7239 size += sizeof(MDB_xcursor);
7241 if ((mc = malloc(size)) != NULL) {
7242 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7243 if (txn->mt_cursors) {
7244 mc->mc_next = txn->mt_cursors[dbi];
7245 txn->mt_cursors[dbi] = mc;
7246 mc->mc_flags |= C_UNTRACK;
7258 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7260 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7263 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7266 if (txn->mt_flags & MDB_TXN_ERROR)
7269 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7273 /* Return the count of duplicate data items for the current key */
7275 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7279 if (mc == NULL || countp == NULL)
7282 if (mc->mc_xcursor == NULL)
7283 return MDB_INCOMPATIBLE;
7285 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7288 if (!(mc->mc_flags & C_INITIALIZED))
7291 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7292 return MDB_NOTFOUND;
7294 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7295 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7298 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7301 *countp = mc->mc_xcursor->mx_db.md_entries;
7307 mdb_cursor_close(MDB_cursor *mc)
7309 if (mc && !mc->mc_backup) {
7310 /* remove from txn, if tracked */
7311 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7312 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7313 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7315 *prev = mc->mc_next;
7322 mdb_cursor_txn(MDB_cursor *mc)
7324 if (!mc) return NULL;
7329 mdb_cursor_dbi(MDB_cursor *mc)
7334 /** Replace the key for a branch node with a new key.
7335 * @param[in] mc Cursor pointing to the node to operate on.
7336 * @param[in] key The new key to use.
7337 * @return 0 on success, non-zero on failure.
7340 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7346 int delta, ksize, oksize;
7347 indx_t ptr, i, numkeys, indx;
7350 indx = mc->mc_ki[mc->mc_top];
7351 mp = mc->mc_pg[mc->mc_top];
7352 node = NODEPTR(mp, indx);
7353 ptr = mp->mp_ptrs[indx];
7357 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7358 k2.mv_data = NODEKEY(node);
7359 k2.mv_size = node->mn_ksize;
7360 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7362 mdb_dkey(&k2, kbuf2),
7368 /* Sizes must be 2-byte aligned. */
7369 ksize = EVEN(key->mv_size);
7370 oksize = EVEN(node->mn_ksize);
7371 delta = ksize - oksize;
7373 /* Shift node contents if EVEN(key length) changed. */
7375 if (delta > 0 && SIZELEFT(mp) < delta) {
7377 /* not enough space left, do a delete and split */
7378 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7379 pgno = NODEPGNO(node);
7380 mdb_node_del(mc, 0);
7381 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7384 numkeys = NUMKEYS(mp);
7385 for (i = 0; i < numkeys; i++) {
7386 if (mp->mp_ptrs[i] <= ptr)
7387 mp->mp_ptrs[i] -= delta;
7390 base = (char *)mp + mp->mp_upper + PAGEBASE;
7391 len = ptr - mp->mp_upper + NODESIZE;
7392 memmove(base - delta, base, len);
7393 mp->mp_upper -= delta;
7395 node = NODEPTR(mp, indx);
7398 /* But even if no shift was needed, update ksize */
7399 if (node->mn_ksize != key->mv_size)
7400 node->mn_ksize = key->mv_size;
7403 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7409 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7411 /** Move a node from csrc to cdst.
7414 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7421 unsigned short flags;
7425 /* Mark src and dst as dirty. */
7426 if ((rc = mdb_page_touch(csrc)) ||
7427 (rc = mdb_page_touch(cdst)))
7430 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7431 key.mv_size = csrc->mc_db->md_pad;
7432 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7434 data.mv_data = NULL;
7438 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7439 mdb_cassert(csrc, !((size_t)srcnode & 1));
7440 srcpg = NODEPGNO(srcnode);
7441 flags = srcnode->mn_flags;
7442 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7443 unsigned int snum = csrc->mc_snum;
7445 /* must find the lowest key below src */
7446 rc = mdb_page_search_lowest(csrc);
7449 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7450 key.mv_size = csrc->mc_db->md_pad;
7451 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7453 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7454 key.mv_size = NODEKSZ(s2);
7455 key.mv_data = NODEKEY(s2);
7457 csrc->mc_snum = snum--;
7458 csrc->mc_top = snum;
7460 key.mv_size = NODEKSZ(srcnode);
7461 key.mv_data = NODEKEY(srcnode);
7463 data.mv_size = NODEDSZ(srcnode);
7464 data.mv_data = NODEDATA(srcnode);
7466 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7467 unsigned int snum = cdst->mc_snum;
7470 /* must find the lowest key below dst */
7471 mdb_cursor_copy(cdst, &mn);
7472 rc = mdb_page_search_lowest(&mn);
7475 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7476 bkey.mv_size = mn.mc_db->md_pad;
7477 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7479 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7480 bkey.mv_size = NODEKSZ(s2);
7481 bkey.mv_data = NODEKEY(s2);
7483 mn.mc_snum = snum--;
7486 rc = mdb_update_key(&mn, &bkey);
7491 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7492 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7493 csrc->mc_ki[csrc->mc_top],
7495 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7496 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7498 /* Add the node to the destination page.
7500 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7501 if (rc != MDB_SUCCESS)
7504 /* Delete the node from the source page.
7506 mdb_node_del(csrc, key.mv_size);
7509 /* Adjust other cursors pointing to mp */
7510 MDB_cursor *m2, *m3;
7511 MDB_dbi dbi = csrc->mc_dbi;
7512 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7514 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7515 if (csrc->mc_flags & C_SUB)
7516 m3 = &m2->mc_xcursor->mx_cursor;
7519 if (m3 == csrc) continue;
7520 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7521 csrc->mc_ki[csrc->mc_top]) {
7522 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7523 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7528 /* Update the parent separators.
7530 if (csrc->mc_ki[csrc->mc_top] == 0) {
7531 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7532 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7533 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7535 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7536 key.mv_size = NODEKSZ(srcnode);
7537 key.mv_data = NODEKEY(srcnode);
7539 DPRINTF(("update separator for source page %"Z"u to [%s]",
7540 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7541 mdb_cursor_copy(csrc, &mn);
7544 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7547 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7549 indx_t ix = csrc->mc_ki[csrc->mc_top];
7550 nullkey.mv_size = 0;
7551 csrc->mc_ki[csrc->mc_top] = 0;
7552 rc = mdb_update_key(csrc, &nullkey);
7553 csrc->mc_ki[csrc->mc_top] = ix;
7554 mdb_cassert(csrc, rc == MDB_SUCCESS);
7558 if (cdst->mc_ki[cdst->mc_top] == 0) {
7559 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7560 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7561 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7563 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7564 key.mv_size = NODEKSZ(srcnode);
7565 key.mv_data = NODEKEY(srcnode);
7567 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7568 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7569 mdb_cursor_copy(cdst, &mn);
7572 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7575 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7577 indx_t ix = cdst->mc_ki[cdst->mc_top];
7578 nullkey.mv_size = 0;
7579 cdst->mc_ki[cdst->mc_top] = 0;
7580 rc = mdb_update_key(cdst, &nullkey);
7581 cdst->mc_ki[cdst->mc_top] = ix;
7582 mdb_cassert(cdst, rc == MDB_SUCCESS);
7589 /** Merge one page into another.
7590 * The nodes from the page pointed to by \b csrc will
7591 * be copied to the page pointed to by \b cdst and then
7592 * the \b csrc page will be freed.
7593 * @param[in] csrc Cursor pointing to the source page.
7594 * @param[in] cdst Cursor pointing to the destination page.
7595 * @return 0 on success, non-zero on failure.
7598 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7600 MDB_page *psrc, *pdst;
7607 psrc = csrc->mc_pg[csrc->mc_top];
7608 pdst = cdst->mc_pg[cdst->mc_top];
7610 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7612 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7613 mdb_cassert(csrc, cdst->mc_snum > 1);
7615 /* Mark dst as dirty. */
7616 if ((rc = mdb_page_touch(cdst)))
7619 /* Move all nodes from src to dst.
7621 j = nkeys = NUMKEYS(pdst);
7622 if (IS_LEAF2(psrc)) {
7623 key.mv_size = csrc->mc_db->md_pad;
7624 key.mv_data = METADATA(psrc);
7625 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7626 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7627 if (rc != MDB_SUCCESS)
7629 key.mv_data = (char *)key.mv_data + key.mv_size;
7632 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7633 srcnode = NODEPTR(psrc, i);
7634 if (i == 0 && IS_BRANCH(psrc)) {
7637 mdb_cursor_copy(csrc, &mn);
7638 /* must find the lowest key below src */
7639 rc = mdb_page_search_lowest(&mn);
7642 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7643 key.mv_size = mn.mc_db->md_pad;
7644 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7646 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7647 key.mv_size = NODEKSZ(s2);
7648 key.mv_data = NODEKEY(s2);
7651 key.mv_size = srcnode->mn_ksize;
7652 key.mv_data = NODEKEY(srcnode);
7655 data.mv_size = NODEDSZ(srcnode);
7656 data.mv_data = NODEDATA(srcnode);
7657 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7658 if (rc != MDB_SUCCESS)
7663 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7664 pdst->mp_pgno, NUMKEYS(pdst),
7665 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7667 /* Unlink the src page from parent and add to free list.
7670 mdb_node_del(csrc, 0);
7671 if (csrc->mc_ki[csrc->mc_top] == 0) {
7673 rc = mdb_update_key(csrc, &key);
7681 psrc = csrc->mc_pg[csrc->mc_top];
7682 /* If not operating on FreeDB, allow this page to be reused
7683 * in this txn. Otherwise just add to free list.
7685 rc = mdb_page_loose(csrc, psrc);
7689 csrc->mc_db->md_leaf_pages--;
7691 csrc->mc_db->md_branch_pages--;
7693 /* Adjust other cursors pointing to mp */
7694 MDB_cursor *m2, *m3;
7695 MDB_dbi dbi = csrc->mc_dbi;
7697 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7698 if (csrc->mc_flags & C_SUB)
7699 m3 = &m2->mc_xcursor->mx_cursor;
7702 if (m3 == csrc) continue;
7703 if (m3->mc_snum < csrc->mc_snum) continue;
7704 if (m3->mc_pg[csrc->mc_top] == psrc) {
7705 m3->mc_pg[csrc->mc_top] = pdst;
7706 m3->mc_ki[csrc->mc_top] += nkeys;
7711 unsigned int snum = cdst->mc_snum;
7712 uint16_t depth = cdst->mc_db->md_depth;
7713 mdb_cursor_pop(cdst);
7714 rc = mdb_rebalance(cdst);
7715 /* Did the tree shrink? */
7716 if (depth > cdst->mc_db->md_depth)
7718 cdst->mc_snum = snum;
7719 cdst->mc_top = snum-1;
7724 /** Copy the contents of a cursor.
7725 * @param[in] csrc The cursor to copy from.
7726 * @param[out] cdst The cursor to copy to.
7729 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7733 cdst->mc_txn = csrc->mc_txn;
7734 cdst->mc_dbi = csrc->mc_dbi;
7735 cdst->mc_db = csrc->mc_db;
7736 cdst->mc_dbx = csrc->mc_dbx;
7737 cdst->mc_snum = csrc->mc_snum;
7738 cdst->mc_top = csrc->mc_top;
7739 cdst->mc_flags = csrc->mc_flags;
7741 for (i=0; i<csrc->mc_snum; i++) {
7742 cdst->mc_pg[i] = csrc->mc_pg[i];
7743 cdst->mc_ki[i] = csrc->mc_ki[i];
7747 /** Rebalance the tree after a delete operation.
7748 * @param[in] mc Cursor pointing to the page where rebalancing
7750 * @return 0 on success, non-zero on failure.
7753 mdb_rebalance(MDB_cursor *mc)
7757 unsigned int ptop, minkeys;
7761 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7762 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7763 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7764 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7765 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7767 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7768 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7769 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7770 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7774 if (mc->mc_snum < 2) {
7775 MDB_page *mp = mc->mc_pg[0];
7777 DPUTS("Can't rebalance a subpage, ignoring");
7780 if (NUMKEYS(mp) == 0) {
7781 DPUTS("tree is completely empty");
7782 mc->mc_db->md_root = P_INVALID;
7783 mc->mc_db->md_depth = 0;
7784 mc->mc_db->md_leaf_pages = 0;
7785 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7788 /* Adjust cursors pointing to mp */
7791 mc->mc_flags &= ~C_INITIALIZED;
7793 MDB_cursor *m2, *m3;
7794 MDB_dbi dbi = mc->mc_dbi;
7796 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7797 if (mc->mc_flags & C_SUB)
7798 m3 = &m2->mc_xcursor->mx_cursor;
7801 if (m3->mc_snum < mc->mc_snum) continue;
7802 if (m3->mc_pg[0] == mp) {
7805 m3->mc_flags &= ~C_INITIALIZED;
7809 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7811 DPUTS("collapsing root page!");
7812 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7815 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7816 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7819 mc->mc_db->md_depth--;
7820 mc->mc_db->md_branch_pages--;
7821 mc->mc_ki[0] = mc->mc_ki[1];
7822 for (i = 1; i<mc->mc_db->md_depth; i++) {
7823 mc->mc_pg[i] = mc->mc_pg[i+1];
7824 mc->mc_ki[i] = mc->mc_ki[i+1];
7827 /* Adjust other cursors pointing to mp */
7828 MDB_cursor *m2, *m3;
7829 MDB_dbi dbi = mc->mc_dbi;
7831 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7832 if (mc->mc_flags & C_SUB)
7833 m3 = &m2->mc_xcursor->mx_cursor;
7836 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7837 if (m3->mc_pg[0] == mp) {
7840 for (i=0; i<m3->mc_snum; i++) {
7841 m3->mc_pg[i] = m3->mc_pg[i+1];
7842 m3->mc_ki[i] = m3->mc_ki[i+1];
7848 DPUTS("root page doesn't need rebalancing");
7852 /* The parent (branch page) must have at least 2 pointers,
7853 * otherwise the tree is invalid.
7855 ptop = mc->mc_top-1;
7856 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7858 /* Leaf page fill factor is below the threshold.
7859 * Try to move keys from left or right neighbor, or
7860 * merge with a neighbor page.
7865 mdb_cursor_copy(mc, &mn);
7866 mn.mc_xcursor = NULL;
7868 oldki = mc->mc_ki[mc->mc_top];
7869 if (mc->mc_ki[ptop] == 0) {
7870 /* We're the leftmost leaf in our parent.
7872 DPUTS("reading right neighbor");
7874 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7875 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7878 mn.mc_ki[mn.mc_top] = 0;
7879 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7881 /* There is at least one neighbor to the left.
7883 DPUTS("reading left neighbor");
7885 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7886 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7889 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7890 mc->mc_ki[mc->mc_top] = 0;
7893 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7894 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7895 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7897 /* If the neighbor page is above threshold and has enough keys,
7898 * move one key from it. Otherwise we should try to merge them.
7899 * (A branch page must never have less than 2 keys.)
7901 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7902 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7903 rc = mdb_node_move(&mn, mc);
7904 if (mc->mc_ki[ptop]) {
7908 if (mc->mc_ki[ptop] == 0) {
7909 rc = mdb_page_merge(&mn, mc);
7911 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7912 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7913 rc = mdb_page_merge(mc, &mn);
7914 mdb_cursor_copy(&mn, mc);
7916 mc->mc_flags &= ~C_EOF;
7918 mc->mc_ki[mc->mc_top] = oldki;
7922 /** Complete a delete operation started by #mdb_cursor_del(). */
7924 mdb_cursor_del0(MDB_cursor *mc)
7931 ki = mc->mc_ki[mc->mc_top];
7932 mdb_node_del(mc, mc->mc_db->md_pad);
7933 mc->mc_db->md_entries--;
7934 rc = mdb_rebalance(mc);
7936 if (rc == MDB_SUCCESS) {
7937 MDB_cursor *m2, *m3;
7938 MDB_dbi dbi = mc->mc_dbi;
7940 mp = mc->mc_pg[mc->mc_top];
7941 nkeys = NUMKEYS(mp);
7943 /* if mc points past last node in page, find next sibling */
7944 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7945 rc = mdb_cursor_sibling(mc, 1);
7946 if (rc == MDB_NOTFOUND) {
7947 mc->mc_flags |= C_EOF;
7952 /* Adjust other cursors pointing to mp */
7953 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7954 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7955 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7957 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7959 if (m3->mc_pg[mc->mc_top] == mp) {
7960 if (m3->mc_ki[mc->mc_top] >= ki) {
7961 m3->mc_flags |= C_DEL;
7962 if (m3->mc_ki[mc->mc_top] > ki)
7963 m3->mc_ki[mc->mc_top]--;
7964 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7965 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7967 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7968 rc = mdb_cursor_sibling(m3, 1);
7969 if (rc == MDB_NOTFOUND) {
7970 m3->mc_flags |= C_EOF;
7976 mc->mc_flags |= C_DEL;
7980 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7985 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7986 MDB_val *key, MDB_val *data)
7988 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7991 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7992 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7994 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7995 /* must ignore any data */
7999 return mdb_del0(txn, dbi, key, data, 0);
8003 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8004 MDB_val *key, MDB_val *data, unsigned flags)
8009 MDB_val rdata, *xdata;
8013 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8015 mdb_cursor_init(&mc, txn, dbi, &mx);
8024 flags |= MDB_NODUPDATA;
8026 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8028 /* let mdb_page_split know about this cursor if needed:
8029 * delete will trigger a rebalance; if it needs to move
8030 * a node from one page to another, it will have to
8031 * update the parent's separator key(s). If the new sepkey
8032 * is larger than the current one, the parent page may
8033 * run out of space, triggering a split. We need this
8034 * cursor to be consistent until the end of the rebalance.
8036 mc.mc_flags |= C_UNTRACK;
8037 mc.mc_next = txn->mt_cursors[dbi];
8038 txn->mt_cursors[dbi] = &mc;
8039 rc = mdb_cursor_del(&mc, flags);
8040 txn->mt_cursors[dbi] = mc.mc_next;
8045 /** Split a page and insert a new node.
8046 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8047 * The cursor will be updated to point to the actual page and index where
8048 * the node got inserted after the split.
8049 * @param[in] newkey The key for the newly inserted node.
8050 * @param[in] newdata The data for the newly inserted node.
8051 * @param[in] newpgno The page number, if the new node is a branch node.
8052 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8053 * @return 0 on success, non-zero on failure.
8056 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8057 unsigned int nflags)
8060 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8063 int i, j, split_indx, nkeys, pmax;
8064 MDB_env *env = mc->mc_txn->mt_env;
8066 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8067 MDB_page *copy = NULL;
8068 MDB_page *mp, *rp, *pp;
8073 mp = mc->mc_pg[mc->mc_top];
8074 newindx = mc->mc_ki[mc->mc_top];
8075 nkeys = NUMKEYS(mp);
8077 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8078 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8079 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8081 /* Create a right sibling. */
8082 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8084 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8086 if (mc->mc_snum < 2) {
8087 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8089 /* shift current top to make room for new parent */
8090 mc->mc_pg[1] = mc->mc_pg[0];
8091 mc->mc_ki[1] = mc->mc_ki[0];
8094 mc->mc_db->md_root = pp->mp_pgno;
8095 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8096 mc->mc_db->md_depth++;
8099 /* Add left (implicit) pointer. */
8100 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8101 /* undo the pre-push */
8102 mc->mc_pg[0] = mc->mc_pg[1];
8103 mc->mc_ki[0] = mc->mc_ki[1];
8104 mc->mc_db->md_root = mp->mp_pgno;
8105 mc->mc_db->md_depth--;
8112 ptop = mc->mc_top-1;
8113 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8116 mc->mc_flags |= C_SPLITTING;
8117 mdb_cursor_copy(mc, &mn);
8118 mn.mc_pg[mn.mc_top] = rp;
8119 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8121 if (nflags & MDB_APPEND) {
8122 mn.mc_ki[mn.mc_top] = 0;
8124 split_indx = newindx;
8128 split_indx = (nkeys+1) / 2;
8133 unsigned int lsize, rsize, ksize;
8134 /* Move half of the keys to the right sibling */
8135 x = mc->mc_ki[mc->mc_top] - split_indx;
8136 ksize = mc->mc_db->md_pad;
8137 split = LEAF2KEY(mp, split_indx, ksize);
8138 rsize = (nkeys - split_indx) * ksize;
8139 lsize = (nkeys - split_indx) * sizeof(indx_t);
8140 mp->mp_lower -= lsize;
8141 rp->mp_lower += lsize;
8142 mp->mp_upper += rsize - lsize;
8143 rp->mp_upper -= rsize - lsize;
8144 sepkey.mv_size = ksize;
8145 if (newindx == split_indx) {
8146 sepkey.mv_data = newkey->mv_data;
8148 sepkey.mv_data = split;
8151 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8152 memcpy(rp->mp_ptrs, split, rsize);
8153 sepkey.mv_data = rp->mp_ptrs;
8154 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8155 memcpy(ins, newkey->mv_data, ksize);
8156 mp->mp_lower += sizeof(indx_t);
8157 mp->mp_upper -= ksize - sizeof(indx_t);
8160 memcpy(rp->mp_ptrs, split, x * ksize);
8161 ins = LEAF2KEY(rp, x, ksize);
8162 memcpy(ins, newkey->mv_data, ksize);
8163 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8164 rp->mp_lower += sizeof(indx_t);
8165 rp->mp_upper -= ksize - sizeof(indx_t);
8166 mc->mc_ki[mc->mc_top] = x;
8167 mc->mc_pg[mc->mc_top] = rp;
8170 int psize, nsize, k;
8171 /* Maximum free space in an empty page */
8172 pmax = env->me_psize - PAGEHDRSZ;
8174 nsize = mdb_leaf_size(env, newkey, newdata);
8176 nsize = mdb_branch_size(env, newkey);
8177 nsize = EVEN(nsize);
8179 /* grab a page to hold a temporary copy */
8180 copy = mdb_page_malloc(mc->mc_txn, 1);
8185 copy->mp_pgno = mp->mp_pgno;
8186 copy->mp_flags = mp->mp_flags;
8187 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8188 copy->mp_upper = env->me_psize - PAGEBASE;
8190 /* prepare to insert */
8191 for (i=0, j=0; i<nkeys; i++) {
8193 copy->mp_ptrs[j++] = 0;
8195 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8198 /* When items are relatively large the split point needs
8199 * to be checked, because being off-by-one will make the
8200 * difference between success or failure in mdb_node_add.
8202 * It's also relevant if a page happens to be laid out
8203 * such that one half of its nodes are all "small" and
8204 * the other half of its nodes are "large." If the new
8205 * item is also "large" and falls on the half with
8206 * "large" nodes, it also may not fit.
8208 * As a final tweak, if the new item goes on the last
8209 * spot on the page (and thus, onto the new page), bias
8210 * the split so the new page is emptier than the old page.
8211 * This yields better packing during sequential inserts.
8213 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8214 /* Find split point */
8216 if (newindx <= split_indx || newindx >= nkeys) {
8218 k = newindx >= nkeys ? nkeys : split_indx+2;
8223 for (; i!=k; i+=j) {
8228 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8229 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8231 if (F_ISSET(node->mn_flags, F_BIGDATA))
8232 psize += sizeof(pgno_t);
8234 psize += NODEDSZ(node);
8236 psize = EVEN(psize);
8238 if (psize > pmax || i == k-j) {
8239 split_indx = i + (j<0);
8244 if (split_indx == newindx) {
8245 sepkey.mv_size = newkey->mv_size;
8246 sepkey.mv_data = newkey->mv_data;
8248 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8249 sepkey.mv_size = node->mn_ksize;
8250 sepkey.mv_data = NODEKEY(node);
8255 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8257 /* Copy separator key to the parent.
8259 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8263 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8268 if (mn.mc_snum == mc->mc_snum) {
8269 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8270 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8271 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8272 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8277 /* Right page might now have changed parent.
8278 * Check if left page also changed parent.
8280 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8281 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8282 for (i=0; i<ptop; i++) {
8283 mc->mc_pg[i] = mn.mc_pg[i];
8284 mc->mc_ki[i] = mn.mc_ki[i];
8286 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8287 if (mn.mc_ki[ptop]) {
8288 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8290 /* find right page's left sibling */
8291 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8292 mdb_cursor_sibling(mc, 0);
8297 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8300 mc->mc_flags ^= C_SPLITTING;
8301 if (rc != MDB_SUCCESS) {
8304 if (nflags & MDB_APPEND) {
8305 mc->mc_pg[mc->mc_top] = rp;
8306 mc->mc_ki[mc->mc_top] = 0;
8307 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8310 for (i=0; i<mc->mc_top; i++)
8311 mc->mc_ki[i] = mn.mc_ki[i];
8312 } else if (!IS_LEAF2(mp)) {
8314 mc->mc_pg[mc->mc_top] = rp;
8319 rkey.mv_data = newkey->mv_data;
8320 rkey.mv_size = newkey->mv_size;
8326 /* Update index for the new key. */
8327 mc->mc_ki[mc->mc_top] = j;
8329 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8330 rkey.mv_data = NODEKEY(node);
8331 rkey.mv_size = node->mn_ksize;
8333 xdata.mv_data = NODEDATA(node);
8334 xdata.mv_size = NODEDSZ(node);
8337 pgno = NODEPGNO(node);
8338 flags = node->mn_flags;
8341 if (!IS_LEAF(mp) && j == 0) {
8342 /* First branch index doesn't need key data. */
8346 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8352 mc->mc_pg[mc->mc_top] = copy;
8357 } while (i != split_indx);
8359 nkeys = NUMKEYS(copy);
8360 for (i=0; i<nkeys; i++)
8361 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8362 mp->mp_lower = copy->mp_lower;
8363 mp->mp_upper = copy->mp_upper;
8364 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8365 env->me_psize - copy->mp_upper - PAGEBASE);
8367 /* reset back to original page */
8368 if (newindx < split_indx) {
8369 mc->mc_pg[mc->mc_top] = mp;
8370 if (nflags & MDB_RESERVE) {
8371 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8372 if (!(node->mn_flags & F_BIGDATA))
8373 newdata->mv_data = NODEDATA(node);
8376 mc->mc_pg[mc->mc_top] = rp;
8378 /* Make sure mc_ki is still valid.
8380 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8381 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8382 for (i=0; i<=ptop; i++) {
8383 mc->mc_pg[i] = mn.mc_pg[i];
8384 mc->mc_ki[i] = mn.mc_ki[i];
8391 /* Adjust other cursors pointing to mp */
8392 MDB_cursor *m2, *m3;
8393 MDB_dbi dbi = mc->mc_dbi;
8394 int fixup = NUMKEYS(mp);
8396 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8397 if (mc->mc_flags & C_SUB)
8398 m3 = &m2->mc_xcursor->mx_cursor;
8403 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8405 if (m3->mc_flags & C_SPLITTING)
8410 for (k=m3->mc_top; k>=0; k--) {
8411 m3->mc_ki[k+1] = m3->mc_ki[k];
8412 m3->mc_pg[k+1] = m3->mc_pg[k];
8414 if (m3->mc_ki[0] >= split_indx) {
8419 m3->mc_pg[0] = mc->mc_pg[0];
8423 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8424 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8425 m3->mc_ki[mc->mc_top]++;
8426 if (m3->mc_ki[mc->mc_top] >= fixup) {
8427 m3->mc_pg[mc->mc_top] = rp;
8428 m3->mc_ki[mc->mc_top] -= fixup;
8429 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8431 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8432 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8437 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8440 if (copy) /* tmp page */
8441 mdb_page_free(env, copy);
8443 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8448 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8449 MDB_val *key, MDB_val *data, unsigned int flags)
8454 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8457 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8460 mdb_cursor_init(&mc, txn, dbi, &mx);
8461 return mdb_cursor_put(&mc, key, data, flags);
8465 #define MDB_WBUF (1024*1024)
8468 /** State needed for a compacting copy. */
8469 typedef struct mdb_copy {
8470 pthread_mutex_t mc_mutex;
8471 pthread_cond_t mc_cond;
8478 pgno_t mc_next_pgno;
8481 volatile int mc_new;
8486 /** Dedicated writer thread for compacting copy. */
8487 static THREAD_RET ESECT
8488 mdb_env_copythr(void *arg)
8492 int toggle = 0, wsize, rc;
8495 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8498 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8501 pthread_mutex_lock(&my->mc_mutex);
8503 pthread_cond_signal(&my->mc_cond);
8506 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8507 if (my->mc_new < 0) {
8512 wsize = my->mc_wlen[toggle];
8513 ptr = my->mc_wbuf[toggle];
8516 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8520 } else if (len > 0) {
8534 /* If there's an overflow page tail, write it too */
8535 if (my->mc_olen[toggle]) {
8536 wsize = my->mc_olen[toggle];
8537 ptr = my->mc_over[toggle];
8538 my->mc_olen[toggle] = 0;
8541 my->mc_wlen[toggle] = 0;
8543 pthread_cond_signal(&my->mc_cond);
8545 pthread_cond_signal(&my->mc_cond);
8546 pthread_mutex_unlock(&my->mc_mutex);
8547 return (THREAD_RET)0;
8551 /** Tell the writer thread there's a buffer ready to write */
8553 mdb_env_cthr_toggle(mdb_copy *my, int st)
8555 int toggle = my->mc_toggle ^ 1;
8556 pthread_mutex_lock(&my->mc_mutex);
8557 if (my->mc_status) {
8558 pthread_mutex_unlock(&my->mc_mutex);
8559 return my->mc_status;
8561 while (my->mc_new == 1)
8562 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8564 my->mc_toggle = toggle;
8565 pthread_cond_signal(&my->mc_cond);
8566 pthread_mutex_unlock(&my->mc_mutex);
8570 /** Depth-first tree traversal for compacting copy. */
8572 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8575 MDB_txn *txn = my->mc_txn;
8577 MDB_page *mo, *mp, *leaf;
8582 /* Empty DB, nothing to do */
8583 if (*pg == P_INVALID)
8590 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8593 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8597 /* Make cursor pages writable */
8598 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8602 for (i=0; i<mc.mc_top; i++) {
8603 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8604 mc.mc_pg[i] = (MDB_page *)ptr;
8605 ptr += my->mc_env->me_psize;
8608 /* This is writable space for a leaf page. Usually not needed. */
8609 leaf = (MDB_page *)ptr;
8611 toggle = my->mc_toggle;
8612 while (mc.mc_snum > 0) {
8614 mp = mc.mc_pg[mc.mc_top];
8618 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8619 for (i=0; i<n; i++) {
8620 ni = NODEPTR(mp, i);
8621 if (ni->mn_flags & F_BIGDATA) {
8625 /* Need writable leaf */
8627 mc.mc_pg[mc.mc_top] = leaf;
8628 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8630 ni = NODEPTR(mp, i);
8633 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8634 rc = mdb_page_get(txn, pg, &omp, NULL);
8637 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8638 rc = mdb_env_cthr_toggle(my, 1);
8641 toggle = my->mc_toggle;
8643 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8644 memcpy(mo, omp, my->mc_env->me_psize);
8645 mo->mp_pgno = my->mc_next_pgno;
8646 my->mc_next_pgno += omp->mp_pages;
8647 my->mc_wlen[toggle] += my->mc_env->me_psize;
8648 if (omp->mp_pages > 1) {
8649 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8650 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8651 rc = mdb_env_cthr_toggle(my, 1);
8654 toggle = my->mc_toggle;
8656 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8657 } else if (ni->mn_flags & F_SUBDATA) {
8660 /* Need writable leaf */
8662 mc.mc_pg[mc.mc_top] = leaf;
8663 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8665 ni = NODEPTR(mp, i);
8668 memcpy(&db, NODEDATA(ni), sizeof(db));
8669 my->mc_toggle = toggle;
8670 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8673 toggle = my->mc_toggle;
8674 memcpy(NODEDATA(ni), &db, sizeof(db));
8679 mc.mc_ki[mc.mc_top]++;
8680 if (mc.mc_ki[mc.mc_top] < n) {
8683 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8685 rc = mdb_page_get(txn, pg, &mp, NULL);
8690 mc.mc_ki[mc.mc_top] = 0;
8691 if (IS_BRANCH(mp)) {
8692 /* Whenever we advance to a sibling branch page,
8693 * we must proceed all the way down to its first leaf.
8695 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8698 mc.mc_pg[mc.mc_top] = mp;
8702 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8703 rc = mdb_env_cthr_toggle(my, 1);
8706 toggle = my->mc_toggle;
8708 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8709 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8710 mo->mp_pgno = my->mc_next_pgno++;
8711 my->mc_wlen[toggle] += my->mc_env->me_psize;
8713 /* Update parent if there is one */
8714 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8715 SETPGNO(ni, mo->mp_pgno);
8716 mdb_cursor_pop(&mc);
8718 /* Otherwise we're done */
8728 /** Copy environment with compaction. */
8730 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8735 MDB_txn *txn = NULL;
8740 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8741 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8742 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8743 if (my.mc_wbuf[0] == NULL)
8746 pthread_mutex_init(&my.mc_mutex, NULL);
8747 pthread_cond_init(&my.mc_cond, NULL);
8748 #ifdef HAVE_MEMALIGN
8749 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8750 if (my.mc_wbuf[0] == NULL)
8753 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8758 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8759 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8764 my.mc_next_pgno = 2;
8770 THREAD_CREATE(thr, mdb_env_copythr, &my);
8772 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8776 mp = (MDB_page *)my.mc_wbuf[0];
8777 memset(mp, 0, 2*env->me_psize);
8779 mp->mp_flags = P_META;
8780 mm = (MDB_meta *)METADATA(mp);
8781 mdb_env_init_meta0(env, mm);
8782 mm->mm_address = env->me_metas[0]->mm_address;
8784 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8786 mp->mp_flags = P_META;
8787 *(MDB_meta *)METADATA(mp) = *mm;
8788 mm = (MDB_meta *)METADATA(mp);
8790 /* Count the number of free pages, subtract from lastpg to find
8791 * number of active pages
8794 MDB_ID freecount = 0;
8797 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8798 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8799 freecount += *(MDB_ID *)data.mv_data;
8800 freecount += txn->mt_dbs[0].md_branch_pages +
8801 txn->mt_dbs[0].md_leaf_pages +
8802 txn->mt_dbs[0].md_overflow_pages;
8804 /* Set metapage 1 */
8805 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8806 mm->mm_dbs[1] = txn->mt_dbs[1];
8807 if (mm->mm_last_pg > 1) {
8808 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8811 mm->mm_dbs[1].md_root = P_INVALID;
8814 my.mc_wlen[0] = env->me_psize * 2;
8816 pthread_mutex_lock(&my.mc_mutex);
8818 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8819 pthread_mutex_unlock(&my.mc_mutex);
8820 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8821 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8822 rc = mdb_env_cthr_toggle(&my, 1);
8823 mdb_env_cthr_toggle(&my, -1);
8824 pthread_mutex_lock(&my.mc_mutex);
8826 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8827 pthread_mutex_unlock(&my.mc_mutex);
8832 CloseHandle(my.mc_cond);
8833 CloseHandle(my.mc_mutex);
8834 _aligned_free(my.mc_wbuf[0]);
8836 pthread_cond_destroy(&my.mc_cond);
8837 pthread_mutex_destroy(&my.mc_mutex);
8838 free(my.mc_wbuf[0]);
8843 /** Copy environment as-is. */
8845 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8847 MDB_txn *txn = NULL;
8848 mdb_mutex_t *wmutex = NULL;
8854 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8858 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8861 /* Do the lock/unlock of the reader mutex before starting the
8862 * write txn. Otherwise other read txns could block writers.
8864 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8869 /* We must start the actual read txn after blocking writers */
8870 mdb_txn_reset0(txn, "reset-stage1");
8872 /* Temporarily block writers until we snapshot the meta pages */
8873 wmutex = MDB_MUTEX(env, w);
8874 if (LOCK_MUTEX(rc, env, wmutex))
8877 rc = mdb_txn_renew0(txn);
8879 UNLOCK_MUTEX(wmutex);
8884 wsize = env->me_psize * 2;
8888 DO_WRITE(rc, fd, ptr, w2, len);
8892 } else if (len > 0) {
8898 /* Non-blocking or async handles are not supported */
8904 UNLOCK_MUTEX(wmutex);
8909 w2 = txn->mt_next_pgno * env->me_psize;
8912 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8919 if (wsize > MAX_WRITE)
8923 DO_WRITE(rc, fd, ptr, w2, len);
8927 } else if (len > 0) {
8944 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8946 if (flags & MDB_CP_COMPACT)
8947 return mdb_env_copyfd1(env, fd);
8949 return mdb_env_copyfd0(env, fd);
8953 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8955 return mdb_env_copyfd2(env, fd, 0);
8959 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8963 HANDLE newfd = INVALID_HANDLE_VALUE;
8965 if (env->me_flags & MDB_NOSUBDIR) {
8966 lpath = (char *)path;
8969 len += sizeof(DATANAME);
8970 lpath = malloc(len);
8973 sprintf(lpath, "%s" DATANAME, path);
8976 /* The destination path must exist, but the destination file must not.
8977 * We don't want the OS to cache the writes, since the source data is
8978 * already in the OS cache.
8981 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8982 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8984 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8986 if (newfd == INVALID_HANDLE_VALUE) {
8991 if (env->me_psize >= env->me_os_psize) {
8993 /* Set O_DIRECT if the file system supports it */
8994 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8995 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8997 #ifdef F_NOCACHE /* __APPLE__ */
8998 rc = fcntl(newfd, F_NOCACHE, 1);
9006 rc = mdb_env_copyfd2(env, newfd, flags);
9009 if (!(env->me_flags & MDB_NOSUBDIR))
9011 if (newfd != INVALID_HANDLE_VALUE)
9012 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9019 mdb_env_copy(MDB_env *env, const char *path)
9021 return mdb_env_copy2(env, path, 0);
9025 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9027 if ((flag & CHANGEABLE) != flag)
9030 env->me_flags |= flag;
9032 env->me_flags &= ~flag;
9037 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9042 *arg = env->me_flags;
9047 mdb_env_set_userctx(MDB_env *env, void *ctx)
9051 env->me_userctx = ctx;
9056 mdb_env_get_userctx(MDB_env *env)
9058 return env ? env->me_userctx : NULL;
9062 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9067 env->me_assert_func = func;
9073 mdb_env_get_path(MDB_env *env, const char **arg)
9078 *arg = env->me_path;
9083 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9092 /** Common code for #mdb_stat() and #mdb_env_stat().
9093 * @param[in] env the environment to operate in.
9094 * @param[in] db the #MDB_db record containing the stats to return.
9095 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9096 * @return 0, this function always succeeds.
9099 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9101 arg->ms_psize = env->me_psize;
9102 arg->ms_depth = db->md_depth;
9103 arg->ms_branch_pages = db->md_branch_pages;
9104 arg->ms_leaf_pages = db->md_leaf_pages;
9105 arg->ms_overflow_pages = db->md_overflow_pages;
9106 arg->ms_entries = db->md_entries;
9112 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9116 if (env == NULL || arg == NULL)
9119 toggle = mdb_env_pick_meta(env);
9121 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9125 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9129 if (env == NULL || arg == NULL)
9132 toggle = mdb_env_pick_meta(env);
9133 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9134 arg->me_mapsize = env->me_mapsize;
9135 arg->me_maxreaders = env->me_maxreaders;
9136 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9138 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9139 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9143 /** Set the default comparison functions for a database.
9144 * Called immediately after a database is opened to set the defaults.
9145 * The user can then override them with #mdb_set_compare() or
9146 * #mdb_set_dupsort().
9147 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9148 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9151 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9153 uint16_t f = txn->mt_dbs[dbi].md_flags;
9155 txn->mt_dbxs[dbi].md_cmp =
9156 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9157 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9159 txn->mt_dbxs[dbi].md_dcmp =
9160 !(f & MDB_DUPSORT) ? 0 :
9161 ((f & MDB_INTEGERDUP)
9162 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9163 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9166 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9172 int rc, dbflag, exact;
9173 unsigned int unused = 0, seq;
9176 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9177 mdb_default_cmp(txn, FREE_DBI);
9180 if ((flags & VALID_FLAGS) != flags)
9182 if (txn->mt_flags & MDB_TXN_ERROR)
9188 if (flags & PERSISTENT_FLAGS) {
9189 uint16_t f2 = flags & PERSISTENT_FLAGS;
9190 /* make sure flag changes get committed */
9191 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9192 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9193 txn->mt_flags |= MDB_TXN_DIRTY;
9196 mdb_default_cmp(txn, MAIN_DBI);
9200 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9201 mdb_default_cmp(txn, MAIN_DBI);
9204 /* Is the DB already open? */
9206 for (i=2; i<txn->mt_numdbs; i++) {
9207 if (!txn->mt_dbxs[i].md_name.mv_size) {
9208 /* Remember this free slot */
9209 if (!unused) unused = i;
9212 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9213 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9219 /* If no free slot and max hit, fail */
9220 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9221 return MDB_DBS_FULL;
9223 /* Cannot mix named databases with some mainDB flags */
9224 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9225 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9227 /* Find the DB info */
9228 dbflag = DB_NEW|DB_VALID;
9231 key.mv_data = (void *)name;
9232 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9233 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9234 if (rc == MDB_SUCCESS) {
9235 /* make sure this is actually a DB */
9236 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9237 if (!(node->mn_flags & F_SUBDATA))
9238 return MDB_INCOMPATIBLE;
9239 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9240 /* Create if requested */
9241 data.mv_size = sizeof(MDB_db);
9242 data.mv_data = &dummy;
9243 memset(&dummy, 0, sizeof(dummy));
9244 dummy.md_root = P_INVALID;
9245 dummy.md_flags = flags & PERSISTENT_FLAGS;
9246 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9250 /* OK, got info, add to table */
9251 if (rc == MDB_SUCCESS) {
9252 unsigned int slot = unused ? unused : txn->mt_numdbs;
9253 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9254 txn->mt_dbxs[slot].md_name.mv_size = len;
9255 txn->mt_dbxs[slot].md_rel = NULL;
9256 txn->mt_dbflags[slot] = dbflag;
9257 /* txn-> and env-> are the same in read txns, use
9258 * tmp variable to avoid undefined assignment
9260 seq = ++txn->mt_env->me_dbiseqs[slot];
9261 txn->mt_dbiseqs[slot] = seq;
9263 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9265 mdb_default_cmp(txn, slot);
9274 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9276 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9279 if (txn->mt_flags & MDB_TXN_ERROR)
9282 if (txn->mt_dbflags[dbi] & DB_STALE) {
9285 /* Stale, must read the DB's root. cursor_init does it for us. */
9286 mdb_cursor_init(&mc, txn, dbi, &mx);
9288 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9291 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9294 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9296 ptr = env->me_dbxs[dbi].md_name.mv_data;
9297 /* If there was no name, this was already closed */
9299 env->me_dbxs[dbi].md_name.mv_data = NULL;
9300 env->me_dbxs[dbi].md_name.mv_size = 0;
9301 env->me_dbflags[dbi] = 0;
9302 env->me_dbiseqs[dbi]++;
9307 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9309 /* We could return the flags for the FREE_DBI too but what's the point? */
9310 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9312 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9316 /** Add all the DB's pages to the free list.
9317 * @param[in] mc Cursor on the DB to free.
9318 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9319 * @return 0 on success, non-zero on failure.
9322 mdb_drop0(MDB_cursor *mc, int subs)
9326 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9327 if (rc == MDB_SUCCESS) {
9328 MDB_txn *txn = mc->mc_txn;
9333 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9334 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9337 mdb_cursor_copy(mc, &mx);
9338 while (mc->mc_snum > 0) {
9339 MDB_page *mp = mc->mc_pg[mc->mc_top];
9340 unsigned n = NUMKEYS(mp);
9342 for (i=0; i<n; i++) {
9343 ni = NODEPTR(mp, i);
9344 if (ni->mn_flags & F_BIGDATA) {
9347 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9348 rc = mdb_page_get(txn, pg, &omp, NULL);
9351 mdb_cassert(mc, IS_OVERFLOW(omp));
9352 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9356 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9357 mdb_xcursor_init1(mc, ni);
9358 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9364 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9366 for (i=0; i<n; i++) {
9368 ni = NODEPTR(mp, i);
9371 mdb_midl_xappend(txn->mt_free_pgs, pg);
9376 mc->mc_ki[mc->mc_top] = i;
9377 rc = mdb_cursor_sibling(mc, 1);
9379 if (rc != MDB_NOTFOUND)
9381 /* no more siblings, go back to beginning
9382 * of previous level.
9386 for (i=1; i<mc->mc_snum; i++) {
9388 mc->mc_pg[i] = mx.mc_pg[i];
9393 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9396 txn->mt_flags |= MDB_TXN_ERROR;
9397 } else if (rc == MDB_NOTFOUND) {
9403 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9405 MDB_cursor *mc, *m2;
9408 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9411 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9414 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9417 rc = mdb_cursor_open(txn, dbi, &mc);
9421 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9422 /* Invalidate the dropped DB's cursors */
9423 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9424 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9428 /* Can't delete the main DB */
9429 if (del && dbi > MAIN_DBI) {
9430 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9432 txn->mt_dbflags[dbi] = DB_STALE;
9433 mdb_dbi_close(txn->mt_env, dbi);
9435 txn->mt_flags |= MDB_TXN_ERROR;
9438 /* reset the DB record, mark it dirty */
9439 txn->mt_dbflags[dbi] |= DB_DIRTY;
9440 txn->mt_dbs[dbi].md_depth = 0;
9441 txn->mt_dbs[dbi].md_branch_pages = 0;
9442 txn->mt_dbs[dbi].md_leaf_pages = 0;
9443 txn->mt_dbs[dbi].md_overflow_pages = 0;
9444 txn->mt_dbs[dbi].md_entries = 0;
9445 txn->mt_dbs[dbi].md_root = P_INVALID;
9447 txn->mt_flags |= MDB_TXN_DIRTY;
9450 mdb_cursor_close(mc);
9454 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9456 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9459 txn->mt_dbxs[dbi].md_cmp = cmp;
9463 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9465 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9468 txn->mt_dbxs[dbi].md_dcmp = cmp;
9472 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9474 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9477 txn->mt_dbxs[dbi].md_rel = rel;
9481 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9483 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9486 txn->mt_dbxs[dbi].md_relctx = ctx;
9491 mdb_env_get_maxkeysize(MDB_env *env)
9493 return ENV_MAXKEY(env);
9497 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9499 unsigned int i, rdrs;
9502 int rc = 0, first = 1;
9506 if (!env->me_txns) {
9507 return func("(no reader locks)\n", ctx);
9509 rdrs = env->me_txns->mti_numreaders;
9510 mr = env->me_txns->mti_readers;
9511 for (i=0; i<rdrs; i++) {
9513 txnid_t txnid = mr[i].mr_txnid;
9514 sprintf(buf, txnid == (txnid_t)-1 ?
9515 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9516 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9519 rc = func(" pid thread txnid\n", ctx);
9523 rc = func(buf, ctx);
9529 rc = func("(no active readers)\n", ctx);
9534 /** Insert pid into list if not already present.
9535 * return -1 if already present.
9538 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9540 /* binary search of pid in list */
9542 unsigned cursor = 1;
9544 unsigned n = ids[0];
9547 unsigned pivot = n >> 1;
9548 cursor = base + pivot + 1;
9549 val = pid - ids[cursor];
9554 } else if ( val > 0 ) {
9559 /* found, so it's a duplicate */
9568 for (n = ids[0]; n > cursor; n--)
9575 mdb_reader_check(MDB_env *env, int *dead)
9581 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9584 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9585 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9587 mdb_mutex_t *rmutex = rlocked ? NULL : MDB_MUTEX(env, r);
9588 unsigned int i, j, rdrs;
9590 MDB_PID_T *pids, pid;
9591 int rc = MDB_SUCCESS, count = 0;
9593 rdrs = env->me_txns->mti_numreaders;
9594 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9598 mr = env->me_txns->mti_readers;
9599 for (i=0; i<rdrs; i++) {
9601 if (pid && pid != env->me_pid) {
9602 if (mdb_pid_insert(pids, pid) == 0) {
9603 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9604 /* Stale reader found */
9607 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9608 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9610 rdrs = 0; /* the above checked all readers */
9612 /* Recheck, a new process may have reused pid */
9613 if (mdb_reader_pid(env, Pidcheck, pid))
9618 if (mr[j].mr_pid == pid) {
9619 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9620 (unsigned) pid, mr[j].mr_txnid));
9625 UNLOCK_MUTEX(rmutex);
9636 #ifdef MDB_ROBUST_SUPPORTED
9637 /** Handle #LOCK_MUTEX0() failure.
9638 * Try to repair the lock file if the mutex owner died.
9639 * @param[in] env the environment handle
9640 * @param[in] mutex LOCK_MUTEX0() mutex
9641 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9642 * @return 0 on success with the mutex locked, or an error code on failure.
9644 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc)
9646 int toggle, rlocked, rc2;
9648 if (rc == MDB_OWNERDEAD) {
9649 /* We own the mutex. Clean up after dead previous owner. */
9651 rlocked = (mutex == MDB_MUTEX(env, r));
9653 /* Keep mti_txnid updated, otherwise next writer can
9654 * overwrite data which latest meta page refers to.
9656 toggle = mdb_env_pick_meta(env);
9657 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9658 /* env is hosed if the dead thread was ours */
9660 env->me_flags |= MDB_FATAL_ERROR;
9665 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9666 (rc ? "this process' env is hosed" : "recovering")));
9667 rc2 = mdb_reader_check0(env, rlocked, NULL);
9669 rc2 = mdb_mutex_consistent(mutex);
9670 if (rc || (rc = rc2)) {
9671 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9672 UNLOCK_MUTEX(mutex);
9678 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9683 #endif /* MDB_ROBUST_SUPPORTED */
projects / platform / upstream / lmdb.git / blob