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 */
1088 unsigned int mt_flags; /**< @ref mdb_txn */
1089 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1090 * Includes ancestor txns' dirty pages not hidden by other txns'
1091 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1092 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1094 unsigned int mt_dirty_room;
1097 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1098 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1099 * raise this on a 64 bit machine.
1101 #define CURSOR_STACK 32
1105 /** Cursors are used for all DB operations.
1106 * A cursor holds a path of (page pointer, key index) from the DB
1107 * root to a position in the DB, plus other state. #MDB_DUPSORT
1108 * cursors include an xcursor to the current data item. Write txns
1109 * track their cursors and keep them up to date when data moves.
1110 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1111 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1114 /** Next cursor on this DB in this txn */
1115 MDB_cursor *mc_next;
1116 /** Backup of the original cursor if this cursor is a shadow */
1117 MDB_cursor *mc_backup;
1118 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1119 struct MDB_xcursor *mc_xcursor;
1120 /** The transaction that owns this cursor */
1122 /** The database handle this cursor operates on */
1124 /** The database record for this cursor */
1126 /** The database auxiliary record for this cursor */
1128 /** The @ref mt_dbflag for this database */
1129 unsigned char *mc_dbflag;
1130 unsigned short mc_snum; /**< number of pushed pages */
1131 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1132 /** @defgroup mdb_cursor Cursor Flags
1134 * Cursor state flags.
1137 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1138 #define C_EOF 0x02 /**< No more data */
1139 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1140 #define C_DEL 0x08 /**< last op was a cursor_del */
1141 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1142 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1144 unsigned int mc_flags; /**< @ref mdb_cursor */
1145 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1146 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1149 /** Context for sorted-dup records.
1150 * We could have gone to a fully recursive design, with arbitrarily
1151 * deep nesting of sub-databases. But for now we only handle these
1152 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1154 typedef struct MDB_xcursor {
1155 /** A sub-cursor for traversing the Dup DB */
1156 MDB_cursor mx_cursor;
1157 /** The database record for this Dup DB */
1159 /** The auxiliary DB record for this Dup DB */
1161 /** The @ref mt_dbflag for this Dup DB */
1162 unsigned char mx_dbflag;
1165 /** State of FreeDB old pages, stored in the MDB_env */
1166 typedef struct MDB_pgstate {
1167 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1168 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1171 /** The database environment. */
1173 HANDLE me_fd; /**< The main data file */
1174 HANDLE me_lfd; /**< The lock file */
1175 HANDLE me_mfd; /**< just for writing the meta pages */
1176 /** Failed to update the meta page. Probably an I/O error. */
1177 #define MDB_FATAL_ERROR 0x80000000U
1178 /** Some fields are initialized. */
1179 #define MDB_ENV_ACTIVE 0x20000000U
1180 /** me_txkey is set */
1181 #define MDB_ENV_TXKEY 0x10000000U
1182 /** fdatasync is unreliable */
1183 #define MDB_FSYNCONLY 0x08000000U
1184 uint32_t me_flags; /**< @ref mdb_env */
1185 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1186 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1187 unsigned int me_maxreaders; /**< size of the reader table */
1188 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1189 volatile int me_close_readers;
1190 MDB_dbi me_numdbs; /**< number of DBs opened */
1191 MDB_dbi me_maxdbs; /**< size of the DB table */
1192 MDB_PID_T me_pid; /**< process ID of this env */
1193 char *me_path; /**< path to the DB files */
1194 char *me_map; /**< the memory map of the data file */
1195 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1196 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1197 void *me_pbuf; /**< scratch area for DUPSORT put() */
1198 MDB_txn *me_txn; /**< current write transaction */
1199 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1200 size_t me_mapsize; /**< size of the data memory map */
1201 off_t me_size; /**< current file size */
1202 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1203 MDB_dbx *me_dbxs; /**< array of static DB info */
1204 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1205 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1206 pthread_key_t me_txkey; /**< thread-key for readers */
1207 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1208 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1209 # define me_pglast me_pgstate.mf_pglast
1210 # define me_pghead me_pgstate.mf_pghead
1211 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1212 /** IDL of pages that became unused in a write txn */
1213 MDB_IDL me_free_pgs;
1214 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1215 MDB_ID2L me_dirty_list;
1216 /** Max number of freelist items that can fit in a single overflow page */
1218 /** Max size of a node on a page */
1219 unsigned int me_nodemax;
1220 #if !(MDB_MAXKEYSIZE)
1221 unsigned int me_maxkey; /**< max size of a key */
1223 int me_live_reader; /**< have liveness lock in reader table */
1225 int me_pidquery; /**< Used in OpenProcess */
1227 #if defined(_WIN32) || defined(MDB_USE_SYSV_SEM)
1228 /* Windows mutexes/SysV semaphores do not reside in shared mem */
1229 mdb_mutex_t me_rmutex;
1230 mdb_mutex_t me_wmutex;
1232 void *me_userctx; /**< User-settable context */
1233 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1236 /** Nested transaction */
1237 typedef struct MDB_ntxn {
1238 MDB_txn mnt_txn; /**< the transaction */
1239 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1242 /** max number of pages to commit in one writev() call */
1243 #define MDB_COMMIT_PAGES 64
1244 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1245 #undef MDB_COMMIT_PAGES
1246 #define MDB_COMMIT_PAGES IOV_MAX
1249 /** max bytes to write in one call */
1250 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1252 /** Check \b txn and \b dbi arguments to a function */
1253 #define TXN_DBI_EXIST(txn, dbi) \
1254 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1256 /** Check for misused \b dbi handles */
1257 #define TXN_DBI_CHANGED(txn, dbi) \
1258 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1260 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1261 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1262 static int mdb_page_touch(MDB_cursor *mc);
1264 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1265 static int mdb_page_search_root(MDB_cursor *mc,
1266 MDB_val *key, int modify);
1267 #define MDB_PS_MODIFY 1
1268 #define MDB_PS_ROOTONLY 2
1269 #define MDB_PS_FIRST 4
1270 #define MDB_PS_LAST 8
1271 static int mdb_page_search(MDB_cursor *mc,
1272 MDB_val *key, int flags);
1273 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1275 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1276 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1277 pgno_t newpgno, unsigned int nflags);
1279 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1280 static int mdb_env_pick_meta(const MDB_env *env);
1281 static int mdb_env_write_meta(MDB_txn *txn);
1282 #if !(defined(_WIN32) || defined(MDB_USE_SYSV_SEM)) /* Drop unused excl arg */
1283 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1285 static void mdb_env_close0(MDB_env *env, int excl);
1287 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1288 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1289 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1290 static void mdb_node_del(MDB_cursor *mc, int ksize);
1291 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1292 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1293 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1294 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1295 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1297 static int mdb_rebalance(MDB_cursor *mc);
1298 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1300 static void mdb_cursor_pop(MDB_cursor *mc);
1301 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1303 static int mdb_cursor_del0(MDB_cursor *mc);
1304 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1305 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1306 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1307 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1308 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1310 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1311 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1313 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1314 static void mdb_xcursor_init0(MDB_cursor *mc);
1315 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1317 static int mdb_drop0(MDB_cursor *mc, int subs);
1318 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1319 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1322 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1326 static SECURITY_DESCRIPTOR mdb_null_sd;
1327 static SECURITY_ATTRIBUTES mdb_all_sa;
1328 static int mdb_sec_inited;
1331 /** Return the library version info. */
1333 mdb_version(int *major, int *minor, int *patch)
1335 if (major) *major = MDB_VERSION_MAJOR;
1336 if (minor) *minor = MDB_VERSION_MINOR;
1337 if (patch) *patch = MDB_VERSION_PATCH;
1338 return MDB_VERSION_STRING;
1341 /** Table of descriptions for LMDB @ref errors */
1342 static char *const mdb_errstr[] = {
1343 "MDB_KEYEXIST: Key/data pair already exists",
1344 "MDB_NOTFOUND: No matching key/data pair found",
1345 "MDB_PAGE_NOTFOUND: Requested page not found",
1346 "MDB_CORRUPTED: Located page was wrong type",
1347 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1348 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1349 "MDB_INVALID: File is not an LMDB file",
1350 "MDB_MAP_FULL: Environment mapsize limit reached",
1351 "MDB_DBS_FULL: Environment maxdbs limit reached",
1352 "MDB_READERS_FULL: Environment maxreaders limit reached",
1353 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1354 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1355 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1356 "MDB_PAGE_FULL: Internal error - page has no more space",
1357 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1358 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1359 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1360 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1361 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1362 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1366 mdb_strerror(int err)
1369 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1370 * This works as long as no function between the call to mdb_strerror
1371 * and the actual use of the message uses more than 4K of stack.
1374 char buf[1024], *ptr = buf;
1378 return ("Successful return: 0");
1380 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1381 i = err - MDB_KEYEXIST;
1382 return mdb_errstr[i];
1386 /* These are the C-runtime error codes we use. The comment indicates
1387 * their numeric value, and the Win32 error they would correspond to
1388 * if the error actually came from a Win32 API. A major mess, we should
1389 * have used LMDB-specific error codes for everything.
1392 case ENOENT: /* 2, FILE_NOT_FOUND */
1393 case EIO: /* 5, ACCESS_DENIED */
1394 case ENOMEM: /* 12, INVALID_ACCESS */
1395 case EACCES: /* 13, INVALID_DATA */
1396 case EBUSY: /* 16, CURRENT_DIRECTORY */
1397 case EINVAL: /* 22, BAD_COMMAND */
1398 case ENOSPC: /* 28, OUT_OF_PAPER */
1399 return strerror(err);
1404 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1405 FORMAT_MESSAGE_IGNORE_INSERTS,
1406 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1409 return strerror(err);
1413 /** assert(3) variant in cursor context */
1414 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1415 /** assert(3) variant in transaction context */
1416 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1417 /** assert(3) variant in environment context */
1418 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1421 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1422 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1425 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1426 const char *func, const char *file, int line)
1429 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1430 file, line, expr_txt, func);
1431 if (env->me_assert_func)
1432 env->me_assert_func(env, buf);
1433 fprintf(stderr, "%s\n", buf);
1437 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1441 /** Return the page number of \b mp which may be sub-page, for debug output */
1443 mdb_dbg_pgno(MDB_page *mp)
1446 COPY_PGNO(ret, mp->mp_pgno);
1450 /** Display a key in hexadecimal and return the address of the result.
1451 * @param[in] key the key to display
1452 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1453 * @return The key in hexadecimal form.
1456 mdb_dkey(MDB_val *key, char *buf)
1459 unsigned char *c = key->mv_data;
1465 if (key->mv_size > DKBUF_MAXKEYSIZE)
1466 return "MDB_MAXKEYSIZE";
1467 /* may want to make this a dynamic check: if the key is mostly
1468 * printable characters, print it as-is instead of converting to hex.
1472 for (i=0; i<key->mv_size; i++)
1473 ptr += sprintf(ptr, "%02x", *c++);
1475 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1481 mdb_leafnode_type(MDB_node *n)
1483 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1484 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1485 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1488 /** Display all the keys in the page. */
1490 mdb_page_list(MDB_page *mp)
1492 pgno_t pgno = mdb_dbg_pgno(mp);
1493 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1495 unsigned int i, nkeys, nsize, total = 0;
1499 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1500 case P_BRANCH: type = "Branch page"; break;
1501 case P_LEAF: type = "Leaf page"; break;
1502 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1503 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1504 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1506 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1507 pgno, mp->mp_pages, state);
1510 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1511 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1514 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1518 nkeys = NUMKEYS(mp);
1519 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1521 for (i=0; i<nkeys; i++) {
1522 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1523 key.mv_size = nsize = mp->mp_pad;
1524 key.mv_data = LEAF2KEY(mp, i, nsize);
1526 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1529 node = NODEPTR(mp, i);
1530 key.mv_size = node->mn_ksize;
1531 key.mv_data = node->mn_data;
1532 nsize = NODESIZE + key.mv_size;
1533 if (IS_BRANCH(mp)) {
1534 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1538 if (F_ISSET(node->mn_flags, F_BIGDATA))
1539 nsize += sizeof(pgno_t);
1541 nsize += NODEDSZ(node);
1543 nsize += sizeof(indx_t);
1544 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1545 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1547 total = EVEN(total);
1549 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1550 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1554 mdb_cursor_chk(MDB_cursor *mc)
1560 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1561 for (i=0; i<mc->mc_top; i++) {
1563 node = NODEPTR(mp, mc->mc_ki[i]);
1564 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1567 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1573 /** Count all the pages in each DB and in the freelist
1574 * and make sure it matches the actual number of pages
1576 * All named DBs must be open for a correct count.
1578 static void mdb_audit(MDB_txn *txn)
1582 MDB_ID freecount, count;
1587 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1588 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1589 freecount += *(MDB_ID *)data.mv_data;
1590 mdb_tassert(txn, rc == MDB_NOTFOUND);
1593 for (i = 0; i<txn->mt_numdbs; i++) {
1595 if (!(txn->mt_dbflags[i] & DB_VALID))
1597 mdb_cursor_init(&mc, txn, i, &mx);
1598 if (txn->mt_dbs[i].md_root == P_INVALID)
1600 count += txn->mt_dbs[i].md_branch_pages +
1601 txn->mt_dbs[i].md_leaf_pages +
1602 txn->mt_dbs[i].md_overflow_pages;
1603 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1604 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1605 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1608 mp = mc.mc_pg[mc.mc_top];
1609 for (j=0; j<NUMKEYS(mp); j++) {
1610 MDB_node *leaf = NODEPTR(mp, j);
1611 if (leaf->mn_flags & F_SUBDATA) {
1613 memcpy(&db, NODEDATA(leaf), sizeof(db));
1614 count += db.md_branch_pages + db.md_leaf_pages +
1615 db.md_overflow_pages;
1619 mdb_tassert(txn, rc == MDB_NOTFOUND);
1622 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1623 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1624 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1630 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1632 return txn->mt_dbxs[dbi].md_cmp(a, b);
1636 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1638 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1641 /** Allocate memory for a page.
1642 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1645 mdb_page_malloc(MDB_txn *txn, unsigned num)
1647 MDB_env *env = txn->mt_env;
1648 MDB_page *ret = env->me_dpages;
1649 size_t psize = env->me_psize, sz = psize, off;
1650 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1651 * For a single page alloc, we init everything after the page header.
1652 * For multi-page, we init the final page; if the caller needed that
1653 * many pages they will be filling in at least up to the last page.
1657 VGMEMP_ALLOC(env, ret, sz);
1658 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1659 env->me_dpages = ret->mp_next;
1662 psize -= off = PAGEHDRSZ;
1667 if ((ret = malloc(sz)) != NULL) {
1668 VGMEMP_ALLOC(env, ret, sz);
1669 if (!(env->me_flags & MDB_NOMEMINIT)) {
1670 memset((char *)ret + off, 0, psize);
1674 txn->mt_flags |= MDB_TXN_ERROR;
1678 /** Free a single page.
1679 * Saves single pages to a list, for future reuse.
1680 * (This is not used for multi-page overflow pages.)
1683 mdb_page_free(MDB_env *env, MDB_page *mp)
1685 mp->mp_next = env->me_dpages;
1686 VGMEMP_FREE(env, mp);
1687 env->me_dpages = mp;
1690 /** Free a dirty page */
1692 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1694 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1695 mdb_page_free(env, dp);
1697 /* large pages just get freed directly */
1698 VGMEMP_FREE(env, dp);
1703 /** Return all dirty pages to dpage list */
1705 mdb_dlist_free(MDB_txn *txn)
1707 MDB_env *env = txn->mt_env;
1708 MDB_ID2L dl = txn->mt_u.dirty_list;
1709 unsigned i, n = dl[0].mid;
1711 for (i = 1; i <= n; i++) {
1712 mdb_dpage_free(env, dl[i].mptr);
1717 /** Loosen or free a single page.
1718 * Saves single pages to a list for future reuse
1719 * in this same txn. It has been pulled from the freeDB
1720 * and already resides on the dirty list, but has been
1721 * deleted. Use these pages first before pulling again
1724 * If the page wasn't dirtied in this txn, just add it
1725 * to this txn's free list.
1728 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1731 pgno_t pgno = mp->mp_pgno;
1732 MDB_txn *txn = mc->mc_txn;
1734 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1735 if (txn->mt_parent) {
1736 MDB_ID2 *dl = txn->mt_u.dirty_list;
1737 /* If txn has a parent, make sure the page is in our
1741 unsigned x = mdb_mid2l_search(dl, pgno);
1742 if (x <= dl[0].mid && dl[x].mid == pgno) {
1743 if (mp != dl[x].mptr) { /* bad cursor? */
1744 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1745 txn->mt_flags |= MDB_TXN_ERROR;
1746 return MDB_CORRUPTED;
1753 /* no parent txn, so it's just ours */
1758 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1760 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1761 txn->mt_loose_pgs = mp;
1762 txn->mt_loose_count++;
1763 mp->mp_flags |= P_LOOSE;
1765 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1773 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1774 * @param[in] mc A cursor handle for the current operation.
1775 * @param[in] pflags Flags of the pages to update:
1776 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1777 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1778 * @return 0 on success, non-zero on failure.
1781 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1783 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1784 MDB_txn *txn = mc->mc_txn;
1790 int rc = MDB_SUCCESS, level;
1792 /* Mark pages seen by cursors */
1793 if (mc->mc_flags & C_UNTRACK)
1794 mc = NULL; /* will find mc in mt_cursors */
1795 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1796 for (; mc; mc=mc->mc_next) {
1797 if (!(mc->mc_flags & C_INITIALIZED))
1799 for (m3 = mc;; m3 = &mx->mx_cursor) {
1801 for (j=0; j<m3->mc_snum; j++) {
1803 if ((mp->mp_flags & Mask) == pflags)
1804 mp->mp_flags ^= P_KEEP;
1806 mx = m3->mc_xcursor;
1807 /* Proceed to mx if it is at a sub-database */
1808 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1810 if (! (mp && (mp->mp_flags & P_LEAF)))
1812 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1813 if (!(leaf->mn_flags & F_SUBDATA))
1822 /* Mark dirty root pages */
1823 for (i=0; i<txn->mt_numdbs; i++) {
1824 if (txn->mt_dbflags[i] & DB_DIRTY) {
1825 pgno_t pgno = txn->mt_dbs[i].md_root;
1826 if (pgno == P_INVALID)
1828 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1830 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1831 dp->mp_flags ^= P_KEEP;
1839 static int mdb_page_flush(MDB_txn *txn, int keep);
1841 /** Spill pages from the dirty list back to disk.
1842 * This is intended to prevent running into #MDB_TXN_FULL situations,
1843 * but note that they may still occur in a few cases:
1844 * 1) our estimate of the txn size could be too small. Currently this
1845 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1846 * 2) child txns may run out of space if their parents dirtied a
1847 * lot of pages and never spilled them. TODO: we probably should do
1848 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1849 * the parent's dirty_room is below a given threshold.
1851 * Otherwise, if not using nested txns, it is expected that apps will
1852 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1853 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1854 * If the txn never references them again, they can be left alone.
1855 * If the txn only reads them, they can be used without any fuss.
1856 * If the txn writes them again, they can be dirtied immediately without
1857 * going thru all of the work of #mdb_page_touch(). Such references are
1858 * handled by #mdb_page_unspill().
1860 * Also note, we never spill DB root pages, nor pages of active cursors,
1861 * because we'll need these back again soon anyway. And in nested txns,
1862 * we can't spill a page in a child txn if it was already spilled in a
1863 * parent txn. That would alter the parent txns' data even though
1864 * the child hasn't committed yet, and we'd have no way to undo it if
1865 * the child aborted.
1867 * @param[in] m0 cursor A cursor handle identifying the transaction and
1868 * database for which we are checking space.
1869 * @param[in] key For a put operation, the key being stored.
1870 * @param[in] data For a put operation, the data being stored.
1871 * @return 0 on success, non-zero on failure.
1874 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1876 MDB_txn *txn = m0->mc_txn;
1878 MDB_ID2L dl = txn->mt_u.dirty_list;
1879 unsigned int i, j, need;
1882 if (m0->mc_flags & C_SUB)
1885 /* Estimate how much space this op will take */
1886 i = m0->mc_db->md_depth;
1887 /* Named DBs also dirty the main DB */
1888 if (m0->mc_dbi > MAIN_DBI)
1889 i += txn->mt_dbs[MAIN_DBI].md_depth;
1890 /* For puts, roughly factor in the key+data size */
1892 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1893 i += i; /* double it for good measure */
1896 if (txn->mt_dirty_room > i)
1899 if (!txn->mt_spill_pgs) {
1900 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1901 if (!txn->mt_spill_pgs)
1904 /* purge deleted slots */
1905 MDB_IDL sl = txn->mt_spill_pgs;
1906 unsigned int num = sl[0];
1908 for (i=1; i<=num; i++) {
1915 /* Preserve pages which may soon be dirtied again */
1916 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1919 /* Less aggressive spill - we originally spilled the entire dirty list,
1920 * with a few exceptions for cursor pages and DB root pages. But this
1921 * turns out to be a lot of wasted effort because in a large txn many
1922 * of those pages will need to be used again. So now we spill only 1/8th
1923 * of the dirty pages. Testing revealed this to be a good tradeoff,
1924 * better than 1/2, 1/4, or 1/10.
1926 if (need < MDB_IDL_UM_MAX / 8)
1927 need = MDB_IDL_UM_MAX / 8;
1929 /* Save the page IDs of all the pages we're flushing */
1930 /* flush from the tail forward, this saves a lot of shifting later on. */
1931 for (i=dl[0].mid; i && need; i--) {
1932 MDB_ID pn = dl[i].mid << 1;
1934 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1936 /* Can't spill twice, make sure it's not already in a parent's
1939 if (txn->mt_parent) {
1941 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1942 if (tx2->mt_spill_pgs) {
1943 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1944 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1945 dp->mp_flags |= P_KEEP;
1953 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1957 mdb_midl_sort(txn->mt_spill_pgs);
1959 /* Flush the spilled part of dirty list */
1960 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1963 /* Reset any dirty pages we kept that page_flush didn't see */
1964 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1967 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1971 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1973 mdb_find_oldest(MDB_txn *txn)
1976 txnid_t mr, oldest = txn->mt_txnid - 1;
1977 if (txn->mt_env->me_txns) {
1978 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1979 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1990 /** Add a page to the txn's dirty list */
1992 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1995 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1997 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1998 insert = mdb_mid2l_append;
2000 insert = mdb_mid2l_insert;
2002 mid.mid = mp->mp_pgno;
2004 rc = insert(txn->mt_u.dirty_list, &mid);
2005 mdb_tassert(txn, rc == 0);
2006 txn->mt_dirty_room--;
2009 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2010 * me_pghead and mt_next_pgno.
2012 * If there are free pages available from older transactions, they
2013 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2014 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2015 * and move me_pglast to say which records were consumed. Only this
2016 * function can create me_pghead and move me_pglast/mt_next_pgno.
2017 * @param[in] mc cursor A cursor handle identifying the transaction and
2018 * database for which we are allocating.
2019 * @param[in] num the number of pages to allocate.
2020 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2021 * will always be satisfied by a single contiguous chunk of memory.
2022 * @return 0 on success, non-zero on failure.
2025 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2027 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2028 /* Get at most <Max_retries> more freeDB records once me_pghead
2029 * has enough pages. If not enough, use new pages from the map.
2030 * If <Paranoid> and mc is updating the freeDB, only get new
2031 * records if me_pghead is empty. Then the freelist cannot play
2032 * catch-up with itself by growing while trying to save it.
2034 enum { Paranoid = 1, Max_retries = 500 };
2036 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2038 int rc, retry = num * 60;
2039 MDB_txn *txn = mc->mc_txn;
2040 MDB_env *env = txn->mt_env;
2041 pgno_t pgno, *mop = env->me_pghead;
2042 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2044 txnid_t oldest = 0, last;
2049 /* If there are any loose pages, just use them */
2050 if (num == 1 && txn->mt_loose_pgs) {
2051 np = txn->mt_loose_pgs;
2052 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2053 txn->mt_loose_count--;
2054 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2062 /* If our dirty list is already full, we can't do anything */
2063 if (txn->mt_dirty_room == 0) {
2068 for (op = MDB_FIRST;; op = MDB_NEXT) {
2073 /* Seek a big enough contiguous page range. Prefer
2074 * pages at the tail, just truncating the list.
2080 if (mop[i-n2] == pgno+n2)
2087 if (op == MDB_FIRST) { /* 1st iteration */
2088 /* Prepare to fetch more and coalesce */
2089 last = env->me_pglast;
2090 oldest = env->me_pgoldest;
2091 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2094 key.mv_data = &last; /* will look up last+1 */
2095 key.mv_size = sizeof(last);
2097 if (Paranoid && mc->mc_dbi == FREE_DBI)
2100 if (Paranoid && retry < 0 && mop_len)
2104 /* Do not fetch more if the record will be too recent */
2105 if (oldest <= last) {
2107 oldest = mdb_find_oldest(txn);
2108 env->me_pgoldest = oldest;
2114 rc = mdb_cursor_get(&m2, &key, NULL, op);
2116 if (rc == MDB_NOTFOUND)
2120 last = *(txnid_t*)key.mv_data;
2121 if (oldest <= last) {
2123 oldest = mdb_find_oldest(txn);
2124 env->me_pgoldest = oldest;
2130 np = m2.mc_pg[m2.mc_top];
2131 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2132 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2135 idl = (MDB_ID *) data.mv_data;
2138 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2143 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2145 mop = env->me_pghead;
2147 env->me_pglast = last;
2149 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2150 last, txn->mt_dbs[FREE_DBI].md_root, i));
2152 DPRINTF(("IDL %"Z"u", idl[j]));
2154 /* Merge in descending sorted order */
2155 mdb_midl_xmerge(mop, idl);
2159 /* Use new pages from the map when nothing suitable in the freeDB */
2161 pgno = txn->mt_next_pgno;
2162 if (pgno + num >= env->me_maxpg) {
2163 DPUTS("DB size maxed out");
2169 if (env->me_flags & MDB_WRITEMAP) {
2170 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2172 if (!(np = mdb_page_malloc(txn, num))) {
2178 mop[0] = mop_len -= num;
2179 /* Move any stragglers down */
2180 for (j = i-num; j < mop_len; )
2181 mop[++j] = mop[++i];
2183 txn->mt_next_pgno = pgno + num;
2186 mdb_page_dirty(txn, np);
2192 txn->mt_flags |= MDB_TXN_ERROR;
2196 /** Copy the used portions of a non-overflow page.
2197 * @param[in] dst page to copy into
2198 * @param[in] src page to copy from
2199 * @param[in] psize size of a page
2202 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2204 enum { Align = sizeof(pgno_t) };
2205 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2207 /* If page isn't full, just copy the used portion. Adjust
2208 * alignment so memcpy may copy words instead of bytes.
2210 if ((unused &= -Align) && !IS_LEAF2(src)) {
2211 upper = (upper + PAGEBASE) & -Align;
2212 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2213 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2216 memcpy(dst, src, psize - unused);
2220 /** Pull a page off the txn's spill list, if present.
2221 * If a page being referenced was spilled to disk in this txn, bring
2222 * it back and make it dirty/writable again.
2223 * @param[in] txn the transaction handle.
2224 * @param[in] mp the page being referenced. It must not be dirty.
2225 * @param[out] ret the writable page, if any. ret is unchanged if
2226 * mp wasn't spilled.
2229 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2231 MDB_env *env = txn->mt_env;
2234 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2236 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2237 if (!tx2->mt_spill_pgs)
2239 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2240 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2243 if (txn->mt_dirty_room == 0)
2244 return MDB_TXN_FULL;
2245 if (IS_OVERFLOW(mp))
2249 if (env->me_flags & MDB_WRITEMAP) {
2252 np = mdb_page_malloc(txn, num);
2256 memcpy(np, mp, num * env->me_psize);
2258 mdb_page_copy(np, mp, env->me_psize);
2261 /* If in current txn, this page is no longer spilled.
2262 * If it happens to be the last page, truncate the spill list.
2263 * Otherwise mark it as deleted by setting the LSB.
2265 if (x == txn->mt_spill_pgs[0])
2266 txn->mt_spill_pgs[0]--;
2268 txn->mt_spill_pgs[x] |= 1;
2269 } /* otherwise, if belonging to a parent txn, the
2270 * page remains spilled until child commits
2273 mdb_page_dirty(txn, np);
2274 np->mp_flags |= P_DIRTY;
2282 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2283 * @param[in] mc cursor pointing to the page to be touched
2284 * @return 0 on success, non-zero on failure.
2287 mdb_page_touch(MDB_cursor *mc)
2289 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2290 MDB_txn *txn = mc->mc_txn;
2291 MDB_cursor *m2, *m3;
2295 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2296 if (txn->mt_flags & MDB_TXN_SPILLS) {
2298 rc = mdb_page_unspill(txn, mp, &np);
2304 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2305 (rc = mdb_page_alloc(mc, 1, &np)))
2308 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2309 mp->mp_pgno, pgno));
2310 mdb_cassert(mc, mp->mp_pgno != pgno);
2311 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2312 /* Update the parent page, if any, to point to the new page */
2314 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2315 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2316 SETPGNO(node, pgno);
2318 mc->mc_db->md_root = pgno;
2320 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2321 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2323 /* If txn has a parent, make sure the page is in our
2327 unsigned x = mdb_mid2l_search(dl, pgno);
2328 if (x <= dl[0].mid && dl[x].mid == pgno) {
2329 if (mp != dl[x].mptr) { /* bad cursor? */
2330 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2331 txn->mt_flags |= MDB_TXN_ERROR;
2332 return MDB_CORRUPTED;
2337 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2339 np = mdb_page_malloc(txn, 1);
2344 rc = mdb_mid2l_insert(dl, &mid);
2345 mdb_cassert(mc, rc == 0);
2350 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2352 np->mp_flags |= P_DIRTY;
2355 /* Adjust cursors pointing to mp */
2356 mc->mc_pg[mc->mc_top] = np;
2357 m2 = txn->mt_cursors[mc->mc_dbi];
2358 if (mc->mc_flags & C_SUB) {
2359 for (; m2; m2=m2->mc_next) {
2360 m3 = &m2->mc_xcursor->mx_cursor;
2361 if (m3->mc_snum < mc->mc_snum) continue;
2362 if (m3->mc_pg[mc->mc_top] == mp)
2363 m3->mc_pg[mc->mc_top] = np;
2366 for (; m2; m2=m2->mc_next) {
2367 if (m2->mc_snum < mc->mc_snum) continue;
2368 if (m2->mc_pg[mc->mc_top] == mp) {
2369 m2->mc_pg[mc->mc_top] = np;
2370 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2372 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2374 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2375 if (!(leaf->mn_flags & F_SUBDATA))
2376 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2384 txn->mt_flags |= MDB_TXN_ERROR;
2389 mdb_env_sync(MDB_env *env, int force)
2392 if (env->me_flags & MDB_RDONLY)
2394 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2395 if (env->me_flags & MDB_WRITEMAP) {
2396 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2397 ? MS_ASYNC : MS_SYNC;
2398 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2401 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2405 #ifdef BROKEN_FDATASYNC
2406 if (env->me_flags & MDB_FSYNCONLY) {
2407 if (fsync(env->me_fd))
2411 if (MDB_FDATASYNC(env->me_fd))
2418 /** Back up parent txn's cursors, then grab the originals for tracking */
2420 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2422 MDB_cursor *mc, *bk;
2427 for (i = src->mt_numdbs; --i >= 0; ) {
2428 if ((mc = src->mt_cursors[i]) != NULL) {
2429 size = sizeof(MDB_cursor);
2431 size += sizeof(MDB_xcursor);
2432 for (; mc; mc = bk->mc_next) {
2438 mc->mc_db = &dst->mt_dbs[i];
2439 /* Kill pointers into src - and dst to reduce abuse: The
2440 * user may not use mc until dst ends. Otherwise we'd...
2442 mc->mc_txn = NULL; /* ...set this to dst */
2443 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2444 if ((mx = mc->mc_xcursor) != NULL) {
2445 *(MDB_xcursor *)(bk+1) = *mx;
2446 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2448 mc->mc_next = dst->mt_cursors[i];
2449 dst->mt_cursors[i] = mc;
2456 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2457 * @param[in] txn the transaction handle.
2458 * @param[in] merge true to keep changes to parent cursors, false to revert.
2459 * @return 0 on success, non-zero on failure.
2462 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2464 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2468 for (i = txn->mt_numdbs; --i >= 0; ) {
2469 for (mc = cursors[i]; mc; mc = next) {
2471 if ((bk = mc->mc_backup) != NULL) {
2473 /* Commit changes to parent txn */
2474 mc->mc_next = bk->mc_next;
2475 mc->mc_backup = bk->mc_backup;
2476 mc->mc_txn = bk->mc_txn;
2477 mc->mc_db = bk->mc_db;
2478 mc->mc_dbflag = bk->mc_dbflag;
2479 if ((mx = mc->mc_xcursor) != NULL)
2480 mx->mx_cursor.mc_txn = bk->mc_txn;
2482 /* Abort nested txn */
2484 if ((mx = mc->mc_xcursor) != NULL)
2485 *mx = *(MDB_xcursor *)(bk+1);
2489 /* Only malloced cursors are permanently tracked. */
2497 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2500 mdb_txn_reset0(MDB_txn *txn, const char *act);
2502 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2508 Pidset = F_SETLK, Pidcheck = F_GETLK
2512 /** Set or check a pid lock. Set returns 0 on success.
2513 * Check returns 0 if the process is certainly dead, nonzero if it may
2514 * be alive (the lock exists or an error happened so we do not know).
2516 * On Windows Pidset is a no-op, we merely check for the existence
2517 * of the process with the given pid. On POSIX we use a single byte
2518 * lock on the lockfile, set at an offset equal to the pid.
2521 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2523 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2526 if (op == Pidcheck) {
2527 h = OpenProcess(env->me_pidquery, FALSE, pid);
2528 /* No documented "no such process" code, but other program use this: */
2530 return ErrCode() != ERROR_INVALID_PARAMETER;
2531 /* A process exists until all handles to it close. Has it exited? */
2532 ret = WaitForSingleObject(h, 0) != 0;
2539 struct flock lock_info;
2540 memset(&lock_info, 0, sizeof(lock_info));
2541 lock_info.l_type = F_WRLCK;
2542 lock_info.l_whence = SEEK_SET;
2543 lock_info.l_start = pid;
2544 lock_info.l_len = 1;
2545 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2546 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2548 } else if ((rc = ErrCode()) == EINTR) {
2556 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2557 * @param[in] txn the transaction handle to initialize
2558 * @return 0 on success, non-zero on failure.
2561 mdb_txn_renew0(MDB_txn *txn)
2563 MDB_env *env = txn->mt_env;
2564 MDB_txninfo *ti = env->me_txns;
2568 int rc, new_notls = 0;
2570 if (txn->mt_flags & MDB_TXN_RDONLY) {
2572 txn->mt_numdbs = env->me_numdbs;
2573 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2575 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2576 txn->mt_txnid = meta->mm_txnid;
2577 txn->mt_u.reader = NULL;
2579 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2580 pthread_getspecific(env->me_txkey);
2582 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2583 return MDB_BAD_RSLOT;
2585 MDB_PID_T pid = env->me_pid;
2586 MDB_THR_T tid = pthread_self();
2587 mdb_mutex_t *rmutex = MDB_MUTEX(env, r);
2589 if (!env->me_live_reader) {
2590 rc = mdb_reader_pid(env, Pidset, pid);
2593 env->me_live_reader = 1;
2596 if (LOCK_MUTEX(rc, env, rmutex))
2598 nr = ti->mti_numreaders;
2599 for (i=0; i<nr; i++)
2600 if (ti->mti_readers[i].mr_pid == 0)
2602 if (i == env->me_maxreaders) {
2603 UNLOCK_MUTEX(rmutex);
2604 return MDB_READERS_FULL;
2606 r = &ti->mti_readers[i];
2607 /* Claim the reader slot, carefully since other code
2608 * uses the reader table un-mutexed: First reset the
2609 * slot, next publish it in mti_numreaders. After
2610 * that, it is safe for mdb_env_close() to touch it.
2611 * When it will be closed, we can finally claim it.
2614 r->mr_txnid = (txnid_t)-1;
2617 ti->mti_numreaders = ++nr;
2618 env->me_close_readers = nr;
2620 UNLOCK_MUTEX(rmutex);
2622 new_notls = (env->me_flags & MDB_NOTLS);
2623 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2628 do /* LY: Retry on a race, ITS#7970. */
2629 r->mr_txnid = ti->mti_txnid;
2630 while(r->mr_txnid != ti->mti_txnid);
2631 txn->mt_txnid = r->mr_txnid;
2632 txn->mt_u.reader = r;
2633 meta = env->me_metas[txn->mt_txnid & 1];
2637 if (LOCK_MUTEX(rc, env, MDB_MUTEX(env, w)))
2639 txn->mt_txnid = ti->mti_txnid;
2640 meta = env->me_metas[txn->mt_txnid & 1];
2642 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2643 txn->mt_txnid = meta->mm_txnid;
2646 txn->mt_numdbs = env->me_numdbs;
2649 if (txn->mt_txnid == mdb_debug_start)
2653 txn->mt_child = NULL;
2654 txn->mt_loose_pgs = NULL;
2655 txn->mt_loose_count = 0;
2656 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2657 txn->mt_u.dirty_list = env->me_dirty_list;
2658 txn->mt_u.dirty_list[0].mid = 0;
2659 txn->mt_free_pgs = env->me_free_pgs;
2660 txn->mt_free_pgs[0] = 0;
2661 txn->mt_spill_pgs = NULL;
2663 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2666 /* Copy the DB info and flags */
2667 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2669 /* Moved to here to avoid a data race in read TXNs */
2670 txn->mt_next_pgno = meta->mm_last_pg+1;
2672 for (i=2; i<txn->mt_numdbs; i++) {
2673 x = env->me_dbflags[i];
2674 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2675 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2677 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2679 if (env->me_maxpg < txn->mt_next_pgno) {
2680 mdb_txn_reset0(txn, "renew0-mapfail");
2682 txn->mt_u.reader->mr_pid = 0;
2683 txn->mt_u.reader = NULL;
2685 return MDB_MAP_RESIZED;
2692 mdb_txn_renew(MDB_txn *txn)
2696 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2699 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2700 DPUTS("environment had fatal error, must shutdown!");
2704 rc = mdb_txn_renew0(txn);
2705 if (rc == MDB_SUCCESS) {
2706 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2707 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2708 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2714 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2718 int rc, size, tsize = sizeof(MDB_txn);
2720 if (env->me_flags & MDB_FATAL_ERROR) {
2721 DPUTS("environment had fatal error, must shutdown!");
2724 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2727 /* Nested transactions: Max 1 child, write txns only, no writemap */
2728 if (parent->mt_child ||
2729 (flags & MDB_RDONLY) ||
2730 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2731 (env->me_flags & MDB_WRITEMAP))
2733 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2735 tsize = sizeof(MDB_ntxn);
2738 if (!(flags & MDB_RDONLY)) {
2740 txn = env->me_txn0; /* just reuse preallocated write txn */
2743 /* child txns use own copy of cursors */
2744 size += env->me_maxdbs * sizeof(MDB_cursor *);
2746 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2748 if ((txn = calloc(1, size)) == NULL) {
2749 DPRINTF(("calloc: %s", strerror(errno)));
2752 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2753 if (flags & MDB_RDONLY) {
2754 txn->mt_flags |= MDB_TXN_RDONLY;
2755 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2756 txn->mt_dbiseqs = env->me_dbiseqs;
2758 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2760 txn->mt_dbiseqs = parent->mt_dbiseqs;
2761 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2763 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2764 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2772 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2773 if (!txn->mt_u.dirty_list ||
2774 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2776 free(txn->mt_u.dirty_list);
2780 txn->mt_txnid = parent->mt_txnid;
2781 txn->mt_dirty_room = parent->mt_dirty_room;
2782 txn->mt_u.dirty_list[0].mid = 0;
2783 txn->mt_spill_pgs = NULL;
2784 txn->mt_next_pgno = parent->mt_next_pgno;
2785 parent->mt_child = txn;
2786 txn->mt_parent = parent;
2787 txn->mt_numdbs = parent->mt_numdbs;
2788 txn->mt_flags = parent->mt_flags;
2789 txn->mt_dbxs = parent->mt_dbxs;
2790 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2791 /* Copy parent's mt_dbflags, but clear DB_NEW */
2792 for (i=0; i<txn->mt_numdbs; i++)
2793 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2795 ntxn = (MDB_ntxn *)txn;
2796 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2797 if (env->me_pghead) {
2798 size = MDB_IDL_SIZEOF(env->me_pghead);
2799 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2801 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2806 rc = mdb_cursor_shadow(parent, txn);
2808 mdb_txn_reset0(txn, "beginchild-fail");
2810 rc = mdb_txn_renew0(txn);
2813 if (txn != env->me_txn0)
2817 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2818 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2819 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2826 mdb_txn_env(MDB_txn *txn)
2828 if(!txn) return NULL;
2833 mdb_txn_id(MDB_txn *txn)
2836 return txn->mt_txnid;
2839 /** Export or close DBI handles opened in this txn. */
2841 mdb_dbis_update(MDB_txn *txn, int keep)
2844 MDB_dbi n = txn->mt_numdbs;
2845 MDB_env *env = txn->mt_env;
2846 unsigned char *tdbflags = txn->mt_dbflags;
2848 for (i = n; --i >= 2;) {
2849 if (tdbflags[i] & DB_NEW) {
2851 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2853 char *ptr = env->me_dbxs[i].md_name.mv_data;
2855 env->me_dbxs[i].md_name.mv_data = NULL;
2856 env->me_dbxs[i].md_name.mv_size = 0;
2857 env->me_dbflags[i] = 0;
2858 env->me_dbiseqs[i]++;
2864 if (keep && env->me_numdbs < n)
2868 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2869 * May be called twice for readonly txns: First reset it, then abort.
2870 * @param[in] txn the transaction handle to reset
2871 * @param[in] act why the transaction is being reset
2874 mdb_txn_reset0(MDB_txn *txn, const char *act)
2876 MDB_env *env = txn->mt_env;
2878 /* Close any DBI handles opened in this txn */
2879 mdb_dbis_update(txn, 0);
2881 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2882 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2883 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2885 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2886 if (txn->mt_u.reader) {
2887 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2888 if (!(env->me_flags & MDB_NOTLS))
2889 txn->mt_u.reader = NULL; /* txn does not own reader */
2891 txn->mt_numdbs = 0; /* close nothing if called again */
2892 txn->mt_dbxs = NULL; /* mark txn as reset */
2894 pgno_t *pghead = env->me_pghead;
2896 mdb_cursors_close(txn, 0);
2897 if (!(env->me_flags & MDB_WRITEMAP)) {
2898 mdb_dlist_free(txn);
2901 if (!txn->mt_parent) {
2902 if (mdb_midl_shrink(&txn->mt_free_pgs))
2903 env->me_free_pgs = txn->mt_free_pgs;
2905 env->me_pghead = NULL;
2909 /* The writer mutex was locked in mdb_txn_begin. */
2911 UNLOCK_MUTEX(MDB_MUTEX(env, w));
2913 txn->mt_parent->mt_child = NULL;
2914 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2915 mdb_midl_free(txn->mt_free_pgs);
2916 mdb_midl_free(txn->mt_spill_pgs);
2917 free(txn->mt_u.dirty_list);
2920 mdb_midl_free(pghead);
2925 mdb_txn_reset(MDB_txn *txn)
2930 /* This call is only valid for read-only txns */
2931 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2934 mdb_txn_reset0(txn, "reset");
2938 mdb_txn_abort(MDB_txn *txn)
2944 mdb_txn_abort(txn->mt_child);
2946 mdb_txn_reset0(txn, "abort");
2947 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2948 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2949 txn->mt_u.reader->mr_pid = 0;
2951 if (txn != txn->mt_env->me_txn0)
2955 /** Save the freelist as of this transaction to the freeDB.
2956 * This changes the freelist. Keep trying until it stabilizes.
2959 mdb_freelist_save(MDB_txn *txn)
2961 /* env->me_pghead[] can grow and shrink during this call.
2962 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2963 * Page numbers cannot disappear from txn->mt_free_pgs[].
2966 MDB_env *env = txn->mt_env;
2967 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2968 txnid_t pglast = 0, head_id = 0;
2969 pgno_t freecnt = 0, *free_pgs, *mop;
2970 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2972 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2974 if (env->me_pghead) {
2975 /* Make sure first page of freeDB is touched and on freelist */
2976 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2977 if (rc && rc != MDB_NOTFOUND)
2981 if (!env->me_pghead && txn->mt_loose_pgs) {
2982 /* Put loose page numbers in mt_free_pgs, since
2983 * we may be unable to return them to me_pghead.
2985 MDB_page *mp = txn->mt_loose_pgs;
2986 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2988 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2989 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2990 txn->mt_loose_pgs = NULL;
2991 txn->mt_loose_count = 0;
2994 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2995 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2996 ? SSIZE_MAX : maxfree_1pg;
2999 /* Come back here after each Put() in case freelist changed */
3004 /* If using records from freeDB which we have not yet
3005 * deleted, delete them and any we reserved for me_pghead.
3007 while (pglast < env->me_pglast) {
3008 rc = mdb_cursor_first(&mc, &key, NULL);
3011 pglast = head_id = *(txnid_t *)key.mv_data;
3012 total_room = head_room = 0;
3013 mdb_tassert(txn, pglast <= env->me_pglast);
3014 rc = mdb_cursor_del(&mc, 0);
3019 /* Save the IDL of pages freed by this txn, to a single record */
3020 if (freecnt < txn->mt_free_pgs[0]) {
3022 /* Make sure last page of freeDB is touched and on freelist */
3023 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3024 if (rc && rc != MDB_NOTFOUND)
3027 free_pgs = txn->mt_free_pgs;
3028 /* Write to last page of freeDB */
3029 key.mv_size = sizeof(txn->mt_txnid);
3030 key.mv_data = &txn->mt_txnid;
3032 freecnt = free_pgs[0];
3033 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3034 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3037 /* Retry if mt_free_pgs[] grew during the Put() */
3038 free_pgs = txn->mt_free_pgs;
3039 } while (freecnt < free_pgs[0]);
3040 mdb_midl_sort(free_pgs);
3041 memcpy(data.mv_data, free_pgs, data.mv_size);
3044 unsigned int i = free_pgs[0];
3045 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
3046 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3048 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3054 mop = env->me_pghead;
3055 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3057 /* Reserve records for me_pghead[]. Split it if multi-page,
3058 * to avoid searching freeDB for a page range. Use keys in
3059 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3061 if (total_room >= mop_len) {
3062 if (total_room == mop_len || --more < 0)
3064 } else if (head_room >= maxfree_1pg && head_id > 1) {
3065 /* Keep current record (overflow page), add a new one */
3069 /* (Re)write {key = head_id, IDL length = head_room} */
3070 total_room -= head_room;
3071 head_room = mop_len - total_room;
3072 if (head_room > maxfree_1pg && head_id > 1) {
3073 /* Overflow multi-page for part of me_pghead */
3074 head_room /= head_id; /* amortize page sizes */
3075 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3076 } else if (head_room < 0) {
3077 /* Rare case, not bothering to delete this record */
3080 key.mv_size = sizeof(head_id);
3081 key.mv_data = &head_id;
3082 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3083 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3086 /* IDL is initially empty, zero out at least the length */
3087 pgs = (pgno_t *)data.mv_data;
3088 j = head_room > clean_limit ? head_room : 0;
3092 total_room += head_room;
3095 /* Return loose page numbers to me_pghead, though usually none are
3096 * left at this point. The pages themselves remain in dirty_list.
3098 if (txn->mt_loose_pgs) {
3099 MDB_page *mp = txn->mt_loose_pgs;
3100 unsigned count = txn->mt_loose_count;
3102 /* Room for loose pages + temp IDL with same */
3103 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3105 mop = env->me_pghead;
3106 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3107 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3108 loose[ ++count ] = mp->mp_pgno;
3110 mdb_midl_sort(loose);
3111 mdb_midl_xmerge(mop, loose);
3112 txn->mt_loose_pgs = NULL;
3113 txn->mt_loose_count = 0;
3117 /* Fill in the reserved me_pghead records */
3123 rc = mdb_cursor_first(&mc, &key, &data);
3124 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3125 txnid_t id = *(txnid_t *)key.mv_data;
3126 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3129 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3131 if (len > mop_len) {
3133 data.mv_size = (len + 1) * sizeof(MDB_ID);
3135 data.mv_data = mop -= len;
3138 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3140 if (rc || !(mop_len -= len))
3147 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3148 * @param[in] txn the transaction that's being committed
3149 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3150 * @return 0 on success, non-zero on failure.
3153 mdb_page_flush(MDB_txn *txn, int keep)
3155 MDB_env *env = txn->mt_env;
3156 MDB_ID2L dl = txn->mt_u.dirty_list;
3157 unsigned psize = env->me_psize, j;
3158 int i, pagecount = dl[0].mid, rc;
3159 size_t size = 0, pos = 0;
3161 MDB_page *dp = NULL;
3165 struct iovec iov[MDB_COMMIT_PAGES];
3166 ssize_t wpos = 0, wsize = 0, wres;
3167 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3173 if (env->me_flags & MDB_WRITEMAP) {
3174 /* Clear dirty flags */
3175 while (++i <= pagecount) {
3177 /* Don't flush this page yet */
3178 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3179 dp->mp_flags &= ~P_KEEP;
3183 dp->mp_flags &= ~P_DIRTY;
3188 /* Write the pages */
3190 if (++i <= pagecount) {
3192 /* Don't flush this page yet */
3193 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3194 dp->mp_flags &= ~P_KEEP;
3199 /* clear dirty flag */
3200 dp->mp_flags &= ~P_DIRTY;
3203 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3208 /* Windows actually supports scatter/gather I/O, but only on
3209 * unbuffered file handles. Since we're relying on the OS page
3210 * cache for all our data, that's self-defeating. So we just
3211 * write pages one at a time. We use the ov structure to set
3212 * the write offset, to at least save the overhead of a Seek
3215 DPRINTF(("committing page %"Z"u", pgno));
3216 memset(&ov, 0, sizeof(ov));
3217 ov.Offset = pos & 0xffffffff;
3218 ov.OffsetHigh = pos >> 16 >> 16;
3219 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3221 DPRINTF(("WriteFile: %d", rc));
3225 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3226 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3228 /* Write previous page(s) */
3229 #ifdef MDB_USE_PWRITEV
3230 wres = pwritev(env->me_fd, iov, n, wpos);
3233 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3235 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3237 DPRINTF(("lseek: %s", strerror(rc)));
3240 wres = writev(env->me_fd, iov, n);
3243 if (wres != wsize) {
3246 DPRINTF(("Write error: %s", strerror(rc)));
3248 rc = EIO; /* TODO: Use which error code? */
3249 DPUTS("short write, filesystem full?");
3260 DPRINTF(("committing page %"Z"u", pgno));
3261 next_pos = pos + size;
3262 iov[n].iov_len = size;
3263 iov[n].iov_base = (char *)dp;
3269 /* MIPS has cache coherency issues, this is a no-op everywhere else
3270 * Note: for any size >= on-chip cache size, entire on-chip cache is
3273 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3275 for (i = keep; ++i <= pagecount; ) {
3277 /* This is a page we skipped above */
3280 dl[j].mid = dp->mp_pgno;
3283 mdb_dpage_free(env, dp);
3288 txn->mt_dirty_room += i - j;
3294 mdb_txn_commit(MDB_txn *txn)
3300 if (txn == NULL || txn->mt_env == NULL)
3303 if (txn->mt_child) {
3304 rc = mdb_txn_commit(txn->mt_child);
3305 txn->mt_child = NULL;
3312 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3313 mdb_dbis_update(txn, 1);
3314 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3319 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3320 DPUTS("error flag is set, can't commit");
3322 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3327 if (txn->mt_parent) {
3328 MDB_txn *parent = txn->mt_parent;
3332 unsigned x, y, len, ps_len;
3334 /* Append our free list to parent's */
3335 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3338 mdb_midl_free(txn->mt_free_pgs);
3339 /* Failures after this must either undo the changes
3340 * to the parent or set MDB_TXN_ERROR in the parent.
3343 parent->mt_next_pgno = txn->mt_next_pgno;
3344 parent->mt_flags = txn->mt_flags;
3346 /* Merge our cursors into parent's and close them */
3347 mdb_cursors_close(txn, 1);
3349 /* Update parent's DB table. */
3350 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3351 parent->mt_numdbs = txn->mt_numdbs;
3352 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3353 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3354 for (i=2; i<txn->mt_numdbs; i++) {
3355 /* preserve parent's DB_NEW status */
3356 x = parent->mt_dbflags[i] & DB_NEW;
3357 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3360 dst = parent->mt_u.dirty_list;
3361 src = txn->mt_u.dirty_list;
3362 /* Remove anything in our dirty list from parent's spill list */
3363 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3365 pspill[0] = (pgno_t)-1;
3366 /* Mark our dirty pages as deleted in parent spill list */
3367 for (i=0, len=src[0].mid; ++i <= len; ) {
3368 MDB_ID pn = src[i].mid << 1;
3369 while (pn > pspill[x])
3371 if (pn == pspill[x]) {
3376 /* Squash deleted pagenums if we deleted any */
3377 for (x=y; ++x <= ps_len; )
3378 if (!(pspill[x] & 1))
3379 pspill[++y] = pspill[x];
3383 /* Find len = length of merging our dirty list with parent's */
3385 dst[0].mid = 0; /* simplify loops */
3386 if (parent->mt_parent) {
3387 len = x + src[0].mid;
3388 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3389 for (i = x; y && i; y--) {
3390 pgno_t yp = src[y].mid;
3391 while (yp < dst[i].mid)
3393 if (yp == dst[i].mid) {
3398 } else { /* Simplify the above for single-ancestor case */
3399 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3401 /* Merge our dirty list with parent's */
3403 for (i = len; y; dst[i--] = src[y--]) {
3404 pgno_t yp = src[y].mid;
3405 while (yp < dst[x].mid)
3406 dst[i--] = dst[x--];
3407 if (yp == dst[x].mid)
3408 free(dst[x--].mptr);
3410 mdb_tassert(txn, i == x);
3412 free(txn->mt_u.dirty_list);
3413 parent->mt_dirty_room = txn->mt_dirty_room;
3414 if (txn->mt_spill_pgs) {
3415 if (parent->mt_spill_pgs) {
3416 /* TODO: Prevent failure here, so parent does not fail */
3417 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3419 parent->mt_flags |= MDB_TXN_ERROR;
3420 mdb_midl_free(txn->mt_spill_pgs);
3421 mdb_midl_sort(parent->mt_spill_pgs);
3423 parent->mt_spill_pgs = txn->mt_spill_pgs;
3427 /* Append our loose page list to parent's */
3428 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3430 *lp = txn->mt_loose_pgs;
3431 parent->mt_loose_count += txn->mt_loose_count;
3433 parent->mt_child = NULL;
3434 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3439 if (txn != env->me_txn) {
3440 DPUTS("attempt to commit unknown transaction");
3445 mdb_cursors_close(txn, 0);
3447 if (!txn->mt_u.dirty_list[0].mid &&
3448 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3451 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3452 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3454 /* Update DB root pointers */
3455 if (txn->mt_numdbs > 2) {
3459 data.mv_size = sizeof(MDB_db);
3461 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3462 for (i = 2; i < txn->mt_numdbs; i++) {
3463 if (txn->mt_dbflags[i] & DB_DIRTY) {
3464 if (TXN_DBI_CHANGED(txn, i)) {
3468 data.mv_data = &txn->mt_dbs[i];
3469 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3476 rc = mdb_freelist_save(txn);
3480 mdb_midl_free(env->me_pghead);
3481 env->me_pghead = NULL;
3482 if (mdb_midl_shrink(&txn->mt_free_pgs))
3483 env->me_free_pgs = txn->mt_free_pgs;
3489 if ((rc = mdb_page_flush(txn, 0)) ||
3490 (rc = mdb_env_sync(env, 0)) ||
3491 (rc = mdb_env_write_meta(txn)))
3494 /* Free P_LOOSE pages left behind in dirty_list */
3495 if (!(env->me_flags & MDB_WRITEMAP))
3496 mdb_dlist_free(txn);
3501 mdb_dbis_update(txn, 1);
3504 UNLOCK_MUTEX(MDB_MUTEX(env, w));
3505 if (txn != env->me_txn0)
3515 /** Read the environment parameters of a DB environment before
3516 * mapping it into memory.
3517 * @param[in] env the environment handle
3518 * @param[out] meta address of where to store the meta information
3519 * @return 0 on success, non-zero on failure.
3522 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3528 enum { Size = sizeof(pbuf) };
3530 /* We don't know the page size yet, so use a minimum value.
3531 * Read both meta pages so we can use the latest one.
3534 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3538 memset(&ov, 0, sizeof(ov));
3540 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3541 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3544 rc = pread(env->me_fd, &pbuf, Size, off);
3547 if (rc == 0 && off == 0)
3549 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3550 DPRINTF(("read: %s", mdb_strerror(rc)));
3554 p = (MDB_page *)&pbuf;
3556 if (!F_ISSET(p->mp_flags, P_META)) {
3557 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3562 if (m->mm_magic != MDB_MAGIC) {
3563 DPUTS("meta has invalid magic");
3567 if (m->mm_version != MDB_DATA_VERSION) {
3568 DPRINTF(("database is version %u, expected version %u",
3569 m->mm_version, MDB_DATA_VERSION));
3570 return MDB_VERSION_MISMATCH;
3573 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3579 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3581 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3583 meta->mm_magic = MDB_MAGIC;
3584 meta->mm_version = MDB_DATA_VERSION;
3585 meta->mm_mapsize = env->me_mapsize;
3586 meta->mm_psize = env->me_psize;
3587 meta->mm_last_pg = 1;
3588 meta->mm_flags = env->me_flags & 0xffff;
3589 meta->mm_flags |= MDB_INTEGERKEY;
3590 meta->mm_dbs[0].md_root = P_INVALID;
3591 meta->mm_dbs[1].md_root = P_INVALID;
3594 /** Write the environment parameters of a freshly created DB environment.
3595 * @param[in] env the environment handle
3596 * @param[in] meta the #MDB_meta to write
3597 * @return 0 on success, non-zero on failure.
3600 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3608 memset(&ov, 0, sizeof(ov));
3609 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3611 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3614 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3615 len = pwrite(fd, ptr, size, pos); \
3616 rc = (len >= 0); } while(0)
3619 DPUTS("writing new meta page");
3621 psize = env->me_psize;
3623 p = calloc(2, psize);
3625 p->mp_flags = P_META;
3626 *(MDB_meta *)METADATA(p) = *meta;
3628 q = (MDB_page *)((char *)p + psize);
3630 q->mp_flags = P_META;
3631 *(MDB_meta *)METADATA(q) = *meta;
3633 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3636 else if ((unsigned) len == psize * 2)
3644 /** Update the environment info to commit a transaction.
3645 * @param[in] txn the transaction that's being committed
3646 * @return 0 on success, non-zero on failure.
3649 mdb_env_write_meta(MDB_txn *txn)
3652 MDB_meta meta, metab, *mp;
3655 int rc, len, toggle;
3664 toggle = txn->mt_txnid & 1;
3665 DPRINTF(("writing meta page %d for root page %"Z"u",
3666 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3669 mp = env->me_metas[toggle];
3670 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3671 /* Persist any increases of mapsize config */
3672 if (mapsize < env->me_mapsize)
3673 mapsize = env->me_mapsize;
3675 if (env->me_flags & MDB_WRITEMAP) {
3676 mp->mm_mapsize = mapsize;
3677 mp->mm_dbs[0] = txn->mt_dbs[0];
3678 mp->mm_dbs[1] = txn->mt_dbs[1];
3679 mp->mm_last_pg = txn->mt_next_pgno - 1;
3680 #if !(defined(_MSC_VER) || defined(__i386__) || defined(__x86_64__))
3681 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3682 __sync_synchronize();
3684 mp->mm_txnid = txn->mt_txnid;
3685 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3686 unsigned meta_size = env->me_psize;
3687 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3690 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3691 if (meta_size < env->me_os_psize)
3692 meta_size += meta_size;
3697 if (MDB_MSYNC(ptr, meta_size, rc)) {
3704 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3705 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3707 meta.mm_mapsize = mapsize;
3708 meta.mm_dbs[0] = txn->mt_dbs[0];
3709 meta.mm_dbs[1] = txn->mt_dbs[1];
3710 meta.mm_last_pg = txn->mt_next_pgno - 1;
3711 meta.mm_txnid = txn->mt_txnid;
3713 off = offsetof(MDB_meta, mm_mapsize);
3714 ptr = (char *)&meta + off;
3715 len = sizeof(MDB_meta) - off;
3717 off += env->me_psize;
3720 /* Write to the SYNC fd */
3721 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3722 env->me_fd : env->me_mfd;
3725 memset(&ov, 0, sizeof(ov));
3727 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3731 rc = pwrite(mfd, ptr, len, off);
3734 rc = rc < 0 ? ErrCode() : EIO;
3735 DPUTS("write failed, disk error?");
3736 /* On a failure, the pagecache still contains the new data.
3737 * Write some old data back, to prevent it from being used.
3738 * Use the non-SYNC fd; we know it will fail anyway.
3740 meta.mm_last_pg = metab.mm_last_pg;
3741 meta.mm_txnid = metab.mm_txnid;
3743 memset(&ov, 0, sizeof(ov));
3745 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3747 r2 = pwrite(env->me_fd, ptr, len, off);
3748 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3751 env->me_flags |= MDB_FATAL_ERROR;
3754 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3755 CACHEFLUSH(env->me_map + off, len, DCACHE);
3757 /* Memory ordering issues are irrelevant; since the entire writer
3758 * is wrapped by wmutex, all of these changes will become visible
3759 * after the wmutex is unlocked. Since the DB is multi-version,
3760 * readers will get consistent data regardless of how fresh or
3761 * how stale their view of these values is.
3764 env->me_txns->mti_txnid = txn->mt_txnid;
3769 /** Check both meta pages to see which one is newer.
3770 * @param[in] env the environment handle
3771 * @return meta toggle (0 or 1).
3774 mdb_env_pick_meta(const MDB_env *env)
3776 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3780 mdb_env_create(MDB_env **env)
3784 e = calloc(1, sizeof(MDB_env));
3788 e->me_maxreaders = DEFAULT_READERS;
3789 e->me_maxdbs = e->me_numdbs = 2;
3790 e->me_fd = INVALID_HANDLE_VALUE;
3791 e->me_lfd = INVALID_HANDLE_VALUE;
3792 e->me_mfd = INVALID_HANDLE_VALUE;
3793 #ifdef MDB_USE_SYSV_SEM
3794 e->me_rmutex.semid = -1;
3795 e->me_wmutex.semid = -1;
3797 e->me_pid = getpid();
3798 GET_PAGESIZE(e->me_os_psize);
3799 VGMEMP_CREATE(e,0,0);
3805 mdb_env_map(MDB_env *env, void *addr)
3808 unsigned int flags = env->me_flags;
3812 LONG sizelo, sizehi;
3815 if (flags & MDB_RDONLY) {
3816 /* Don't set explicit map size, use whatever exists */
3821 msize = env->me_mapsize;
3822 sizelo = msize & 0xffffffff;
3823 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3825 /* Windows won't create mappings for zero length files.
3826 * and won't map more than the file size.
3827 * Just set the maxsize right now.
3829 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3830 || !SetEndOfFile(env->me_fd)
3831 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3835 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3836 PAGE_READWRITE : PAGE_READONLY,
3837 sizehi, sizelo, NULL);
3840 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3841 FILE_MAP_WRITE : FILE_MAP_READ,
3843 rc = env->me_map ? 0 : ErrCode();
3848 int prot = PROT_READ;
3849 if (flags & MDB_WRITEMAP) {
3851 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3854 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3856 if (env->me_map == MAP_FAILED) {
3861 if (flags & MDB_NORDAHEAD) {
3862 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3864 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3866 #ifdef POSIX_MADV_RANDOM
3867 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3868 #endif /* POSIX_MADV_RANDOM */
3869 #endif /* MADV_RANDOM */
3873 /* Can happen because the address argument to mmap() is just a
3874 * hint. mmap() can pick another, e.g. if the range is in use.
3875 * The MAP_FIXED flag would prevent that, but then mmap could
3876 * instead unmap existing pages to make room for the new map.
3878 if (addr && env->me_map != addr)
3879 return EBUSY; /* TODO: Make a new MDB_* error code? */
3881 p = (MDB_page *)env->me_map;
3882 env->me_metas[0] = METADATA(p);
3883 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3889 mdb_env_set_mapsize(MDB_env *env, size_t size)
3891 /* If env is already open, caller is responsible for making
3892 * sure there are no active txns.
3900 meta = env->me_metas[mdb_env_pick_meta(env)];
3902 size = meta->mm_mapsize;
3904 /* Silently round up to minimum if the size is too small */
3905 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3909 munmap(env->me_map, env->me_mapsize);
3910 env->me_mapsize = size;
3911 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3912 rc = mdb_env_map(env, old);
3916 env->me_mapsize = size;
3918 env->me_maxpg = env->me_mapsize / env->me_psize;
3923 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3927 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3932 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3934 if (env->me_map || readers < 1)
3936 env->me_maxreaders = readers;
3941 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3943 if (!env || !readers)
3945 *readers = env->me_maxreaders;
3950 mdb_fsize(HANDLE fd, size_t *size)
3953 LARGE_INTEGER fsize;
3955 if (!GetFileSizeEx(fd, &fsize))
3958 *size = fsize.QuadPart;
3970 #ifdef BROKEN_FDATASYNC
3971 #include <sys/utsname.h>
3972 #include <sys/vfs.h>
3975 /** Further setup required for opening an LMDB environment
3978 mdb_env_open2(MDB_env *env)
3980 unsigned int flags = env->me_flags;
3981 int i, newenv = 0, rc;
3985 /* See if we should use QueryLimited */
3987 if ((rc & 0xff) > 5)
3988 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3990 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3993 #ifdef BROKEN_FDATASYNC
3994 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3995 * https://lkml.org/lkml/2012/9/3/83
3996 * Kernels after 3.6-rc6 are known good.
3997 * https://lkml.org/lkml/2012/9/10/556
3998 * See if the DB is on ext3/ext4, then check for new enough kernel
3999 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4004 fstatfs(env->me_fd, &st);
4005 while (st.f_type == 0xEF53) {
4009 if (uts.release[0] < '3') {
4010 if (!strncmp(uts.release, "2.6.32.", 7)) {
4011 i = atoi(uts.release+7);
4013 break; /* 2.6.32.60 and newer is OK */
4014 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4015 i = atoi(uts.release+7);
4017 break; /* 2.6.34.15 and newer is OK */
4019 } else if (uts.release[0] == '3') {
4020 i = atoi(uts.release+2);
4022 break; /* 3.6 and newer is OK */
4024 i = atoi(uts.release+4);
4026 break; /* 3.5.4 and newer is OK */
4027 } else if (i == 2) {
4028 i = atoi(uts.release+4);
4030 break; /* 3.2.30 and newer is OK */
4032 } else { /* 4.x and newer is OK */
4035 env->me_flags |= MDB_FSYNCONLY;
4041 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4044 DPUTS("new mdbenv");
4046 env->me_psize = env->me_os_psize;
4047 if (env->me_psize > MAX_PAGESIZE)
4048 env->me_psize = MAX_PAGESIZE;
4049 memset(&meta, 0, sizeof(meta));
4050 mdb_env_init_meta0(env, &meta);
4051 meta.mm_mapsize = DEFAULT_MAPSIZE;
4053 env->me_psize = meta.mm_psize;
4056 /* Was a mapsize configured? */
4057 if (!env->me_mapsize) {
4058 env->me_mapsize = meta.mm_mapsize;
4061 /* Make sure mapsize >= committed data size. Even when using
4062 * mm_mapsize, which could be broken in old files (ITS#7789).
4064 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4065 if (env->me_mapsize < minsize)
4066 env->me_mapsize = minsize;
4068 meta.mm_mapsize = env->me_mapsize;
4070 if (newenv && !(flags & MDB_FIXEDMAP)) {
4071 /* mdb_env_map() may grow the datafile. Write the metapages
4072 * first, so the file will be valid if initialization fails.
4073 * Except with FIXEDMAP, since we do not yet know mm_address.
4074 * We could fill in mm_address later, but then a different
4075 * program might end up doing that - one with a memory layout
4076 * and map address which does not suit the main program.
4078 rc = mdb_env_init_meta(env, &meta);
4084 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4089 if (flags & MDB_FIXEDMAP)
4090 meta.mm_address = env->me_map;
4091 i = mdb_env_init_meta(env, &meta);
4092 if (i != MDB_SUCCESS) {
4097 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4098 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4100 #if !(MDB_MAXKEYSIZE)
4101 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4103 env->me_maxpg = env->me_mapsize / env->me_psize;
4107 int toggle = mdb_env_pick_meta(env);
4108 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4110 DPRINTF(("opened database version %u, pagesize %u",
4111 env->me_metas[0]->mm_version, env->me_psize));
4112 DPRINTF(("using meta page %d", toggle));
4113 DPRINTF(("depth: %u", db->md_depth));
4114 DPRINTF(("entries: %"Z"u", db->md_entries));
4115 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4116 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4117 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4118 DPRINTF(("root: %"Z"u", db->md_root));
4126 /** Release a reader thread's slot in the reader lock table.
4127 * This function is called automatically when a thread exits.
4128 * @param[in] ptr This points to the slot in the reader lock table.
4131 mdb_env_reader_dest(void *ptr)
4133 MDB_reader *reader = ptr;
4139 /** Junk for arranging thread-specific callbacks on Windows. This is
4140 * necessarily platform and compiler-specific. Windows supports up
4141 * to 1088 keys. Let's assume nobody opens more than 64 environments
4142 * in a single process, for now. They can override this if needed.
4144 #ifndef MAX_TLS_KEYS
4145 #define MAX_TLS_KEYS 64
4147 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4148 static int mdb_tls_nkeys;
4150 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4154 case DLL_PROCESS_ATTACH: break;
4155 case DLL_THREAD_ATTACH: break;
4156 case DLL_THREAD_DETACH:
4157 for (i=0; i<mdb_tls_nkeys; i++) {
4158 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4160 mdb_env_reader_dest(r);
4164 case DLL_PROCESS_DETACH: break;
4169 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4171 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4175 /* Force some symbol references.
4176 * _tls_used forces the linker to create the TLS directory if not already done
4177 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4179 #pragma comment(linker, "/INCLUDE:_tls_used")
4180 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4181 #pragma const_seg(".CRT$XLB")
4182 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4183 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4186 #pragma comment(linker, "/INCLUDE:__tls_used")
4187 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4188 #pragma data_seg(".CRT$XLB")
4189 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4191 #endif /* WIN 32/64 */
4192 #endif /* !__GNUC__ */
4195 /** Downgrade the exclusive lock on the region back to shared */
4197 mdb_env_share_locks(MDB_env *env, int *excl)
4199 int rc = 0, toggle = mdb_env_pick_meta(env);
4201 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4206 /* First acquire a shared lock. The Unlock will
4207 * then release the existing exclusive lock.
4209 memset(&ov, 0, sizeof(ov));
4210 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4213 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4219 struct flock lock_info;
4220 /* The shared lock replaces the existing lock */
4221 memset((void *)&lock_info, 0, sizeof(lock_info));
4222 lock_info.l_type = F_RDLCK;
4223 lock_info.l_whence = SEEK_SET;
4224 lock_info.l_start = 0;
4225 lock_info.l_len = 1;
4226 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4227 (rc = ErrCode()) == EINTR) ;
4228 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4235 /** Try to get exclusive lock, otherwise shared.
4236 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4239 mdb_env_excl_lock(MDB_env *env, int *excl)
4243 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4247 memset(&ov, 0, sizeof(ov));
4248 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4255 struct flock lock_info;
4256 memset((void *)&lock_info, 0, sizeof(lock_info));
4257 lock_info.l_type = F_WRLCK;
4258 lock_info.l_whence = SEEK_SET;
4259 lock_info.l_start = 0;
4260 lock_info.l_len = 1;
4261 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4262 (rc = ErrCode()) == EINTR) ;
4266 # ifdef MDB_USE_SYSV_SEM
4267 if (*excl < 0) /* always true when !MDB_USE_SYSV_SEM */
4270 lock_info.l_type = F_RDLCK;
4271 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4272 (rc = ErrCode()) == EINTR) ;
4282 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4284 * @(#) $Revision: 5.1 $
4285 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4286 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4288 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4292 * Please do not copyright this code. This code is in the public domain.
4294 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4295 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4296 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4297 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4298 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4299 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4300 * PERFORMANCE OF THIS SOFTWARE.
4303 * chongo <Landon Curt Noll> /\oo/\
4304 * http://www.isthe.com/chongo/
4306 * Share and Enjoy! :-)
4309 typedef unsigned long long mdb_hash_t;
4310 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4312 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4313 * @param[in] val value to hash
4314 * @param[in] hval initial value for hash
4315 * @return 64 bit hash
4317 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4318 * hval arg on the first call.
4321 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4323 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4324 unsigned char *end = s + val->mv_size;
4326 * FNV-1a hash each octet of the string
4329 /* xor the bottom with the current octet */
4330 hval ^= (mdb_hash_t)*s++;
4332 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4333 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4334 (hval << 7) + (hval << 8) + (hval << 40);
4336 /* return our new hash value */
4340 /** Hash the string and output the encoded hash.
4341 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4342 * very short name limits. We don't care about the encoding being reversible,
4343 * we just want to preserve as many bits of the input as possible in a
4344 * small printable string.
4345 * @param[in] str string to hash
4346 * @param[out] encbuf an array of 11 chars to hold the hash
4348 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4351 mdb_pack85(unsigned long l, char *out)
4355 for (i=0; i<5; i++) {
4356 *out++ = mdb_a85[l % 85];
4362 mdb_hash_enc(MDB_val *val, char *encbuf)
4364 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4366 mdb_pack85(h, encbuf);
4367 mdb_pack85(h>>32, encbuf+5);
4372 /** Open and/or initialize the lock region for the environment.
4373 * @param[in] env The LMDB environment.
4374 * @param[in] lpath The pathname of the file used for the lock region.
4375 * @param[in] mode The Unix permissions for the file, if we create it.
4376 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4377 * @return 0 on success, non-zero on failure.
4380 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4383 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4385 # define MDB_ERRCODE_ROFS EROFS
4386 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4387 # define MDB_CLOEXEC O_CLOEXEC
4390 # define MDB_CLOEXEC 0
4393 #ifdef MDB_USE_SYSV_SEM
4401 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4402 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4403 FILE_ATTRIBUTE_NORMAL, NULL);
4405 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4407 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4409 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4414 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4415 /* Lose record locks when exec*() */
4416 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4417 fcntl(env->me_lfd, F_SETFD, fdflags);
4420 if (!(env->me_flags & MDB_NOTLS)) {
4421 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4424 env->me_flags |= MDB_ENV_TXKEY;
4426 /* Windows TLS callbacks need help finding their TLS info. */
4427 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4431 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4435 /* Try to get exclusive lock. If we succeed, then
4436 * nobody is using the lock region and we should initialize it.
4438 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4441 size = GetFileSize(env->me_lfd, NULL);
4443 size = lseek(env->me_lfd, 0, SEEK_END);
4444 if (size == -1) goto fail_errno;
4446 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4447 if (size < rsize && *excl > 0) {
4449 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4450 || !SetEndOfFile(env->me_lfd))
4453 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4457 size = rsize - sizeof(MDB_txninfo);
4458 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4463 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4465 if (!mh) goto fail_errno;
4466 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4468 if (!env->me_txns) goto fail_errno;
4470 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4472 if (m == MAP_FAILED) goto fail_errno;
4478 BY_HANDLE_FILE_INFORMATION stbuf;
4487 if (!mdb_sec_inited) {
4488 InitializeSecurityDescriptor(&mdb_null_sd,
4489 SECURITY_DESCRIPTOR_REVISION);
4490 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4491 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4492 mdb_all_sa.bInheritHandle = FALSE;
4493 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4496 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4497 idbuf.volume = stbuf.dwVolumeSerialNumber;
4498 idbuf.nhigh = stbuf.nFileIndexHigh;
4499 idbuf.nlow = stbuf.nFileIndexLow;
4500 val.mv_data = &idbuf;
4501 val.mv_size = sizeof(idbuf);
4502 mdb_hash_enc(&val, encbuf);
4503 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4504 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4505 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4506 if (!env->me_rmutex) goto fail_errno;
4507 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4508 if (!env->me_wmutex) goto fail_errno;
4509 #elif defined(MDB_USE_SYSV_SEM)
4510 unsigned short vals[2] = {1, 1};
4511 semid = semget(IPC_PRIVATE, 2, mode);
4515 if (semctl(semid, 0, SETALL, semu) < 0)
4517 env->me_txns->mti_semid = semid;
4518 #else /* MDB_USE_SYSV_SEM */
4519 pthread_mutexattr_t mattr;
4521 if ((rc = pthread_mutexattr_init(&mattr))
4522 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4523 #ifdef MDB_ROBUST_SUPPORTED
4524 || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
4526 || (rc = pthread_mutex_init(&env->me_txns->mti_rmutex, &mattr))
4527 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4529 pthread_mutexattr_destroy(&mattr);
4530 #endif /* _WIN32 || MDB_USE_SYSV_SEM */
4532 env->me_txns->mti_magic = MDB_MAGIC;
4533 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4534 env->me_txns->mti_txnid = 0;
4535 env->me_txns->mti_numreaders = 0;
4538 #ifdef MDB_USE_SYSV_SEM
4539 struct semid_ds buf;
4541 if (env->me_txns->mti_magic != MDB_MAGIC) {
4542 DPUTS("lock region has invalid magic");
4546 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4547 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4548 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4549 rc = MDB_VERSION_MISMATCH;
4553 if (rc && rc != EACCES && rc != EAGAIN) {
4557 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4558 if (!env->me_rmutex) goto fail_errno;
4559 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4560 if (!env->me_wmutex) goto fail_errno;
4561 #elif defined(MDB_USE_SYSV_SEM)
4562 semid = env->me_txns->mti_semid;
4564 /* check for read access */
4565 if (semctl(semid, 0, IPC_STAT, semu) < 0)
4567 /* check for write access */
4568 if (semctl(semid, 0, IPC_SET, semu) < 0)
4572 #ifdef MDB_USE_SYSV_SEM
4573 env->me_rmutex.semid = semid;
4574 env->me_wmutex.semid = semid;
4575 env->me_rmutex.semnum = 0;
4576 env->me_wmutex.semnum = 1;
4577 env->me_rmutex.locked = &env->me_txns->mti_rlocked;
4578 env->me_wmutex.locked = &env->me_txns->mti_wlocked;
4589 /** The name of the lock file in the DB environment */
4590 #define LOCKNAME "/lock.mdb"
4591 /** The name of the data file in the DB environment */
4592 #define DATANAME "/data.mdb"
4593 /** The suffix of the lock file when no subdir is used */
4594 #define LOCKSUFF "-lock"
4595 /** Only a subset of the @ref mdb_env flags can be changed
4596 * at runtime. Changing other flags requires closing the
4597 * environment and re-opening it with the new flags.
4599 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4600 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
4601 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4603 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4604 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4608 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4610 int oflags, rc, len, excl = -1;
4611 char *lpath, *dpath;
4613 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4617 if (flags & MDB_NOSUBDIR) {
4618 rc = len + sizeof(LOCKSUFF) + len + 1;
4620 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4625 if (flags & MDB_NOSUBDIR) {
4626 dpath = lpath + len + sizeof(LOCKSUFF);
4627 sprintf(lpath, "%s" LOCKSUFF, path);
4628 strcpy(dpath, path);
4630 dpath = lpath + len + sizeof(LOCKNAME);
4631 sprintf(lpath, "%s" LOCKNAME, path);
4632 sprintf(dpath, "%s" DATANAME, path);
4636 flags |= env->me_flags;
4637 if (flags & MDB_RDONLY) {
4638 /* silently ignore WRITEMAP when we're only getting read access */
4639 flags &= ~MDB_WRITEMAP;
4641 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4642 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4645 env->me_flags = flags |= MDB_ENV_ACTIVE;
4649 env->me_path = strdup(path);
4650 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4651 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4652 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4653 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4658 /* For RDONLY, get lockfile after we know datafile exists */
4659 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4660 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4666 if (F_ISSET(flags, MDB_RDONLY)) {
4667 oflags = GENERIC_READ;
4668 len = OPEN_EXISTING;
4670 oflags = GENERIC_READ|GENERIC_WRITE;
4673 mode = FILE_ATTRIBUTE_NORMAL;
4674 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4675 NULL, len, mode, NULL);
4677 if (F_ISSET(flags, MDB_RDONLY))
4680 oflags = O_RDWR | O_CREAT;
4682 env->me_fd = open(dpath, oflags, mode);
4684 if (env->me_fd == INVALID_HANDLE_VALUE) {
4689 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4690 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4695 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4696 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4697 env->me_mfd = env->me_fd;
4699 /* Synchronous fd for meta writes. Needed even with
4700 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4703 len = OPEN_EXISTING;
4704 env->me_mfd = CreateFile(dpath, oflags,
4705 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4706 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4709 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4711 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4716 DPRINTF(("opened dbenv %p", (void *) env));
4718 rc = mdb_env_share_locks(env, &excl);
4722 if (!((flags & MDB_RDONLY) ||
4723 (env->me_pbuf = calloc(1, env->me_psize))))
4725 if (!(flags & MDB_RDONLY)) {
4727 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4728 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4729 txn = calloc(1, size);
4731 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4732 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4733 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4734 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4736 txn->mt_dbxs = env->me_dbxs;
4746 mdb_env_close0(env, excl);
4752 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4754 mdb_env_close0(MDB_env *env, int excl)
4758 if (!(env->me_flags & MDB_ENV_ACTIVE))
4761 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4763 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4764 free(env->me_dbxs[i].md_name.mv_data);
4769 free(env->me_dbiseqs);
4770 free(env->me_dbflags);
4772 free(env->me_dirty_list);
4774 mdb_midl_free(env->me_free_pgs);
4776 if (env->me_flags & MDB_ENV_TXKEY) {
4777 pthread_key_delete(env->me_txkey);
4779 /* Delete our key from the global list */
4780 for (i=0; i<mdb_tls_nkeys; i++)
4781 if (mdb_tls_keys[i] == env->me_txkey) {
4782 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4790 munmap(env->me_map, env->me_mapsize);
4792 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4793 (void) close(env->me_mfd);
4794 if (env->me_fd != INVALID_HANDLE_VALUE)
4795 (void) close(env->me_fd);
4797 MDB_PID_T pid = env->me_pid;
4798 /* Clearing readers is done in this function because
4799 * me_txkey with its destructor must be disabled first.
4801 * We skip the the reader mutex, so we touch only
4802 * data owned by this process (me_close_readers and
4803 * our readers), and clear each reader atomically.
4805 for (i = env->me_close_readers; --i >= 0; )
4806 if (env->me_txns->mti_readers[i].mr_pid == pid)
4807 env->me_txns->mti_readers[i].mr_pid = 0;
4809 if (env->me_rmutex) {
4810 CloseHandle(env->me_rmutex);
4811 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4813 /* Windows automatically destroys the mutexes when
4814 * the last handle closes.
4816 #elif defined(MDB_USE_SYSV_SEM)
4817 if (env->me_rmutex.semid != -1) {
4818 /* If we have the filelock: If we are the
4819 * only remaining user, clean up semaphores.
4822 mdb_env_excl_lock(env, &excl);
4824 semctl(env->me_rmutex.semid, 0, IPC_RMID);
4827 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4829 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4832 /* Unlock the lockfile. Windows would have unlocked it
4833 * after closing anyway, but not necessarily at once.
4835 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4838 (void) close(env->me_lfd);
4841 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4845 mdb_env_close(MDB_env *env)
4852 VGMEMP_DESTROY(env);
4853 while ((dp = env->me_dpages) != NULL) {
4854 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4855 env->me_dpages = dp->mp_next;
4859 mdb_env_close0(env, 0);
4863 /** Compare two items pointing at aligned size_t's */
4865 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4867 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4868 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4871 /** Compare two items pointing at aligned unsigned int's */
4873 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4875 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4876 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4879 /** Compare two items pointing at unsigned ints of unknown alignment.
4880 * Nodes and keys are guaranteed to be 2-byte aligned.
4883 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4885 #if BYTE_ORDER == LITTLE_ENDIAN
4886 unsigned short *u, *c;
4889 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4890 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4893 } while(!x && u > (unsigned short *)a->mv_data);
4896 unsigned short *u, *c, *end;
4899 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4900 u = (unsigned short *)a->mv_data;
4901 c = (unsigned short *)b->mv_data;
4904 } while(!x && u < end);
4909 /** Compare two items pointing at size_t's of unknown alignment. */
4910 #ifdef MISALIGNED_OK
4911 # define mdb_cmp_clong mdb_cmp_long
4913 # define mdb_cmp_clong mdb_cmp_cint
4916 /** Compare two items lexically */
4918 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4925 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4931 diff = memcmp(a->mv_data, b->mv_data, len);
4932 return diff ? diff : len_diff<0 ? -1 : len_diff;
4935 /** Compare two items in reverse byte order */
4937 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4939 const unsigned char *p1, *p2, *p1_lim;
4943 p1_lim = (const unsigned char *)a->mv_data;
4944 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4945 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4947 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4953 while (p1 > p1_lim) {
4954 diff = *--p1 - *--p2;
4958 return len_diff<0 ? -1 : len_diff;
4961 /** Search for key within a page, using binary search.
4962 * Returns the smallest entry larger or equal to the key.
4963 * If exactp is non-null, stores whether the found entry was an exact match
4964 * in *exactp (1 or 0).
4965 * Updates the cursor index with the index of the found entry.
4966 * If no entry larger or equal to the key is found, returns NULL.
4969 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4971 unsigned int i = 0, nkeys;
4974 MDB_page *mp = mc->mc_pg[mc->mc_top];
4975 MDB_node *node = NULL;
4980 nkeys = NUMKEYS(mp);
4982 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4983 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4986 low = IS_LEAF(mp) ? 0 : 1;
4988 cmp = mc->mc_dbx->md_cmp;
4990 /* Branch pages have no data, so if using integer keys,
4991 * alignment is guaranteed. Use faster mdb_cmp_int.
4993 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4994 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
5001 nodekey.mv_size = mc->mc_db->md_pad;
5002 node = NODEPTR(mp, 0); /* fake */
5003 while (low <= high) {
5004 i = (low + high) >> 1;
5005 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5006 rc = cmp(key, &nodekey);
5007 DPRINTF(("found leaf index %u [%s], rc = %i",
5008 i, DKEY(&nodekey), rc));
5017 while (low <= high) {
5018 i = (low + high) >> 1;
5020 node = NODEPTR(mp, i);
5021 nodekey.mv_size = NODEKSZ(node);
5022 nodekey.mv_data = NODEKEY(node);
5024 rc = cmp(key, &nodekey);
5027 DPRINTF(("found leaf index %u [%s], rc = %i",
5028 i, DKEY(&nodekey), rc));
5030 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5031 i, DKEY(&nodekey), NODEPGNO(node), rc));
5042 if (rc > 0) { /* Found entry is less than the key. */
5043 i++; /* Skip to get the smallest entry larger than key. */
5045 node = NODEPTR(mp, i);
5048 *exactp = (rc == 0 && nkeys > 0);
5049 /* store the key index */
5050 mc->mc_ki[mc->mc_top] = i;
5052 /* There is no entry larger or equal to the key. */
5055 /* nodeptr is fake for LEAF2 */
5061 mdb_cursor_adjust(MDB_cursor *mc, func)
5065 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5066 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5073 /** Pop a page off the top of the cursor's stack. */
5075 mdb_cursor_pop(MDB_cursor *mc)
5079 MDB_page *top = mc->mc_pg[mc->mc_top];
5085 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5086 DDBI(mc), (void *) mc));
5090 /** Push a page onto the top of the cursor's stack. */
5092 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5094 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5095 DDBI(mc), (void *) mc));
5097 if (mc->mc_snum >= CURSOR_STACK) {
5098 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5099 return MDB_CURSOR_FULL;
5102 mc->mc_top = mc->mc_snum++;
5103 mc->mc_pg[mc->mc_top] = mp;
5104 mc->mc_ki[mc->mc_top] = 0;
5109 /** Find the address of the page corresponding to a given page number.
5110 * @param[in] txn the transaction for this access.
5111 * @param[in] pgno the page number for the page to retrieve.
5112 * @param[out] ret address of a pointer where the page's address will be stored.
5113 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5114 * @return 0 on success, non-zero on failure.
5117 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5119 MDB_env *env = txn->mt_env;
5123 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
5127 MDB_ID2L dl = tx2->mt_u.dirty_list;
5129 /* Spilled pages were dirtied in this txn and flushed
5130 * because the dirty list got full. Bring this page
5131 * back in from the map (but don't unspill it here,
5132 * leave that unless page_touch happens again).
5134 if (tx2->mt_spill_pgs) {
5135 MDB_ID pn = pgno << 1;
5136 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5137 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5138 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5143 unsigned x = mdb_mid2l_search(dl, pgno);
5144 if (x <= dl[0].mid && dl[x].mid == pgno) {
5150 } while ((tx2 = tx2->mt_parent) != NULL);
5153 if (pgno < txn->mt_next_pgno) {
5155 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5157 DPRINTF(("page %"Z"u not found", pgno));
5158 txn->mt_flags |= MDB_TXN_ERROR;
5159 return MDB_PAGE_NOTFOUND;
5169 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5170 * The cursor is at the root page, set up the rest of it.
5173 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5175 MDB_page *mp = mc->mc_pg[mc->mc_top];
5179 while (IS_BRANCH(mp)) {
5183 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5184 mdb_cassert(mc, NUMKEYS(mp) > 1);
5185 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5187 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5189 if (flags & MDB_PS_LAST)
5190 i = NUMKEYS(mp) - 1;
5193 node = mdb_node_search(mc, key, &exact);
5195 i = NUMKEYS(mp) - 1;
5197 i = mc->mc_ki[mc->mc_top];
5199 mdb_cassert(mc, i > 0);
5203 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5206 mdb_cassert(mc, i < NUMKEYS(mp));
5207 node = NODEPTR(mp, i);
5209 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5212 mc->mc_ki[mc->mc_top] = i;
5213 if ((rc = mdb_cursor_push(mc, mp)))
5216 if (flags & MDB_PS_MODIFY) {
5217 if ((rc = mdb_page_touch(mc)) != 0)
5219 mp = mc->mc_pg[mc->mc_top];
5224 DPRINTF(("internal error, index points to a %02X page!?",
5226 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5227 return MDB_CORRUPTED;
5230 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5231 key ? DKEY(key) : "null"));
5232 mc->mc_flags |= C_INITIALIZED;
5233 mc->mc_flags &= ~C_EOF;
5238 /** Search for the lowest key under the current branch page.
5239 * This just bypasses a NUMKEYS check in the current page
5240 * before calling mdb_page_search_root(), because the callers
5241 * are all in situations where the current page is known to
5245 mdb_page_search_lowest(MDB_cursor *mc)
5247 MDB_page *mp = mc->mc_pg[mc->mc_top];
5248 MDB_node *node = NODEPTR(mp, 0);
5251 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5254 mc->mc_ki[mc->mc_top] = 0;
5255 if ((rc = mdb_cursor_push(mc, mp)))
5257 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5260 /** Search for the page a given key should be in.
5261 * Push it and its parent pages on the cursor stack.
5262 * @param[in,out] mc the cursor for this operation.
5263 * @param[in] key the key to search for, or NULL for first/last page.
5264 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5265 * are touched (updated with new page numbers).
5266 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5267 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5268 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5269 * @return 0 on success, non-zero on failure.
5272 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5277 /* Make sure the txn is still viable, then find the root from
5278 * the txn's db table and set it as the root of the cursor's stack.
5280 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5281 DPUTS("transaction has failed, must abort");
5284 /* Make sure we're using an up-to-date root */
5285 if (*mc->mc_dbflag & DB_STALE) {
5287 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5289 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5290 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5297 MDB_node *leaf = mdb_node_search(&mc2,
5298 &mc->mc_dbx->md_name, &exact);
5300 return MDB_NOTFOUND;
5301 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5304 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5306 /* The txn may not know this DBI, or another process may
5307 * have dropped and recreated the DB with other flags.
5309 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5310 return MDB_INCOMPATIBLE;
5311 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5313 *mc->mc_dbflag &= ~DB_STALE;
5315 root = mc->mc_db->md_root;
5317 if (root == P_INVALID) { /* Tree is empty. */
5318 DPUTS("tree is empty");
5319 return MDB_NOTFOUND;
5323 mdb_cassert(mc, root > 1);
5324 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5325 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5331 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5332 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5334 if (flags & MDB_PS_MODIFY) {
5335 if ((rc = mdb_page_touch(mc)))
5339 if (flags & MDB_PS_ROOTONLY)
5342 return mdb_page_search_root(mc, key, flags);
5346 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5348 MDB_txn *txn = mc->mc_txn;
5349 pgno_t pg = mp->mp_pgno;
5350 unsigned x = 0, ovpages = mp->mp_pages;
5351 MDB_env *env = txn->mt_env;
5352 MDB_IDL sl = txn->mt_spill_pgs;
5353 MDB_ID pn = pg << 1;
5356 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5357 /* If the page is dirty or on the spill list we just acquired it,
5358 * so we should give it back to our current free list, if any.
5359 * Otherwise put it onto the list of pages we freed in this txn.
5361 * Won't create me_pghead: me_pglast must be inited along with it.
5362 * Unsupported in nested txns: They would need to hide the page
5363 * range in ancestor txns' dirty and spilled lists.
5365 if (env->me_pghead &&
5367 ((mp->mp_flags & P_DIRTY) ||
5368 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5372 MDB_ID2 *dl, ix, iy;
5373 rc = mdb_midl_need(&env->me_pghead, ovpages);
5376 if (!(mp->mp_flags & P_DIRTY)) {
5377 /* This page is no longer spilled */
5384 /* Remove from dirty list */
5385 dl = txn->mt_u.dirty_list;
5387 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5393 mdb_cassert(mc, x > 1);
5395 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5396 txn->mt_flags |= MDB_TXN_ERROR;
5397 return MDB_CORRUPTED;
5400 if (!(env->me_flags & MDB_WRITEMAP))
5401 mdb_dpage_free(env, mp);
5403 /* Insert in me_pghead */
5404 mop = env->me_pghead;
5405 j = mop[0] + ovpages;
5406 for (i = mop[0]; i && mop[i] < pg; i--)
5412 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5416 mc->mc_db->md_overflow_pages -= ovpages;
5420 /** Return the data associated with a given node.
5421 * @param[in] txn The transaction for this operation.
5422 * @param[in] leaf The node being read.
5423 * @param[out] data Updated to point to the node's data.
5424 * @return 0 on success, non-zero on failure.
5427 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5429 MDB_page *omp; /* overflow page */
5433 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5434 data->mv_size = NODEDSZ(leaf);
5435 data->mv_data = NODEDATA(leaf);
5439 /* Read overflow data.
5441 data->mv_size = NODEDSZ(leaf);
5442 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5443 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5444 DPRINTF(("read overflow page %"Z"u failed", pgno));
5447 data->mv_data = METADATA(omp);
5453 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5454 MDB_val *key, MDB_val *data)
5461 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5463 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5466 if (txn->mt_flags & MDB_TXN_ERROR)
5469 mdb_cursor_init(&mc, txn, dbi, &mx);
5470 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5473 /** Find a sibling for a page.
5474 * Replaces the page at the top of the cursor's stack with the
5475 * specified sibling, if one exists.
5476 * @param[in] mc The cursor for this operation.
5477 * @param[in] move_right Non-zero if the right sibling is requested,
5478 * otherwise the left sibling.
5479 * @return 0 on success, non-zero on failure.
5482 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5488 if (mc->mc_snum < 2) {
5489 return MDB_NOTFOUND; /* root has no siblings */
5493 DPRINTF(("parent page is page %"Z"u, index %u",
5494 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5496 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5497 : (mc->mc_ki[mc->mc_top] == 0)) {
5498 DPRINTF(("no more keys left, moving to %s sibling",
5499 move_right ? "right" : "left"));
5500 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5501 /* undo cursor_pop before returning */
5508 mc->mc_ki[mc->mc_top]++;
5510 mc->mc_ki[mc->mc_top]--;
5511 DPRINTF(("just moving to %s index key %u",
5512 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5514 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5516 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5517 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5518 /* mc will be inconsistent if caller does mc_snum++ as above */
5519 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5523 mdb_cursor_push(mc, mp);
5525 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5530 /** Move the cursor to the next data item. */
5532 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5538 if (mc->mc_flags & C_EOF) {
5539 return MDB_NOTFOUND;
5542 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5544 mp = mc->mc_pg[mc->mc_top];
5546 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5547 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5548 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5549 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5550 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5551 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5552 if (rc == MDB_SUCCESS)
5553 MDB_GET_KEY(leaf, key);
5558 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5559 if (op == MDB_NEXT_DUP)
5560 return MDB_NOTFOUND;
5564 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5565 mdb_dbg_pgno(mp), (void *) mc));
5566 if (mc->mc_flags & C_DEL)
5569 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5570 DPUTS("=====> move to next sibling page");
5571 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5572 mc->mc_flags |= C_EOF;
5575 mp = mc->mc_pg[mc->mc_top];
5576 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5578 mc->mc_ki[mc->mc_top]++;
5581 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5582 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5585 key->mv_size = mc->mc_db->md_pad;
5586 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5590 mdb_cassert(mc, IS_LEAF(mp));
5591 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5593 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5594 mdb_xcursor_init1(mc, leaf);
5597 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5600 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5601 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5602 if (rc != MDB_SUCCESS)
5607 MDB_GET_KEY(leaf, key);
5611 /** Move the cursor to the previous data item. */
5613 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5619 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5621 mp = mc->mc_pg[mc->mc_top];
5623 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5624 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5625 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5626 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5627 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5628 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5629 if (rc == MDB_SUCCESS) {
5630 MDB_GET_KEY(leaf, key);
5631 mc->mc_flags &= ~C_EOF;
5637 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5638 if (op == MDB_PREV_DUP)
5639 return MDB_NOTFOUND;
5643 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5644 mdb_dbg_pgno(mp), (void *) mc));
5646 if (mc->mc_ki[mc->mc_top] == 0) {
5647 DPUTS("=====> move to prev sibling page");
5648 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5651 mp = mc->mc_pg[mc->mc_top];
5652 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5653 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5655 mc->mc_ki[mc->mc_top]--;
5657 mc->mc_flags &= ~C_EOF;
5659 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5660 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5663 key->mv_size = mc->mc_db->md_pad;
5664 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5668 mdb_cassert(mc, IS_LEAF(mp));
5669 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5671 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5672 mdb_xcursor_init1(mc, leaf);
5675 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5678 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5679 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5680 if (rc != MDB_SUCCESS)
5685 MDB_GET_KEY(leaf, key);
5689 /** Set the cursor on a specific data item. */
5691 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5692 MDB_cursor_op op, int *exactp)
5696 MDB_node *leaf = NULL;
5699 if (key->mv_size == 0)
5700 return MDB_BAD_VALSIZE;
5703 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5705 /* See if we're already on the right page */
5706 if (mc->mc_flags & C_INITIALIZED) {
5709 mp = mc->mc_pg[mc->mc_top];
5711 mc->mc_ki[mc->mc_top] = 0;
5712 return MDB_NOTFOUND;
5714 if (mp->mp_flags & P_LEAF2) {
5715 nodekey.mv_size = mc->mc_db->md_pad;
5716 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5718 leaf = NODEPTR(mp, 0);
5719 MDB_GET_KEY2(leaf, nodekey);
5721 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5723 /* Probably happens rarely, but first node on the page
5724 * was the one we wanted.
5726 mc->mc_ki[mc->mc_top] = 0;
5733 unsigned int nkeys = NUMKEYS(mp);
5735 if (mp->mp_flags & P_LEAF2) {
5736 nodekey.mv_data = LEAF2KEY(mp,
5737 nkeys-1, nodekey.mv_size);
5739 leaf = NODEPTR(mp, nkeys-1);
5740 MDB_GET_KEY2(leaf, nodekey);
5742 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5744 /* last node was the one we wanted */
5745 mc->mc_ki[mc->mc_top] = nkeys-1;
5751 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5752 /* This is definitely the right page, skip search_page */
5753 if (mp->mp_flags & P_LEAF2) {
5754 nodekey.mv_data = LEAF2KEY(mp,
5755 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5757 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5758 MDB_GET_KEY2(leaf, nodekey);
5760 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5762 /* current node was the one we wanted */
5772 /* If any parents have right-sibs, search.
5773 * Otherwise, there's nothing further.
5775 for (i=0; i<mc->mc_top; i++)
5777 NUMKEYS(mc->mc_pg[i])-1)
5779 if (i == mc->mc_top) {
5780 /* There are no other pages */
5781 mc->mc_ki[mc->mc_top] = nkeys;
5782 return MDB_NOTFOUND;
5786 /* There are no other pages */
5787 mc->mc_ki[mc->mc_top] = 0;
5788 if (op == MDB_SET_RANGE && !exactp) {
5792 return MDB_NOTFOUND;
5796 rc = mdb_page_search(mc, key, 0);
5797 if (rc != MDB_SUCCESS)
5800 mp = mc->mc_pg[mc->mc_top];
5801 mdb_cassert(mc, IS_LEAF(mp));
5804 leaf = mdb_node_search(mc, key, exactp);
5805 if (exactp != NULL && !*exactp) {
5806 /* MDB_SET specified and not an exact match. */
5807 return MDB_NOTFOUND;
5811 DPUTS("===> inexact leaf not found, goto sibling");
5812 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5813 return rc; /* no entries matched */
5814 mp = mc->mc_pg[mc->mc_top];
5815 mdb_cassert(mc, IS_LEAF(mp));
5816 leaf = NODEPTR(mp, 0);
5820 mc->mc_flags |= C_INITIALIZED;
5821 mc->mc_flags &= ~C_EOF;
5824 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5825 key->mv_size = mc->mc_db->md_pad;
5826 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5831 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5832 mdb_xcursor_init1(mc, leaf);
5835 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5836 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5837 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5840 if (op == MDB_GET_BOTH) {
5846 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5847 if (rc != MDB_SUCCESS)
5850 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5852 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5854 rc = mc->mc_dbx->md_dcmp(data, &d2);
5856 if (op == MDB_GET_BOTH || rc > 0)
5857 return MDB_NOTFOUND;
5864 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5865 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5870 /* The key already matches in all other cases */
5871 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5872 MDB_GET_KEY(leaf, key);
5873 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5878 /** Move the cursor to the first item in the database. */
5880 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5886 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5888 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5889 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5890 if (rc != MDB_SUCCESS)
5893 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5895 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5896 mc->mc_flags |= C_INITIALIZED;
5897 mc->mc_flags &= ~C_EOF;
5899 mc->mc_ki[mc->mc_top] = 0;
5901 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5902 key->mv_size = mc->mc_db->md_pad;
5903 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5908 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5909 mdb_xcursor_init1(mc, leaf);
5910 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5914 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5918 MDB_GET_KEY(leaf, key);
5922 /** Move the cursor to the last item in the database. */
5924 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5930 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5932 if (!(mc->mc_flags & C_EOF)) {
5934 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5935 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5936 if (rc != MDB_SUCCESS)
5939 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5942 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5943 mc->mc_flags |= C_INITIALIZED|C_EOF;
5944 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5946 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5947 key->mv_size = mc->mc_db->md_pad;
5948 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5953 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5954 mdb_xcursor_init1(mc, leaf);
5955 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5959 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5964 MDB_GET_KEY(leaf, key);
5969 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5974 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5979 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5983 case MDB_GET_CURRENT:
5984 if (!(mc->mc_flags & C_INITIALIZED)) {
5987 MDB_page *mp = mc->mc_pg[mc->mc_top];
5988 int nkeys = NUMKEYS(mp);
5989 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5990 mc->mc_ki[mc->mc_top] = nkeys;
5996 key->mv_size = mc->mc_db->md_pad;
5997 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5999 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6000 MDB_GET_KEY(leaf, key);
6002 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6003 if (mc->mc_flags & C_DEL)
6004 mdb_xcursor_init1(mc, leaf);
6005 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6007 rc = mdb_node_read(mc->mc_txn, leaf, data);
6014 case MDB_GET_BOTH_RANGE:
6019 if (mc->mc_xcursor == NULL) {
6020 rc = MDB_INCOMPATIBLE;
6030 rc = mdb_cursor_set(mc, key, data, op,
6031 op == MDB_SET_RANGE ? NULL : &exact);
6034 case MDB_GET_MULTIPLE:
6035 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6039 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6040 rc = MDB_INCOMPATIBLE;
6044 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6045 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6048 case MDB_NEXT_MULTIPLE:
6053 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6054 rc = MDB_INCOMPATIBLE;
6057 if (!(mc->mc_flags & C_INITIALIZED))
6058 rc = mdb_cursor_first(mc, key, data);
6060 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6061 if (rc == MDB_SUCCESS) {
6062 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6065 mx = &mc->mc_xcursor->mx_cursor;
6066 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6068 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6069 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6077 case MDB_NEXT_NODUP:
6078 if (!(mc->mc_flags & C_INITIALIZED))
6079 rc = mdb_cursor_first(mc, key, data);
6081 rc = mdb_cursor_next(mc, key, data, op);
6085 case MDB_PREV_NODUP:
6086 if (!(mc->mc_flags & C_INITIALIZED)) {
6087 rc = mdb_cursor_last(mc, key, data);
6090 mc->mc_flags |= C_INITIALIZED;
6091 mc->mc_ki[mc->mc_top]++;
6093 rc = mdb_cursor_prev(mc, key, data, op);
6096 rc = mdb_cursor_first(mc, key, data);
6099 mfunc = mdb_cursor_first;
6101 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6105 if (mc->mc_xcursor == NULL) {
6106 rc = MDB_INCOMPATIBLE;
6110 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6111 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6112 MDB_GET_KEY(leaf, key);
6113 rc = mdb_node_read(mc->mc_txn, leaf, data);
6117 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6121 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6124 rc = mdb_cursor_last(mc, key, data);
6127 mfunc = mdb_cursor_last;
6130 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6135 if (mc->mc_flags & C_DEL)
6136 mc->mc_flags ^= C_DEL;
6141 /** Touch all the pages in the cursor stack. Set mc_top.
6142 * Makes sure all the pages are writable, before attempting a write operation.
6143 * @param[in] mc The cursor to operate on.
6146 mdb_cursor_touch(MDB_cursor *mc)
6148 int rc = MDB_SUCCESS;
6150 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6153 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6155 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6156 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6159 *mc->mc_dbflag |= DB_DIRTY;
6164 rc = mdb_page_touch(mc);
6165 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6166 mc->mc_top = mc->mc_snum-1;
6171 /** Do not spill pages to disk if txn is getting full, may fail instead */
6172 #define MDB_NOSPILL 0x8000
6175 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6179 MDB_node *leaf = NULL;
6182 MDB_val xdata, *rdata, dkey, olddata;
6184 int do_sub = 0, insert_key, insert_data;
6185 unsigned int mcount = 0, dcount = 0, nospill;
6188 unsigned int nflags;
6191 if (mc == NULL || key == NULL)
6194 env = mc->mc_txn->mt_env;
6196 /* Check this first so counter will always be zero on any
6199 if (flags & MDB_MULTIPLE) {
6200 dcount = data[1].mv_size;
6201 data[1].mv_size = 0;
6202 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6203 return MDB_INCOMPATIBLE;
6206 nospill = flags & MDB_NOSPILL;
6207 flags &= ~MDB_NOSPILL;
6209 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6210 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6212 if (key->mv_size-1 >= ENV_MAXKEY(env))
6213 return MDB_BAD_VALSIZE;
6215 #if SIZE_MAX > MAXDATASIZE
6216 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6217 return MDB_BAD_VALSIZE;
6219 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6220 return MDB_BAD_VALSIZE;
6223 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6224 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6228 if (flags == MDB_CURRENT) {
6229 if (!(mc->mc_flags & C_INITIALIZED))
6232 } else if (mc->mc_db->md_root == P_INVALID) {
6233 /* new database, cursor has nothing to point to */
6236 mc->mc_flags &= ~C_INITIALIZED;
6241 if (flags & MDB_APPEND) {
6243 rc = mdb_cursor_last(mc, &k2, &d2);
6245 rc = mc->mc_dbx->md_cmp(key, &k2);
6248 mc->mc_ki[mc->mc_top]++;
6250 /* new key is <= last key */
6255 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6257 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6258 DPRINTF(("duplicate key [%s]", DKEY(key)));
6260 return MDB_KEYEXIST;
6262 if (rc && rc != MDB_NOTFOUND)
6266 if (mc->mc_flags & C_DEL)
6267 mc->mc_flags ^= C_DEL;
6269 /* Cursor is positioned, check for room in the dirty list */
6271 if (flags & MDB_MULTIPLE) {
6273 xdata.mv_size = data->mv_size * dcount;
6277 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6281 if (rc == MDB_NO_ROOT) {
6283 /* new database, write a root leaf page */
6284 DPUTS("allocating new root leaf page");
6285 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6288 mdb_cursor_push(mc, np);
6289 mc->mc_db->md_root = np->mp_pgno;
6290 mc->mc_db->md_depth++;
6291 *mc->mc_dbflag |= DB_DIRTY;
6292 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6294 np->mp_flags |= P_LEAF2;
6295 mc->mc_flags |= C_INITIALIZED;
6297 /* make sure all cursor pages are writable */
6298 rc2 = mdb_cursor_touch(mc);
6303 insert_key = insert_data = rc;
6305 /* The key does not exist */
6306 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6307 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6308 LEAFSIZE(key, data) > env->me_nodemax)
6310 /* Too big for a node, insert in sub-DB. Set up an empty
6311 * "old sub-page" for prep_subDB to expand to a full page.
6313 fp_flags = P_LEAF|P_DIRTY;
6315 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6316 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6317 olddata.mv_size = PAGEHDRSZ;
6321 /* there's only a key anyway, so this is a no-op */
6322 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6324 unsigned int ksize = mc->mc_db->md_pad;
6325 if (key->mv_size != ksize)
6326 return MDB_BAD_VALSIZE;
6327 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6328 memcpy(ptr, key->mv_data, ksize);
6330 /* if overwriting slot 0 of leaf, need to
6331 * update branch key if there is a parent page
6333 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6334 unsigned short top = mc->mc_top;
6336 /* slot 0 is always an empty key, find real slot */
6337 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6339 if (mc->mc_ki[mc->mc_top])
6340 rc2 = mdb_update_key(mc, key);
6351 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6352 olddata.mv_size = NODEDSZ(leaf);
6353 olddata.mv_data = NODEDATA(leaf);
6356 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6357 /* Prepare (sub-)page/sub-DB to accept the new item,
6358 * if needed. fp: old sub-page or a header faking
6359 * it. mp: new (sub-)page. offset: growth in page
6360 * size. xdata: node data with new page or DB.
6362 unsigned i, offset = 0;
6363 mp = fp = xdata.mv_data = env->me_pbuf;
6364 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6366 /* Was a single item before, must convert now */
6367 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6368 /* Just overwrite the current item */
6369 if (flags == MDB_CURRENT)
6372 #if UINT_MAX < SIZE_MAX
6373 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6374 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6376 /* does data match? */
6377 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6378 if (flags & MDB_NODUPDATA)
6379 return MDB_KEYEXIST;
6384 /* Back up original data item */
6385 dkey.mv_size = olddata.mv_size;
6386 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6388 /* Make sub-page header for the dup items, with dummy body */
6389 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6390 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6391 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6392 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6393 fp->mp_flags |= P_LEAF2;
6394 fp->mp_pad = data->mv_size;
6395 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6397 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6398 (dkey.mv_size & 1) + (data->mv_size & 1);
6400 fp->mp_upper = xdata.mv_size - PAGEBASE;
6401 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6402 } else if (leaf->mn_flags & F_SUBDATA) {
6403 /* Data is on sub-DB, just store it */
6404 flags |= F_DUPDATA|F_SUBDATA;
6407 /* Data is on sub-page */
6408 fp = olddata.mv_data;
6411 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6412 offset = EVEN(NODESIZE + sizeof(indx_t) +
6416 offset = fp->mp_pad;
6417 if (SIZELEFT(fp) < offset) {
6418 offset *= 4; /* space for 4 more */
6421 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6423 fp->mp_flags |= P_DIRTY;
6424 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6425 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6429 xdata.mv_size = olddata.mv_size + offset;
6432 fp_flags = fp->mp_flags;
6433 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6434 /* Too big for a sub-page, convert to sub-DB */
6435 fp_flags &= ~P_SUBP;
6437 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6438 fp_flags |= P_LEAF2;
6439 dummy.md_pad = fp->mp_pad;
6440 dummy.md_flags = MDB_DUPFIXED;
6441 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6442 dummy.md_flags |= MDB_INTEGERKEY;
6448 dummy.md_branch_pages = 0;
6449 dummy.md_leaf_pages = 1;
6450 dummy.md_overflow_pages = 0;
6451 dummy.md_entries = NUMKEYS(fp);
6452 xdata.mv_size = sizeof(MDB_db);
6453 xdata.mv_data = &dummy;
6454 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6456 offset = env->me_psize - olddata.mv_size;
6457 flags |= F_DUPDATA|F_SUBDATA;
6458 dummy.md_root = mp->mp_pgno;
6461 mp->mp_flags = fp_flags | P_DIRTY;
6462 mp->mp_pad = fp->mp_pad;
6463 mp->mp_lower = fp->mp_lower;
6464 mp->mp_upper = fp->mp_upper + offset;
6465 if (fp_flags & P_LEAF2) {
6466 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6468 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6469 olddata.mv_size - fp->mp_upper - PAGEBASE);
6470 for (i=0; i<NUMKEYS(fp); i++)
6471 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6479 mdb_node_del(mc, 0);
6483 /* overflow page overwrites need special handling */
6484 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6487 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6489 memcpy(&pg, olddata.mv_data, sizeof(pg));
6490 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6492 ovpages = omp->mp_pages;
6494 /* Is the ov page large enough? */
6495 if (ovpages >= dpages) {
6496 if (!(omp->mp_flags & P_DIRTY) &&
6497 (level || (env->me_flags & MDB_WRITEMAP)))
6499 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6502 level = 0; /* dirty in this txn or clean */
6505 if (omp->mp_flags & P_DIRTY) {
6506 /* yes, overwrite it. Note in this case we don't
6507 * bother to try shrinking the page if the new data
6508 * is smaller than the overflow threshold.
6511 /* It is writable only in a parent txn */
6512 size_t sz = (size_t) env->me_psize * ovpages, off;
6513 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6519 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6520 mdb_cassert(mc, rc2 == 0);
6521 if (!(flags & MDB_RESERVE)) {
6522 /* Copy end of page, adjusting alignment so
6523 * compiler may copy words instead of bytes.
6525 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6526 memcpy((size_t *)((char *)np + off),
6527 (size_t *)((char *)omp + off), sz - off);
6530 memcpy(np, omp, sz); /* Copy beginning of page */
6533 SETDSZ(leaf, data->mv_size);
6534 if (F_ISSET(flags, MDB_RESERVE))
6535 data->mv_data = METADATA(omp);
6537 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6541 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6543 } else if (data->mv_size == olddata.mv_size) {
6544 /* same size, just replace it. Note that we could
6545 * also reuse this node if the new data is smaller,
6546 * but instead we opt to shrink the node in that case.
6548 if (F_ISSET(flags, MDB_RESERVE))
6549 data->mv_data = olddata.mv_data;
6550 else if (!(mc->mc_flags & C_SUB))
6551 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6553 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6558 mdb_node_del(mc, 0);
6564 nflags = flags & NODE_ADD_FLAGS;
6565 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6566 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6567 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6568 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6570 nflags |= MDB_SPLIT_REPLACE;
6571 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6573 /* There is room already in this leaf page. */
6574 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6575 if (rc == 0 && insert_key) {
6576 /* Adjust other cursors pointing to mp */
6577 MDB_cursor *m2, *m3;
6578 MDB_dbi dbi = mc->mc_dbi;
6579 unsigned i = mc->mc_top;
6580 MDB_page *mp = mc->mc_pg[i];
6582 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6583 if (mc->mc_flags & C_SUB)
6584 m3 = &m2->mc_xcursor->mx_cursor;
6587 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6588 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6595 if (rc == MDB_SUCCESS) {
6596 /* Now store the actual data in the child DB. Note that we're
6597 * storing the user data in the keys field, so there are strict
6598 * size limits on dupdata. The actual data fields of the child
6599 * DB are all zero size.
6607 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6608 if (flags & MDB_CURRENT) {
6609 xflags = MDB_CURRENT|MDB_NOSPILL;
6611 mdb_xcursor_init1(mc, leaf);
6612 xflags = (flags & MDB_NODUPDATA) ?
6613 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6615 /* converted, write the original data first */
6617 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6621 /* Adjust other cursors pointing to mp */
6623 unsigned i = mc->mc_top;
6624 MDB_page *mp = mc->mc_pg[i];
6626 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6627 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6628 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6629 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6630 mdb_xcursor_init1(m2, leaf);
6634 /* we've done our job */
6637 ecount = mc->mc_xcursor->mx_db.md_entries;
6638 if (flags & MDB_APPENDDUP)
6639 xflags |= MDB_APPEND;
6640 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6641 if (flags & F_SUBDATA) {
6642 void *db = NODEDATA(leaf);
6643 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6645 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6647 /* Increment count unless we just replaced an existing item. */
6649 mc->mc_db->md_entries++;
6651 /* Invalidate txn if we created an empty sub-DB */
6654 /* If we succeeded and the key didn't exist before,
6655 * make sure the cursor is marked valid.
6657 mc->mc_flags |= C_INITIALIZED;
6659 if (flags & MDB_MULTIPLE) {
6662 /* let caller know how many succeeded, if any */
6663 data[1].mv_size = mcount;
6664 if (mcount < dcount) {
6665 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6666 insert_key = insert_data = 0;
6673 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6676 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6681 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6687 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6688 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6690 if (!(mc->mc_flags & C_INITIALIZED))
6693 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6694 return MDB_NOTFOUND;
6696 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6699 rc = mdb_cursor_touch(mc);
6703 mp = mc->mc_pg[mc->mc_top];
6706 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6708 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6709 if (flags & MDB_NODUPDATA) {
6710 /* mdb_cursor_del0() will subtract the final entry */
6711 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6713 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6714 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6716 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6719 /* If sub-DB still has entries, we're done */
6720 if (mc->mc_xcursor->mx_db.md_entries) {
6721 if (leaf->mn_flags & F_SUBDATA) {
6722 /* update subDB info */
6723 void *db = NODEDATA(leaf);
6724 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6727 /* shrink fake page */
6728 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6729 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6730 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6731 /* fix other sub-DB cursors pointed at this fake page */
6732 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6733 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6734 if (m2->mc_pg[mc->mc_top] == mp &&
6735 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6736 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6739 mc->mc_db->md_entries--;
6740 mc->mc_flags |= C_DEL;
6743 /* otherwise fall thru and delete the sub-DB */
6746 if (leaf->mn_flags & F_SUBDATA) {
6747 /* add all the child DB's pages to the free list */
6748 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6754 /* add overflow pages to free list */
6755 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6759 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6760 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6761 (rc = mdb_ovpage_free(mc, omp)))
6766 return mdb_cursor_del0(mc);
6769 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6773 /** Allocate and initialize new pages for a database.
6774 * @param[in] mc a cursor on the database being added to.
6775 * @param[in] flags flags defining what type of page is being allocated.
6776 * @param[in] num the number of pages to allocate. This is usually 1,
6777 * unless allocating overflow pages for a large record.
6778 * @param[out] mp Address of a page, or NULL on failure.
6779 * @return 0 on success, non-zero on failure.
6782 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6787 if ((rc = mdb_page_alloc(mc, num, &np)))
6789 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6790 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6791 np->mp_flags = flags | P_DIRTY;
6792 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6793 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6796 mc->mc_db->md_branch_pages++;
6797 else if (IS_LEAF(np))
6798 mc->mc_db->md_leaf_pages++;
6799 else if (IS_OVERFLOW(np)) {
6800 mc->mc_db->md_overflow_pages += num;
6808 /** Calculate the size of a leaf node.
6809 * The size depends on the environment's page size; if a data item
6810 * is too large it will be put onto an overflow page and the node
6811 * size will only include the key and not the data. Sizes are always
6812 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6813 * of the #MDB_node headers.
6814 * @param[in] env The environment handle.
6815 * @param[in] key The key for the node.
6816 * @param[in] data The data for the node.
6817 * @return The number of bytes needed to store the node.
6820 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6824 sz = LEAFSIZE(key, data);
6825 if (sz > env->me_nodemax) {
6826 /* put on overflow page */
6827 sz -= data->mv_size - sizeof(pgno_t);
6830 return EVEN(sz + sizeof(indx_t));
6833 /** Calculate the size of a branch node.
6834 * The size should depend on the environment's page size but since
6835 * we currently don't support spilling large keys onto overflow
6836 * pages, it's simply the size of the #MDB_node header plus the
6837 * size of the key. Sizes are always rounded up to an even number
6838 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6839 * @param[in] env The environment handle.
6840 * @param[in] key The key for the node.
6841 * @return The number of bytes needed to store the node.
6844 mdb_branch_size(MDB_env *env, MDB_val *key)
6849 if (sz > env->me_nodemax) {
6850 /* put on overflow page */
6851 /* not implemented */
6852 /* sz -= key->size - sizeof(pgno_t); */
6855 return sz + sizeof(indx_t);
6858 /** Add a node to the page pointed to by the cursor.
6859 * @param[in] mc The cursor for this operation.
6860 * @param[in] indx The index on the page where the new node should be added.
6861 * @param[in] key The key for the new node.
6862 * @param[in] data The data for the new node, if any.
6863 * @param[in] pgno The page number, if adding a branch node.
6864 * @param[in] flags Flags for the node.
6865 * @return 0 on success, non-zero on failure. Possible errors are:
6867 * <li>ENOMEM - failed to allocate overflow pages for the node.
6868 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6869 * should never happen since all callers already calculate the
6870 * page's free space before calling this function.
6874 mdb_node_add(MDB_cursor *mc, indx_t indx,
6875 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6878 size_t node_size = NODESIZE;
6882 MDB_page *mp = mc->mc_pg[mc->mc_top];
6883 MDB_page *ofp = NULL; /* overflow page */
6886 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6888 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6889 IS_LEAF(mp) ? "leaf" : "branch",
6890 IS_SUBP(mp) ? "sub-" : "",
6891 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6892 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6895 /* Move higher keys up one slot. */
6896 int ksize = mc->mc_db->md_pad, dif;
6897 char *ptr = LEAF2KEY(mp, indx, ksize);
6898 dif = NUMKEYS(mp) - indx;
6900 memmove(ptr+ksize, ptr, dif*ksize);
6901 /* insert new key */
6902 memcpy(ptr, key->mv_data, ksize);
6904 /* Just using these for counting */
6905 mp->mp_lower += sizeof(indx_t);
6906 mp->mp_upper -= ksize - sizeof(indx_t);
6910 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6912 node_size += key->mv_size;
6914 mdb_cassert(mc, data);
6915 if (F_ISSET(flags, F_BIGDATA)) {
6916 /* Data already on overflow page. */
6917 node_size += sizeof(pgno_t);
6918 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6919 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6921 /* Put data on overflow page. */
6922 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6923 data->mv_size, node_size+data->mv_size));
6924 node_size = EVEN(node_size + sizeof(pgno_t));
6925 if ((ssize_t)node_size > room)
6927 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6929 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6933 node_size += data->mv_size;
6936 node_size = EVEN(node_size);
6937 if ((ssize_t)node_size > room)
6941 /* Move higher pointers up one slot. */
6942 for (i = NUMKEYS(mp); i > indx; i--)
6943 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6945 /* Adjust free space offsets. */
6946 ofs = mp->mp_upper - node_size;
6947 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6948 mp->mp_ptrs[indx] = ofs;
6950 mp->mp_lower += sizeof(indx_t);
6952 /* Write the node data. */
6953 node = NODEPTR(mp, indx);
6954 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6955 node->mn_flags = flags;
6957 SETDSZ(node,data->mv_size);
6962 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6965 mdb_cassert(mc, key);
6967 if (F_ISSET(flags, F_BIGDATA))
6968 memcpy(node->mn_data + key->mv_size, data->mv_data,
6970 else if (F_ISSET(flags, MDB_RESERVE))
6971 data->mv_data = node->mn_data + key->mv_size;
6973 memcpy(node->mn_data + key->mv_size, data->mv_data,
6976 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6978 if (F_ISSET(flags, MDB_RESERVE))
6979 data->mv_data = METADATA(ofp);
6981 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6988 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6989 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6990 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6991 DPRINTF(("node size = %"Z"u", node_size));
6992 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6993 return MDB_PAGE_FULL;
6996 /** Delete the specified node from a page.
6997 * @param[in] mc Cursor pointing to the node to delete.
6998 * @param[in] ksize The size of a node. Only used if the page is
6999 * part of a #MDB_DUPFIXED database.
7002 mdb_node_del(MDB_cursor *mc, int ksize)
7004 MDB_page *mp = mc->mc_pg[mc->mc_top];
7005 indx_t indx = mc->mc_ki[mc->mc_top];
7007 indx_t i, j, numkeys, ptr;
7011 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7012 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7013 numkeys = NUMKEYS(mp);
7014 mdb_cassert(mc, indx < numkeys);
7017 int x = numkeys - 1 - indx;
7018 base = LEAF2KEY(mp, indx, ksize);
7020 memmove(base, base + ksize, x * ksize);
7021 mp->mp_lower -= sizeof(indx_t);
7022 mp->mp_upper += ksize - sizeof(indx_t);
7026 node = NODEPTR(mp, indx);
7027 sz = NODESIZE + node->mn_ksize;
7029 if (F_ISSET(node->mn_flags, F_BIGDATA))
7030 sz += sizeof(pgno_t);
7032 sz += NODEDSZ(node);
7036 ptr = mp->mp_ptrs[indx];
7037 for (i = j = 0; i < numkeys; i++) {
7039 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7040 if (mp->mp_ptrs[i] < ptr)
7041 mp->mp_ptrs[j] += sz;
7046 base = (char *)mp + mp->mp_upper + PAGEBASE;
7047 memmove(base + sz, base, ptr - mp->mp_upper);
7049 mp->mp_lower -= sizeof(indx_t);
7053 /** Compact the main page after deleting a node on a subpage.
7054 * @param[in] mp The main page to operate on.
7055 * @param[in] indx The index of the subpage on the main page.
7058 mdb_node_shrink(MDB_page *mp, indx_t indx)
7064 indx_t i, numkeys, ptr;
7066 node = NODEPTR(mp, indx);
7067 sp = (MDB_page *)NODEDATA(node);
7068 delta = SIZELEFT(sp);
7069 xp = (MDB_page *)((char *)sp + delta);
7071 /* shift subpage upward */
7073 nsize = NUMKEYS(sp) * sp->mp_pad;
7075 return; /* do not make the node uneven-sized */
7076 memmove(METADATA(xp), METADATA(sp), nsize);
7079 numkeys = NUMKEYS(sp);
7080 for (i=numkeys-1; i>=0; i--)
7081 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7083 xp->mp_upper = sp->mp_lower;
7084 xp->mp_lower = sp->mp_lower;
7085 xp->mp_flags = sp->mp_flags;
7086 xp->mp_pad = sp->mp_pad;
7087 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7089 nsize = NODEDSZ(node) - delta;
7090 SETDSZ(node, nsize);
7092 /* shift lower nodes upward */
7093 ptr = mp->mp_ptrs[indx];
7094 numkeys = NUMKEYS(mp);
7095 for (i = 0; i < numkeys; i++) {
7096 if (mp->mp_ptrs[i] <= ptr)
7097 mp->mp_ptrs[i] += delta;
7100 base = (char *)mp + mp->mp_upper + PAGEBASE;
7101 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7102 mp->mp_upper += delta;
7105 /** Initial setup of a sorted-dups cursor.
7106 * Sorted duplicates are implemented as a sub-database for the given key.
7107 * The duplicate data items are actually keys of the sub-database.
7108 * Operations on the duplicate data items are performed using a sub-cursor
7109 * initialized when the sub-database is first accessed. This function does
7110 * the preliminary setup of the sub-cursor, filling in the fields that
7111 * depend only on the parent DB.
7112 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7115 mdb_xcursor_init0(MDB_cursor *mc)
7117 MDB_xcursor *mx = mc->mc_xcursor;
7119 mx->mx_cursor.mc_xcursor = NULL;
7120 mx->mx_cursor.mc_txn = mc->mc_txn;
7121 mx->mx_cursor.mc_db = &mx->mx_db;
7122 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7123 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7124 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7125 mx->mx_cursor.mc_snum = 0;
7126 mx->mx_cursor.mc_top = 0;
7127 mx->mx_cursor.mc_flags = C_SUB;
7128 mx->mx_dbx.md_name.mv_size = 0;
7129 mx->mx_dbx.md_name.mv_data = NULL;
7130 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7131 mx->mx_dbx.md_dcmp = NULL;
7132 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7135 /** Final setup of a sorted-dups cursor.
7136 * Sets up the fields that depend on the data from the main cursor.
7137 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7138 * @param[in] node The data containing the #MDB_db record for the
7139 * sorted-dup database.
7142 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7144 MDB_xcursor *mx = mc->mc_xcursor;
7146 if (node->mn_flags & F_SUBDATA) {
7147 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7148 mx->mx_cursor.mc_pg[0] = 0;
7149 mx->mx_cursor.mc_snum = 0;
7150 mx->mx_cursor.mc_top = 0;
7151 mx->mx_cursor.mc_flags = C_SUB;
7153 MDB_page *fp = NODEDATA(node);
7154 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7155 mx->mx_db.md_flags = 0;
7156 mx->mx_db.md_depth = 1;
7157 mx->mx_db.md_branch_pages = 0;
7158 mx->mx_db.md_leaf_pages = 1;
7159 mx->mx_db.md_overflow_pages = 0;
7160 mx->mx_db.md_entries = NUMKEYS(fp);
7161 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7162 mx->mx_cursor.mc_snum = 1;
7163 mx->mx_cursor.mc_top = 0;
7164 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7165 mx->mx_cursor.mc_pg[0] = fp;
7166 mx->mx_cursor.mc_ki[0] = 0;
7167 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7168 mx->mx_db.md_flags = MDB_DUPFIXED;
7169 mx->mx_db.md_pad = fp->mp_pad;
7170 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7171 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7174 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7175 mx->mx_db.md_root));
7176 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7177 #if UINT_MAX < SIZE_MAX
7178 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7179 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7183 /** Initialize a cursor for a given transaction and database. */
7185 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7188 mc->mc_backup = NULL;
7191 mc->mc_db = &txn->mt_dbs[dbi];
7192 mc->mc_dbx = &txn->mt_dbxs[dbi];
7193 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7198 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7199 mdb_tassert(txn, mx != NULL);
7200 mc->mc_xcursor = mx;
7201 mdb_xcursor_init0(mc);
7203 mc->mc_xcursor = NULL;
7205 if (*mc->mc_dbflag & DB_STALE) {
7206 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7211 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7214 size_t size = sizeof(MDB_cursor);
7216 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7219 if (txn->mt_flags & MDB_TXN_ERROR)
7222 /* Allow read access to the freelist */
7223 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7226 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7227 size += sizeof(MDB_xcursor);
7229 if ((mc = malloc(size)) != NULL) {
7230 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7231 if (txn->mt_cursors) {
7232 mc->mc_next = txn->mt_cursors[dbi];
7233 txn->mt_cursors[dbi] = mc;
7234 mc->mc_flags |= C_UNTRACK;
7246 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7248 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7251 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7254 if (txn->mt_flags & MDB_TXN_ERROR)
7257 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7261 /* Return the count of duplicate data items for the current key */
7263 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7267 if (mc == NULL || countp == NULL)
7270 if (mc->mc_xcursor == NULL)
7271 return MDB_INCOMPATIBLE;
7273 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7276 if (!(mc->mc_flags & C_INITIALIZED))
7279 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7280 return MDB_NOTFOUND;
7282 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7283 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7286 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7289 *countp = mc->mc_xcursor->mx_db.md_entries;
7295 mdb_cursor_close(MDB_cursor *mc)
7297 if (mc && !mc->mc_backup) {
7298 /* remove from txn, if tracked */
7299 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7300 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7301 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7303 *prev = mc->mc_next;
7310 mdb_cursor_txn(MDB_cursor *mc)
7312 if (!mc) return NULL;
7317 mdb_cursor_dbi(MDB_cursor *mc)
7322 /** Replace the key for a branch node with a new key.
7323 * @param[in] mc Cursor pointing to the node to operate on.
7324 * @param[in] key The new key to use.
7325 * @return 0 on success, non-zero on failure.
7328 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7334 int delta, ksize, oksize;
7335 indx_t ptr, i, numkeys, indx;
7338 indx = mc->mc_ki[mc->mc_top];
7339 mp = mc->mc_pg[mc->mc_top];
7340 node = NODEPTR(mp, indx);
7341 ptr = mp->mp_ptrs[indx];
7345 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7346 k2.mv_data = NODEKEY(node);
7347 k2.mv_size = node->mn_ksize;
7348 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7350 mdb_dkey(&k2, kbuf2),
7356 /* Sizes must be 2-byte aligned. */
7357 ksize = EVEN(key->mv_size);
7358 oksize = EVEN(node->mn_ksize);
7359 delta = ksize - oksize;
7361 /* Shift node contents if EVEN(key length) changed. */
7363 if (delta > 0 && SIZELEFT(mp) < delta) {
7365 /* not enough space left, do a delete and split */
7366 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7367 pgno = NODEPGNO(node);
7368 mdb_node_del(mc, 0);
7369 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7372 numkeys = NUMKEYS(mp);
7373 for (i = 0; i < numkeys; i++) {
7374 if (mp->mp_ptrs[i] <= ptr)
7375 mp->mp_ptrs[i] -= delta;
7378 base = (char *)mp + mp->mp_upper + PAGEBASE;
7379 len = ptr - mp->mp_upper + NODESIZE;
7380 memmove(base - delta, base, len);
7381 mp->mp_upper -= delta;
7383 node = NODEPTR(mp, indx);
7386 /* But even if no shift was needed, update ksize */
7387 if (node->mn_ksize != key->mv_size)
7388 node->mn_ksize = key->mv_size;
7391 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7397 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7399 /** Move a node from csrc to cdst.
7402 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7409 unsigned short flags;
7413 /* Mark src and dst as dirty. */
7414 if ((rc = mdb_page_touch(csrc)) ||
7415 (rc = mdb_page_touch(cdst)))
7418 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7419 key.mv_size = csrc->mc_db->md_pad;
7420 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7422 data.mv_data = NULL;
7426 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7427 mdb_cassert(csrc, !((size_t)srcnode & 1));
7428 srcpg = NODEPGNO(srcnode);
7429 flags = srcnode->mn_flags;
7430 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7431 unsigned int snum = csrc->mc_snum;
7433 /* must find the lowest key below src */
7434 rc = mdb_page_search_lowest(csrc);
7437 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7438 key.mv_size = csrc->mc_db->md_pad;
7439 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7441 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7442 key.mv_size = NODEKSZ(s2);
7443 key.mv_data = NODEKEY(s2);
7445 csrc->mc_snum = snum--;
7446 csrc->mc_top = snum;
7448 key.mv_size = NODEKSZ(srcnode);
7449 key.mv_data = NODEKEY(srcnode);
7451 data.mv_size = NODEDSZ(srcnode);
7452 data.mv_data = NODEDATA(srcnode);
7454 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7455 unsigned int snum = cdst->mc_snum;
7458 /* must find the lowest key below dst */
7459 mdb_cursor_copy(cdst, &mn);
7460 rc = mdb_page_search_lowest(&mn);
7463 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7464 bkey.mv_size = mn.mc_db->md_pad;
7465 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7467 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7468 bkey.mv_size = NODEKSZ(s2);
7469 bkey.mv_data = NODEKEY(s2);
7471 mn.mc_snum = snum--;
7474 rc = mdb_update_key(&mn, &bkey);
7479 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7480 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7481 csrc->mc_ki[csrc->mc_top],
7483 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7484 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7486 /* Add the node to the destination page.
7488 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7489 if (rc != MDB_SUCCESS)
7492 /* Delete the node from the source page.
7494 mdb_node_del(csrc, key.mv_size);
7497 /* Adjust other cursors pointing to mp */
7498 MDB_cursor *m2, *m3;
7499 MDB_dbi dbi = csrc->mc_dbi;
7500 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7502 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7503 if (csrc->mc_flags & C_SUB)
7504 m3 = &m2->mc_xcursor->mx_cursor;
7507 if (m3 == csrc) continue;
7508 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7509 csrc->mc_ki[csrc->mc_top]) {
7510 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7511 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7516 /* Update the parent separators.
7518 if (csrc->mc_ki[csrc->mc_top] == 0) {
7519 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7520 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7521 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7523 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7524 key.mv_size = NODEKSZ(srcnode);
7525 key.mv_data = NODEKEY(srcnode);
7527 DPRINTF(("update separator for source page %"Z"u to [%s]",
7528 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7529 mdb_cursor_copy(csrc, &mn);
7532 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7535 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7537 indx_t ix = csrc->mc_ki[csrc->mc_top];
7538 nullkey.mv_size = 0;
7539 csrc->mc_ki[csrc->mc_top] = 0;
7540 rc = mdb_update_key(csrc, &nullkey);
7541 csrc->mc_ki[csrc->mc_top] = ix;
7542 mdb_cassert(csrc, rc == MDB_SUCCESS);
7546 if (cdst->mc_ki[cdst->mc_top] == 0) {
7547 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7548 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7549 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7551 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7552 key.mv_size = NODEKSZ(srcnode);
7553 key.mv_data = NODEKEY(srcnode);
7555 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7556 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7557 mdb_cursor_copy(cdst, &mn);
7560 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7563 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7565 indx_t ix = cdst->mc_ki[cdst->mc_top];
7566 nullkey.mv_size = 0;
7567 cdst->mc_ki[cdst->mc_top] = 0;
7568 rc = mdb_update_key(cdst, &nullkey);
7569 cdst->mc_ki[cdst->mc_top] = ix;
7570 mdb_cassert(cdst, rc == MDB_SUCCESS);
7577 /** Merge one page into another.
7578 * The nodes from the page pointed to by \b csrc will
7579 * be copied to the page pointed to by \b cdst and then
7580 * the \b csrc page will be freed.
7581 * @param[in] csrc Cursor pointing to the source page.
7582 * @param[in] cdst Cursor pointing to the destination page.
7583 * @return 0 on success, non-zero on failure.
7586 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7588 MDB_page *psrc, *pdst;
7595 psrc = csrc->mc_pg[csrc->mc_top];
7596 pdst = cdst->mc_pg[cdst->mc_top];
7598 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7600 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7601 mdb_cassert(csrc, cdst->mc_snum > 1);
7603 /* Mark dst as dirty. */
7604 if ((rc = mdb_page_touch(cdst)))
7607 /* Move all nodes from src to dst.
7609 j = nkeys = NUMKEYS(pdst);
7610 if (IS_LEAF2(psrc)) {
7611 key.mv_size = csrc->mc_db->md_pad;
7612 key.mv_data = METADATA(psrc);
7613 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7614 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7615 if (rc != MDB_SUCCESS)
7617 key.mv_data = (char *)key.mv_data + key.mv_size;
7620 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7621 srcnode = NODEPTR(psrc, i);
7622 if (i == 0 && IS_BRANCH(psrc)) {
7625 mdb_cursor_copy(csrc, &mn);
7626 /* must find the lowest key below src */
7627 rc = mdb_page_search_lowest(&mn);
7630 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7631 key.mv_size = mn.mc_db->md_pad;
7632 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7634 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7635 key.mv_size = NODEKSZ(s2);
7636 key.mv_data = NODEKEY(s2);
7639 key.mv_size = srcnode->mn_ksize;
7640 key.mv_data = NODEKEY(srcnode);
7643 data.mv_size = NODEDSZ(srcnode);
7644 data.mv_data = NODEDATA(srcnode);
7645 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7646 if (rc != MDB_SUCCESS)
7651 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7652 pdst->mp_pgno, NUMKEYS(pdst),
7653 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7655 /* Unlink the src page from parent and add to free list.
7658 mdb_node_del(csrc, 0);
7659 if (csrc->mc_ki[csrc->mc_top] == 0) {
7661 rc = mdb_update_key(csrc, &key);
7669 psrc = csrc->mc_pg[csrc->mc_top];
7670 /* If not operating on FreeDB, allow this page to be reused
7671 * in this txn. Otherwise just add to free list.
7673 rc = mdb_page_loose(csrc, psrc);
7677 csrc->mc_db->md_leaf_pages--;
7679 csrc->mc_db->md_branch_pages--;
7681 /* Adjust other cursors pointing to mp */
7682 MDB_cursor *m2, *m3;
7683 MDB_dbi dbi = csrc->mc_dbi;
7685 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7686 if (csrc->mc_flags & C_SUB)
7687 m3 = &m2->mc_xcursor->mx_cursor;
7690 if (m3 == csrc) continue;
7691 if (m3->mc_snum < csrc->mc_snum) continue;
7692 if (m3->mc_pg[csrc->mc_top] == psrc) {
7693 m3->mc_pg[csrc->mc_top] = pdst;
7694 m3->mc_ki[csrc->mc_top] += nkeys;
7699 unsigned int snum = cdst->mc_snum;
7700 uint16_t depth = cdst->mc_db->md_depth;
7701 mdb_cursor_pop(cdst);
7702 rc = mdb_rebalance(cdst);
7703 /* Did the tree shrink? */
7704 if (depth > cdst->mc_db->md_depth)
7706 cdst->mc_snum = snum;
7707 cdst->mc_top = snum-1;
7712 /** Copy the contents of a cursor.
7713 * @param[in] csrc The cursor to copy from.
7714 * @param[out] cdst The cursor to copy to.
7717 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7721 cdst->mc_txn = csrc->mc_txn;
7722 cdst->mc_dbi = csrc->mc_dbi;
7723 cdst->mc_db = csrc->mc_db;
7724 cdst->mc_dbx = csrc->mc_dbx;
7725 cdst->mc_snum = csrc->mc_snum;
7726 cdst->mc_top = csrc->mc_top;
7727 cdst->mc_flags = csrc->mc_flags;
7729 for (i=0; i<csrc->mc_snum; i++) {
7730 cdst->mc_pg[i] = csrc->mc_pg[i];
7731 cdst->mc_ki[i] = csrc->mc_ki[i];
7735 /** Rebalance the tree after a delete operation.
7736 * @param[in] mc Cursor pointing to the page where rebalancing
7738 * @return 0 on success, non-zero on failure.
7741 mdb_rebalance(MDB_cursor *mc)
7745 unsigned int ptop, minkeys;
7749 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7750 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7751 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7752 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7753 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7755 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7756 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7757 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7758 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7762 if (mc->mc_snum < 2) {
7763 MDB_page *mp = mc->mc_pg[0];
7765 DPUTS("Can't rebalance a subpage, ignoring");
7768 if (NUMKEYS(mp) == 0) {
7769 DPUTS("tree is completely empty");
7770 mc->mc_db->md_root = P_INVALID;
7771 mc->mc_db->md_depth = 0;
7772 mc->mc_db->md_leaf_pages = 0;
7773 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7776 /* Adjust cursors pointing to mp */
7779 mc->mc_flags &= ~C_INITIALIZED;
7781 MDB_cursor *m2, *m3;
7782 MDB_dbi dbi = mc->mc_dbi;
7784 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7785 if (mc->mc_flags & C_SUB)
7786 m3 = &m2->mc_xcursor->mx_cursor;
7789 if (m3->mc_snum < mc->mc_snum) continue;
7790 if (m3->mc_pg[0] == mp) {
7793 m3->mc_flags &= ~C_INITIALIZED;
7797 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7799 DPUTS("collapsing root page!");
7800 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7803 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7804 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7807 mc->mc_db->md_depth--;
7808 mc->mc_db->md_branch_pages--;
7809 mc->mc_ki[0] = mc->mc_ki[1];
7810 for (i = 1; i<mc->mc_db->md_depth; i++) {
7811 mc->mc_pg[i] = mc->mc_pg[i+1];
7812 mc->mc_ki[i] = mc->mc_ki[i+1];
7815 /* Adjust other cursors pointing to mp */
7816 MDB_cursor *m2, *m3;
7817 MDB_dbi dbi = mc->mc_dbi;
7819 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7820 if (mc->mc_flags & C_SUB)
7821 m3 = &m2->mc_xcursor->mx_cursor;
7824 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7825 if (m3->mc_pg[0] == mp) {
7828 for (i=0; i<m3->mc_snum; i++) {
7829 m3->mc_pg[i] = m3->mc_pg[i+1];
7830 m3->mc_ki[i] = m3->mc_ki[i+1];
7836 DPUTS("root page doesn't need rebalancing");
7840 /* The parent (branch page) must have at least 2 pointers,
7841 * otherwise the tree is invalid.
7843 ptop = mc->mc_top-1;
7844 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7846 /* Leaf page fill factor is below the threshold.
7847 * Try to move keys from left or right neighbor, or
7848 * merge with a neighbor page.
7853 mdb_cursor_copy(mc, &mn);
7854 mn.mc_xcursor = NULL;
7856 oldki = mc->mc_ki[mc->mc_top];
7857 if (mc->mc_ki[ptop] == 0) {
7858 /* We're the leftmost leaf in our parent.
7860 DPUTS("reading right neighbor");
7862 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7863 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7866 mn.mc_ki[mn.mc_top] = 0;
7867 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7869 /* There is at least one neighbor to the left.
7871 DPUTS("reading left neighbor");
7873 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7874 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7877 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7878 mc->mc_ki[mc->mc_top] = 0;
7881 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7882 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7883 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7885 /* If the neighbor page is above threshold and has enough keys,
7886 * move one key from it. Otherwise we should try to merge them.
7887 * (A branch page must never have less than 2 keys.)
7889 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7890 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7891 rc = mdb_node_move(&mn, mc);
7892 if (mc->mc_ki[ptop]) {
7896 if (mc->mc_ki[ptop] == 0) {
7897 rc = mdb_page_merge(&mn, mc);
7899 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7900 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7901 rc = mdb_page_merge(mc, &mn);
7902 mdb_cursor_copy(&mn, mc);
7904 mc->mc_flags &= ~C_EOF;
7906 mc->mc_ki[mc->mc_top] = oldki;
7910 /** Complete a delete operation started by #mdb_cursor_del(). */
7912 mdb_cursor_del0(MDB_cursor *mc)
7919 ki = mc->mc_ki[mc->mc_top];
7920 mdb_node_del(mc, mc->mc_db->md_pad);
7921 mc->mc_db->md_entries--;
7922 rc = mdb_rebalance(mc);
7924 if (rc == MDB_SUCCESS) {
7925 MDB_cursor *m2, *m3;
7926 MDB_dbi dbi = mc->mc_dbi;
7928 mp = mc->mc_pg[mc->mc_top];
7929 nkeys = NUMKEYS(mp);
7931 /* if mc points past last node in page, find next sibling */
7932 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7933 rc = mdb_cursor_sibling(mc, 1);
7934 if (rc == MDB_NOTFOUND) {
7935 mc->mc_flags |= C_EOF;
7940 /* Adjust other cursors pointing to mp */
7941 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7942 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7943 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7945 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7947 if (m3->mc_pg[mc->mc_top] == mp) {
7948 if (m3->mc_ki[mc->mc_top] >= ki) {
7949 m3->mc_flags |= C_DEL;
7950 if (m3->mc_ki[mc->mc_top] > ki)
7951 m3->mc_ki[mc->mc_top]--;
7952 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7953 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7955 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7956 rc = mdb_cursor_sibling(m3, 1);
7957 if (rc == MDB_NOTFOUND) {
7958 m3->mc_flags |= C_EOF;
7964 mc->mc_flags |= C_DEL;
7968 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7973 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7974 MDB_val *key, MDB_val *data)
7976 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7979 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7980 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7982 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7983 /* must ignore any data */
7987 return mdb_del0(txn, dbi, key, data, 0);
7991 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7992 MDB_val *key, MDB_val *data, unsigned flags)
7997 MDB_val rdata, *xdata;
8001 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8003 mdb_cursor_init(&mc, txn, dbi, &mx);
8012 flags |= MDB_NODUPDATA;
8014 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8016 /* let mdb_page_split know about this cursor if needed:
8017 * delete will trigger a rebalance; if it needs to move
8018 * a node from one page to another, it will have to
8019 * update the parent's separator key(s). If the new sepkey
8020 * is larger than the current one, the parent page may
8021 * run out of space, triggering a split. We need this
8022 * cursor to be consistent until the end of the rebalance.
8024 mc.mc_flags |= C_UNTRACK;
8025 mc.mc_next = txn->mt_cursors[dbi];
8026 txn->mt_cursors[dbi] = &mc;
8027 rc = mdb_cursor_del(&mc, flags);
8028 txn->mt_cursors[dbi] = mc.mc_next;
8033 /** Split a page and insert a new node.
8034 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8035 * The cursor will be updated to point to the actual page and index where
8036 * the node got inserted after the split.
8037 * @param[in] newkey The key for the newly inserted node.
8038 * @param[in] newdata The data for the newly inserted node.
8039 * @param[in] newpgno The page number, if the new node is a branch node.
8040 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8041 * @return 0 on success, non-zero on failure.
8044 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8045 unsigned int nflags)
8048 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8051 int i, j, split_indx, nkeys, pmax;
8052 MDB_env *env = mc->mc_txn->mt_env;
8054 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8055 MDB_page *copy = NULL;
8056 MDB_page *mp, *rp, *pp;
8061 mp = mc->mc_pg[mc->mc_top];
8062 newindx = mc->mc_ki[mc->mc_top];
8063 nkeys = NUMKEYS(mp);
8065 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8066 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8067 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8069 /* Create a right sibling. */
8070 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8072 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8074 if (mc->mc_snum < 2) {
8075 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8077 /* shift current top to make room for new parent */
8078 mc->mc_pg[1] = mc->mc_pg[0];
8079 mc->mc_ki[1] = mc->mc_ki[0];
8082 mc->mc_db->md_root = pp->mp_pgno;
8083 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8084 mc->mc_db->md_depth++;
8087 /* Add left (implicit) pointer. */
8088 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8089 /* undo the pre-push */
8090 mc->mc_pg[0] = mc->mc_pg[1];
8091 mc->mc_ki[0] = mc->mc_ki[1];
8092 mc->mc_db->md_root = mp->mp_pgno;
8093 mc->mc_db->md_depth--;
8100 ptop = mc->mc_top-1;
8101 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8104 mc->mc_flags |= C_SPLITTING;
8105 mdb_cursor_copy(mc, &mn);
8106 mn.mc_pg[mn.mc_top] = rp;
8107 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8109 if (nflags & MDB_APPEND) {
8110 mn.mc_ki[mn.mc_top] = 0;
8112 split_indx = newindx;
8116 split_indx = (nkeys+1) / 2;
8121 unsigned int lsize, rsize, ksize;
8122 /* Move half of the keys to the right sibling */
8123 x = mc->mc_ki[mc->mc_top] - split_indx;
8124 ksize = mc->mc_db->md_pad;
8125 split = LEAF2KEY(mp, split_indx, ksize);
8126 rsize = (nkeys - split_indx) * ksize;
8127 lsize = (nkeys - split_indx) * sizeof(indx_t);
8128 mp->mp_lower -= lsize;
8129 rp->mp_lower += lsize;
8130 mp->mp_upper += rsize - lsize;
8131 rp->mp_upper -= rsize - lsize;
8132 sepkey.mv_size = ksize;
8133 if (newindx == split_indx) {
8134 sepkey.mv_data = newkey->mv_data;
8136 sepkey.mv_data = split;
8139 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8140 memcpy(rp->mp_ptrs, split, rsize);
8141 sepkey.mv_data = rp->mp_ptrs;
8142 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8143 memcpy(ins, newkey->mv_data, ksize);
8144 mp->mp_lower += sizeof(indx_t);
8145 mp->mp_upper -= ksize - sizeof(indx_t);
8148 memcpy(rp->mp_ptrs, split, x * ksize);
8149 ins = LEAF2KEY(rp, x, ksize);
8150 memcpy(ins, newkey->mv_data, ksize);
8151 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8152 rp->mp_lower += sizeof(indx_t);
8153 rp->mp_upper -= ksize - sizeof(indx_t);
8154 mc->mc_ki[mc->mc_top] = x;
8155 mc->mc_pg[mc->mc_top] = rp;
8158 int psize, nsize, k;
8159 /* Maximum free space in an empty page */
8160 pmax = env->me_psize - PAGEHDRSZ;
8162 nsize = mdb_leaf_size(env, newkey, newdata);
8164 nsize = mdb_branch_size(env, newkey);
8165 nsize = EVEN(nsize);
8167 /* grab a page to hold a temporary copy */
8168 copy = mdb_page_malloc(mc->mc_txn, 1);
8173 copy->mp_pgno = mp->mp_pgno;
8174 copy->mp_flags = mp->mp_flags;
8175 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8176 copy->mp_upper = env->me_psize - PAGEBASE;
8178 /* prepare to insert */
8179 for (i=0, j=0; i<nkeys; i++) {
8181 copy->mp_ptrs[j++] = 0;
8183 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8186 /* When items are relatively large the split point needs
8187 * to be checked, because being off-by-one will make the
8188 * difference between success or failure in mdb_node_add.
8190 * It's also relevant if a page happens to be laid out
8191 * such that one half of its nodes are all "small" and
8192 * the other half of its nodes are "large." If the new
8193 * item is also "large" and falls on the half with
8194 * "large" nodes, it also may not fit.
8196 * As a final tweak, if the new item goes on the last
8197 * spot on the page (and thus, onto the new page), bias
8198 * the split so the new page is emptier than the old page.
8199 * This yields better packing during sequential inserts.
8201 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8202 /* Find split point */
8204 if (newindx <= split_indx || newindx >= nkeys) {
8206 k = newindx >= nkeys ? nkeys : split_indx+2;
8211 for (; i!=k; i+=j) {
8216 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8217 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8219 if (F_ISSET(node->mn_flags, F_BIGDATA))
8220 psize += sizeof(pgno_t);
8222 psize += NODEDSZ(node);
8224 psize = EVEN(psize);
8226 if (psize > pmax || i == k-j) {
8227 split_indx = i + (j<0);
8232 if (split_indx == newindx) {
8233 sepkey.mv_size = newkey->mv_size;
8234 sepkey.mv_data = newkey->mv_data;
8236 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8237 sepkey.mv_size = node->mn_ksize;
8238 sepkey.mv_data = NODEKEY(node);
8243 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8245 /* Copy separator key to the parent.
8247 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8251 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8256 if (mn.mc_snum == mc->mc_snum) {
8257 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8258 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8259 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8260 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8265 /* Right page might now have changed parent.
8266 * Check if left page also changed parent.
8268 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8269 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8270 for (i=0; i<ptop; i++) {
8271 mc->mc_pg[i] = mn.mc_pg[i];
8272 mc->mc_ki[i] = mn.mc_ki[i];
8274 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8275 if (mn.mc_ki[ptop]) {
8276 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8278 /* find right page's left sibling */
8279 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8280 mdb_cursor_sibling(mc, 0);
8285 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8288 mc->mc_flags ^= C_SPLITTING;
8289 if (rc != MDB_SUCCESS) {
8292 if (nflags & MDB_APPEND) {
8293 mc->mc_pg[mc->mc_top] = rp;
8294 mc->mc_ki[mc->mc_top] = 0;
8295 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8298 for (i=0; i<mc->mc_top; i++)
8299 mc->mc_ki[i] = mn.mc_ki[i];
8300 } else if (!IS_LEAF2(mp)) {
8302 mc->mc_pg[mc->mc_top] = rp;
8307 rkey.mv_data = newkey->mv_data;
8308 rkey.mv_size = newkey->mv_size;
8314 /* Update index for the new key. */
8315 mc->mc_ki[mc->mc_top] = j;
8317 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8318 rkey.mv_data = NODEKEY(node);
8319 rkey.mv_size = node->mn_ksize;
8321 xdata.mv_data = NODEDATA(node);
8322 xdata.mv_size = NODEDSZ(node);
8325 pgno = NODEPGNO(node);
8326 flags = node->mn_flags;
8329 if (!IS_LEAF(mp) && j == 0) {
8330 /* First branch index doesn't need key data. */
8334 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8340 mc->mc_pg[mc->mc_top] = copy;
8345 } while (i != split_indx);
8347 nkeys = NUMKEYS(copy);
8348 for (i=0; i<nkeys; i++)
8349 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8350 mp->mp_lower = copy->mp_lower;
8351 mp->mp_upper = copy->mp_upper;
8352 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8353 env->me_psize - copy->mp_upper - PAGEBASE);
8355 /* reset back to original page */
8356 if (newindx < split_indx) {
8357 mc->mc_pg[mc->mc_top] = mp;
8358 if (nflags & MDB_RESERVE) {
8359 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8360 if (!(node->mn_flags & F_BIGDATA))
8361 newdata->mv_data = NODEDATA(node);
8364 mc->mc_pg[mc->mc_top] = rp;
8366 /* Make sure mc_ki is still valid.
8368 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8369 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8370 for (i=0; i<=ptop; i++) {
8371 mc->mc_pg[i] = mn.mc_pg[i];
8372 mc->mc_ki[i] = mn.mc_ki[i];
8379 /* Adjust other cursors pointing to mp */
8380 MDB_cursor *m2, *m3;
8381 MDB_dbi dbi = mc->mc_dbi;
8382 int fixup = NUMKEYS(mp);
8384 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8385 if (mc->mc_flags & C_SUB)
8386 m3 = &m2->mc_xcursor->mx_cursor;
8391 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8393 if (m3->mc_flags & C_SPLITTING)
8398 for (k=m3->mc_top; k>=0; k--) {
8399 m3->mc_ki[k+1] = m3->mc_ki[k];
8400 m3->mc_pg[k+1] = m3->mc_pg[k];
8402 if (m3->mc_ki[0] >= split_indx) {
8407 m3->mc_pg[0] = mc->mc_pg[0];
8411 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8412 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8413 m3->mc_ki[mc->mc_top]++;
8414 if (m3->mc_ki[mc->mc_top] >= fixup) {
8415 m3->mc_pg[mc->mc_top] = rp;
8416 m3->mc_ki[mc->mc_top] -= fixup;
8417 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8419 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8420 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8425 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8428 if (copy) /* tmp page */
8429 mdb_page_free(env, copy);
8431 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8436 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8437 MDB_val *key, MDB_val *data, unsigned int flags)
8442 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8445 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8448 mdb_cursor_init(&mc, txn, dbi, &mx);
8449 return mdb_cursor_put(&mc, key, data, flags);
8453 #define MDB_WBUF (1024*1024)
8456 /** State needed for a compacting copy. */
8457 typedef struct mdb_copy {
8458 pthread_mutex_t mc_mutex;
8459 pthread_cond_t mc_cond;
8466 pgno_t mc_next_pgno;
8469 volatile int mc_new;
8474 /** Dedicated writer thread for compacting copy. */
8475 static THREAD_RET ESECT
8476 mdb_env_copythr(void *arg)
8480 int toggle = 0, wsize, rc;
8483 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8486 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8489 pthread_mutex_lock(&my->mc_mutex);
8491 pthread_cond_signal(&my->mc_cond);
8494 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8495 if (my->mc_new < 0) {
8500 wsize = my->mc_wlen[toggle];
8501 ptr = my->mc_wbuf[toggle];
8504 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8508 } else if (len > 0) {
8522 /* If there's an overflow page tail, write it too */
8523 if (my->mc_olen[toggle]) {
8524 wsize = my->mc_olen[toggle];
8525 ptr = my->mc_over[toggle];
8526 my->mc_olen[toggle] = 0;
8529 my->mc_wlen[toggle] = 0;
8531 pthread_cond_signal(&my->mc_cond);
8533 pthread_cond_signal(&my->mc_cond);
8534 pthread_mutex_unlock(&my->mc_mutex);
8535 return (THREAD_RET)0;
8539 /** Tell the writer thread there's a buffer ready to write */
8541 mdb_env_cthr_toggle(mdb_copy *my, int st)
8543 int toggle = my->mc_toggle ^ 1;
8544 pthread_mutex_lock(&my->mc_mutex);
8545 if (my->mc_status) {
8546 pthread_mutex_unlock(&my->mc_mutex);
8547 return my->mc_status;
8549 while (my->mc_new == 1)
8550 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8552 my->mc_toggle = toggle;
8553 pthread_cond_signal(&my->mc_cond);
8554 pthread_mutex_unlock(&my->mc_mutex);
8558 /** Depth-first tree traversal for compacting copy. */
8560 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8563 MDB_txn *txn = my->mc_txn;
8565 MDB_page *mo, *mp, *leaf;
8570 /* Empty DB, nothing to do */
8571 if (*pg == P_INVALID)
8578 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8581 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8585 /* Make cursor pages writable */
8586 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8590 for (i=0; i<mc.mc_top; i++) {
8591 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8592 mc.mc_pg[i] = (MDB_page *)ptr;
8593 ptr += my->mc_env->me_psize;
8596 /* This is writable space for a leaf page. Usually not needed. */
8597 leaf = (MDB_page *)ptr;
8599 toggle = my->mc_toggle;
8600 while (mc.mc_snum > 0) {
8602 mp = mc.mc_pg[mc.mc_top];
8606 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8607 for (i=0; i<n; i++) {
8608 ni = NODEPTR(mp, i);
8609 if (ni->mn_flags & F_BIGDATA) {
8613 /* Need writable leaf */
8615 mc.mc_pg[mc.mc_top] = leaf;
8616 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8618 ni = NODEPTR(mp, i);
8621 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8622 rc = mdb_page_get(txn, pg, &omp, NULL);
8625 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8626 rc = mdb_env_cthr_toggle(my, 1);
8629 toggle = my->mc_toggle;
8631 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8632 memcpy(mo, omp, my->mc_env->me_psize);
8633 mo->mp_pgno = my->mc_next_pgno;
8634 my->mc_next_pgno += omp->mp_pages;
8635 my->mc_wlen[toggle] += my->mc_env->me_psize;
8636 if (omp->mp_pages > 1) {
8637 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8638 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8639 rc = mdb_env_cthr_toggle(my, 1);
8642 toggle = my->mc_toggle;
8644 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8645 } else if (ni->mn_flags & F_SUBDATA) {
8648 /* Need writable leaf */
8650 mc.mc_pg[mc.mc_top] = leaf;
8651 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8653 ni = NODEPTR(mp, i);
8656 memcpy(&db, NODEDATA(ni), sizeof(db));
8657 my->mc_toggle = toggle;
8658 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8661 toggle = my->mc_toggle;
8662 memcpy(NODEDATA(ni), &db, sizeof(db));
8667 mc.mc_ki[mc.mc_top]++;
8668 if (mc.mc_ki[mc.mc_top] < n) {
8671 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8673 rc = mdb_page_get(txn, pg, &mp, NULL);
8678 mc.mc_ki[mc.mc_top] = 0;
8679 if (IS_BRANCH(mp)) {
8680 /* Whenever we advance to a sibling branch page,
8681 * we must proceed all the way down to its first leaf.
8683 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8686 mc.mc_pg[mc.mc_top] = mp;
8690 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8691 rc = mdb_env_cthr_toggle(my, 1);
8694 toggle = my->mc_toggle;
8696 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8697 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8698 mo->mp_pgno = my->mc_next_pgno++;
8699 my->mc_wlen[toggle] += my->mc_env->me_psize;
8701 /* Update parent if there is one */
8702 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8703 SETPGNO(ni, mo->mp_pgno);
8704 mdb_cursor_pop(&mc);
8706 /* Otherwise we're done */
8716 /** Copy environment with compaction. */
8718 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8723 MDB_txn *txn = NULL;
8728 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8729 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8730 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8731 if (my.mc_wbuf[0] == NULL)
8734 pthread_mutex_init(&my.mc_mutex, NULL);
8735 pthread_cond_init(&my.mc_cond, NULL);
8736 #ifdef HAVE_MEMALIGN
8737 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8738 if (my.mc_wbuf[0] == NULL)
8741 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8746 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8747 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8752 my.mc_next_pgno = 2;
8758 THREAD_CREATE(thr, mdb_env_copythr, &my);
8760 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8764 mp = (MDB_page *)my.mc_wbuf[0];
8765 memset(mp, 0, 2*env->me_psize);
8767 mp->mp_flags = P_META;
8768 mm = (MDB_meta *)METADATA(mp);
8769 mdb_env_init_meta0(env, mm);
8770 mm->mm_address = env->me_metas[0]->mm_address;
8772 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8774 mp->mp_flags = P_META;
8775 *(MDB_meta *)METADATA(mp) = *mm;
8776 mm = (MDB_meta *)METADATA(mp);
8778 /* Count the number of free pages, subtract from lastpg to find
8779 * number of active pages
8782 MDB_ID freecount = 0;
8785 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8786 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8787 freecount += *(MDB_ID *)data.mv_data;
8788 freecount += txn->mt_dbs[0].md_branch_pages +
8789 txn->mt_dbs[0].md_leaf_pages +
8790 txn->mt_dbs[0].md_overflow_pages;
8792 /* Set metapage 1 */
8793 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8794 mm->mm_dbs[1] = txn->mt_dbs[1];
8795 if (mm->mm_last_pg > 1) {
8796 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8799 mm->mm_dbs[1].md_root = P_INVALID;
8802 my.mc_wlen[0] = env->me_psize * 2;
8804 pthread_mutex_lock(&my.mc_mutex);
8806 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8807 pthread_mutex_unlock(&my.mc_mutex);
8808 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8809 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8810 rc = mdb_env_cthr_toggle(&my, 1);
8811 mdb_env_cthr_toggle(&my, -1);
8812 pthread_mutex_lock(&my.mc_mutex);
8814 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8815 pthread_mutex_unlock(&my.mc_mutex);
8820 CloseHandle(my.mc_cond);
8821 CloseHandle(my.mc_mutex);
8822 _aligned_free(my.mc_wbuf[0]);
8824 pthread_cond_destroy(&my.mc_cond);
8825 pthread_mutex_destroy(&my.mc_mutex);
8826 free(my.mc_wbuf[0]);
8831 /** Copy environment as-is. */
8833 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8835 MDB_txn *txn = NULL;
8836 mdb_mutex_t *wmutex = NULL;
8842 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8846 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8849 /* Do the lock/unlock of the reader mutex before starting the
8850 * write txn. Otherwise other read txns could block writers.
8852 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8857 /* We must start the actual read txn after blocking writers */
8858 mdb_txn_reset0(txn, "reset-stage1");
8860 /* Temporarily block writers until we snapshot the meta pages */
8861 wmutex = MDB_MUTEX(env, w);
8862 if (LOCK_MUTEX(rc, env, wmutex))
8865 rc = mdb_txn_renew0(txn);
8867 UNLOCK_MUTEX(wmutex);
8872 wsize = env->me_psize * 2;
8876 DO_WRITE(rc, fd, ptr, w2, len);
8880 } else if (len > 0) {
8886 /* Non-blocking or async handles are not supported */
8892 UNLOCK_MUTEX(wmutex);
8897 w2 = txn->mt_next_pgno * env->me_psize;
8900 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8907 if (wsize > MAX_WRITE)
8911 DO_WRITE(rc, fd, ptr, w2, len);
8915 } else if (len > 0) {
8932 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8934 if (flags & MDB_CP_COMPACT)
8935 return mdb_env_copyfd1(env, fd);
8937 return mdb_env_copyfd0(env, fd);
8941 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8943 return mdb_env_copyfd2(env, fd, 0);
8947 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8951 HANDLE newfd = INVALID_HANDLE_VALUE;
8953 if (env->me_flags & MDB_NOSUBDIR) {
8954 lpath = (char *)path;
8957 len += sizeof(DATANAME);
8958 lpath = malloc(len);
8961 sprintf(lpath, "%s" DATANAME, path);
8964 /* The destination path must exist, but the destination file must not.
8965 * We don't want the OS to cache the writes, since the source data is
8966 * already in the OS cache.
8969 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8970 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8972 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8974 if (newfd == INVALID_HANDLE_VALUE) {
8979 if (env->me_psize >= env->me_os_psize) {
8981 /* Set O_DIRECT if the file system supports it */
8982 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8983 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8985 #ifdef F_NOCACHE /* __APPLE__ */
8986 rc = fcntl(newfd, F_NOCACHE, 1);
8994 rc = mdb_env_copyfd2(env, newfd, flags);
8997 if (!(env->me_flags & MDB_NOSUBDIR))
8999 if (newfd != INVALID_HANDLE_VALUE)
9000 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9007 mdb_env_copy(MDB_env *env, const char *path)
9009 return mdb_env_copy2(env, path, 0);
9013 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9015 if ((flag & CHANGEABLE) != flag)
9018 env->me_flags |= flag;
9020 env->me_flags &= ~flag;
9025 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9030 *arg = env->me_flags;
9035 mdb_env_set_userctx(MDB_env *env, void *ctx)
9039 env->me_userctx = ctx;
9044 mdb_env_get_userctx(MDB_env *env)
9046 return env ? env->me_userctx : NULL;
9050 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9055 env->me_assert_func = func;
9061 mdb_env_get_path(MDB_env *env, const char **arg)
9066 *arg = env->me_path;
9071 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9080 /** Common code for #mdb_stat() and #mdb_env_stat().
9081 * @param[in] env the environment to operate in.
9082 * @param[in] db the #MDB_db record containing the stats to return.
9083 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9084 * @return 0, this function always succeeds.
9087 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9089 arg->ms_psize = env->me_psize;
9090 arg->ms_depth = db->md_depth;
9091 arg->ms_branch_pages = db->md_branch_pages;
9092 arg->ms_leaf_pages = db->md_leaf_pages;
9093 arg->ms_overflow_pages = db->md_overflow_pages;
9094 arg->ms_entries = db->md_entries;
9100 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9104 if (env == NULL || arg == NULL)
9107 toggle = mdb_env_pick_meta(env);
9109 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9113 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9117 if (env == NULL || arg == NULL)
9120 toggle = mdb_env_pick_meta(env);
9121 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9122 arg->me_mapsize = env->me_mapsize;
9123 arg->me_maxreaders = env->me_maxreaders;
9124 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9126 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9127 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9131 /** Set the default comparison functions for a database.
9132 * Called immediately after a database is opened to set the defaults.
9133 * The user can then override them with #mdb_set_compare() or
9134 * #mdb_set_dupsort().
9135 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9136 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9139 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9141 uint16_t f = txn->mt_dbs[dbi].md_flags;
9143 txn->mt_dbxs[dbi].md_cmp =
9144 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9145 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9147 txn->mt_dbxs[dbi].md_dcmp =
9148 !(f & MDB_DUPSORT) ? 0 :
9149 ((f & MDB_INTEGERDUP)
9150 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9151 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9154 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9160 int rc, dbflag, exact;
9161 unsigned int unused = 0, seq;
9164 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9165 mdb_default_cmp(txn, FREE_DBI);
9168 if ((flags & VALID_FLAGS) != flags)
9170 if (txn->mt_flags & MDB_TXN_ERROR)
9176 if (flags & PERSISTENT_FLAGS) {
9177 uint16_t f2 = flags & PERSISTENT_FLAGS;
9178 /* make sure flag changes get committed */
9179 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9180 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9181 txn->mt_flags |= MDB_TXN_DIRTY;
9184 mdb_default_cmp(txn, MAIN_DBI);
9188 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9189 mdb_default_cmp(txn, MAIN_DBI);
9192 /* Is the DB already open? */
9194 for (i=2; i<txn->mt_numdbs; i++) {
9195 if (!txn->mt_dbxs[i].md_name.mv_size) {
9196 /* Remember this free slot */
9197 if (!unused) unused = i;
9200 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9201 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9207 /* If no free slot and max hit, fail */
9208 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9209 return MDB_DBS_FULL;
9211 /* Cannot mix named databases with some mainDB flags */
9212 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9213 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9215 /* Find the DB info */
9216 dbflag = DB_NEW|DB_VALID;
9219 key.mv_data = (void *)name;
9220 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9221 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9222 if (rc == MDB_SUCCESS) {
9223 /* make sure this is actually a DB */
9224 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9225 if (!(node->mn_flags & F_SUBDATA))
9226 return MDB_INCOMPATIBLE;
9227 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9228 /* Create if requested */
9229 data.mv_size = sizeof(MDB_db);
9230 data.mv_data = &dummy;
9231 memset(&dummy, 0, sizeof(dummy));
9232 dummy.md_root = P_INVALID;
9233 dummy.md_flags = flags & PERSISTENT_FLAGS;
9234 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9238 /* OK, got info, add to table */
9239 if (rc == MDB_SUCCESS) {
9240 unsigned int slot = unused ? unused : txn->mt_numdbs;
9241 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9242 txn->mt_dbxs[slot].md_name.mv_size = len;
9243 txn->mt_dbxs[slot].md_rel = NULL;
9244 txn->mt_dbflags[slot] = dbflag;
9245 /* txn-> and env-> are the same in read txns, use
9246 * tmp variable to avoid undefined assignment
9248 seq = ++txn->mt_env->me_dbiseqs[slot];
9249 txn->mt_dbiseqs[slot] = seq;
9251 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9253 mdb_default_cmp(txn, slot);
9262 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9264 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9267 if (txn->mt_flags & MDB_TXN_ERROR)
9270 if (txn->mt_dbflags[dbi] & DB_STALE) {
9273 /* Stale, must read the DB's root. cursor_init does it for us. */
9274 mdb_cursor_init(&mc, txn, dbi, &mx);
9276 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9279 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9282 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9284 ptr = env->me_dbxs[dbi].md_name.mv_data;
9285 /* If there was no name, this was already closed */
9287 env->me_dbxs[dbi].md_name.mv_data = NULL;
9288 env->me_dbxs[dbi].md_name.mv_size = 0;
9289 env->me_dbflags[dbi] = 0;
9290 env->me_dbiseqs[dbi]++;
9295 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9297 /* We could return the flags for the FREE_DBI too but what's the point? */
9298 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9300 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9304 /** Add all the DB's pages to the free list.
9305 * @param[in] mc Cursor on the DB to free.
9306 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9307 * @return 0 on success, non-zero on failure.
9310 mdb_drop0(MDB_cursor *mc, int subs)
9314 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9315 if (rc == MDB_SUCCESS) {
9316 MDB_txn *txn = mc->mc_txn;
9321 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9322 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9325 mdb_cursor_copy(mc, &mx);
9326 while (mc->mc_snum > 0) {
9327 MDB_page *mp = mc->mc_pg[mc->mc_top];
9328 unsigned n = NUMKEYS(mp);
9330 for (i=0; i<n; i++) {
9331 ni = NODEPTR(mp, i);
9332 if (ni->mn_flags & F_BIGDATA) {
9335 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9336 rc = mdb_page_get(txn, pg, &omp, NULL);
9339 mdb_cassert(mc, IS_OVERFLOW(omp));
9340 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9344 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9345 mdb_xcursor_init1(mc, ni);
9346 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9352 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9354 for (i=0; i<n; i++) {
9356 ni = NODEPTR(mp, i);
9359 mdb_midl_xappend(txn->mt_free_pgs, pg);
9364 mc->mc_ki[mc->mc_top] = i;
9365 rc = mdb_cursor_sibling(mc, 1);
9367 if (rc != MDB_NOTFOUND)
9369 /* no more siblings, go back to beginning
9370 * of previous level.
9374 for (i=1; i<mc->mc_snum; i++) {
9376 mc->mc_pg[i] = mx.mc_pg[i];
9381 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9384 txn->mt_flags |= MDB_TXN_ERROR;
9385 } else if (rc == MDB_NOTFOUND) {
9391 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9393 MDB_cursor *mc, *m2;
9396 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9399 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9402 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9405 rc = mdb_cursor_open(txn, dbi, &mc);
9409 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9410 /* Invalidate the dropped DB's cursors */
9411 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9412 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9416 /* Can't delete the main DB */
9417 if (del && dbi > MAIN_DBI) {
9418 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9420 txn->mt_dbflags[dbi] = DB_STALE;
9421 mdb_dbi_close(txn->mt_env, dbi);
9423 txn->mt_flags |= MDB_TXN_ERROR;
9426 /* reset the DB record, mark it dirty */
9427 txn->mt_dbflags[dbi] |= DB_DIRTY;
9428 txn->mt_dbs[dbi].md_depth = 0;
9429 txn->mt_dbs[dbi].md_branch_pages = 0;
9430 txn->mt_dbs[dbi].md_leaf_pages = 0;
9431 txn->mt_dbs[dbi].md_overflow_pages = 0;
9432 txn->mt_dbs[dbi].md_entries = 0;
9433 txn->mt_dbs[dbi].md_root = P_INVALID;
9435 txn->mt_flags |= MDB_TXN_DIRTY;
9438 mdb_cursor_close(mc);
9442 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9444 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9447 txn->mt_dbxs[dbi].md_cmp = cmp;
9451 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9453 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9456 txn->mt_dbxs[dbi].md_dcmp = cmp;
9460 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9462 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9465 txn->mt_dbxs[dbi].md_rel = rel;
9469 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9471 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9474 txn->mt_dbxs[dbi].md_relctx = ctx;
9479 mdb_env_get_maxkeysize(MDB_env *env)
9481 return ENV_MAXKEY(env);
9485 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9487 unsigned int i, rdrs;
9490 int rc = 0, first = 1;
9494 if (!env->me_txns) {
9495 return func("(no reader locks)\n", ctx);
9497 rdrs = env->me_txns->mti_numreaders;
9498 mr = env->me_txns->mti_readers;
9499 for (i=0; i<rdrs; i++) {
9501 txnid_t txnid = mr[i].mr_txnid;
9502 sprintf(buf, txnid == (txnid_t)-1 ?
9503 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9504 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9507 rc = func(" pid thread txnid\n", ctx);
9511 rc = func(buf, ctx);
9517 rc = func("(no active readers)\n", ctx);
9522 /** Insert pid into list if not already present.
9523 * return -1 if already present.
9526 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9528 /* binary search of pid in list */
9530 unsigned cursor = 1;
9532 unsigned n = ids[0];
9535 unsigned pivot = n >> 1;
9536 cursor = base + pivot + 1;
9537 val = pid - ids[cursor];
9542 } else if ( val > 0 ) {
9547 /* found, so it's a duplicate */
9556 for (n = ids[0]; n > cursor; n--)
9563 mdb_reader_check(MDB_env *env, int *dead)
9569 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
9572 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
9573 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
9575 mdb_mutex_t *rmutex = rlocked ? NULL : MDB_MUTEX(env, r);
9576 unsigned int i, j, rdrs;
9578 MDB_PID_T *pids, pid;
9579 int rc = MDB_SUCCESS, count = 0;
9581 rdrs = env->me_txns->mti_numreaders;
9582 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9586 mr = env->me_txns->mti_readers;
9587 for (i=0; i<rdrs; i++) {
9589 if (pid && pid != env->me_pid) {
9590 if (mdb_pid_insert(pids, pid) == 0) {
9591 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9592 /* Stale reader found */
9595 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
9596 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
9598 rdrs = 0; /* the above checked all readers */
9600 /* Recheck, a new process may have reused pid */
9601 if (mdb_reader_pid(env, Pidcheck, pid))
9606 if (mr[j].mr_pid == pid) {
9607 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9608 (unsigned) pid, mr[j].mr_txnid));
9613 UNLOCK_MUTEX(rmutex);
9624 #ifdef MDB_ROBUST_SUPPORTED
9625 /** Handle #LOCK_MUTEX0() failure.
9626 * Try to repair the lock file if the mutex owner died.
9627 * @param[in] env the environment handle
9628 * @param[in] mutex LOCK_MUTEX0() mutex
9629 * @param[in] rc LOCK_MUTEX0() error (nonzero)
9630 * @return 0 on success with the mutex locked, or an error code on failure.
9632 static int mdb_mutex_failed(MDB_env *env, mdb_mutex_t *mutex, int rc)
9634 int toggle, rlocked, rc2;
9636 if (rc == MDB_OWNERDEAD) {
9637 /* We own the mutex. Clean up after dead previous owner. */
9639 rlocked = (mutex == MDB_MUTEX(env, r));
9641 /* Keep mti_txnid updated, otherwise next writer can
9642 * overwrite data which latest meta page refers to.
9644 toggle = mdb_env_pick_meta(env);
9645 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
9646 /* env is hosed if the dead thread was ours */
9648 env->me_flags |= MDB_FATAL_ERROR;
9653 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
9654 (rc ? "this process' env is hosed" : "recovering")));
9655 rc2 = mdb_reader_check0(env, rlocked, NULL);
9657 rc2 = mdb_mutex_consistent(mutex);
9658 if (rc || (rc = rc2)) {
9659 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
9660 UNLOCK_MUTEX(mutex);
9666 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
9671 #endif /* MDB_ROBUST_SUPPORTED */
projects / openldap / blob