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)
92 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
93 #include <netinet/in.h>
94 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
97 #if defined(__APPLE__) || defined (BSD)
98 # define MDB_USE_POSIX_SEM 1
99 # define MDB_FDATASYNC fsync
100 #elif defined(ANDROID)
101 # define MDB_FDATASYNC fsync
106 #ifdef MDB_USE_POSIX_SEM
107 # define MDB_USE_HASH 1
108 #include <semaphore.h>
113 #include <valgrind/memcheck.h>
114 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
115 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
116 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
117 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
118 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
120 #define VGMEMP_CREATE(h,r,z)
121 #define VGMEMP_ALLOC(h,a,s)
122 #define VGMEMP_FREE(h,a)
123 #define VGMEMP_DESTROY(h)
124 #define VGMEMP_DEFINED(a,s)
128 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
129 /* Solaris just defines one or the other */
130 # define LITTLE_ENDIAN 1234
131 # define BIG_ENDIAN 4321
132 # ifdef _LITTLE_ENDIAN
133 # define BYTE_ORDER LITTLE_ENDIAN
135 # define BYTE_ORDER BIG_ENDIAN
138 # define BYTE_ORDER __BYTE_ORDER
142 #ifndef LITTLE_ENDIAN
143 #define LITTLE_ENDIAN __LITTLE_ENDIAN
146 #define BIG_ENDIAN __BIG_ENDIAN
149 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
150 #define MISALIGNED_OK 1
156 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
157 # error "Unknown or unsupported endianness (BYTE_ORDER)"
158 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
159 # error "Two's complement, reasonably sized integer types, please"
163 /** Put infrequently used env functions in separate section */
165 # define ESECT __attribute__ ((section("__TEXT,text_env")))
167 # define ESECT __attribute__ ((section("text_env")))
173 /** @defgroup internal LMDB Internals
176 /** @defgroup compat Compatibility Macros
177 * A bunch of macros to minimize the amount of platform-specific ifdefs
178 * needed throughout the rest of the code. When the features this library
179 * needs are similar enough to POSIX to be hidden in a one-or-two line
180 * replacement, this macro approach is used.
184 /** Features under development */
189 /** Wrapper around __func__, which is a C99 feature */
190 #if __STDC_VERSION__ >= 199901L
191 # define mdb_func_ __func__
192 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
193 # define mdb_func_ __FUNCTION__
195 /* If a debug message says <mdb_unknown>(), update the #if statements above */
196 # define mdb_func_ "<mdb_unknown>"
200 #define MDB_USE_HASH 1
201 #define MDB_PIDLOCK 0
202 #define THREAD_RET DWORD
203 #define pthread_t HANDLE
204 #define pthread_mutex_t HANDLE
205 #define pthread_cond_t HANDLE
206 #define pthread_key_t DWORD
207 #define pthread_self() GetCurrentThreadId()
208 #define pthread_key_create(x,y) \
209 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
210 #define pthread_key_delete(x) TlsFree(x)
211 #define pthread_getspecific(x) TlsGetValue(x)
212 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
213 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
214 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
215 #define pthread_cond_signal(x) SetEvent(*x)
216 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
217 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
218 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
219 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
220 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
221 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
222 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
223 #define getpid() GetCurrentProcessId()
224 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
225 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
226 #define ErrCode() GetLastError()
227 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
228 #define close(fd) (CloseHandle(fd) ? 0 : -1)
229 #define munmap(ptr,len) UnmapViewOfFile(ptr)
230 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
231 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
233 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
237 #define THREAD_RET void *
238 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
239 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
240 #define Z "z" /**< printf format modifier for size_t */
242 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
243 #define MDB_PIDLOCK 1
245 #ifdef MDB_USE_POSIX_SEM
247 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
248 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
249 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
250 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
253 mdb_sem_wait(sem_t *sem)
256 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
261 /** Lock the reader mutex.
263 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
264 /** Unlock the reader mutex.
266 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
268 /** Lock the writer mutex.
269 * Only a single write transaction is allowed at a time. Other writers
270 * will block waiting for this mutex.
272 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
273 /** Unlock the writer mutex.
275 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
276 #endif /* MDB_USE_POSIX_SEM */
278 /** Get the error code for the last failed system function.
280 #define ErrCode() errno
282 /** An abstraction for a file handle.
283 * On POSIX systems file handles are small integers. On Windows
284 * they're opaque pointers.
288 /** A value for an invalid file handle.
289 * Mainly used to initialize file variables and signify that they are
292 #define INVALID_HANDLE_VALUE (-1)
294 /** Get the size of a memory page for the system.
295 * This is the basic size that the platform's memory manager uses, and is
296 * fundamental to the use of memory-mapped files.
298 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
301 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
304 #define MNAME_LEN (sizeof(pthread_mutex_t))
310 /** A flag for opening a file and requesting synchronous data writes.
311 * This is only used when writing a meta page. It's not strictly needed;
312 * we could just do a normal write and then immediately perform a flush.
313 * But if this flag is available it saves us an extra system call.
315 * @note If O_DSYNC is undefined but exists in /usr/include,
316 * preferably set some compiler flag to get the definition.
317 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
320 # define MDB_DSYNC O_DSYNC
324 /** Function for flushing the data of a file. Define this to fsync
325 * if fdatasync() is not supported.
327 #ifndef MDB_FDATASYNC
328 # define MDB_FDATASYNC fdatasync
332 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
343 /** A page number in the database.
344 * Note that 64 bit page numbers are overkill, since pages themselves
345 * already represent 12-13 bits of addressable memory, and the OS will
346 * always limit applications to a maximum of 63 bits of address space.
348 * @note In the #MDB_node structure, we only store 48 bits of this value,
349 * which thus limits us to only 60 bits of addressable data.
351 typedef MDB_ID pgno_t;
353 /** A transaction ID.
354 * See struct MDB_txn.mt_txnid for details.
356 typedef MDB_ID txnid_t;
358 /** @defgroup debug Debug Macros
362 /** Enable debug output. Needs variable argument macros (a C99 feature).
363 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
364 * read from and written to the database (used for free space management).
370 static int mdb_debug;
371 static txnid_t mdb_debug_start;
373 /** Print a debug message with printf formatting.
374 * Requires double parenthesis around 2 or more args.
376 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
377 # define DPRINTF0(fmt, ...) \
378 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
380 # define DPRINTF(args) ((void) 0)
382 /** Print a debug string.
383 * The string is printed literally, with no format processing.
385 #define DPUTS(arg) DPRINTF(("%s", arg))
386 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
388 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
391 /** @brief The maximum size of a database page.
393 * It is 32k or 64k, since value-PAGEBASE must fit in
394 * #MDB_page.%mp_upper.
396 * LMDB will use database pages < OS pages if needed.
397 * That causes more I/O in write transactions: The OS must
398 * know (read) the whole page before writing a partial page.
400 * Note that we don't currently support Huge pages. On Linux,
401 * regular data files cannot use Huge pages, and in general
402 * Huge pages aren't actually pageable. We rely on the OS
403 * demand-pager to read our data and page it out when memory
404 * pressure from other processes is high. So until OSs have
405 * actual paging support for Huge pages, they're not viable.
407 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
409 /** The minimum number of keys required in a database page.
410 * Setting this to a larger value will place a smaller bound on the
411 * maximum size of a data item. Data items larger than this size will
412 * be pushed into overflow pages instead of being stored directly in
413 * the B-tree node. This value used to default to 4. With a page size
414 * of 4096 bytes that meant that any item larger than 1024 bytes would
415 * go into an overflow page. That also meant that on average 2-3KB of
416 * each overflow page was wasted space. The value cannot be lower than
417 * 2 because then there would no longer be a tree structure. With this
418 * value, items larger than 2KB will go into overflow pages, and on
419 * average only 1KB will be wasted.
421 #define MDB_MINKEYS 2
423 /** A stamp that identifies a file as an LMDB file.
424 * There's nothing special about this value other than that it is easily
425 * recognizable, and it will reflect any byte order mismatches.
427 #define MDB_MAGIC 0xBEEFC0DE
429 /** The version number for a database's datafile format. */
430 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
431 /** The version number for a database's lockfile format. */
432 #define MDB_LOCK_VERSION 1
434 /** @brief The max size of a key we can write, or 0 for dynamic max.
436 * Define this as 0 to compute the max from the page size. 511
437 * is default for backwards compat: liblmdb <= 0.9.10 can break
438 * when modifying a DB with keys/dupsort data bigger than its max.
439 * #MDB_DEVEL sets the default to 0.
441 * Data items in an #MDB_DUPSORT database are also limited to
442 * this size, since they're actually keys of a sub-DB. Keys and
443 * #MDB_DUPSORT data items must fit on a node in a regular page.
445 #ifndef MDB_MAXKEYSIZE
446 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
449 /** The maximum size of a key we can write to the environment. */
451 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
453 #define ENV_MAXKEY(env) ((env)->me_maxkey)
456 /** @brief The maximum size of a data item.
458 * We only store a 32 bit value for node sizes.
460 #define MAXDATASIZE 0xffffffffUL
463 /** Key size which fits in a #DKBUF.
466 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
469 * This is used for printing a hex dump of a key's contents.
471 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
472 /** Display a key in hex.
474 * Invoke a function to display a key in hex.
476 #define DKEY(x) mdb_dkey(x, kbuf)
482 /** An invalid page number.
483 * Mainly used to denote an empty tree.
485 #define P_INVALID (~(pgno_t)0)
487 /** Test if the flags \b f are set in a flag word \b w. */
488 #define F_ISSET(w, f) (((w) & (f)) == (f))
490 /** Round \b n up to an even number. */
491 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
493 /** Used for offsets within a single page.
494 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
497 typedef uint16_t indx_t;
499 /** Default size of memory map.
500 * This is certainly too small for any actual applications. Apps should always set
501 * the size explicitly using #mdb_env_set_mapsize().
503 #define DEFAULT_MAPSIZE 1048576
505 /** @defgroup readers Reader Lock Table
506 * Readers don't acquire any locks for their data access. Instead, they
507 * simply record their transaction ID in the reader table. The reader
508 * mutex is needed just to find an empty slot in the reader table. The
509 * slot's address is saved in thread-specific data so that subsequent read
510 * transactions started by the same thread need no further locking to proceed.
512 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
514 * No reader table is used if the database is on a read-only filesystem, or
515 * if #MDB_NOLOCK is set.
517 * Since the database uses multi-version concurrency control, readers don't
518 * actually need any locking. This table is used to keep track of which
519 * readers are using data from which old transactions, so that we'll know
520 * when a particular old transaction is no longer in use. Old transactions
521 * that have discarded any data pages can then have those pages reclaimed
522 * for use by a later write transaction.
524 * The lock table is constructed such that reader slots are aligned with the
525 * processor's cache line size. Any slot is only ever used by one thread.
526 * This alignment guarantees that there will be no contention or cache
527 * thrashing as threads update their own slot info, and also eliminates
528 * any need for locking when accessing a slot.
530 * A writer thread will scan every slot in the table to determine the oldest
531 * outstanding reader transaction. Any freed pages older than this will be
532 * reclaimed by the writer. The writer doesn't use any locks when scanning
533 * this table. This means that there's no guarantee that the writer will
534 * see the most up-to-date reader info, but that's not required for correct
535 * operation - all we need is to know the upper bound on the oldest reader,
536 * we don't care at all about the newest reader. So the only consequence of
537 * reading stale information here is that old pages might hang around a
538 * while longer before being reclaimed. That's actually good anyway, because
539 * the longer we delay reclaiming old pages, the more likely it is that a
540 * string of contiguous pages can be found after coalescing old pages from
541 * many old transactions together.
544 /** Number of slots in the reader table.
545 * This value was chosen somewhat arbitrarily. 126 readers plus a
546 * couple mutexes fit exactly into 8KB on my development machine.
547 * Applications should set the table size using #mdb_env_set_maxreaders().
549 #define DEFAULT_READERS 126
551 /** The size of a CPU cache line in bytes. We want our lock structures
552 * aligned to this size to avoid false cache line sharing in the
554 * This value works for most CPUs. For Itanium this should be 128.
560 /** The information we store in a single slot of the reader table.
561 * In addition to a transaction ID, we also record the process and
562 * thread ID that owns a slot, so that we can detect stale information,
563 * e.g. threads or processes that went away without cleaning up.
564 * @note We currently don't check for stale records. We simply re-init
565 * the table when we know that we're the only process opening the
568 typedef struct MDB_rxbody {
569 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
570 * Multiple readers that start at the same time will probably have the
571 * same ID here. Again, it's not important to exclude them from
572 * anything; all we need to know is which version of the DB they
573 * started from so we can avoid overwriting any data used in that
574 * particular version.
577 /** The process ID of the process owning this reader txn. */
579 /** The thread ID of the thread owning this txn. */
583 /** The actual reader record, with cacheline padding. */
584 typedef struct MDB_reader {
587 /** shorthand for mrb_txnid */
588 #define mr_txnid mru.mrx.mrb_txnid
589 #define mr_pid mru.mrx.mrb_pid
590 #define mr_tid mru.mrx.mrb_tid
591 /** cache line alignment */
592 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
596 /** The header for the reader table.
597 * The table resides in a memory-mapped file. (This is a different file
598 * than is used for the main database.)
600 * For POSIX the actual mutexes reside in the shared memory of this
601 * mapped file. On Windows, mutexes are named objects allocated by the
602 * kernel; we store the mutex names in this mapped file so that other
603 * processes can grab them. This same approach is also used on
604 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
605 * process-shared POSIX mutexes. For these cases where a named object
606 * is used, the object name is derived from a 64 bit FNV hash of the
607 * environment pathname. As such, naming collisions are extremely
608 * unlikely. If a collision occurs, the results are unpredictable.
610 typedef struct MDB_txbody {
611 /** Stamp identifying this as an LMDB file. It must be set
614 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
616 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
617 char mtb_rmname[MNAME_LEN];
619 /** Mutex protecting access to this table.
620 * This is the reader lock that #LOCK_MUTEX_R acquires.
622 pthread_mutex_t mtb_mutex;
624 /** The ID of the last transaction committed to the database.
625 * This is recorded here only for convenience; the value can always
626 * be determined by reading the main database meta pages.
629 /** The number of slots that have been used in the reader table.
630 * This always records the maximum count, it is not decremented
631 * when readers release their slots.
633 unsigned mtb_numreaders;
636 /** The actual reader table definition. */
637 typedef struct MDB_txninfo {
640 #define mti_magic mt1.mtb.mtb_magic
641 #define mti_format mt1.mtb.mtb_format
642 #define mti_mutex mt1.mtb.mtb_mutex
643 #define mti_rmname mt1.mtb.mtb_rmname
644 #define mti_txnid mt1.mtb.mtb_txnid
645 #define mti_numreaders mt1.mtb.mtb_numreaders
646 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
649 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
650 char mt2_wmname[MNAME_LEN];
651 #define mti_wmname mt2.mt2_wmname
653 pthread_mutex_t mt2_wmutex;
654 #define mti_wmutex mt2.mt2_wmutex
656 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
658 MDB_reader mti_readers[1];
661 /** Lockfile format signature: version, features and field layout */
662 #define MDB_LOCK_FORMAT \
664 ((MDB_LOCK_VERSION) \
665 /* Flags which describe functionality */ \
666 + (((MDB_PIDLOCK) != 0) << 16)))
669 /** Common header for all page types.
670 * Overflow records occupy a number of contiguous pages with no
671 * headers on any page after the first.
673 typedef struct MDB_page {
674 #define mp_pgno mp_p.p_pgno
675 #define mp_next mp_p.p_next
677 pgno_t p_pgno; /**< page number */
678 struct MDB_page *p_next; /**< for in-memory list of freed pages */
681 /** @defgroup mdb_page Page Flags
683 * Flags for the page headers.
686 #define P_BRANCH 0x01 /**< branch page */
687 #define P_LEAF 0x02 /**< leaf page */
688 #define P_OVERFLOW 0x04 /**< overflow page */
689 #define P_META 0x08 /**< meta page */
690 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
691 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
692 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
693 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
694 #define P_KEEP 0x8000 /**< leave this page alone during spill */
696 uint16_t mp_flags; /**< @ref mdb_page */
697 #define mp_lower mp_pb.pb.pb_lower
698 #define mp_upper mp_pb.pb.pb_upper
699 #define mp_pages mp_pb.pb_pages
702 indx_t pb_lower; /**< lower bound of free space */
703 indx_t pb_upper; /**< upper bound of free space */
705 uint32_t pb_pages; /**< number of overflow pages */
707 indx_t mp_ptrs[1]; /**< dynamic size */
710 /** Size of the page header, excluding dynamic data at the end */
711 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
713 /** Address of first usable data byte in a page, after the header */
714 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
716 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
717 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
719 /** Number of nodes on a page */
720 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
722 /** The amount of space remaining in the page */
723 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
725 /** The percentage of space used in the page, in tenths of a percent. */
726 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
727 ((env)->me_psize - PAGEHDRSZ))
728 /** The minimum page fill factor, in tenths of a percent.
729 * Pages emptier than this are candidates for merging.
731 #define FILL_THRESHOLD 250
733 /** Test if a page is a leaf page */
734 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
735 /** Test if a page is a LEAF2 page */
736 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
737 /** Test if a page is a branch page */
738 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
739 /** Test if a page is an overflow page */
740 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
741 /** Test if a page is a sub page */
742 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
744 /** The number of overflow pages needed to store the given size. */
745 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
747 /** Link in #MDB_txn.%mt_loose_pages list */
748 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
750 /** Header for a single key/data pair within a page.
751 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
752 * We guarantee 2-byte alignment for 'MDB_node's.
754 typedef struct MDB_node {
755 /** lo and hi are used for data size on leaf nodes and for
756 * child pgno on branch nodes. On 64 bit platforms, flags
757 * is also used for pgno. (Branch nodes have no flags).
758 * They are in host byte order in case that lets some
759 * accesses be optimized into a 32-bit word access.
761 #if BYTE_ORDER == LITTLE_ENDIAN
762 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
764 unsigned short mn_hi, mn_lo;
766 /** @defgroup mdb_node Node Flags
768 * Flags for node headers.
771 #define F_BIGDATA 0x01 /**< data put on overflow page */
772 #define F_SUBDATA 0x02 /**< data is a sub-database */
773 #define F_DUPDATA 0x04 /**< data has duplicates */
775 /** valid flags for #mdb_node_add() */
776 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
779 unsigned short mn_flags; /**< @ref mdb_node */
780 unsigned short mn_ksize; /**< key size */
781 char mn_data[1]; /**< key and data are appended here */
784 /** Size of the node header, excluding dynamic data at the end */
785 #define NODESIZE offsetof(MDB_node, mn_data)
787 /** Bit position of top word in page number, for shifting mn_flags */
788 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
790 /** Size of a node in a branch page with a given key.
791 * This is just the node header plus the key, there is no data.
793 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
795 /** Size of a node in a leaf page with a given key and data.
796 * This is node header plus key plus data size.
798 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
800 /** Address of node \b i in page \b p */
801 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
803 /** Address of the key for the node */
804 #define NODEKEY(node) (void *)((node)->mn_data)
806 /** Address of the data for a node */
807 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
809 /** Get the page number pointed to by a branch node */
810 #define NODEPGNO(node) \
811 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
812 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
813 /** Set the page number in a branch node */
814 #define SETPGNO(node,pgno) do { \
815 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
816 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
818 /** Get the size of the data in a leaf node */
819 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
820 /** Set the size of the data for a leaf node */
821 #define SETDSZ(node,size) do { \
822 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
823 /** The size of a key in a node */
824 #define NODEKSZ(node) ((node)->mn_ksize)
826 /** Copy a page number from src to dst */
828 #define COPY_PGNO(dst,src) dst = src
830 #if SIZE_MAX > 4294967295UL
831 #define COPY_PGNO(dst,src) do { \
832 unsigned short *s, *d; \
833 s = (unsigned short *)&(src); \
834 d = (unsigned short *)&(dst); \
841 #define COPY_PGNO(dst,src) do { \
842 unsigned short *s, *d; \
843 s = (unsigned short *)&(src); \
844 d = (unsigned short *)&(dst); \
850 /** The address of a key in a LEAF2 page.
851 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
852 * There are no node headers, keys are stored contiguously.
854 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
856 /** Set the \b node's key into \b keyptr, if requested. */
857 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
858 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
860 /** Set the \b node's key into \b key. */
861 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
863 /** Information about a single database in the environment. */
864 typedef struct MDB_db {
865 uint32_t md_pad; /**< also ksize for LEAF2 pages */
866 uint16_t md_flags; /**< @ref mdb_dbi_open */
867 uint16_t md_depth; /**< depth of this tree */
868 pgno_t md_branch_pages; /**< number of internal pages */
869 pgno_t md_leaf_pages; /**< number of leaf pages */
870 pgno_t md_overflow_pages; /**< number of overflow pages */
871 size_t md_entries; /**< number of data items */
872 pgno_t md_root; /**< the root page of this tree */
875 /** mdb_dbi_open flags */
876 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
877 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
878 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
879 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
881 /** Handle for the DB used to track free pages. */
883 /** Handle for the default DB. */
886 /** Meta page content.
887 * A meta page is the start point for accessing a database snapshot.
888 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
890 typedef struct MDB_meta {
891 /** Stamp identifying this as an LMDB file. It must be set
894 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
896 void *mm_address; /**< address for fixed mapping */
897 size_t mm_mapsize; /**< size of mmap region */
898 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
899 /** The size of pages used in this DB */
900 #define mm_psize mm_dbs[0].md_pad
901 /** Any persistent environment flags. @ref mdb_env */
902 #define mm_flags mm_dbs[0].md_flags
903 pgno_t mm_last_pg; /**< last used page in file */
904 txnid_t mm_txnid; /**< txnid that committed this page */
907 /** Buffer for a stack-allocated meta page.
908 * The members define size and alignment, and silence type
909 * aliasing warnings. They are not used directly; that could
910 * mean incorrectly using several union members in parallel.
912 typedef union MDB_metabuf {
915 char mm_pad[PAGEHDRSZ];
920 /** Auxiliary DB info.
921 * The information here is mostly static/read-only. There is
922 * only a single copy of this record in the environment.
924 typedef struct MDB_dbx {
925 MDB_val md_name; /**< name of the database */
926 MDB_cmp_func *md_cmp; /**< function for comparing keys */
927 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
928 MDB_rel_func *md_rel; /**< user relocate function */
929 void *md_relctx; /**< user-provided context for md_rel */
932 /** A database transaction.
933 * Every operation requires a transaction handle.
936 MDB_txn *mt_parent; /**< parent of a nested txn */
937 MDB_txn *mt_child; /**< nested txn under this txn */
938 pgno_t mt_next_pgno; /**< next unallocated page */
939 /** The ID of this transaction. IDs are integers incrementing from 1.
940 * Only committed write transactions increment the ID. If a transaction
941 * aborts, the ID may be re-used by the next writer.
944 MDB_env *mt_env; /**< the DB environment */
945 /** The list of pages that became unused during this transaction.
948 /** The list of loose pages that became unused and may be reused
949 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
951 MDB_page *mt_loose_pgs;
952 /** The sorted list of dirty pages we temporarily wrote to disk
953 * because the dirty list was full. page numbers in here are
954 * shifted left by 1, deleted slots have the LSB set.
956 MDB_IDL mt_spill_pgs;
958 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
960 /** For read txns: This thread/txn's reader table slot, or NULL. */
963 /** Array of records for each DB known in the environment. */
965 /** Array of MDB_db records for each known DB */
967 /** Array of sequence numbers for each DB handle */
968 unsigned int *mt_dbiseqs;
969 /** @defgroup mt_dbflag Transaction DB Flags
973 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
974 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
975 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
976 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
978 /** In write txns, array of cursors for each DB */
979 MDB_cursor **mt_cursors;
980 /** Array of flags for each DB */
981 unsigned char *mt_dbflags;
982 /** Number of DB records in use. This number only ever increments;
983 * we don't decrement it when individual DB handles are closed.
987 /** @defgroup mdb_txn Transaction Flags
991 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
992 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
993 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
994 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
996 unsigned int mt_flags; /**< @ref mdb_txn */
997 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
998 * Includes ancestor txns' dirty pages not hidden by other txns'
999 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1000 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1002 unsigned int mt_dirty_room;
1005 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1006 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1007 * raise this on a 64 bit machine.
1009 #define CURSOR_STACK 32
1013 /** Cursors are used for all DB operations.
1014 * A cursor holds a path of (page pointer, key index) from the DB
1015 * root to a position in the DB, plus other state. #MDB_DUPSORT
1016 * cursors include an xcursor to the current data item. Write txns
1017 * track their cursors and keep them up to date when data moves.
1018 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1019 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1022 /** Next cursor on this DB in this txn */
1023 MDB_cursor *mc_next;
1024 /** Backup of the original cursor if this cursor is a shadow */
1025 MDB_cursor *mc_backup;
1026 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1027 struct MDB_xcursor *mc_xcursor;
1028 /** The transaction that owns this cursor */
1030 /** The database handle this cursor operates on */
1032 /** The database record for this cursor */
1034 /** The database auxiliary record for this cursor */
1036 /** The @ref mt_dbflag for this database */
1037 unsigned char *mc_dbflag;
1038 unsigned short mc_snum; /**< number of pushed pages */
1039 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1040 /** @defgroup mdb_cursor Cursor Flags
1042 * Cursor state flags.
1045 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1046 #define C_EOF 0x02 /**< No more data */
1047 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1048 #define C_DEL 0x08 /**< last op was a cursor_del */
1049 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1050 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1052 unsigned int mc_flags; /**< @ref mdb_cursor */
1053 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1054 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1057 /** Context for sorted-dup records.
1058 * We could have gone to a fully recursive design, with arbitrarily
1059 * deep nesting of sub-databases. But for now we only handle these
1060 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1062 typedef struct MDB_xcursor {
1063 /** A sub-cursor for traversing the Dup DB */
1064 MDB_cursor mx_cursor;
1065 /** The database record for this Dup DB */
1067 /** The auxiliary DB record for this Dup DB */
1069 /** The @ref mt_dbflag for this Dup DB */
1070 unsigned char mx_dbflag;
1073 /** State of FreeDB old pages, stored in the MDB_env */
1074 typedef struct MDB_pgstate {
1075 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1076 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1079 /** The database environment. */
1081 HANDLE me_fd; /**< The main data file */
1082 HANDLE me_lfd; /**< The lock file */
1083 HANDLE me_mfd; /**< just for writing the meta pages */
1084 /** Failed to update the meta page. Probably an I/O error. */
1085 #define MDB_FATAL_ERROR 0x80000000U
1086 /** Some fields are initialized. */
1087 #define MDB_ENV_ACTIVE 0x20000000U
1088 /** me_txkey is set */
1089 #define MDB_ENV_TXKEY 0x10000000U
1090 uint32_t me_flags; /**< @ref mdb_env */
1091 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1092 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1093 unsigned int me_maxreaders; /**< size of the reader table */
1094 unsigned int me_numreaders; /**< max numreaders set by this env */
1095 MDB_dbi me_numdbs; /**< number of DBs opened */
1096 MDB_dbi me_maxdbs; /**< size of the DB table */
1097 MDB_PID_T me_pid; /**< process ID of this env */
1098 char *me_path; /**< path to the DB files */
1099 char *me_map; /**< the memory map of the data file */
1100 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1101 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1102 void *me_pbuf; /**< scratch area for DUPSORT put() */
1103 MDB_txn *me_txn; /**< current write transaction */
1104 size_t me_mapsize; /**< size of the data memory map */
1105 off_t me_size; /**< current file size */
1106 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1107 MDB_dbx *me_dbxs; /**< array of static DB info */
1108 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1109 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1110 pthread_key_t me_txkey; /**< thread-key for readers */
1111 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1112 # define me_pglast me_pgstate.mf_pglast
1113 # define me_pghead me_pgstate.mf_pghead
1114 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1115 /** IDL of pages that became unused in a write txn */
1116 MDB_IDL me_free_pgs;
1117 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1118 MDB_ID2L me_dirty_list;
1119 /** Max number of freelist items that can fit in a single overflow page */
1121 /** Max size of a node on a page */
1122 unsigned int me_nodemax;
1123 #if !(MDB_MAXKEYSIZE)
1124 unsigned int me_maxkey; /**< max size of a key */
1126 int me_live_reader; /**< have liveness lock in reader table */
1128 int me_pidquery; /**< Used in OpenProcess */
1129 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1131 #elif defined(MDB_USE_POSIX_SEM)
1132 sem_t *me_rmutex; /* Shared mutexes are not supported */
1135 void *me_userctx; /**< User-settable context */
1136 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1139 /** Nested transaction */
1140 typedef struct MDB_ntxn {
1141 MDB_txn mnt_txn; /**< the transaction */
1142 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1145 /** max number of pages to commit in one writev() call */
1146 #define MDB_COMMIT_PAGES 64
1147 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1148 #undef MDB_COMMIT_PAGES
1149 #define MDB_COMMIT_PAGES IOV_MAX
1152 /** max bytes to write in one call */
1153 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1155 /** Check \b txn and \b dbi arguments to a function */
1156 #define TXN_DBI_EXIST(txn, dbi) \
1157 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1159 /** Check for misused \b dbi handles */
1160 #define TXN_DBI_CHANGED(txn, dbi) \
1161 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1163 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1164 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1165 static int mdb_page_touch(MDB_cursor *mc);
1167 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1168 static int mdb_page_search_root(MDB_cursor *mc,
1169 MDB_val *key, int modify);
1170 #define MDB_PS_MODIFY 1
1171 #define MDB_PS_ROOTONLY 2
1172 #define MDB_PS_FIRST 4
1173 #define MDB_PS_LAST 8
1174 static int mdb_page_search(MDB_cursor *mc,
1175 MDB_val *key, int flags);
1176 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1178 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1179 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1180 pgno_t newpgno, unsigned int nflags);
1182 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1183 static int mdb_env_pick_meta(const MDB_env *env);
1184 static int mdb_env_write_meta(MDB_txn *txn);
1185 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1186 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1188 static void mdb_env_close0(MDB_env *env, int excl);
1190 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1191 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1192 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1193 static void mdb_node_del(MDB_cursor *mc, int ksize);
1194 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1195 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1196 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1197 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1198 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1200 static int mdb_rebalance(MDB_cursor *mc);
1201 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1203 static void mdb_cursor_pop(MDB_cursor *mc);
1204 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1206 static int mdb_cursor_del0(MDB_cursor *mc);
1207 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1208 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1209 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1210 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1211 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1213 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1214 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1216 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1217 static void mdb_xcursor_init0(MDB_cursor *mc);
1218 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1220 static int mdb_drop0(MDB_cursor *mc, int subs);
1221 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1224 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1228 static SECURITY_DESCRIPTOR mdb_null_sd;
1229 static SECURITY_ATTRIBUTES mdb_all_sa;
1230 static int mdb_sec_inited;
1233 /** Return the library version info. */
1235 mdb_version(int *major, int *minor, int *patch)
1237 if (major) *major = MDB_VERSION_MAJOR;
1238 if (minor) *minor = MDB_VERSION_MINOR;
1239 if (patch) *patch = MDB_VERSION_PATCH;
1240 return MDB_VERSION_STRING;
1243 /** Table of descriptions for LMDB @ref errors */
1244 static char *const mdb_errstr[] = {
1245 "MDB_KEYEXIST: Key/data pair already exists",
1246 "MDB_NOTFOUND: No matching key/data pair found",
1247 "MDB_PAGE_NOTFOUND: Requested page not found",
1248 "MDB_CORRUPTED: Located page was wrong type",
1249 "MDB_PANIC: Update of meta page failed",
1250 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1251 "MDB_INVALID: File is not an LMDB file",
1252 "MDB_MAP_FULL: Environment mapsize limit reached",
1253 "MDB_DBS_FULL: Environment maxdbs limit reached",
1254 "MDB_READERS_FULL: Environment maxreaders limit reached",
1255 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1256 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1257 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1258 "MDB_PAGE_FULL: Internal error - page has no more space",
1259 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1260 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1261 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1262 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1263 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1264 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1268 mdb_strerror(int err)
1271 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1272 * This works as long as no function between the call to mdb_strerror
1273 * and the actual use of the message uses more than 4K of stack.
1276 char buf[1024], *ptr = buf;
1280 return ("Successful return: 0");
1282 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1283 i = err - MDB_KEYEXIST;
1284 return mdb_errstr[i];
1288 /* These are the C-runtime error codes we use. The comment indicates
1289 * their numeric value, and the Win32 error they would correspond to
1290 * if the error actually came from a Win32 API. A major mess, we should
1291 * have used LMDB-specific error codes for everything.
1294 case ENOENT: /* 2, FILE_NOT_FOUND */
1295 case EIO: /* 5, ACCESS_DENIED */
1296 case ENOMEM: /* 12, INVALID_ACCESS */
1297 case EACCES: /* 13, INVALID_DATA */
1298 case EBUSY: /* 16, CURRENT_DIRECTORY */
1299 case EINVAL: /* 22, BAD_COMMAND */
1300 case ENOSPC: /* 28, OUT_OF_PAPER */
1301 return strerror(err);
1306 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1307 FORMAT_MESSAGE_IGNORE_INSERTS,
1308 NULL, err, 0, ptr, sizeof(buf), pad);
1311 return strerror(err);
1315 /** assert(3) variant in cursor context */
1316 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1317 /** assert(3) variant in transaction context */
1318 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1319 /** assert(3) variant in environment context */
1320 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1323 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1324 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1327 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1328 const char *func, const char *file, int line)
1331 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1332 file, line, expr_txt, func);
1333 if (env->me_assert_func)
1334 env->me_assert_func(env, buf);
1335 fprintf(stderr, "%s\n", buf);
1339 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1343 /** Return the page number of \b mp which may be sub-page, for debug output */
1345 mdb_dbg_pgno(MDB_page *mp)
1348 COPY_PGNO(ret, mp->mp_pgno);
1352 /** Display a key in hexadecimal and return the address of the result.
1353 * @param[in] key the key to display
1354 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1355 * @return The key in hexadecimal form.
1358 mdb_dkey(MDB_val *key, char *buf)
1361 unsigned char *c = key->mv_data;
1367 if (key->mv_size > DKBUF_MAXKEYSIZE)
1368 return "MDB_MAXKEYSIZE";
1369 /* may want to make this a dynamic check: if the key is mostly
1370 * printable characters, print it as-is instead of converting to hex.
1374 for (i=0; i<key->mv_size; i++)
1375 ptr += sprintf(ptr, "%02x", *c++);
1377 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1383 mdb_leafnode_type(MDB_node *n)
1385 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1386 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1387 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1390 /** Display all the keys in the page. */
1392 mdb_page_list(MDB_page *mp)
1394 pgno_t pgno = mdb_dbg_pgno(mp);
1395 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1397 unsigned int i, nkeys, nsize, total = 0;
1401 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1402 case P_BRANCH: type = "Branch page"; break;
1403 case P_LEAF: type = "Leaf page"; break;
1404 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1405 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1406 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1408 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1409 pgno, mp->mp_pages, state);
1412 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1413 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1416 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1420 nkeys = NUMKEYS(mp);
1421 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1423 for (i=0; i<nkeys; i++) {
1424 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1425 key.mv_size = nsize = mp->mp_pad;
1426 key.mv_data = LEAF2KEY(mp, i, nsize);
1428 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1431 node = NODEPTR(mp, i);
1432 key.mv_size = node->mn_ksize;
1433 key.mv_data = node->mn_data;
1434 nsize = NODESIZE + key.mv_size;
1435 if (IS_BRANCH(mp)) {
1436 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1440 if (F_ISSET(node->mn_flags, F_BIGDATA))
1441 nsize += sizeof(pgno_t);
1443 nsize += NODEDSZ(node);
1445 nsize += sizeof(indx_t);
1446 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1447 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1449 total = EVEN(total);
1451 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1452 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1456 mdb_cursor_chk(MDB_cursor *mc)
1462 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1463 for (i=0; i<mc->mc_top; i++) {
1465 node = NODEPTR(mp, mc->mc_ki[i]);
1466 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1469 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1475 /** Count all the pages in each DB and in the freelist
1476 * and make sure it matches the actual number of pages
1478 * All named DBs must be open for a correct count.
1480 static void mdb_audit(MDB_txn *txn)
1484 MDB_ID freecount, count;
1489 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1490 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1491 freecount += *(MDB_ID *)data.mv_data;
1492 mdb_tassert(txn, rc == MDB_NOTFOUND);
1495 for (i = 0; i<txn->mt_numdbs; i++) {
1497 if (!(txn->mt_dbflags[i] & DB_VALID))
1499 mdb_cursor_init(&mc, txn, i, &mx);
1500 if (txn->mt_dbs[i].md_root == P_INVALID)
1502 count += txn->mt_dbs[i].md_branch_pages +
1503 txn->mt_dbs[i].md_leaf_pages +
1504 txn->mt_dbs[i].md_overflow_pages;
1505 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1506 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1507 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1510 mp = mc.mc_pg[mc.mc_top];
1511 for (j=0; j<NUMKEYS(mp); j++) {
1512 MDB_node *leaf = NODEPTR(mp, j);
1513 if (leaf->mn_flags & F_SUBDATA) {
1515 memcpy(&db, NODEDATA(leaf), sizeof(db));
1516 count += db.md_branch_pages + db.md_leaf_pages +
1517 db.md_overflow_pages;
1521 mdb_tassert(txn, rc == MDB_NOTFOUND);
1524 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1525 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1526 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1532 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1534 return txn->mt_dbxs[dbi].md_cmp(a, b);
1538 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1540 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1543 /** Allocate memory for a page.
1544 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1547 mdb_page_malloc(MDB_txn *txn, unsigned num)
1549 MDB_env *env = txn->mt_env;
1550 MDB_page *ret = env->me_dpages;
1551 size_t psize = env->me_psize, sz = psize, off;
1552 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1553 * For a single page alloc, we init everything after the page header.
1554 * For multi-page, we init the final page; if the caller needed that
1555 * many pages they will be filling in at least up to the last page.
1559 VGMEMP_ALLOC(env, ret, sz);
1560 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1561 env->me_dpages = ret->mp_next;
1564 psize -= off = PAGEHDRSZ;
1569 if ((ret = malloc(sz)) != NULL) {
1570 VGMEMP_ALLOC(env, ret, sz);
1571 if (!(env->me_flags & MDB_NOMEMINIT)) {
1572 memset((char *)ret + off, 0, psize);
1576 txn->mt_flags |= MDB_TXN_ERROR;
1580 /** Free a single page.
1581 * Saves single pages to a list, for future reuse.
1582 * (This is not used for multi-page overflow pages.)
1585 mdb_page_free(MDB_env *env, MDB_page *mp)
1587 mp->mp_next = env->me_dpages;
1588 VGMEMP_FREE(env, mp);
1589 env->me_dpages = mp;
1592 /** Free a dirty page */
1594 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1596 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1597 mdb_page_free(env, dp);
1599 /* large pages just get freed directly */
1600 VGMEMP_FREE(env, dp);
1605 /** Return all dirty pages to dpage list */
1607 mdb_dlist_free(MDB_txn *txn)
1609 MDB_env *env = txn->mt_env;
1610 MDB_ID2L dl = txn->mt_u.dirty_list;
1611 unsigned i, n = dl[0].mid;
1613 for (i = 1; i <= n; i++) {
1614 mdb_dpage_free(env, dl[i].mptr);
1619 /** Loosen or free a single page.
1620 * Saves single pages to a list for future reuse
1621 * in this same txn. It has been pulled from the freeDB
1622 * and already resides on the dirty list, but has been
1623 * deleted. Use these pages first before pulling again
1626 * If the page wasn't dirtied in this txn, just add it
1627 * to this txn's free list.
1630 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1633 pgno_t pgno = mp->mp_pgno;
1635 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1636 if (mc->mc_txn->mt_parent) {
1637 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1638 /* If txn has a parent, make sure the page is in our
1642 unsigned x = mdb_mid2l_search(dl, pgno);
1643 if (x <= dl[0].mid && dl[x].mid == pgno) {
1644 if (mp != dl[x].mptr) { /* bad cursor? */
1645 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1646 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1647 return MDB_CORRUPTED;
1654 /* no parent txn, so it's just ours */
1659 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1661 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1662 mc->mc_txn->mt_loose_pgs = mp;
1663 mp->mp_flags |= P_LOOSE;
1665 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1673 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1674 * @param[in] mc A cursor handle for the current operation.
1675 * @param[in] pflags Flags of the pages to update:
1676 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1677 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1678 * @return 0 on success, non-zero on failure.
1681 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1683 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1684 MDB_txn *txn = mc->mc_txn;
1690 int rc = MDB_SUCCESS, level;
1692 /* Mark pages seen by cursors */
1693 if (mc->mc_flags & C_UNTRACK)
1694 mc = NULL; /* will find mc in mt_cursors */
1695 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1696 for (; mc; mc=mc->mc_next) {
1697 if (!(mc->mc_flags & C_INITIALIZED))
1699 for (m3 = mc;; m3 = &mx->mx_cursor) {
1701 for (j=0; j<m3->mc_snum; j++) {
1703 if ((mp->mp_flags & Mask) == pflags)
1704 mp->mp_flags ^= P_KEEP;
1706 mx = m3->mc_xcursor;
1707 /* Proceed to mx if it is at a sub-database */
1708 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1710 if (! (mp && (mp->mp_flags & P_LEAF)))
1712 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1713 if (!(leaf->mn_flags & F_SUBDATA))
1722 /* Mark dirty root pages */
1723 for (i=0; i<txn->mt_numdbs; i++) {
1724 if (txn->mt_dbflags[i] & DB_DIRTY) {
1725 pgno_t pgno = txn->mt_dbs[i].md_root;
1726 if (pgno == P_INVALID)
1728 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1730 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1731 dp->mp_flags ^= P_KEEP;
1739 static int mdb_page_flush(MDB_txn *txn, int keep);
1741 /** Spill pages from the dirty list back to disk.
1742 * This is intended to prevent running into #MDB_TXN_FULL situations,
1743 * but note that they may still occur in a few cases:
1744 * 1) our estimate of the txn size could be too small. Currently this
1745 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1746 * 2) child txns may run out of space if their parents dirtied a
1747 * lot of pages and never spilled them. TODO: we probably should do
1748 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1749 * the parent's dirty_room is below a given threshold.
1751 * Otherwise, if not using nested txns, it is expected that apps will
1752 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1753 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1754 * If the txn never references them again, they can be left alone.
1755 * If the txn only reads them, they can be used without any fuss.
1756 * If the txn writes them again, they can be dirtied immediately without
1757 * going thru all of the work of #mdb_page_touch(). Such references are
1758 * handled by #mdb_page_unspill().
1760 * Also note, we never spill DB root pages, nor pages of active cursors,
1761 * because we'll need these back again soon anyway. And in nested txns,
1762 * we can't spill a page in a child txn if it was already spilled in a
1763 * parent txn. That would alter the parent txns' data even though
1764 * the child hasn't committed yet, and we'd have no way to undo it if
1765 * the child aborted.
1767 * @param[in] m0 cursor A cursor handle identifying the transaction and
1768 * database for which we are checking space.
1769 * @param[in] key For a put operation, the key being stored.
1770 * @param[in] data For a put operation, the data being stored.
1771 * @return 0 on success, non-zero on failure.
1774 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1776 MDB_txn *txn = m0->mc_txn;
1778 MDB_ID2L dl = txn->mt_u.dirty_list;
1779 unsigned int i, j, need;
1782 if (m0->mc_flags & C_SUB)
1785 /* Estimate how much space this op will take */
1786 i = m0->mc_db->md_depth;
1787 /* Named DBs also dirty the main DB */
1788 if (m0->mc_dbi > MAIN_DBI)
1789 i += txn->mt_dbs[MAIN_DBI].md_depth;
1790 /* For puts, roughly factor in the key+data size */
1792 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1793 i += i; /* double it for good measure */
1796 if (txn->mt_dirty_room > i)
1799 if (!txn->mt_spill_pgs) {
1800 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1801 if (!txn->mt_spill_pgs)
1804 /* purge deleted slots */
1805 MDB_IDL sl = txn->mt_spill_pgs;
1806 unsigned int num = sl[0];
1808 for (i=1; i<=num; i++) {
1815 /* Preserve pages which may soon be dirtied again */
1816 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1819 /* Less aggressive spill - we originally spilled the entire dirty list,
1820 * with a few exceptions for cursor pages and DB root pages. But this
1821 * turns out to be a lot of wasted effort because in a large txn many
1822 * of those pages will need to be used again. So now we spill only 1/8th
1823 * of the dirty pages. Testing revealed this to be a good tradeoff,
1824 * better than 1/2, 1/4, or 1/10.
1826 if (need < MDB_IDL_UM_MAX / 8)
1827 need = MDB_IDL_UM_MAX / 8;
1829 /* Save the page IDs of all the pages we're flushing */
1830 /* flush from the tail forward, this saves a lot of shifting later on. */
1831 for (i=dl[0].mid; i && need; i--) {
1832 MDB_ID pn = dl[i].mid << 1;
1834 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1836 /* Can't spill twice, make sure it's not already in a parent's
1839 if (txn->mt_parent) {
1841 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1842 if (tx2->mt_spill_pgs) {
1843 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1844 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1845 dp->mp_flags |= P_KEEP;
1853 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1857 mdb_midl_sort(txn->mt_spill_pgs);
1859 /* Flush the spilled part of dirty list */
1860 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1863 /* Reset any dirty pages we kept that page_flush didn't see */
1864 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1867 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1871 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1873 mdb_find_oldest(MDB_txn *txn)
1876 txnid_t mr, oldest = txn->mt_txnid - 1;
1877 if (txn->mt_env->me_txns) {
1878 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1879 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1890 /** Add a page to the txn's dirty list */
1892 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1895 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1897 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1898 insert = mdb_mid2l_append;
1900 insert = mdb_mid2l_insert;
1902 mid.mid = mp->mp_pgno;
1904 rc = insert(txn->mt_u.dirty_list, &mid);
1905 mdb_tassert(txn, rc == 0);
1906 txn->mt_dirty_room--;
1909 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1910 * me_pghead and mt_next_pgno.
1912 * If there are free pages available from older transactions, they
1913 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1914 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1915 * and move me_pglast to say which records were consumed. Only this
1916 * function can create me_pghead and move me_pglast/mt_next_pgno.
1917 * @param[in] mc cursor A cursor handle identifying the transaction and
1918 * database for which we are allocating.
1919 * @param[in] num the number of pages to allocate.
1920 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1921 * will always be satisfied by a single contiguous chunk of memory.
1922 * @return 0 on success, non-zero on failure.
1925 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1927 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1928 /* Get at most <Max_retries> more freeDB records once me_pghead
1929 * has enough pages. If not enough, use new pages from the map.
1930 * If <Paranoid> and mc is updating the freeDB, only get new
1931 * records if me_pghead is empty. Then the freelist cannot play
1932 * catch-up with itself by growing while trying to save it.
1934 enum { Paranoid = 1, Max_retries = 500 };
1936 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1938 int rc, retry = num * 20;
1939 MDB_txn *txn = mc->mc_txn;
1940 MDB_env *env = txn->mt_env;
1941 pgno_t pgno, *mop = env->me_pghead;
1942 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1944 txnid_t oldest = 0, last;
1948 /* If there are any loose pages, just use them */
1949 if (num == 1 && txn->mt_loose_pgs) {
1950 np = txn->mt_loose_pgs;
1951 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1952 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1960 /* If our dirty list is already full, we can't do anything */
1961 if (txn->mt_dirty_room == 0) {
1966 for (op = MDB_FIRST;; op = MDB_NEXT) {
1969 pgno_t *idl, old_id, new_id;
1971 /* Seek a big enough contiguous page range. Prefer
1972 * pages at the tail, just truncating the list.
1978 if (mop[i-n2] == pgno+n2)
1985 if (op == MDB_FIRST) { /* 1st iteration */
1986 /* Prepare to fetch more and coalesce */
1987 oldest = mdb_find_oldest(txn);
1988 last = env->me_pglast;
1989 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1992 key.mv_data = &last; /* will look up last+1 */
1993 key.mv_size = sizeof(last);
1995 if (Paranoid && mc->mc_dbi == FREE_DBI)
1998 if (Paranoid && retry < 0 && mop_len)
2002 /* Do not fetch more if the record will be too recent */
2005 rc = mdb_cursor_get(&m2, &key, NULL, op);
2007 if (rc == MDB_NOTFOUND)
2011 last = *(txnid_t*)key.mv_data;
2014 np = m2.mc_pg[m2.mc_top];
2015 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2016 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2019 idl = (MDB_ID *) data.mv_data;
2022 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2027 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2029 mop = env->me_pghead;
2031 env->me_pglast = last;
2033 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2034 last, txn->mt_dbs[FREE_DBI].md_root, i));
2036 DPRINTF(("IDL %"Z"u", idl[k]));
2038 /* Merge in descending sorted order */
2041 mop[0] = (pgno_t)-1;
2045 for (; old_id < new_id; old_id = mop[--j])
2052 /* Use new pages from the map when nothing suitable in the freeDB */
2054 pgno = txn->mt_next_pgno;
2055 if (pgno + num >= env->me_maxpg) {
2056 DPUTS("DB size maxed out");
2062 if (env->me_flags & MDB_WRITEMAP) {
2063 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2065 if (!(np = mdb_page_malloc(txn, num))) {
2071 mop[0] = mop_len -= num;
2072 /* Move any stragglers down */
2073 for (j = i-num; j < mop_len; )
2074 mop[++j] = mop[++i];
2076 txn->mt_next_pgno = pgno + num;
2079 mdb_page_dirty(txn, np);
2085 txn->mt_flags |= MDB_TXN_ERROR;
2089 /** Copy the used portions of a non-overflow page.
2090 * @param[in] dst page to copy into
2091 * @param[in] src page to copy from
2092 * @param[in] psize size of a page
2095 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2097 enum { Align = sizeof(pgno_t) };
2098 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2100 /* If page isn't full, just copy the used portion. Adjust
2101 * alignment so memcpy may copy words instead of bytes.
2103 if ((unused &= -Align) && !IS_LEAF2(src)) {
2104 upper = (upper + PAGEBASE) & -Align;
2105 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2106 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2109 memcpy(dst, src, psize - unused);
2113 /** Pull a page off the txn's spill list, if present.
2114 * If a page being referenced was spilled to disk in this txn, bring
2115 * it back and make it dirty/writable again.
2116 * @param[in] txn the transaction handle.
2117 * @param[in] mp the page being referenced. It must not be dirty.
2118 * @param[out] ret the writable page, if any. ret is unchanged if
2119 * mp wasn't spilled.
2122 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2124 MDB_env *env = txn->mt_env;
2127 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2129 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2130 if (!tx2->mt_spill_pgs)
2132 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2133 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2136 if (txn->mt_dirty_room == 0)
2137 return MDB_TXN_FULL;
2138 if (IS_OVERFLOW(mp))
2142 if (env->me_flags & MDB_WRITEMAP) {
2145 np = mdb_page_malloc(txn, num);
2149 memcpy(np, mp, num * env->me_psize);
2151 mdb_page_copy(np, mp, env->me_psize);
2154 /* If in current txn, this page is no longer spilled.
2155 * If it happens to be the last page, truncate the spill list.
2156 * Otherwise mark it as deleted by setting the LSB.
2158 if (x == txn->mt_spill_pgs[0])
2159 txn->mt_spill_pgs[0]--;
2161 txn->mt_spill_pgs[x] |= 1;
2162 } /* otherwise, if belonging to a parent txn, the
2163 * page remains spilled until child commits
2166 mdb_page_dirty(txn, np);
2167 np->mp_flags |= P_DIRTY;
2175 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2176 * @param[in] mc cursor pointing to the page to be touched
2177 * @return 0 on success, non-zero on failure.
2180 mdb_page_touch(MDB_cursor *mc)
2182 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2183 MDB_txn *txn = mc->mc_txn;
2184 MDB_cursor *m2, *m3;
2188 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2189 if (txn->mt_flags & MDB_TXN_SPILLS) {
2191 rc = mdb_page_unspill(txn, mp, &np);
2197 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2198 (rc = mdb_page_alloc(mc, 1, &np)))
2201 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2202 mp->mp_pgno, pgno));
2203 mdb_cassert(mc, mp->mp_pgno != pgno);
2204 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2205 /* Update the parent page, if any, to point to the new page */
2207 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2208 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2209 SETPGNO(node, pgno);
2211 mc->mc_db->md_root = pgno;
2213 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2214 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2216 /* If txn has a parent, make sure the page is in our
2220 unsigned x = mdb_mid2l_search(dl, pgno);
2221 if (x <= dl[0].mid && dl[x].mid == pgno) {
2222 if (mp != dl[x].mptr) { /* bad cursor? */
2223 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2224 txn->mt_flags |= MDB_TXN_ERROR;
2225 return MDB_CORRUPTED;
2230 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2232 np = mdb_page_malloc(txn, 1);
2237 rc = mdb_mid2l_insert(dl, &mid);
2238 mdb_cassert(mc, rc == 0);
2243 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2245 np->mp_flags |= P_DIRTY;
2248 /* Adjust cursors pointing to mp */
2249 mc->mc_pg[mc->mc_top] = np;
2250 m2 = txn->mt_cursors[mc->mc_dbi];
2251 if (mc->mc_flags & C_SUB) {
2252 for (; m2; m2=m2->mc_next) {
2253 m3 = &m2->mc_xcursor->mx_cursor;
2254 if (m3->mc_snum < mc->mc_snum) continue;
2255 if (m3->mc_pg[mc->mc_top] == mp)
2256 m3->mc_pg[mc->mc_top] = np;
2259 for (; m2; m2=m2->mc_next) {
2260 if (m2->mc_snum < mc->mc_snum) continue;
2261 if (m2->mc_pg[mc->mc_top] == mp) {
2262 m2->mc_pg[mc->mc_top] = np;
2263 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2265 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2267 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2268 if (!(leaf->mn_flags & F_SUBDATA))
2269 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2277 txn->mt_flags |= MDB_TXN_ERROR;
2282 mdb_env_sync(MDB_env *env, int force)
2285 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2286 if (env->me_flags & MDB_WRITEMAP) {
2287 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2288 ? MS_ASYNC : MS_SYNC;
2289 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2292 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2296 if (MDB_FDATASYNC(env->me_fd))
2303 /** Back up parent txn's cursors, then grab the originals for tracking */
2305 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2307 MDB_cursor *mc, *bk;
2312 for (i = src->mt_numdbs; --i >= 0; ) {
2313 if ((mc = src->mt_cursors[i]) != NULL) {
2314 size = sizeof(MDB_cursor);
2316 size += sizeof(MDB_xcursor);
2317 for (; mc; mc = bk->mc_next) {
2323 mc->mc_db = &dst->mt_dbs[i];
2324 /* Kill pointers into src - and dst to reduce abuse: The
2325 * user may not use mc until dst ends. Otherwise we'd...
2327 mc->mc_txn = NULL; /* ...set this to dst */
2328 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2329 if ((mx = mc->mc_xcursor) != NULL) {
2330 *(MDB_xcursor *)(bk+1) = *mx;
2331 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2333 mc->mc_next = dst->mt_cursors[i];
2334 dst->mt_cursors[i] = mc;
2341 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2342 * @param[in] txn the transaction handle.
2343 * @param[in] merge true to keep changes to parent cursors, false to revert.
2344 * @return 0 on success, non-zero on failure.
2347 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2349 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2353 for (i = txn->mt_numdbs; --i >= 0; ) {
2354 for (mc = cursors[i]; mc; mc = next) {
2356 if ((bk = mc->mc_backup) != NULL) {
2358 /* Commit changes to parent txn */
2359 mc->mc_next = bk->mc_next;
2360 mc->mc_backup = bk->mc_backup;
2361 mc->mc_txn = bk->mc_txn;
2362 mc->mc_db = bk->mc_db;
2363 mc->mc_dbflag = bk->mc_dbflag;
2364 if ((mx = mc->mc_xcursor) != NULL)
2365 mx->mx_cursor.mc_txn = bk->mc_txn;
2367 /* Abort nested txn */
2369 if ((mx = mc->mc_xcursor) != NULL)
2370 *mx = *(MDB_xcursor *)(bk+1);
2374 /* Only malloced cursors are permanently tracked. */
2382 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2385 mdb_txn_reset0(MDB_txn *txn, const char *act);
2387 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2393 Pidset = F_SETLK, Pidcheck = F_GETLK
2397 /** Set or check a pid lock. Set returns 0 on success.
2398 * Check returns 0 if the process is certainly dead, nonzero if it may
2399 * be alive (the lock exists or an error happened so we do not know).
2401 * On Windows Pidset is a no-op, we merely check for the existence
2402 * of the process with the given pid. On POSIX we use a single byte
2403 * lock on the lockfile, set at an offset equal to the pid.
2406 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2408 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2411 if (op == Pidcheck) {
2412 h = OpenProcess(env->me_pidquery, FALSE, pid);
2413 /* No documented "no such process" code, but other program use this: */
2415 return ErrCode() != ERROR_INVALID_PARAMETER;
2416 /* A process exists until all handles to it close. Has it exited? */
2417 ret = WaitForSingleObject(h, 0) != 0;
2424 struct flock lock_info;
2425 memset(&lock_info, 0, sizeof(lock_info));
2426 lock_info.l_type = F_WRLCK;
2427 lock_info.l_whence = SEEK_SET;
2428 lock_info.l_start = pid;
2429 lock_info.l_len = 1;
2430 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2431 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2433 } else if ((rc = ErrCode()) == EINTR) {
2441 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2442 * @param[in] txn the transaction handle to initialize
2443 * @return 0 on success, non-zero on failure.
2446 mdb_txn_renew0(MDB_txn *txn)
2448 MDB_env *env = txn->mt_env;
2449 MDB_txninfo *ti = env->me_txns;
2453 int rc, new_notls = 0;
2456 txn->mt_numdbs = env->me_numdbs;
2457 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2459 if (txn->mt_flags & MDB_TXN_RDONLY) {
2461 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2462 txn->mt_txnid = meta->mm_txnid;
2463 txn->mt_u.reader = NULL;
2465 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2466 pthread_getspecific(env->me_txkey);
2468 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2469 return MDB_BAD_RSLOT;
2471 MDB_PID_T pid = env->me_pid;
2472 MDB_THR_T tid = pthread_self();
2474 if (!env->me_live_reader) {
2475 rc = mdb_reader_pid(env, Pidset, pid);
2478 env->me_live_reader = 1;
2482 nr = ti->mti_numreaders;
2483 for (i=0; i<nr; i++)
2484 if (ti->mti_readers[i].mr_pid == 0)
2486 if (i == env->me_maxreaders) {
2487 UNLOCK_MUTEX_R(env);
2488 return MDB_READERS_FULL;
2490 ti->mti_readers[i].mr_pid = pid;
2491 ti->mti_readers[i].mr_tid = tid;
2493 ti->mti_numreaders = ++nr;
2494 /* Save numreaders for un-mutexed mdb_env_close() */
2495 env->me_numreaders = nr;
2496 UNLOCK_MUTEX_R(env);
2498 r = &ti->mti_readers[i];
2499 new_notls = (env->me_flags & MDB_NOTLS);
2500 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2505 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2506 txn->mt_u.reader = r;
2507 meta = env->me_metas[txn->mt_txnid & 1];
2513 txn->mt_txnid = ti->mti_txnid;
2514 meta = env->me_metas[txn->mt_txnid & 1];
2516 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2517 txn->mt_txnid = meta->mm_txnid;
2521 if (txn->mt_txnid == mdb_debug_start)
2524 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2525 txn->mt_u.dirty_list = env->me_dirty_list;
2526 txn->mt_u.dirty_list[0].mid = 0;
2527 txn->mt_free_pgs = env->me_free_pgs;
2528 txn->mt_free_pgs[0] = 0;
2529 txn->mt_spill_pgs = NULL;
2531 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2534 /* Copy the DB info and flags */
2535 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2537 /* Moved to here to avoid a data race in read TXNs */
2538 txn->mt_next_pgno = meta->mm_last_pg+1;
2540 for (i=2; i<txn->mt_numdbs; i++) {
2541 x = env->me_dbflags[i];
2542 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2543 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2545 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2547 if (env->me_maxpg < txn->mt_next_pgno) {
2548 mdb_txn_reset0(txn, "renew0-mapfail");
2550 txn->mt_u.reader->mr_pid = 0;
2551 txn->mt_u.reader = NULL;
2553 return MDB_MAP_RESIZED;
2560 mdb_txn_renew(MDB_txn *txn)
2564 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2567 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2568 DPUTS("environment had fatal error, must shutdown!");
2572 rc = mdb_txn_renew0(txn);
2573 if (rc == MDB_SUCCESS) {
2574 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2575 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2576 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2582 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2586 int rc, size, tsize = sizeof(MDB_txn);
2588 if (env->me_flags & MDB_FATAL_ERROR) {
2589 DPUTS("environment had fatal error, must shutdown!");
2592 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2595 /* Nested transactions: Max 1 child, write txns only, no writemap */
2596 if (parent->mt_child ||
2597 (flags & MDB_RDONLY) ||
2598 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2599 (env->me_flags & MDB_WRITEMAP))
2601 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2603 tsize = sizeof(MDB_ntxn);
2605 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2606 if (!(flags & MDB_RDONLY)) {
2607 size += env->me_maxdbs * sizeof(MDB_cursor *);
2608 /* child txns use parent's dbiseqs */
2610 size += env->me_maxdbs * sizeof(unsigned int);
2613 if ((txn = calloc(1, size)) == NULL) {
2614 DPRINTF(("calloc: %s", strerror(errno)));
2617 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2618 if (flags & MDB_RDONLY) {
2619 txn->mt_flags |= MDB_TXN_RDONLY;
2620 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2621 txn->mt_dbiseqs = env->me_dbiseqs;
2623 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2625 txn->mt_dbiseqs = parent->mt_dbiseqs;
2626 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2628 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2629 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2636 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2637 if (!txn->mt_u.dirty_list ||
2638 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2640 free(txn->mt_u.dirty_list);
2644 txn->mt_txnid = parent->mt_txnid;
2645 txn->mt_dirty_room = parent->mt_dirty_room;
2646 txn->mt_u.dirty_list[0].mid = 0;
2647 txn->mt_spill_pgs = NULL;
2648 txn->mt_next_pgno = parent->mt_next_pgno;
2649 parent->mt_child = txn;
2650 txn->mt_parent = parent;
2651 txn->mt_numdbs = parent->mt_numdbs;
2652 txn->mt_flags = parent->mt_flags;
2653 txn->mt_dbxs = parent->mt_dbxs;
2654 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2655 /* Copy parent's mt_dbflags, but clear DB_NEW */
2656 for (i=0; i<txn->mt_numdbs; i++)
2657 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2659 ntxn = (MDB_ntxn *)txn;
2660 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2661 if (env->me_pghead) {
2662 size = MDB_IDL_SIZEOF(env->me_pghead);
2663 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2665 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2670 rc = mdb_cursor_shadow(parent, txn);
2672 mdb_txn_reset0(txn, "beginchild-fail");
2674 rc = mdb_txn_renew0(txn);
2680 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2681 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2682 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2689 mdb_txn_env(MDB_txn *txn)
2691 if(!txn) return NULL;
2695 /** Export or close DBI handles opened in this txn. */
2697 mdb_dbis_update(MDB_txn *txn, int keep)
2700 MDB_dbi n = txn->mt_numdbs;
2701 MDB_env *env = txn->mt_env;
2702 unsigned char *tdbflags = txn->mt_dbflags;
2704 for (i = n; --i >= 2;) {
2705 if (tdbflags[i] & DB_NEW) {
2707 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2709 char *ptr = env->me_dbxs[i].md_name.mv_data;
2711 env->me_dbxs[i].md_name.mv_data = NULL;
2712 env->me_dbxs[i].md_name.mv_size = 0;
2713 env->me_dbflags[i] = 0;
2714 env->me_dbiseqs[i]++;
2720 if (keep && env->me_numdbs < n)
2724 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2725 * May be called twice for readonly txns: First reset it, then abort.
2726 * @param[in] txn the transaction handle to reset
2727 * @param[in] act why the transaction is being reset
2730 mdb_txn_reset0(MDB_txn *txn, const char *act)
2732 MDB_env *env = txn->mt_env;
2734 /* Close any DBI handles opened in this txn */
2735 mdb_dbis_update(txn, 0);
2737 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2738 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2739 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2741 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2742 if (txn->mt_u.reader) {
2743 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2744 if (!(env->me_flags & MDB_NOTLS))
2745 txn->mt_u.reader = NULL; /* txn does not own reader */
2747 txn->mt_numdbs = 0; /* close nothing if called again */
2748 txn->mt_dbxs = NULL; /* mark txn as reset */
2750 mdb_cursors_close(txn, 0);
2752 if (!(env->me_flags & MDB_WRITEMAP)) {
2753 mdb_dlist_free(txn);
2755 mdb_midl_free(env->me_pghead);
2757 if (txn->mt_parent) {
2758 txn->mt_parent->mt_child = NULL;
2759 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2760 mdb_midl_free(txn->mt_free_pgs);
2761 mdb_midl_free(txn->mt_spill_pgs);
2762 free(txn->mt_u.dirty_list);
2766 if (mdb_midl_shrink(&txn->mt_free_pgs))
2767 env->me_free_pgs = txn->mt_free_pgs;
2768 env->me_pghead = NULL;
2772 /* The writer mutex was locked in mdb_txn_begin. */
2774 UNLOCK_MUTEX_W(env);
2779 mdb_txn_reset(MDB_txn *txn)
2784 /* This call is only valid for read-only txns */
2785 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2788 mdb_txn_reset0(txn, "reset");
2792 mdb_txn_abort(MDB_txn *txn)
2798 mdb_txn_abort(txn->mt_child);
2800 mdb_txn_reset0(txn, "abort");
2801 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2802 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2803 txn->mt_u.reader->mr_pid = 0;
2808 /** Save the freelist as of this transaction to the freeDB.
2809 * This changes the freelist. Keep trying until it stabilizes.
2812 mdb_freelist_save(MDB_txn *txn)
2814 /* env->me_pghead[] can grow and shrink during this call.
2815 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2816 * Page numbers cannot disappear from txn->mt_free_pgs[].
2819 MDB_env *env = txn->mt_env;
2820 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2821 txnid_t pglast = 0, head_id = 0;
2822 pgno_t freecnt = 0, *free_pgs, *mop;
2823 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2825 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2827 if (env->me_pghead) {
2828 /* Make sure first page of freeDB is touched and on freelist */
2829 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2830 if (rc && rc != MDB_NOTFOUND)
2834 /* Dispose of loose pages. Usually they will have all
2835 * been used up by the time we get here.
2837 if (txn->mt_loose_pgs) {
2838 MDB_page *mp = txn->mt_loose_pgs;
2839 /* Just return them to freeDB */
2840 if (env->me_pghead) {
2842 mop = env->me_pghead;
2843 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2844 pgno_t pg = mp->mp_pgno;
2846 for (i = mop[0]; i && mop[i] < pg; i--)
2852 /* Oh well, they were wasted. Put on freelist */
2853 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2854 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
2857 txn->mt_loose_pgs = NULL;
2860 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2861 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2862 ? SSIZE_MAX : maxfree_1pg;
2865 /* Come back here after each Put() in case freelist changed */
2870 /* If using records from freeDB which we have not yet
2871 * deleted, delete them and any we reserved for me_pghead.
2873 while (pglast < env->me_pglast) {
2874 rc = mdb_cursor_first(&mc, &key, NULL);
2877 pglast = head_id = *(txnid_t *)key.mv_data;
2878 total_room = head_room = 0;
2879 mdb_tassert(txn, pglast <= env->me_pglast);
2880 rc = mdb_cursor_del(&mc, 0);
2885 /* Save the IDL of pages freed by this txn, to a single record */
2886 if (freecnt < txn->mt_free_pgs[0]) {
2888 /* Make sure last page of freeDB is touched and on freelist */
2889 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2890 if (rc && rc != MDB_NOTFOUND)
2893 free_pgs = txn->mt_free_pgs;
2894 /* Write to last page of freeDB */
2895 key.mv_size = sizeof(txn->mt_txnid);
2896 key.mv_data = &txn->mt_txnid;
2898 freecnt = free_pgs[0];
2899 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2900 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2903 /* Retry if mt_free_pgs[] grew during the Put() */
2904 free_pgs = txn->mt_free_pgs;
2905 } while (freecnt < free_pgs[0]);
2906 mdb_midl_sort(free_pgs);
2907 memcpy(data.mv_data, free_pgs, data.mv_size);
2910 unsigned int i = free_pgs[0];
2911 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2912 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2914 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2920 mop = env->me_pghead;
2921 mop_len = mop ? mop[0] : 0;
2923 /* Reserve records for me_pghead[]. Split it if multi-page,
2924 * to avoid searching freeDB for a page range. Use keys in
2925 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2927 if (total_room >= mop_len) {
2928 if (total_room == mop_len || --more < 0)
2930 } else if (head_room >= maxfree_1pg && head_id > 1) {
2931 /* Keep current record (overflow page), add a new one */
2935 /* (Re)write {key = head_id, IDL length = head_room} */
2936 total_room -= head_room;
2937 head_room = mop_len - total_room;
2938 if (head_room > maxfree_1pg && head_id > 1) {
2939 /* Overflow multi-page for part of me_pghead */
2940 head_room /= head_id; /* amortize page sizes */
2941 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2942 } else if (head_room < 0) {
2943 /* Rare case, not bothering to delete this record */
2946 key.mv_size = sizeof(head_id);
2947 key.mv_data = &head_id;
2948 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2949 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2952 /* IDL is initially empty, zero out at least the length */
2953 pgs = (pgno_t *)data.mv_data;
2954 j = head_room > clean_limit ? head_room : 0;
2958 total_room += head_room;
2961 /* Fill in the reserved me_pghead records */
2967 rc = mdb_cursor_first(&mc, &key, &data);
2968 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2969 txnid_t id = *(txnid_t *)key.mv_data;
2970 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2973 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2975 if (len > mop_len) {
2977 data.mv_size = (len + 1) * sizeof(MDB_ID);
2979 data.mv_data = mop -= len;
2982 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2984 if (rc || !(mop_len -= len))
2991 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2992 * @param[in] txn the transaction that's being committed
2993 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2994 * @return 0 on success, non-zero on failure.
2997 mdb_page_flush(MDB_txn *txn, int keep)
2999 MDB_env *env = txn->mt_env;
3000 MDB_ID2L dl = txn->mt_u.dirty_list;
3001 unsigned psize = env->me_psize, j;
3002 int i, pagecount = dl[0].mid, rc;
3003 size_t size = 0, pos = 0;
3005 MDB_page *dp = NULL;
3009 struct iovec iov[MDB_COMMIT_PAGES];
3010 ssize_t wpos = 0, wsize = 0, wres;
3011 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3017 if (env->me_flags & MDB_WRITEMAP) {
3018 /* Clear dirty flags */
3019 while (++i <= pagecount) {
3021 /* Don't flush this page yet */
3022 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3023 dp->mp_flags &= ~P_KEEP;
3027 dp->mp_flags &= ~P_DIRTY;
3032 /* Write the pages */
3034 if (++i <= pagecount) {
3036 /* Don't flush this page yet */
3037 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3038 dp->mp_flags &= ~P_KEEP;
3043 /* clear dirty flag */
3044 dp->mp_flags &= ~P_DIRTY;
3047 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3052 /* Windows actually supports scatter/gather I/O, but only on
3053 * unbuffered file handles. Since we're relying on the OS page
3054 * cache for all our data, that's self-defeating. So we just
3055 * write pages one at a time. We use the ov structure to set
3056 * the write offset, to at least save the overhead of a Seek
3059 DPRINTF(("committing page %"Z"u", pgno));
3060 memset(&ov, 0, sizeof(ov));
3061 ov.Offset = pos & 0xffffffff;
3062 ov.OffsetHigh = pos >> 16 >> 16;
3063 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3065 DPRINTF(("WriteFile: %d", rc));
3069 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3070 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3072 /* Write previous page(s) */
3073 #ifdef MDB_USE_PWRITEV
3074 wres = pwritev(env->me_fd, iov, n, wpos);
3077 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3079 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3081 DPRINTF(("lseek: %s", strerror(rc)));
3084 wres = writev(env->me_fd, iov, n);
3087 if (wres != wsize) {
3090 DPRINTF(("Write error: %s", strerror(rc)));
3092 rc = EIO; /* TODO: Use which error code? */
3093 DPUTS("short write, filesystem full?");
3104 DPRINTF(("committing page %"Z"u", pgno));
3105 next_pos = pos + size;
3106 iov[n].iov_len = size;
3107 iov[n].iov_base = (char *)dp;
3113 for (i = keep; ++i <= pagecount; ) {
3115 /* This is a page we skipped above */
3118 dl[j].mid = dp->mp_pgno;
3121 mdb_dpage_free(env, dp);
3126 txn->mt_dirty_room += i - j;
3132 mdb_txn_commit(MDB_txn *txn)
3138 if (txn == NULL || txn->mt_env == NULL)
3141 if (txn->mt_child) {
3142 rc = mdb_txn_commit(txn->mt_child);
3143 txn->mt_child = NULL;
3150 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3151 mdb_dbis_update(txn, 1);
3152 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3157 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3158 DPUTS("error flag is set, can't commit");
3160 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3165 if (txn->mt_parent) {
3166 MDB_txn *parent = txn->mt_parent;
3170 unsigned x, y, len, ps_len;
3172 /* Append our free list to parent's */
3173 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3176 mdb_midl_free(txn->mt_free_pgs);
3177 /* Failures after this must either undo the changes
3178 * to the parent or set MDB_TXN_ERROR in the parent.
3181 parent->mt_next_pgno = txn->mt_next_pgno;
3182 parent->mt_flags = txn->mt_flags;
3184 /* Merge our cursors into parent's and close them */
3185 mdb_cursors_close(txn, 1);
3187 /* Update parent's DB table. */
3188 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3189 parent->mt_numdbs = txn->mt_numdbs;
3190 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3191 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3192 for (i=2; i<txn->mt_numdbs; i++) {
3193 /* preserve parent's DB_NEW status */
3194 x = parent->mt_dbflags[i] & DB_NEW;
3195 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3198 dst = parent->mt_u.dirty_list;
3199 src = txn->mt_u.dirty_list;
3200 /* Remove anything in our dirty list from parent's spill list */
3201 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3203 pspill[0] = (pgno_t)-1;
3204 /* Mark our dirty pages as deleted in parent spill list */
3205 for (i=0, len=src[0].mid; ++i <= len; ) {
3206 MDB_ID pn = src[i].mid << 1;
3207 while (pn > pspill[x])
3209 if (pn == pspill[x]) {
3214 /* Squash deleted pagenums if we deleted any */
3215 for (x=y; ++x <= ps_len; )
3216 if (!(pspill[x] & 1))
3217 pspill[++y] = pspill[x];
3221 /* Find len = length of merging our dirty list with parent's */
3223 dst[0].mid = 0; /* simplify loops */
3224 if (parent->mt_parent) {
3225 len = x + src[0].mid;
3226 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3227 for (i = x; y && i; y--) {
3228 pgno_t yp = src[y].mid;
3229 while (yp < dst[i].mid)
3231 if (yp == dst[i].mid) {
3236 } else { /* Simplify the above for single-ancestor case */
3237 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3239 /* Merge our dirty list with parent's */
3241 for (i = len; y; dst[i--] = src[y--]) {
3242 pgno_t yp = src[y].mid;
3243 while (yp < dst[x].mid)
3244 dst[i--] = dst[x--];
3245 if (yp == dst[x].mid)
3246 free(dst[x--].mptr);
3248 mdb_tassert(txn, i == x);
3250 free(txn->mt_u.dirty_list);
3251 parent->mt_dirty_room = txn->mt_dirty_room;
3252 if (txn->mt_spill_pgs) {
3253 if (parent->mt_spill_pgs) {
3254 /* TODO: Prevent failure here, so parent does not fail */
3255 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3257 parent->mt_flags |= MDB_TXN_ERROR;
3258 mdb_midl_free(txn->mt_spill_pgs);
3259 mdb_midl_sort(parent->mt_spill_pgs);
3261 parent->mt_spill_pgs = txn->mt_spill_pgs;
3265 /* Append our loose page list to parent's */
3266 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3268 *lp = txn->mt_loose_pgs;
3270 parent->mt_child = NULL;
3271 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3276 if (txn != env->me_txn) {
3277 DPUTS("attempt to commit unknown transaction");
3282 mdb_cursors_close(txn, 0);
3284 if (!txn->mt_u.dirty_list[0].mid &&
3285 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3288 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3289 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3291 /* Update DB root pointers */
3292 if (txn->mt_numdbs > 2) {
3296 data.mv_size = sizeof(MDB_db);
3298 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3299 for (i = 2; i < txn->mt_numdbs; i++) {
3300 if (txn->mt_dbflags[i] & DB_DIRTY) {
3301 if (TXN_DBI_CHANGED(txn, i)) {
3305 data.mv_data = &txn->mt_dbs[i];
3306 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3313 rc = mdb_freelist_save(txn);
3317 mdb_midl_free(env->me_pghead);
3318 env->me_pghead = NULL;
3319 if (mdb_midl_shrink(&txn->mt_free_pgs))
3320 env->me_free_pgs = txn->mt_free_pgs;
3326 if ((rc = mdb_page_flush(txn, 0)) ||
3327 (rc = mdb_env_sync(env, 0)) ||
3328 (rc = mdb_env_write_meta(txn)))
3334 mdb_dbis_update(txn, 1);
3337 UNLOCK_MUTEX_W(env);
3347 /** Read the environment parameters of a DB environment before
3348 * mapping it into memory.
3349 * @param[in] env the environment handle
3350 * @param[out] meta address of where to store the meta information
3351 * @return 0 on success, non-zero on failure.
3354 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3360 enum { Size = sizeof(pbuf) };
3362 /* We don't know the page size yet, so use a minimum value.
3363 * Read both meta pages so we can use the latest one.
3366 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3370 memset(&ov, 0, sizeof(ov));
3372 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3373 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3376 rc = pread(env->me_fd, &pbuf, Size, off);
3379 if (rc == 0 && off == 0)
3381 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3382 DPRINTF(("read: %s", mdb_strerror(rc)));
3386 p = (MDB_page *)&pbuf;
3388 if (!F_ISSET(p->mp_flags, P_META)) {
3389 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3394 if (m->mm_magic != MDB_MAGIC) {
3395 DPUTS("meta has invalid magic");
3399 if (m->mm_version != MDB_DATA_VERSION) {
3400 DPRINTF(("database is version %u, expected version %u",
3401 m->mm_version, MDB_DATA_VERSION));
3402 return MDB_VERSION_MISMATCH;
3405 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3412 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3414 meta->mm_magic = MDB_MAGIC;
3415 meta->mm_version = MDB_DATA_VERSION;
3416 meta->mm_mapsize = env->me_mapsize;
3417 meta->mm_psize = env->me_psize;
3418 meta->mm_last_pg = 1;
3419 meta->mm_flags = env->me_flags & 0xffff;
3420 meta->mm_flags |= MDB_INTEGERKEY;
3421 meta->mm_dbs[0].md_root = P_INVALID;
3422 meta->mm_dbs[1].md_root = P_INVALID;
3425 /** Write the environment parameters of a freshly created DB environment.
3426 * @param[in] env the environment handle
3427 * @param[out] meta address of where to store the meta information
3428 * @return 0 on success, non-zero on failure.
3431 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3439 memset(&ov, 0, sizeof(ov));
3440 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3442 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3445 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3446 len = pwrite(fd, ptr, size, pos); \
3447 rc = (len >= 0); } while(0)
3450 DPUTS("writing new meta page");
3452 psize = env->me_psize;
3454 mdb_env_init_meta0(env, meta);
3456 p = calloc(2, psize);
3458 p->mp_flags = P_META;
3459 *(MDB_meta *)METADATA(p) = *meta;
3461 q = (MDB_page *)((char *)p + psize);
3463 q->mp_flags = P_META;
3464 *(MDB_meta *)METADATA(q) = *meta;
3466 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3469 else if ((unsigned) len == psize * 2)
3477 /** Update the environment info to commit a transaction.
3478 * @param[in] txn the transaction that's being committed
3479 * @return 0 on success, non-zero on failure.
3482 mdb_env_write_meta(MDB_txn *txn)
3485 MDB_meta meta, metab, *mp;
3488 int rc, len, toggle;
3497 toggle = txn->mt_txnid & 1;
3498 DPRINTF(("writing meta page %d for root page %"Z"u",
3499 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3502 mp = env->me_metas[toggle];
3503 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3504 /* Persist any increases of mapsize config */
3505 if (mapsize < env->me_mapsize)
3506 mapsize = env->me_mapsize;
3508 if (env->me_flags & MDB_WRITEMAP) {
3509 mp->mm_mapsize = mapsize;
3510 mp->mm_dbs[0] = txn->mt_dbs[0];
3511 mp->mm_dbs[1] = txn->mt_dbs[1];
3512 mp->mm_last_pg = txn->mt_next_pgno - 1;
3513 mp->mm_txnid = txn->mt_txnid;
3514 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3515 unsigned meta_size = env->me_psize;
3516 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3519 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3520 if (meta_size < env->me_os_psize)
3521 meta_size += meta_size;
3526 if (MDB_MSYNC(ptr, meta_size, rc)) {
3533 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3534 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3536 meta.mm_mapsize = mapsize;
3537 meta.mm_dbs[0] = txn->mt_dbs[0];
3538 meta.mm_dbs[1] = txn->mt_dbs[1];
3539 meta.mm_last_pg = txn->mt_next_pgno - 1;
3540 meta.mm_txnid = txn->mt_txnid;
3542 off = offsetof(MDB_meta, mm_mapsize);
3543 ptr = (char *)&meta + off;
3544 len = sizeof(MDB_meta) - off;
3546 off += env->me_psize;
3549 /* Write to the SYNC fd */
3550 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3551 env->me_fd : env->me_mfd;
3554 memset(&ov, 0, sizeof(ov));
3556 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3560 rc = pwrite(mfd, ptr, len, off);
3563 rc = rc < 0 ? ErrCode() : EIO;
3564 DPUTS("write failed, disk error?");
3565 /* On a failure, the pagecache still contains the new data.
3566 * Write some old data back, to prevent it from being used.
3567 * Use the non-SYNC fd; we know it will fail anyway.
3569 meta.mm_last_pg = metab.mm_last_pg;
3570 meta.mm_txnid = metab.mm_txnid;
3572 memset(&ov, 0, sizeof(ov));
3574 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3576 r2 = pwrite(env->me_fd, ptr, len, off);
3577 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3580 env->me_flags |= MDB_FATAL_ERROR;
3584 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3585 if (!(env->me_flags & MDB_WRITEMAP)) {
3586 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3588 /* Memory ordering issues are irrelevant; since the entire writer
3589 * is wrapped by wmutex, all of these changes will become visible
3590 * after the wmutex is unlocked. Since the DB is multi-version,
3591 * readers will get consistent data regardless of how fresh or
3592 * how stale their view of these values is.
3595 env->me_txns->mti_txnid = txn->mt_txnid;
3600 /** Check both meta pages to see which one is newer.
3601 * @param[in] env the environment handle
3602 * @return meta toggle (0 or 1).
3605 mdb_env_pick_meta(const MDB_env *env)
3607 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3611 mdb_env_create(MDB_env **env)
3615 e = calloc(1, sizeof(MDB_env));
3619 e->me_maxreaders = DEFAULT_READERS;
3620 e->me_maxdbs = e->me_numdbs = 2;
3621 e->me_fd = INVALID_HANDLE_VALUE;
3622 e->me_lfd = INVALID_HANDLE_VALUE;
3623 e->me_mfd = INVALID_HANDLE_VALUE;
3624 #ifdef MDB_USE_POSIX_SEM
3625 e->me_rmutex = SEM_FAILED;
3626 e->me_wmutex = SEM_FAILED;
3628 e->me_pid = getpid();
3629 GET_PAGESIZE(e->me_os_psize);
3630 VGMEMP_CREATE(e,0,0);
3636 mdb_env_map(MDB_env *env, void *addr)
3639 unsigned int flags = env->me_flags;
3643 LONG sizelo, sizehi;
3646 if (flags & MDB_RDONLY) {
3647 /* Don't set explicit map size, use whatever exists */
3652 msize = env->me_mapsize;
3653 sizelo = msize & 0xffffffff;
3654 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3656 /* Windows won't create mappings for zero length files.
3657 * and won't map more than the file size.
3658 * Just set the maxsize right now.
3660 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3661 || !SetEndOfFile(env->me_fd)
3662 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3666 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3667 PAGE_READWRITE : PAGE_READONLY,
3668 sizehi, sizelo, NULL);
3671 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3672 FILE_MAP_WRITE : FILE_MAP_READ,
3674 rc = env->me_map ? 0 : ErrCode();
3679 int prot = PROT_READ;
3680 if (flags & MDB_WRITEMAP) {
3682 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3685 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3687 if (env->me_map == MAP_FAILED) {
3692 if (flags & MDB_NORDAHEAD) {
3693 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3695 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3697 #ifdef POSIX_MADV_RANDOM
3698 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3699 #endif /* POSIX_MADV_RANDOM */
3700 #endif /* MADV_RANDOM */
3704 /* Can happen because the address argument to mmap() is just a
3705 * hint. mmap() can pick another, e.g. if the range is in use.
3706 * The MAP_FIXED flag would prevent that, but then mmap could
3707 * instead unmap existing pages to make room for the new map.
3709 if (addr && env->me_map != addr)
3710 return EBUSY; /* TODO: Make a new MDB_* error code? */
3712 p = (MDB_page *)env->me_map;
3713 env->me_metas[0] = METADATA(p);
3714 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3720 mdb_env_set_mapsize(MDB_env *env, size_t size)
3722 /* If env is already open, caller is responsible for making
3723 * sure there are no active txns.
3731 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3732 else if (size < env->me_mapsize) {
3733 /* If the configured size is smaller, make sure it's
3734 * still big enough. Silently round up to minimum if not.
3736 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3740 munmap(env->me_map, env->me_mapsize);
3741 env->me_mapsize = size;
3742 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3743 rc = mdb_env_map(env, old);
3747 env->me_mapsize = size;
3749 env->me_maxpg = env->me_mapsize / env->me_psize;
3754 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3758 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3763 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3765 if (env->me_map || readers < 1)
3767 env->me_maxreaders = readers;
3772 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3774 if (!env || !readers)
3776 *readers = env->me_maxreaders;
3780 /** Further setup required for opening an LMDB environment
3783 mdb_env_open2(MDB_env *env)
3785 unsigned int flags = env->me_flags;
3786 int i, newenv = 0, rc;
3790 /* See if we should use QueryLimited */
3792 if ((rc & 0xff) > 5)
3793 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3795 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3798 memset(&meta, 0, sizeof(meta));
3800 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3803 DPUTS("new mdbenv");
3805 env->me_psize = env->me_os_psize;
3806 if (env->me_psize > MAX_PAGESIZE)
3807 env->me_psize = MAX_PAGESIZE;
3809 env->me_psize = meta.mm_psize;
3812 /* Was a mapsize configured? */
3813 if (!env->me_mapsize) {
3814 /* If this is a new environment, take the default,
3815 * else use the size recorded in the existing env.
3817 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3818 } else if (env->me_mapsize < meta.mm_mapsize) {
3819 /* If the configured size is smaller, make sure it's
3820 * still big enough. Silently round up to minimum if not.
3822 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3823 if (env->me_mapsize < minsize)
3824 env->me_mapsize = minsize;
3827 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3832 if (flags & MDB_FIXEDMAP)
3833 meta.mm_address = env->me_map;
3834 i = mdb_env_init_meta(env, &meta);
3835 if (i != MDB_SUCCESS) {
3840 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3841 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3843 #if !(MDB_MAXKEYSIZE)
3844 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3846 env->me_maxpg = env->me_mapsize / env->me_psize;
3850 int toggle = mdb_env_pick_meta(env);
3851 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3853 DPRINTF(("opened database version %u, pagesize %u",
3854 env->me_metas[0]->mm_version, env->me_psize));
3855 DPRINTF(("using meta page %d", toggle));
3856 DPRINTF(("depth: %u", db->md_depth));
3857 DPRINTF(("entries: %"Z"u", db->md_entries));
3858 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3859 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3860 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3861 DPRINTF(("root: %"Z"u", db->md_root));
3869 /** Release a reader thread's slot in the reader lock table.
3870 * This function is called automatically when a thread exits.
3871 * @param[in] ptr This points to the slot in the reader lock table.
3874 mdb_env_reader_dest(void *ptr)
3876 MDB_reader *reader = ptr;
3882 /** Junk for arranging thread-specific callbacks on Windows. This is
3883 * necessarily platform and compiler-specific. Windows supports up
3884 * to 1088 keys. Let's assume nobody opens more than 64 environments
3885 * in a single process, for now. They can override this if needed.
3887 #ifndef MAX_TLS_KEYS
3888 #define MAX_TLS_KEYS 64
3890 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3891 static int mdb_tls_nkeys;
3893 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3897 case DLL_PROCESS_ATTACH: break;
3898 case DLL_THREAD_ATTACH: break;
3899 case DLL_THREAD_DETACH:
3900 for (i=0; i<mdb_tls_nkeys; i++) {
3901 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3903 mdb_env_reader_dest(r);
3907 case DLL_PROCESS_DETACH: break;
3912 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3914 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3918 /* Force some symbol references.
3919 * _tls_used forces the linker to create the TLS directory if not already done
3920 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3922 #pragma comment(linker, "/INCLUDE:_tls_used")
3923 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3924 #pragma const_seg(".CRT$XLB")
3925 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3926 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3929 #pragma comment(linker, "/INCLUDE:__tls_used")
3930 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3931 #pragma data_seg(".CRT$XLB")
3932 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3934 #endif /* WIN 32/64 */
3935 #endif /* !__GNUC__ */
3938 /** Downgrade the exclusive lock on the region back to shared */
3940 mdb_env_share_locks(MDB_env *env, int *excl)
3942 int rc = 0, toggle = mdb_env_pick_meta(env);
3944 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3949 /* First acquire a shared lock. The Unlock will
3950 * then release the existing exclusive lock.
3952 memset(&ov, 0, sizeof(ov));
3953 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3956 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3962 struct flock lock_info;
3963 /* The shared lock replaces the existing lock */
3964 memset((void *)&lock_info, 0, sizeof(lock_info));
3965 lock_info.l_type = F_RDLCK;
3966 lock_info.l_whence = SEEK_SET;
3967 lock_info.l_start = 0;
3968 lock_info.l_len = 1;
3969 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3970 (rc = ErrCode()) == EINTR) ;
3971 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3978 /** Try to get exlusive lock, otherwise shared.
3979 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3982 mdb_env_excl_lock(MDB_env *env, int *excl)
3986 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3990 memset(&ov, 0, sizeof(ov));
3991 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3998 struct flock lock_info;
3999 memset((void *)&lock_info, 0, sizeof(lock_info));
4000 lock_info.l_type = F_WRLCK;
4001 lock_info.l_whence = SEEK_SET;
4002 lock_info.l_start = 0;
4003 lock_info.l_len = 1;
4004 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4005 (rc = ErrCode()) == EINTR) ;
4009 # ifdef MDB_USE_POSIX_SEM
4010 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4013 lock_info.l_type = F_RDLCK;
4014 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4015 (rc = ErrCode()) == EINTR) ;
4025 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4027 * @(#) $Revision: 5.1 $
4028 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4029 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4031 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4035 * Please do not copyright this code. This code is in the public domain.
4037 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4038 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4039 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4040 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4041 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4042 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4043 * PERFORMANCE OF THIS SOFTWARE.
4046 * chongo <Landon Curt Noll> /\oo/\
4047 * http://www.isthe.com/chongo/
4049 * Share and Enjoy! :-)
4052 typedef unsigned long long mdb_hash_t;
4053 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4055 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4056 * @param[in] val value to hash
4057 * @param[in] hval initial value for hash
4058 * @return 64 bit hash
4060 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4061 * hval arg on the first call.
4064 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4066 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4067 unsigned char *end = s + val->mv_size;
4069 * FNV-1a hash each octet of the string
4072 /* xor the bottom with the current octet */
4073 hval ^= (mdb_hash_t)*s++;
4075 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4076 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4077 (hval << 7) + (hval << 8) + (hval << 40);
4079 /* return our new hash value */
4083 /** Hash the string and output the encoded hash.
4084 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4085 * very short name limits. We don't care about the encoding being reversible,
4086 * we just want to preserve as many bits of the input as possible in a
4087 * small printable string.
4088 * @param[in] str string to hash
4089 * @param[out] encbuf an array of 11 chars to hold the hash
4091 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4094 mdb_pack85(unsigned long l, char *out)
4098 for (i=0; i<5; i++) {
4099 *out++ = mdb_a85[l % 85];
4105 mdb_hash_enc(MDB_val *val, char *encbuf)
4107 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4109 mdb_pack85(h, encbuf);
4110 mdb_pack85(h>>32, encbuf+5);
4115 /** Open and/or initialize the lock region for the environment.
4116 * @param[in] env The LMDB environment.
4117 * @param[in] lpath The pathname of the file used for the lock region.
4118 * @param[in] mode The Unix permissions for the file, if we create it.
4119 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4120 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4121 * @return 0 on success, non-zero on failure.
4124 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4127 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4129 # define MDB_ERRCODE_ROFS EROFS
4130 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4131 # define MDB_CLOEXEC O_CLOEXEC
4134 # define MDB_CLOEXEC 0
4141 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4142 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4143 FILE_ATTRIBUTE_NORMAL, NULL);
4145 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4147 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4149 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4154 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4155 /* Lose record locks when exec*() */
4156 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4157 fcntl(env->me_lfd, F_SETFD, fdflags);
4160 if (!(env->me_flags & MDB_NOTLS)) {
4161 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4164 env->me_flags |= MDB_ENV_TXKEY;
4166 /* Windows TLS callbacks need help finding their TLS info. */
4167 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4171 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4175 /* Try to get exclusive lock. If we succeed, then
4176 * nobody is using the lock region and we should initialize it.
4178 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4181 size = GetFileSize(env->me_lfd, NULL);
4183 size = lseek(env->me_lfd, 0, SEEK_END);
4184 if (size == -1) goto fail_errno;
4186 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4187 if (size < rsize && *excl > 0) {
4189 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4190 || !SetEndOfFile(env->me_lfd))
4193 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4197 size = rsize - sizeof(MDB_txninfo);
4198 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4203 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4205 if (!mh) goto fail_errno;
4206 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4208 if (!env->me_txns) goto fail_errno;
4210 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4212 if (m == MAP_FAILED) goto fail_errno;
4218 BY_HANDLE_FILE_INFORMATION stbuf;
4227 if (!mdb_sec_inited) {
4228 InitializeSecurityDescriptor(&mdb_null_sd,
4229 SECURITY_DESCRIPTOR_REVISION);
4230 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4231 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4232 mdb_all_sa.bInheritHandle = FALSE;
4233 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4236 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4237 idbuf.volume = stbuf.dwVolumeSerialNumber;
4238 idbuf.nhigh = stbuf.nFileIndexHigh;
4239 idbuf.nlow = stbuf.nFileIndexLow;
4240 val.mv_data = &idbuf;
4241 val.mv_size = sizeof(idbuf);
4242 mdb_hash_enc(&val, encbuf);
4243 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4244 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4245 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4246 if (!env->me_rmutex) goto fail_errno;
4247 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4248 if (!env->me_wmutex) goto fail_errno;
4249 #elif defined(MDB_USE_POSIX_SEM)
4258 #if defined(__NetBSD__)
4259 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4261 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4262 idbuf.dev = stbuf.st_dev;
4263 idbuf.ino = stbuf.st_ino;
4264 val.mv_data = &idbuf;
4265 val.mv_size = sizeof(idbuf);
4266 mdb_hash_enc(&val, encbuf);
4267 #ifdef MDB_SHORT_SEMNAMES
4268 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4270 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4271 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4272 /* Clean up after a previous run, if needed: Try to
4273 * remove both semaphores before doing anything else.
4275 sem_unlink(env->me_txns->mti_rmname);
4276 sem_unlink(env->me_txns->mti_wmname);
4277 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4278 O_CREAT|O_EXCL, mode, 1);
4279 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4280 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4281 O_CREAT|O_EXCL, mode, 1);
4282 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4283 #else /* MDB_USE_POSIX_SEM */
4284 pthread_mutexattr_t mattr;
4286 if ((rc = pthread_mutexattr_init(&mattr))
4287 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4288 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4289 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4291 pthread_mutexattr_destroy(&mattr);
4292 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4294 env->me_txns->mti_magic = MDB_MAGIC;
4295 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4296 env->me_txns->mti_txnid = 0;
4297 env->me_txns->mti_numreaders = 0;
4300 if (env->me_txns->mti_magic != MDB_MAGIC) {
4301 DPUTS("lock region has invalid magic");
4305 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4306 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4307 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4308 rc = MDB_VERSION_MISMATCH;
4312 if (rc && rc != EACCES && rc != EAGAIN) {
4316 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4317 if (!env->me_rmutex) goto fail_errno;
4318 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4319 if (!env->me_wmutex) goto fail_errno;
4320 #elif defined(MDB_USE_POSIX_SEM)
4321 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4322 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4323 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4324 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4335 /** The name of the lock file in the DB environment */
4336 #define LOCKNAME "/lock.mdb"
4337 /** The name of the data file in the DB environment */
4338 #define DATANAME "/data.mdb"
4339 /** The suffix of the lock file when no subdir is used */
4340 #define LOCKSUFF "-lock"
4341 /** Only a subset of the @ref mdb_env flags can be changed
4342 * at runtime. Changing other flags requires closing the
4343 * environment and re-opening it with the new flags.
4345 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4346 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4347 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4349 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4350 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4354 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4356 int oflags, rc, len, excl = -1;
4357 char *lpath, *dpath;
4359 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4363 if (flags & MDB_NOSUBDIR) {
4364 rc = len + sizeof(LOCKSUFF) + len + 1;
4366 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4371 if (flags & MDB_NOSUBDIR) {
4372 dpath = lpath + len + sizeof(LOCKSUFF);
4373 sprintf(lpath, "%s" LOCKSUFF, path);
4374 strcpy(dpath, path);
4376 dpath = lpath + len + sizeof(LOCKNAME);
4377 sprintf(lpath, "%s" LOCKNAME, path);
4378 sprintf(dpath, "%s" DATANAME, path);
4382 flags |= env->me_flags;
4383 if (flags & MDB_RDONLY) {
4384 /* silently ignore WRITEMAP when we're only getting read access */
4385 flags &= ~MDB_WRITEMAP;
4387 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4388 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4391 env->me_flags = flags |= MDB_ENV_ACTIVE;
4395 env->me_path = strdup(path);
4396 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4397 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4398 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4399 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4404 /* For RDONLY, get lockfile after we know datafile exists */
4405 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4406 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4412 if (F_ISSET(flags, MDB_RDONLY)) {
4413 oflags = GENERIC_READ;
4414 len = OPEN_EXISTING;
4416 oflags = GENERIC_READ|GENERIC_WRITE;
4419 mode = FILE_ATTRIBUTE_NORMAL;
4420 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4421 NULL, len, mode, NULL);
4423 if (F_ISSET(flags, MDB_RDONLY))
4426 oflags = O_RDWR | O_CREAT;
4428 env->me_fd = open(dpath, oflags, mode);
4430 if (env->me_fd == INVALID_HANDLE_VALUE) {
4435 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4436 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4441 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4442 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4443 env->me_mfd = env->me_fd;
4445 /* Synchronous fd for meta writes. Needed even with
4446 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4449 len = OPEN_EXISTING;
4450 env->me_mfd = CreateFile(dpath, oflags,
4451 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4452 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4455 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4457 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4462 DPRINTF(("opened dbenv %p", (void *) env));
4464 rc = mdb_env_share_locks(env, &excl);
4468 if (!((flags & MDB_RDONLY) ||
4469 (env->me_pbuf = calloc(1, env->me_psize))))
4475 mdb_env_close0(env, excl);
4481 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4483 mdb_env_close0(MDB_env *env, int excl)
4487 if (!(env->me_flags & MDB_ENV_ACTIVE))
4490 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4491 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4492 free(env->me_dbxs[i].md_name.mv_data);
4495 free(env->me_dbiseqs);
4496 free(env->me_dbflags);
4499 free(env->me_dirty_list);
4500 mdb_midl_free(env->me_free_pgs);
4502 if (env->me_flags & MDB_ENV_TXKEY) {
4503 pthread_key_delete(env->me_txkey);
4505 /* Delete our key from the global list */
4506 for (i=0; i<mdb_tls_nkeys; i++)
4507 if (mdb_tls_keys[i] == env->me_txkey) {
4508 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4516 munmap(env->me_map, env->me_mapsize);
4518 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4519 (void) close(env->me_mfd);
4520 if (env->me_fd != INVALID_HANDLE_VALUE)
4521 (void) close(env->me_fd);
4523 MDB_PID_T pid = env->me_pid;
4524 /* Clearing readers is done in this function because
4525 * me_txkey with its destructor must be disabled first.
4527 for (i = env->me_numreaders; --i >= 0; )
4528 if (env->me_txns->mti_readers[i].mr_pid == pid)
4529 env->me_txns->mti_readers[i].mr_pid = 0;
4531 if (env->me_rmutex) {
4532 CloseHandle(env->me_rmutex);
4533 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4535 /* Windows automatically destroys the mutexes when
4536 * the last handle closes.
4538 #elif defined(MDB_USE_POSIX_SEM)
4539 if (env->me_rmutex != SEM_FAILED) {
4540 sem_close(env->me_rmutex);
4541 if (env->me_wmutex != SEM_FAILED)
4542 sem_close(env->me_wmutex);
4543 /* If we have the filelock: If we are the
4544 * only remaining user, clean up semaphores.
4547 mdb_env_excl_lock(env, &excl);
4549 sem_unlink(env->me_txns->mti_rmname);
4550 sem_unlink(env->me_txns->mti_wmname);
4554 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4556 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4559 /* Unlock the lockfile. Windows would have unlocked it
4560 * after closing anyway, but not necessarily at once.
4562 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4565 (void) close(env->me_lfd);
4568 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4573 mdb_env_close(MDB_env *env)
4580 VGMEMP_DESTROY(env);
4581 while ((dp = env->me_dpages) != NULL) {
4582 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4583 env->me_dpages = dp->mp_next;
4587 mdb_env_close0(env, 0);
4591 /** Compare two items pointing at aligned size_t's */
4593 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4595 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4596 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4599 /** Compare two items pointing at aligned unsigned int's */
4601 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4603 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4604 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4607 /** Compare two items pointing at unsigned ints of unknown alignment.
4608 * Nodes and keys are guaranteed to be 2-byte aligned.
4611 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4613 #if BYTE_ORDER == LITTLE_ENDIAN
4614 unsigned short *u, *c;
4617 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4618 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4621 } while(!x && u > (unsigned short *)a->mv_data);
4624 unsigned short *u, *c, *end;
4627 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4628 u = (unsigned short *)a->mv_data;
4629 c = (unsigned short *)b->mv_data;
4632 } while(!x && u < end);
4637 /** Compare two items pointing at size_t's of unknown alignment. */
4638 #ifdef MISALIGNED_OK
4639 # define mdb_cmp_clong mdb_cmp_long
4641 # define mdb_cmp_clong mdb_cmp_cint
4644 /** Compare two items lexically */
4646 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4653 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4659 diff = memcmp(a->mv_data, b->mv_data, len);
4660 return diff ? diff : len_diff<0 ? -1 : len_diff;
4663 /** Compare two items in reverse byte order */
4665 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4667 const unsigned char *p1, *p2, *p1_lim;
4671 p1_lim = (const unsigned char *)a->mv_data;
4672 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4673 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4675 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4681 while (p1 > p1_lim) {
4682 diff = *--p1 - *--p2;
4686 return len_diff<0 ? -1 : len_diff;
4689 /** Search for key within a page, using binary search.
4690 * Returns the smallest entry larger or equal to the key.
4691 * If exactp is non-null, stores whether the found entry was an exact match
4692 * in *exactp (1 or 0).
4693 * Updates the cursor index with the index of the found entry.
4694 * If no entry larger or equal to the key is found, returns NULL.
4697 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4699 unsigned int i = 0, nkeys;
4702 MDB_page *mp = mc->mc_pg[mc->mc_top];
4703 MDB_node *node = NULL;
4708 nkeys = NUMKEYS(mp);
4710 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4711 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4714 low = IS_LEAF(mp) ? 0 : 1;
4716 cmp = mc->mc_dbx->md_cmp;
4718 /* Branch pages have no data, so if using integer keys,
4719 * alignment is guaranteed. Use faster mdb_cmp_int.
4721 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4722 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4729 nodekey.mv_size = mc->mc_db->md_pad;
4730 node = NODEPTR(mp, 0); /* fake */
4731 while (low <= high) {
4732 i = (low + high) >> 1;
4733 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4734 rc = cmp(key, &nodekey);
4735 DPRINTF(("found leaf index %u [%s], rc = %i",
4736 i, DKEY(&nodekey), rc));
4745 while (low <= high) {
4746 i = (low + high) >> 1;
4748 node = NODEPTR(mp, i);
4749 nodekey.mv_size = NODEKSZ(node);
4750 nodekey.mv_data = NODEKEY(node);
4752 rc = cmp(key, &nodekey);
4755 DPRINTF(("found leaf index %u [%s], rc = %i",
4756 i, DKEY(&nodekey), rc));
4758 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4759 i, DKEY(&nodekey), NODEPGNO(node), rc));
4770 if (rc > 0) { /* Found entry is less than the key. */
4771 i++; /* Skip to get the smallest entry larger than key. */
4773 node = NODEPTR(mp, i);
4776 *exactp = (rc == 0 && nkeys > 0);
4777 /* store the key index */
4778 mc->mc_ki[mc->mc_top] = i;
4780 /* There is no entry larger or equal to the key. */
4783 /* nodeptr is fake for LEAF2 */
4789 mdb_cursor_adjust(MDB_cursor *mc, func)
4793 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4794 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4801 /** Pop a page off the top of the cursor's stack. */
4803 mdb_cursor_pop(MDB_cursor *mc)
4807 MDB_page *top = mc->mc_pg[mc->mc_top];
4813 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4814 DDBI(mc), (void *) mc));
4818 /** Push a page onto the top of the cursor's stack. */
4820 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4822 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4823 DDBI(mc), (void *) mc));
4825 if (mc->mc_snum >= CURSOR_STACK) {
4826 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4827 return MDB_CURSOR_FULL;
4830 mc->mc_top = mc->mc_snum++;
4831 mc->mc_pg[mc->mc_top] = mp;
4832 mc->mc_ki[mc->mc_top] = 0;
4837 /** Find the address of the page corresponding to a given page number.
4838 * @param[in] txn the transaction for this access.
4839 * @param[in] pgno the page number for the page to retrieve.
4840 * @param[out] ret address of a pointer where the page's address will be stored.
4841 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4842 * @return 0 on success, non-zero on failure.
4845 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4847 MDB_env *env = txn->mt_env;
4851 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4855 MDB_ID2L dl = tx2->mt_u.dirty_list;
4857 /* Spilled pages were dirtied in this txn and flushed
4858 * because the dirty list got full. Bring this page
4859 * back in from the map (but don't unspill it here,
4860 * leave that unless page_touch happens again).
4862 if (tx2->mt_spill_pgs) {
4863 MDB_ID pn = pgno << 1;
4864 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4865 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4866 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4871 unsigned x = mdb_mid2l_search(dl, pgno);
4872 if (x <= dl[0].mid && dl[x].mid == pgno) {
4878 } while ((tx2 = tx2->mt_parent) != NULL);
4881 if (pgno < txn->mt_next_pgno) {
4883 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4885 DPRINTF(("page %"Z"u not found", pgno));
4886 txn->mt_flags |= MDB_TXN_ERROR;
4887 return MDB_PAGE_NOTFOUND;
4897 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4898 * The cursor is at the root page, set up the rest of it.
4901 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4903 MDB_page *mp = mc->mc_pg[mc->mc_top];
4907 while (IS_BRANCH(mp)) {
4911 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4912 mdb_cassert(mc, NUMKEYS(mp) > 1);
4913 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4915 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4917 if (flags & MDB_PS_LAST)
4918 i = NUMKEYS(mp) - 1;
4921 node = mdb_node_search(mc, key, &exact);
4923 i = NUMKEYS(mp) - 1;
4925 i = mc->mc_ki[mc->mc_top];
4927 mdb_cassert(mc, i > 0);
4931 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4934 mdb_cassert(mc, i < NUMKEYS(mp));
4935 node = NODEPTR(mp, i);
4937 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4940 mc->mc_ki[mc->mc_top] = i;
4941 if ((rc = mdb_cursor_push(mc, mp)))
4944 if (flags & MDB_PS_MODIFY) {
4945 if ((rc = mdb_page_touch(mc)) != 0)
4947 mp = mc->mc_pg[mc->mc_top];
4952 DPRINTF(("internal error, index points to a %02X page!?",
4954 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4955 return MDB_CORRUPTED;
4958 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4959 key ? DKEY(key) : "null"));
4960 mc->mc_flags |= C_INITIALIZED;
4961 mc->mc_flags &= ~C_EOF;
4966 /** Search for the lowest key under the current branch page.
4967 * This just bypasses a NUMKEYS check in the current page
4968 * before calling mdb_page_search_root(), because the callers
4969 * are all in situations where the current page is known to
4973 mdb_page_search_lowest(MDB_cursor *mc)
4975 MDB_page *mp = mc->mc_pg[mc->mc_top];
4976 MDB_node *node = NODEPTR(mp, 0);
4979 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4982 mc->mc_ki[mc->mc_top] = 0;
4983 if ((rc = mdb_cursor_push(mc, mp)))
4985 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4988 /** Search for the page a given key should be in.
4989 * Push it and its parent pages on the cursor stack.
4990 * @param[in,out] mc the cursor for this operation.
4991 * @param[in] key the key to search for, or NULL for first/last page.
4992 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4993 * are touched (updated with new page numbers).
4994 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4995 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4996 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4997 * @return 0 on success, non-zero on failure.
5000 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5005 /* Make sure the txn is still viable, then find the root from
5006 * the txn's db table and set it as the root of the cursor's stack.
5008 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5009 DPUTS("transaction has failed, must abort");
5012 /* Make sure we're using an up-to-date root */
5013 if (*mc->mc_dbflag & DB_STALE) {
5015 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5017 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5018 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5025 MDB_node *leaf = mdb_node_search(&mc2,
5026 &mc->mc_dbx->md_name, &exact);
5028 return MDB_NOTFOUND;
5029 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5032 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5034 /* The txn may not know this DBI, or another process may
5035 * have dropped and recreated the DB with other flags.
5037 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5038 return MDB_INCOMPATIBLE;
5039 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5041 *mc->mc_dbflag &= ~DB_STALE;
5043 root = mc->mc_db->md_root;
5045 if (root == P_INVALID) { /* Tree is empty. */
5046 DPUTS("tree is empty");
5047 return MDB_NOTFOUND;
5051 mdb_cassert(mc, root > 1);
5052 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5053 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5059 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5060 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5062 if (flags & MDB_PS_MODIFY) {
5063 if ((rc = mdb_page_touch(mc)))
5067 if (flags & MDB_PS_ROOTONLY)
5070 return mdb_page_search_root(mc, key, flags);
5074 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5076 MDB_txn *txn = mc->mc_txn;
5077 pgno_t pg = mp->mp_pgno;
5078 unsigned x = 0, ovpages = mp->mp_pages;
5079 MDB_env *env = txn->mt_env;
5080 MDB_IDL sl = txn->mt_spill_pgs;
5081 MDB_ID pn = pg << 1;
5084 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5085 /* If the page is dirty or on the spill list we just acquired it,
5086 * so we should give it back to our current free list, if any.
5087 * Otherwise put it onto the list of pages we freed in this txn.
5089 * Won't create me_pghead: me_pglast must be inited along with it.
5090 * Unsupported in nested txns: They would need to hide the page
5091 * range in ancestor txns' dirty and spilled lists.
5093 if (env->me_pghead &&
5095 ((mp->mp_flags & P_DIRTY) ||
5096 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5100 MDB_ID2 *dl, ix, iy;
5101 rc = mdb_midl_need(&env->me_pghead, ovpages);
5104 if (!(mp->mp_flags & P_DIRTY)) {
5105 /* This page is no longer spilled */
5112 /* Remove from dirty list */
5113 dl = txn->mt_u.dirty_list;
5115 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5121 mdb_cassert(mc, x > 1);
5123 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5124 txn->mt_flags |= MDB_TXN_ERROR;
5125 return MDB_CORRUPTED;
5128 if (!(env->me_flags & MDB_WRITEMAP))
5129 mdb_dpage_free(env, mp);
5131 /* Insert in me_pghead */
5132 mop = env->me_pghead;
5133 j = mop[0] + ovpages;
5134 for (i = mop[0]; i && mop[i] < pg; i--)
5140 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5144 mc->mc_db->md_overflow_pages -= ovpages;
5148 /** Return the data associated with a given node.
5149 * @param[in] txn The transaction for this operation.
5150 * @param[in] leaf The node being read.
5151 * @param[out] data Updated to point to the node's data.
5152 * @return 0 on success, non-zero on failure.
5155 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5157 MDB_page *omp; /* overflow page */
5161 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5162 data->mv_size = NODEDSZ(leaf);
5163 data->mv_data = NODEDATA(leaf);
5167 /* Read overflow data.
5169 data->mv_size = NODEDSZ(leaf);
5170 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5171 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5172 DPRINTF(("read overflow page %"Z"u failed", pgno));
5175 data->mv_data = METADATA(omp);
5181 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5182 MDB_val *key, MDB_val *data)
5189 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5191 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5194 if (txn->mt_flags & MDB_TXN_ERROR)
5197 mdb_cursor_init(&mc, txn, dbi, &mx);
5198 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5201 /** Find a sibling for a page.
5202 * Replaces the page at the top of the cursor's stack with the
5203 * specified sibling, if one exists.
5204 * @param[in] mc The cursor for this operation.
5205 * @param[in] move_right Non-zero if the right sibling is requested,
5206 * otherwise the left sibling.
5207 * @return 0 on success, non-zero on failure.
5210 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5216 if (mc->mc_snum < 2) {
5217 return MDB_NOTFOUND; /* root has no siblings */
5221 DPRINTF(("parent page is page %"Z"u, index %u",
5222 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5224 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5225 : (mc->mc_ki[mc->mc_top] == 0)) {
5226 DPRINTF(("no more keys left, moving to %s sibling",
5227 move_right ? "right" : "left"));
5228 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5229 /* undo cursor_pop before returning */
5236 mc->mc_ki[mc->mc_top]++;
5238 mc->mc_ki[mc->mc_top]--;
5239 DPRINTF(("just moving to %s index key %u",
5240 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5242 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5244 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5245 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5246 /* mc will be inconsistent if caller does mc_snum++ as above */
5247 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5251 mdb_cursor_push(mc, mp);
5253 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5258 /** Move the cursor to the next data item. */
5260 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5266 if (mc->mc_flags & C_EOF) {
5267 return MDB_NOTFOUND;
5270 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5272 mp = mc->mc_pg[mc->mc_top];
5274 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5275 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5276 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5277 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5278 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5279 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5280 if (rc == MDB_SUCCESS)
5281 MDB_GET_KEY(leaf, key);
5286 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5287 if (op == MDB_NEXT_DUP)
5288 return MDB_NOTFOUND;
5292 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5293 mdb_dbg_pgno(mp), (void *) mc));
5294 if (mc->mc_flags & C_DEL)
5297 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5298 DPUTS("=====> move to next sibling page");
5299 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5300 mc->mc_flags |= C_EOF;
5303 mp = mc->mc_pg[mc->mc_top];
5304 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5306 mc->mc_ki[mc->mc_top]++;
5309 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5310 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5313 key->mv_size = mc->mc_db->md_pad;
5314 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5318 mdb_cassert(mc, IS_LEAF(mp));
5319 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5321 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5322 mdb_xcursor_init1(mc, leaf);
5325 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5328 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5329 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5330 if (rc != MDB_SUCCESS)
5335 MDB_GET_KEY(leaf, key);
5339 /** Move the cursor to the previous data item. */
5341 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5347 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5349 mp = mc->mc_pg[mc->mc_top];
5351 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5352 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5353 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5354 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5355 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5356 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5357 if (rc == MDB_SUCCESS) {
5358 MDB_GET_KEY(leaf, key);
5359 mc->mc_flags &= ~C_EOF;
5364 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5365 if (op == MDB_PREV_DUP)
5366 return MDB_NOTFOUND;
5371 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5372 mdb_dbg_pgno(mp), (void *) mc));
5374 if (mc->mc_ki[mc->mc_top] == 0) {
5375 DPUTS("=====> move to prev sibling page");
5376 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5379 mp = mc->mc_pg[mc->mc_top];
5380 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5381 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5383 mc->mc_ki[mc->mc_top]--;
5385 mc->mc_flags &= ~C_EOF;
5387 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5388 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5391 key->mv_size = mc->mc_db->md_pad;
5392 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5396 mdb_cassert(mc, IS_LEAF(mp));
5397 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5399 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5400 mdb_xcursor_init1(mc, leaf);
5403 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5406 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5407 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5408 if (rc != MDB_SUCCESS)
5413 MDB_GET_KEY(leaf, key);
5417 /** Set the cursor on a specific data item. */
5419 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5420 MDB_cursor_op op, int *exactp)
5424 MDB_node *leaf = NULL;
5427 if (key->mv_size == 0)
5428 return MDB_BAD_VALSIZE;
5431 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5433 /* See if we're already on the right page */
5434 if (mc->mc_flags & C_INITIALIZED) {
5437 mp = mc->mc_pg[mc->mc_top];
5439 mc->mc_ki[mc->mc_top] = 0;
5440 return MDB_NOTFOUND;
5442 if (mp->mp_flags & P_LEAF2) {
5443 nodekey.mv_size = mc->mc_db->md_pad;
5444 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5446 leaf = NODEPTR(mp, 0);
5447 MDB_GET_KEY2(leaf, nodekey);
5449 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5451 /* Probably happens rarely, but first node on the page
5452 * was the one we wanted.
5454 mc->mc_ki[mc->mc_top] = 0;
5461 unsigned int nkeys = NUMKEYS(mp);
5463 if (mp->mp_flags & P_LEAF2) {
5464 nodekey.mv_data = LEAF2KEY(mp,
5465 nkeys-1, nodekey.mv_size);
5467 leaf = NODEPTR(mp, nkeys-1);
5468 MDB_GET_KEY2(leaf, nodekey);
5470 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5472 /* last node was the one we wanted */
5473 mc->mc_ki[mc->mc_top] = nkeys-1;
5479 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5480 /* This is definitely the right page, skip search_page */
5481 if (mp->mp_flags & P_LEAF2) {
5482 nodekey.mv_data = LEAF2KEY(mp,
5483 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5485 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5486 MDB_GET_KEY2(leaf, nodekey);
5488 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5490 /* current node was the one we wanted */
5500 /* If any parents have right-sibs, search.
5501 * Otherwise, there's nothing further.
5503 for (i=0; i<mc->mc_top; i++)
5505 NUMKEYS(mc->mc_pg[i])-1)
5507 if (i == mc->mc_top) {
5508 /* There are no other pages */
5509 mc->mc_ki[mc->mc_top] = nkeys;
5510 return MDB_NOTFOUND;
5514 /* There are no other pages */
5515 mc->mc_ki[mc->mc_top] = 0;
5516 if (op == MDB_SET_RANGE && !exactp) {
5520 return MDB_NOTFOUND;
5524 rc = mdb_page_search(mc, key, 0);
5525 if (rc != MDB_SUCCESS)
5528 mp = mc->mc_pg[mc->mc_top];
5529 mdb_cassert(mc, IS_LEAF(mp));
5532 leaf = mdb_node_search(mc, key, exactp);
5533 if (exactp != NULL && !*exactp) {
5534 /* MDB_SET specified and not an exact match. */
5535 return MDB_NOTFOUND;
5539 DPUTS("===> inexact leaf not found, goto sibling");
5540 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5541 return rc; /* no entries matched */
5542 mp = mc->mc_pg[mc->mc_top];
5543 mdb_cassert(mc, IS_LEAF(mp));
5544 leaf = NODEPTR(mp, 0);
5548 mc->mc_flags |= C_INITIALIZED;
5549 mc->mc_flags &= ~C_EOF;
5552 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5553 key->mv_size = mc->mc_db->md_pad;
5554 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5559 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5560 mdb_xcursor_init1(mc, leaf);
5563 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5564 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5565 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5568 if (op == MDB_GET_BOTH) {
5574 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5575 if (rc != MDB_SUCCESS)
5578 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5580 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5582 rc = mc->mc_dbx->md_dcmp(data, &d2);
5584 if (op == MDB_GET_BOTH || rc > 0)
5585 return MDB_NOTFOUND;
5592 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5593 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5598 /* The key already matches in all other cases */
5599 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5600 MDB_GET_KEY(leaf, key);
5601 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5606 /** Move the cursor to the first item in the database. */
5608 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5614 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5616 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5617 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5618 if (rc != MDB_SUCCESS)
5621 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5623 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5624 mc->mc_flags |= C_INITIALIZED;
5625 mc->mc_flags &= ~C_EOF;
5627 mc->mc_ki[mc->mc_top] = 0;
5629 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5630 key->mv_size = mc->mc_db->md_pad;
5631 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5636 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5637 mdb_xcursor_init1(mc, leaf);
5638 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5642 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5646 MDB_GET_KEY(leaf, key);
5650 /** Move the cursor to the last item in the database. */
5652 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5658 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5660 if (!(mc->mc_flags & C_EOF)) {
5662 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5663 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5664 if (rc != MDB_SUCCESS)
5667 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5670 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5671 mc->mc_flags |= C_INITIALIZED|C_EOF;
5672 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5674 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5675 key->mv_size = mc->mc_db->md_pad;
5676 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5681 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5682 mdb_xcursor_init1(mc, leaf);
5683 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5687 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5692 MDB_GET_KEY(leaf, key);
5697 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5702 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5707 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5711 case MDB_GET_CURRENT:
5712 if (!(mc->mc_flags & C_INITIALIZED)) {
5715 MDB_page *mp = mc->mc_pg[mc->mc_top];
5716 int nkeys = NUMKEYS(mp);
5717 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5718 mc->mc_ki[mc->mc_top] = nkeys;
5724 key->mv_size = mc->mc_db->md_pad;
5725 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5727 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5728 MDB_GET_KEY(leaf, key);
5730 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5731 if (mc->mc_flags & C_DEL)
5732 mdb_xcursor_init1(mc, leaf);
5733 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5735 rc = mdb_node_read(mc->mc_txn, leaf, data);
5742 case MDB_GET_BOTH_RANGE:
5747 if (mc->mc_xcursor == NULL) {
5748 rc = MDB_INCOMPATIBLE;
5758 rc = mdb_cursor_set(mc, key, data, op,
5759 op == MDB_SET_RANGE ? NULL : &exact);
5762 case MDB_GET_MULTIPLE:
5763 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5767 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5768 rc = MDB_INCOMPATIBLE;
5772 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5773 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5776 case MDB_NEXT_MULTIPLE:
5781 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5782 rc = MDB_INCOMPATIBLE;
5785 if (!(mc->mc_flags & C_INITIALIZED))
5786 rc = mdb_cursor_first(mc, key, data);
5788 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5789 if (rc == MDB_SUCCESS) {
5790 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5793 mx = &mc->mc_xcursor->mx_cursor;
5794 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5796 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5797 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5805 case MDB_NEXT_NODUP:
5806 if (!(mc->mc_flags & C_INITIALIZED))
5807 rc = mdb_cursor_first(mc, key, data);
5809 rc = mdb_cursor_next(mc, key, data, op);
5813 case MDB_PREV_NODUP:
5814 if (!(mc->mc_flags & C_INITIALIZED)) {
5815 rc = mdb_cursor_last(mc, key, data);
5818 mc->mc_flags |= C_INITIALIZED;
5819 mc->mc_ki[mc->mc_top]++;
5821 rc = mdb_cursor_prev(mc, key, data, op);
5824 rc = mdb_cursor_first(mc, key, data);
5827 mfunc = mdb_cursor_first;
5829 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5833 if (mc->mc_xcursor == NULL) {
5834 rc = MDB_INCOMPATIBLE;
5838 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5839 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5840 MDB_GET_KEY(leaf, key);
5842 rc = mdb_node_read(mc->mc_txn, leaf, data);
5846 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5850 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5853 rc = mdb_cursor_last(mc, key, data);
5856 mfunc = mdb_cursor_last;
5859 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5864 if (mc->mc_flags & C_DEL)
5865 mc->mc_flags ^= C_DEL;
5870 /** Touch all the pages in the cursor stack. Set mc_top.
5871 * Makes sure all the pages are writable, before attempting a write operation.
5872 * @param[in] mc The cursor to operate on.
5875 mdb_cursor_touch(MDB_cursor *mc)
5877 int rc = MDB_SUCCESS;
5879 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5882 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5884 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5885 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5888 *mc->mc_dbflag |= DB_DIRTY;
5893 rc = mdb_page_touch(mc);
5894 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5895 mc->mc_top = mc->mc_snum-1;
5900 /** Do not spill pages to disk if txn is getting full, may fail instead */
5901 #define MDB_NOSPILL 0x8000
5904 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5907 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5909 MDB_node *leaf = NULL;
5912 MDB_val xdata, *rdata, dkey, olddata;
5914 int do_sub = 0, insert_key, insert_data;
5915 unsigned int mcount = 0, dcount = 0, nospill;
5918 unsigned int nflags;
5921 if (mc == NULL || key == NULL)
5924 env = mc->mc_txn->mt_env;
5926 /* Check this first so counter will always be zero on any
5929 if (flags & MDB_MULTIPLE) {
5930 dcount = data[1].mv_size;
5931 data[1].mv_size = 0;
5932 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5933 return MDB_INCOMPATIBLE;
5936 nospill = flags & MDB_NOSPILL;
5937 flags &= ~MDB_NOSPILL;
5939 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5940 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5942 if (key->mv_size-1 >= ENV_MAXKEY(env))
5943 return MDB_BAD_VALSIZE;
5945 #if SIZE_MAX > MAXDATASIZE
5946 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5947 return MDB_BAD_VALSIZE;
5949 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5950 return MDB_BAD_VALSIZE;
5953 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5954 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5958 if (flags == MDB_CURRENT) {
5959 if (!(mc->mc_flags & C_INITIALIZED))
5962 } else if (mc->mc_db->md_root == P_INVALID) {
5963 /* new database, cursor has nothing to point to */
5966 mc->mc_flags &= ~C_INITIALIZED;
5971 if (flags & MDB_APPEND) {
5973 rc = mdb_cursor_last(mc, &k2, &d2);
5975 rc = mc->mc_dbx->md_cmp(key, &k2);
5978 mc->mc_ki[mc->mc_top]++;
5980 /* new key is <= last key */
5985 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5987 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5988 DPRINTF(("duplicate key [%s]", DKEY(key)));
5990 return MDB_KEYEXIST;
5992 if (rc && rc != MDB_NOTFOUND)
5996 if (mc->mc_flags & C_DEL)
5997 mc->mc_flags ^= C_DEL;
5999 /* Cursor is positioned, check for room in the dirty list */
6001 if (flags & MDB_MULTIPLE) {
6003 xdata.mv_size = data->mv_size * dcount;
6007 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6011 if (rc == MDB_NO_ROOT) {
6013 /* new database, write a root leaf page */
6014 DPUTS("allocating new root leaf page");
6015 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6018 mdb_cursor_push(mc, np);
6019 mc->mc_db->md_root = np->mp_pgno;
6020 mc->mc_db->md_depth++;
6021 *mc->mc_dbflag |= DB_DIRTY;
6022 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6024 np->mp_flags |= P_LEAF2;
6025 mc->mc_flags |= C_INITIALIZED;
6027 /* make sure all cursor pages are writable */
6028 rc2 = mdb_cursor_touch(mc);
6033 insert_key = insert_data = rc;
6035 /* The key does not exist */
6036 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6037 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6038 LEAFSIZE(key, data) > env->me_nodemax)
6040 /* Too big for a node, insert in sub-DB. Set up an empty
6041 * "old sub-page" for prep_subDB to expand to a full page.
6043 fp_flags = P_LEAF|P_DIRTY;
6045 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6046 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6047 olddata.mv_size = PAGEHDRSZ;
6051 /* there's only a key anyway, so this is a no-op */
6052 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6054 unsigned int ksize = mc->mc_db->md_pad;
6055 if (key->mv_size != ksize)
6056 return MDB_BAD_VALSIZE;
6057 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6058 memcpy(ptr, key->mv_data, ksize);
6060 /* if overwriting slot 0 of leaf, need to
6061 * update branch key if there is a parent page
6063 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6064 unsigned short top = mc->mc_top;
6066 /* slot 0 is always an empty key, find real slot */
6067 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6069 if (mc->mc_ki[mc->mc_top])
6070 rc2 = mdb_update_key(mc, key);
6081 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6082 olddata.mv_size = NODEDSZ(leaf);
6083 olddata.mv_data = NODEDATA(leaf);
6086 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6087 /* Prepare (sub-)page/sub-DB to accept the new item,
6088 * if needed. fp: old sub-page or a header faking
6089 * it. mp: new (sub-)page. offset: growth in page
6090 * size. xdata: node data with new page or DB.
6092 unsigned i, offset = 0;
6093 mp = fp = xdata.mv_data = env->me_pbuf;
6094 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6096 /* Was a single item before, must convert now */
6097 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6098 /* Just overwrite the current item */
6099 if (flags == MDB_CURRENT)
6102 #if UINT_MAX < SIZE_MAX
6103 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6104 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6106 /* does data match? */
6107 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6108 if (flags & MDB_NODUPDATA)
6109 return MDB_KEYEXIST;
6114 /* Back up original data item */
6115 dkey.mv_size = olddata.mv_size;
6116 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6118 /* Make sub-page header for the dup items, with dummy body */
6119 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6120 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6121 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6122 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6123 fp->mp_flags |= P_LEAF2;
6124 fp->mp_pad = data->mv_size;
6125 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6127 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6128 (dkey.mv_size & 1) + (data->mv_size & 1);
6130 fp->mp_upper = xdata.mv_size - PAGEBASE;
6131 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6132 } else if (leaf->mn_flags & F_SUBDATA) {
6133 /* Data is on sub-DB, just store it */
6134 flags |= F_DUPDATA|F_SUBDATA;
6137 /* Data is on sub-page */
6138 fp = olddata.mv_data;
6141 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6142 offset = EVEN(NODESIZE + sizeof(indx_t) +
6146 offset = fp->mp_pad;
6147 if (SIZELEFT(fp) < offset) {
6148 offset *= 4; /* space for 4 more */
6151 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6153 fp->mp_flags |= P_DIRTY;
6154 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6155 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6159 xdata.mv_size = olddata.mv_size + offset;
6162 fp_flags = fp->mp_flags;
6163 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6164 /* Too big for a sub-page, convert to sub-DB */
6165 fp_flags &= ~P_SUBP;
6167 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6168 fp_flags |= P_LEAF2;
6169 dummy.md_pad = fp->mp_pad;
6170 dummy.md_flags = MDB_DUPFIXED;
6171 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6172 dummy.md_flags |= MDB_INTEGERKEY;
6178 dummy.md_branch_pages = 0;
6179 dummy.md_leaf_pages = 1;
6180 dummy.md_overflow_pages = 0;
6181 dummy.md_entries = NUMKEYS(fp);
6182 xdata.mv_size = sizeof(MDB_db);
6183 xdata.mv_data = &dummy;
6184 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6186 offset = env->me_psize - olddata.mv_size;
6187 flags |= F_DUPDATA|F_SUBDATA;
6188 dummy.md_root = mp->mp_pgno;
6191 mp->mp_flags = fp_flags | P_DIRTY;
6192 mp->mp_pad = fp->mp_pad;
6193 mp->mp_lower = fp->mp_lower;
6194 mp->mp_upper = fp->mp_upper + offset;
6195 if (fp_flags & P_LEAF2) {
6196 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6198 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6199 olddata.mv_size - fp->mp_upper - PAGEBASE);
6200 for (i=0; i<NUMKEYS(fp); i++)
6201 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6209 mdb_node_del(mc, 0);
6213 /* overflow page overwrites need special handling */
6214 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6217 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6219 memcpy(&pg, olddata.mv_data, sizeof(pg));
6220 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6222 ovpages = omp->mp_pages;
6224 /* Is the ov page large enough? */
6225 if (ovpages >= dpages) {
6226 if (!(omp->mp_flags & P_DIRTY) &&
6227 (level || (env->me_flags & MDB_WRITEMAP)))
6229 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6232 level = 0; /* dirty in this txn or clean */
6235 if (omp->mp_flags & P_DIRTY) {
6236 /* yes, overwrite it. Note in this case we don't
6237 * bother to try shrinking the page if the new data
6238 * is smaller than the overflow threshold.
6241 /* It is writable only in a parent txn */
6242 size_t sz = (size_t) env->me_psize * ovpages, off;
6243 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6249 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6250 mdb_cassert(mc, rc2 == 0);
6251 if (!(flags & MDB_RESERVE)) {
6252 /* Copy end of page, adjusting alignment so
6253 * compiler may copy words instead of bytes.
6255 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6256 memcpy((size_t *)((char *)np + off),
6257 (size_t *)((char *)omp + off), sz - off);
6260 memcpy(np, omp, sz); /* Copy beginning of page */
6263 SETDSZ(leaf, data->mv_size);
6264 if (F_ISSET(flags, MDB_RESERVE))
6265 data->mv_data = METADATA(omp);
6267 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6271 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6273 } else if (data->mv_size == olddata.mv_size) {
6274 /* same size, just replace it. Note that we could
6275 * also reuse this node if the new data is smaller,
6276 * but instead we opt to shrink the node in that case.
6278 if (F_ISSET(flags, MDB_RESERVE))
6279 data->mv_data = olddata.mv_data;
6280 else if (!(mc->mc_flags & C_SUB))
6281 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6283 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6288 mdb_node_del(mc, 0);
6294 nflags = flags & NODE_ADD_FLAGS;
6295 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6296 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6297 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6298 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6300 nflags |= MDB_SPLIT_REPLACE;
6301 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6303 /* There is room already in this leaf page. */
6304 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6305 if (rc == 0 && insert_key) {
6306 /* Adjust other cursors pointing to mp */
6307 MDB_cursor *m2, *m3;
6308 MDB_dbi dbi = mc->mc_dbi;
6309 unsigned i = mc->mc_top;
6310 MDB_page *mp = mc->mc_pg[i];
6312 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6313 if (mc->mc_flags & C_SUB)
6314 m3 = &m2->mc_xcursor->mx_cursor;
6317 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6318 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6325 if (rc == MDB_SUCCESS) {
6326 /* Now store the actual data in the child DB. Note that we're
6327 * storing the user data in the keys field, so there are strict
6328 * size limits on dupdata. The actual data fields of the child
6329 * DB are all zero size.
6337 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6338 if (flags & MDB_CURRENT) {
6339 xflags = MDB_CURRENT|MDB_NOSPILL;
6341 mdb_xcursor_init1(mc, leaf);
6342 xflags = (flags & MDB_NODUPDATA) ?
6343 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6345 /* converted, write the original data first */
6347 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6351 /* Adjust other cursors pointing to mp */
6353 unsigned i = mc->mc_top;
6354 MDB_page *mp = mc->mc_pg[i];
6356 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6357 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6358 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6359 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6360 mdb_xcursor_init1(m2, leaf);
6364 /* we've done our job */
6367 ecount = mc->mc_xcursor->mx_db.md_entries;
6368 if (flags & MDB_APPENDDUP)
6369 xflags |= MDB_APPEND;
6370 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6371 if (flags & F_SUBDATA) {
6372 void *db = NODEDATA(leaf);
6373 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6375 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6377 /* Increment count unless we just replaced an existing item. */
6379 mc->mc_db->md_entries++;
6381 /* Invalidate txn if we created an empty sub-DB */
6384 /* If we succeeded and the key didn't exist before,
6385 * make sure the cursor is marked valid.
6387 mc->mc_flags |= C_INITIALIZED;
6389 if (flags & MDB_MULTIPLE) {
6392 /* let caller know how many succeeded, if any */
6393 data[1].mv_size = mcount;
6394 if (mcount < dcount) {
6395 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6396 insert_key = insert_data = 0;
6403 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6406 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6411 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6417 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6418 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6420 if (!(mc->mc_flags & C_INITIALIZED))
6423 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6424 return MDB_NOTFOUND;
6426 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6429 rc = mdb_cursor_touch(mc);
6433 mp = mc->mc_pg[mc->mc_top];
6436 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6438 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6439 if (flags & MDB_NODUPDATA) {
6440 /* mdb_cursor_del0() will subtract the final entry */
6441 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6443 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6444 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6446 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6449 /* If sub-DB still has entries, we're done */
6450 if (mc->mc_xcursor->mx_db.md_entries) {
6451 if (leaf->mn_flags & F_SUBDATA) {
6452 /* update subDB info */
6453 void *db = NODEDATA(leaf);
6454 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6457 /* shrink fake page */
6458 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6459 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6460 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6461 /* fix other sub-DB cursors pointed at this fake page */
6462 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6463 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6464 if (m2->mc_pg[mc->mc_top] == mp &&
6465 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6466 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6469 mc->mc_db->md_entries--;
6470 mc->mc_flags |= C_DEL;
6473 /* otherwise fall thru and delete the sub-DB */
6476 if (leaf->mn_flags & F_SUBDATA) {
6477 /* add all the child DB's pages to the free list */
6478 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6484 /* add overflow pages to free list */
6485 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6489 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6490 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6491 (rc = mdb_ovpage_free(mc, omp)))
6496 return mdb_cursor_del0(mc);
6499 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6503 /** Allocate and initialize new pages for a database.
6504 * @param[in] mc a cursor on the database being added to.
6505 * @param[in] flags flags defining what type of page is being allocated.
6506 * @param[in] num the number of pages to allocate. This is usually 1,
6507 * unless allocating overflow pages for a large record.
6508 * @param[out] mp Address of a page, or NULL on failure.
6509 * @return 0 on success, non-zero on failure.
6512 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6517 if ((rc = mdb_page_alloc(mc, num, &np)))
6519 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6520 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6521 np->mp_flags = flags | P_DIRTY;
6522 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6523 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6526 mc->mc_db->md_branch_pages++;
6527 else if (IS_LEAF(np))
6528 mc->mc_db->md_leaf_pages++;
6529 else if (IS_OVERFLOW(np)) {
6530 mc->mc_db->md_overflow_pages += num;
6538 /** Calculate the size of a leaf node.
6539 * The size depends on the environment's page size; if a data item
6540 * is too large it will be put onto an overflow page and the node
6541 * size will only include the key and not the data. Sizes are always
6542 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6543 * of the #MDB_node headers.
6544 * @param[in] env The environment handle.
6545 * @param[in] key The key for the node.
6546 * @param[in] data The data for the node.
6547 * @return The number of bytes needed to store the node.
6550 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6554 sz = LEAFSIZE(key, data);
6555 if (sz > env->me_nodemax) {
6556 /* put on overflow page */
6557 sz -= data->mv_size - sizeof(pgno_t);
6560 return EVEN(sz + sizeof(indx_t));
6563 /** Calculate the size of a branch node.
6564 * The size should depend on the environment's page size but since
6565 * we currently don't support spilling large keys onto overflow
6566 * pages, it's simply the size of the #MDB_node header plus the
6567 * size of the key. Sizes are always rounded up to an even number
6568 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6569 * @param[in] env The environment handle.
6570 * @param[in] key The key for the node.
6571 * @return The number of bytes needed to store the node.
6574 mdb_branch_size(MDB_env *env, MDB_val *key)
6579 if (sz > env->me_nodemax) {
6580 /* put on overflow page */
6581 /* not implemented */
6582 /* sz -= key->size - sizeof(pgno_t); */
6585 return sz + sizeof(indx_t);
6588 /** Add a node to the page pointed to by the cursor.
6589 * @param[in] mc The cursor for this operation.
6590 * @param[in] indx The index on the page where the new node should be added.
6591 * @param[in] key The key for the new node.
6592 * @param[in] data The data for the new node, if any.
6593 * @param[in] pgno The page number, if adding a branch node.
6594 * @param[in] flags Flags for the node.
6595 * @return 0 on success, non-zero on failure. Possible errors are:
6597 * <li>ENOMEM - failed to allocate overflow pages for the node.
6598 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6599 * should never happen since all callers already calculate the
6600 * page's free space before calling this function.
6604 mdb_node_add(MDB_cursor *mc, indx_t indx,
6605 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6608 size_t node_size = NODESIZE;
6612 MDB_page *mp = mc->mc_pg[mc->mc_top];
6613 MDB_page *ofp = NULL; /* overflow page */
6616 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6618 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6619 IS_LEAF(mp) ? "leaf" : "branch",
6620 IS_SUBP(mp) ? "sub-" : "",
6621 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6622 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6625 /* Move higher keys up one slot. */
6626 int ksize = mc->mc_db->md_pad, dif;
6627 char *ptr = LEAF2KEY(mp, indx, ksize);
6628 dif = NUMKEYS(mp) - indx;
6630 memmove(ptr+ksize, ptr, dif*ksize);
6631 /* insert new key */
6632 memcpy(ptr, key->mv_data, ksize);
6634 /* Just using these for counting */
6635 mp->mp_lower += sizeof(indx_t);
6636 mp->mp_upper -= ksize - sizeof(indx_t);
6640 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6642 node_size += key->mv_size;
6644 mdb_cassert(mc, data);
6645 if (F_ISSET(flags, F_BIGDATA)) {
6646 /* Data already on overflow page. */
6647 node_size += sizeof(pgno_t);
6648 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6649 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6651 /* Put data on overflow page. */
6652 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6653 data->mv_size, node_size+data->mv_size));
6654 node_size = EVEN(node_size + sizeof(pgno_t));
6655 if ((ssize_t)node_size > room)
6657 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6659 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6663 node_size += data->mv_size;
6666 node_size = EVEN(node_size);
6667 if ((ssize_t)node_size > room)
6671 /* Move higher pointers up one slot. */
6672 for (i = NUMKEYS(mp); i > indx; i--)
6673 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6675 /* Adjust free space offsets. */
6676 ofs = mp->mp_upper - node_size;
6677 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6678 mp->mp_ptrs[indx] = ofs;
6680 mp->mp_lower += sizeof(indx_t);
6682 /* Write the node data. */
6683 node = NODEPTR(mp, indx);
6684 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6685 node->mn_flags = flags;
6687 SETDSZ(node,data->mv_size);
6692 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6695 mdb_cassert(mc, key);
6697 if (F_ISSET(flags, F_BIGDATA))
6698 memcpy(node->mn_data + key->mv_size, data->mv_data,
6700 else if (F_ISSET(flags, MDB_RESERVE))
6701 data->mv_data = node->mn_data + key->mv_size;
6703 memcpy(node->mn_data + key->mv_size, data->mv_data,
6706 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6708 if (F_ISSET(flags, MDB_RESERVE))
6709 data->mv_data = METADATA(ofp);
6711 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6718 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6719 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6720 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6721 DPRINTF(("node size = %"Z"u", node_size));
6722 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6723 return MDB_PAGE_FULL;
6726 /** Delete the specified node from a page.
6727 * @param[in] mc Cursor pointing to the node to delete.
6728 * @param[in] ksize The size of a node. Only used if the page is
6729 * part of a #MDB_DUPFIXED database.
6732 mdb_node_del(MDB_cursor *mc, int ksize)
6734 MDB_page *mp = mc->mc_pg[mc->mc_top];
6735 indx_t indx = mc->mc_ki[mc->mc_top];
6737 indx_t i, j, numkeys, ptr;
6741 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6742 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6743 numkeys = NUMKEYS(mp);
6744 mdb_cassert(mc, indx < numkeys);
6747 int x = numkeys - 1 - indx;
6748 base = LEAF2KEY(mp, indx, ksize);
6750 memmove(base, base + ksize, x * ksize);
6751 mp->mp_lower -= sizeof(indx_t);
6752 mp->mp_upper += ksize - sizeof(indx_t);
6756 node = NODEPTR(mp, indx);
6757 sz = NODESIZE + node->mn_ksize;
6759 if (F_ISSET(node->mn_flags, F_BIGDATA))
6760 sz += sizeof(pgno_t);
6762 sz += NODEDSZ(node);
6766 ptr = mp->mp_ptrs[indx];
6767 for (i = j = 0; i < numkeys; i++) {
6769 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6770 if (mp->mp_ptrs[i] < ptr)
6771 mp->mp_ptrs[j] += sz;
6776 base = (char *)mp + mp->mp_upper + PAGEBASE;
6777 memmove(base + sz, base, ptr - mp->mp_upper);
6779 mp->mp_lower -= sizeof(indx_t);
6783 /** Compact the main page after deleting a node on a subpage.
6784 * @param[in] mp The main page to operate on.
6785 * @param[in] indx The index of the subpage on the main page.
6788 mdb_node_shrink(MDB_page *mp, indx_t indx)
6794 indx_t i, numkeys, ptr;
6796 node = NODEPTR(mp, indx);
6797 sp = (MDB_page *)NODEDATA(node);
6798 delta = SIZELEFT(sp);
6799 xp = (MDB_page *)((char *)sp + delta);
6801 /* shift subpage upward */
6803 nsize = NUMKEYS(sp) * sp->mp_pad;
6805 return; /* do not make the node uneven-sized */
6806 memmove(METADATA(xp), METADATA(sp), nsize);
6809 numkeys = NUMKEYS(sp);
6810 for (i=numkeys-1; i>=0; i--)
6811 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6813 xp->mp_upper = sp->mp_lower;
6814 xp->mp_lower = sp->mp_lower;
6815 xp->mp_flags = sp->mp_flags;
6816 xp->mp_pad = sp->mp_pad;
6817 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6819 nsize = NODEDSZ(node) - delta;
6820 SETDSZ(node, nsize);
6822 /* shift lower nodes upward */
6823 ptr = mp->mp_ptrs[indx];
6824 numkeys = NUMKEYS(mp);
6825 for (i = 0; i < numkeys; i++) {
6826 if (mp->mp_ptrs[i] <= ptr)
6827 mp->mp_ptrs[i] += delta;
6830 base = (char *)mp + mp->mp_upper + PAGEBASE;
6831 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6832 mp->mp_upper += delta;
6835 /** Initial setup of a sorted-dups cursor.
6836 * Sorted duplicates are implemented as a sub-database for the given key.
6837 * The duplicate data items are actually keys of the sub-database.
6838 * Operations on the duplicate data items are performed using a sub-cursor
6839 * initialized when the sub-database is first accessed. This function does
6840 * the preliminary setup of the sub-cursor, filling in the fields that
6841 * depend only on the parent DB.
6842 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6845 mdb_xcursor_init0(MDB_cursor *mc)
6847 MDB_xcursor *mx = mc->mc_xcursor;
6849 mx->mx_cursor.mc_xcursor = NULL;
6850 mx->mx_cursor.mc_txn = mc->mc_txn;
6851 mx->mx_cursor.mc_db = &mx->mx_db;
6852 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6853 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6854 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6855 mx->mx_cursor.mc_snum = 0;
6856 mx->mx_cursor.mc_top = 0;
6857 mx->mx_cursor.mc_flags = C_SUB;
6858 mx->mx_dbx.md_name.mv_size = 0;
6859 mx->mx_dbx.md_name.mv_data = NULL;
6860 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6861 mx->mx_dbx.md_dcmp = NULL;
6862 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6865 /** Final setup of a sorted-dups cursor.
6866 * Sets up the fields that depend on the data from the main cursor.
6867 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6868 * @param[in] node The data containing the #MDB_db record for the
6869 * sorted-dup database.
6872 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6874 MDB_xcursor *mx = mc->mc_xcursor;
6876 if (node->mn_flags & F_SUBDATA) {
6877 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6878 mx->mx_cursor.mc_pg[0] = 0;
6879 mx->mx_cursor.mc_snum = 0;
6880 mx->mx_cursor.mc_top = 0;
6881 mx->mx_cursor.mc_flags = C_SUB;
6883 MDB_page *fp = NODEDATA(node);
6884 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6885 mx->mx_db.md_flags = 0;
6886 mx->mx_db.md_depth = 1;
6887 mx->mx_db.md_branch_pages = 0;
6888 mx->mx_db.md_leaf_pages = 1;
6889 mx->mx_db.md_overflow_pages = 0;
6890 mx->mx_db.md_entries = NUMKEYS(fp);
6891 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6892 mx->mx_cursor.mc_snum = 1;
6893 mx->mx_cursor.mc_top = 0;
6894 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6895 mx->mx_cursor.mc_pg[0] = fp;
6896 mx->mx_cursor.mc_ki[0] = 0;
6897 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6898 mx->mx_db.md_flags = MDB_DUPFIXED;
6899 mx->mx_db.md_pad = fp->mp_pad;
6900 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6901 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6904 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6905 mx->mx_db.md_root));
6906 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6907 #if UINT_MAX < SIZE_MAX
6908 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6909 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6913 /** Initialize a cursor for a given transaction and database. */
6915 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6918 mc->mc_backup = NULL;
6921 mc->mc_db = &txn->mt_dbs[dbi];
6922 mc->mc_dbx = &txn->mt_dbxs[dbi];
6923 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6928 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6929 mdb_tassert(txn, mx != NULL);
6930 mc->mc_xcursor = mx;
6931 mdb_xcursor_init0(mc);
6933 mc->mc_xcursor = NULL;
6935 if (*mc->mc_dbflag & DB_STALE) {
6936 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6941 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6944 size_t size = sizeof(MDB_cursor);
6946 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6949 if (txn->mt_flags & MDB_TXN_ERROR)
6952 /* Allow read access to the freelist */
6953 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6956 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6957 size += sizeof(MDB_xcursor);
6959 if ((mc = malloc(size)) != NULL) {
6960 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6961 if (txn->mt_cursors) {
6962 mc->mc_next = txn->mt_cursors[dbi];
6963 txn->mt_cursors[dbi] = mc;
6964 mc->mc_flags |= C_UNTRACK;
6976 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6978 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6981 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6984 if (txn->mt_flags & MDB_TXN_ERROR)
6987 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6991 /* Return the count of duplicate data items for the current key */
6993 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6997 if (mc == NULL || countp == NULL)
7000 if (mc->mc_xcursor == NULL)
7001 return MDB_INCOMPATIBLE;
7003 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7006 if (!(mc->mc_flags & C_INITIALIZED))
7009 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7010 return MDB_NOTFOUND;
7012 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7013 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7016 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7019 *countp = mc->mc_xcursor->mx_db.md_entries;
7025 mdb_cursor_close(MDB_cursor *mc)
7027 if (mc && !mc->mc_backup) {
7028 /* remove from txn, if tracked */
7029 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7030 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7031 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7033 *prev = mc->mc_next;
7040 mdb_cursor_txn(MDB_cursor *mc)
7042 if (!mc) return NULL;
7047 mdb_cursor_dbi(MDB_cursor *mc)
7052 /** Replace the key for a branch node with a new key.
7053 * @param[in] mc Cursor pointing to the node to operate on.
7054 * @param[in] key The new key to use.
7055 * @return 0 on success, non-zero on failure.
7058 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7064 int delta, ksize, oksize;
7065 indx_t ptr, i, numkeys, indx;
7068 indx = mc->mc_ki[mc->mc_top];
7069 mp = mc->mc_pg[mc->mc_top];
7070 node = NODEPTR(mp, indx);
7071 ptr = mp->mp_ptrs[indx];
7075 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7076 k2.mv_data = NODEKEY(node);
7077 k2.mv_size = node->mn_ksize;
7078 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7080 mdb_dkey(&k2, kbuf2),
7086 /* Sizes must be 2-byte aligned. */
7087 ksize = EVEN(key->mv_size);
7088 oksize = EVEN(node->mn_ksize);
7089 delta = ksize - oksize;
7091 /* Shift node contents if EVEN(key length) changed. */
7093 if (delta > 0 && SIZELEFT(mp) < delta) {
7095 /* not enough space left, do a delete and split */
7096 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7097 pgno = NODEPGNO(node);
7098 mdb_node_del(mc, 0);
7099 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7102 numkeys = NUMKEYS(mp);
7103 for (i = 0; i < numkeys; i++) {
7104 if (mp->mp_ptrs[i] <= ptr)
7105 mp->mp_ptrs[i] -= delta;
7108 base = (char *)mp + mp->mp_upper + PAGEBASE;
7109 len = ptr - mp->mp_upper + NODESIZE;
7110 memmove(base - delta, base, len);
7111 mp->mp_upper -= delta;
7113 node = NODEPTR(mp, indx);
7116 /* But even if no shift was needed, update ksize */
7117 if (node->mn_ksize != key->mv_size)
7118 node->mn_ksize = key->mv_size;
7121 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7127 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7129 /** Move a node from csrc to cdst.
7132 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7139 unsigned short flags;
7143 /* Mark src and dst as dirty. */
7144 if ((rc = mdb_page_touch(csrc)) ||
7145 (rc = mdb_page_touch(cdst)))
7148 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7149 key.mv_size = csrc->mc_db->md_pad;
7150 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7152 data.mv_data = NULL;
7156 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7157 mdb_cassert(csrc, !((size_t)srcnode & 1));
7158 srcpg = NODEPGNO(srcnode);
7159 flags = srcnode->mn_flags;
7160 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7161 unsigned int snum = csrc->mc_snum;
7163 /* must find the lowest key below src */
7164 rc = mdb_page_search_lowest(csrc);
7167 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7168 key.mv_size = csrc->mc_db->md_pad;
7169 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7171 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7172 key.mv_size = NODEKSZ(s2);
7173 key.mv_data = NODEKEY(s2);
7175 csrc->mc_snum = snum--;
7176 csrc->mc_top = snum;
7178 key.mv_size = NODEKSZ(srcnode);
7179 key.mv_data = NODEKEY(srcnode);
7181 data.mv_size = NODEDSZ(srcnode);
7182 data.mv_data = NODEDATA(srcnode);
7184 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7185 unsigned int snum = cdst->mc_snum;
7188 /* must find the lowest key below dst */
7189 mdb_cursor_copy(cdst, &mn);
7190 rc = mdb_page_search_lowest(&mn);
7193 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7194 bkey.mv_size = mn.mc_db->md_pad;
7195 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7197 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7198 bkey.mv_size = NODEKSZ(s2);
7199 bkey.mv_data = NODEKEY(s2);
7201 mn.mc_snum = snum--;
7204 rc = mdb_update_key(&mn, &bkey);
7209 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7210 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7211 csrc->mc_ki[csrc->mc_top],
7213 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7214 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7216 /* Add the node to the destination page.
7218 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7219 if (rc != MDB_SUCCESS)
7222 /* Delete the node from the source page.
7224 mdb_node_del(csrc, key.mv_size);
7227 /* Adjust other cursors pointing to mp */
7228 MDB_cursor *m2, *m3;
7229 MDB_dbi dbi = csrc->mc_dbi;
7230 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7232 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7233 if (csrc->mc_flags & C_SUB)
7234 m3 = &m2->mc_xcursor->mx_cursor;
7237 if (m3 == csrc) continue;
7238 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7239 csrc->mc_ki[csrc->mc_top]) {
7240 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7241 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7246 /* Update the parent separators.
7248 if (csrc->mc_ki[csrc->mc_top] == 0) {
7249 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7250 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7251 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7253 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7254 key.mv_size = NODEKSZ(srcnode);
7255 key.mv_data = NODEKEY(srcnode);
7257 DPRINTF(("update separator for source page %"Z"u to [%s]",
7258 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7259 mdb_cursor_copy(csrc, &mn);
7262 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7265 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7267 indx_t ix = csrc->mc_ki[csrc->mc_top];
7268 nullkey.mv_size = 0;
7269 csrc->mc_ki[csrc->mc_top] = 0;
7270 rc = mdb_update_key(csrc, &nullkey);
7271 csrc->mc_ki[csrc->mc_top] = ix;
7272 mdb_cassert(csrc, rc == MDB_SUCCESS);
7276 if (cdst->mc_ki[cdst->mc_top] == 0) {
7277 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7278 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7279 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7281 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7282 key.mv_size = NODEKSZ(srcnode);
7283 key.mv_data = NODEKEY(srcnode);
7285 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7286 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7287 mdb_cursor_copy(cdst, &mn);
7290 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7293 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7295 indx_t ix = cdst->mc_ki[cdst->mc_top];
7296 nullkey.mv_size = 0;
7297 cdst->mc_ki[cdst->mc_top] = 0;
7298 rc = mdb_update_key(cdst, &nullkey);
7299 cdst->mc_ki[cdst->mc_top] = ix;
7300 mdb_cassert(csrc, rc == MDB_SUCCESS);
7307 /** Merge one page into another.
7308 * The nodes from the page pointed to by \b csrc will
7309 * be copied to the page pointed to by \b cdst and then
7310 * the \b csrc page will be freed.
7311 * @param[in] csrc Cursor pointing to the source page.
7312 * @param[in] cdst Cursor pointing to the destination page.
7313 * @return 0 on success, non-zero on failure.
7316 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7318 MDB_page *psrc, *pdst;
7325 psrc = csrc->mc_pg[csrc->mc_top];
7326 pdst = cdst->mc_pg[cdst->mc_top];
7328 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7330 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7331 mdb_cassert(csrc, cdst->mc_snum > 1);
7333 /* Mark dst as dirty. */
7334 if ((rc = mdb_page_touch(cdst)))
7337 /* Move all nodes from src to dst.
7339 j = nkeys = NUMKEYS(pdst);
7340 if (IS_LEAF2(psrc)) {
7341 key.mv_size = csrc->mc_db->md_pad;
7342 key.mv_data = METADATA(psrc);
7343 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7344 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7345 if (rc != MDB_SUCCESS)
7347 key.mv_data = (char *)key.mv_data + key.mv_size;
7350 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7351 srcnode = NODEPTR(psrc, i);
7352 if (i == 0 && IS_BRANCH(psrc)) {
7355 mdb_cursor_copy(csrc, &mn);
7356 /* must find the lowest key below src */
7357 rc = mdb_page_search_lowest(&mn);
7360 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7361 key.mv_size = mn.mc_db->md_pad;
7362 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7364 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7365 key.mv_size = NODEKSZ(s2);
7366 key.mv_data = NODEKEY(s2);
7369 key.mv_size = srcnode->mn_ksize;
7370 key.mv_data = NODEKEY(srcnode);
7373 data.mv_size = NODEDSZ(srcnode);
7374 data.mv_data = NODEDATA(srcnode);
7375 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7376 if (rc != MDB_SUCCESS)
7381 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7382 pdst->mp_pgno, NUMKEYS(pdst),
7383 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7385 /* Unlink the src page from parent and add to free list.
7388 mdb_node_del(csrc, 0);
7389 if (csrc->mc_ki[csrc->mc_top] == 0) {
7391 rc = mdb_update_key(csrc, &key);
7399 psrc = csrc->mc_pg[csrc->mc_top];
7400 /* If not operating on FreeDB, allow this page to be reused
7401 * in this txn. Otherwise just add to free list.
7403 rc = mdb_page_loose(csrc, psrc);
7407 csrc->mc_db->md_leaf_pages--;
7409 csrc->mc_db->md_branch_pages--;
7411 /* Adjust other cursors pointing to mp */
7412 MDB_cursor *m2, *m3;
7413 MDB_dbi dbi = csrc->mc_dbi;
7415 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7416 if (csrc->mc_flags & C_SUB)
7417 m3 = &m2->mc_xcursor->mx_cursor;
7420 if (m3 == csrc) continue;
7421 if (m3->mc_snum < csrc->mc_snum) continue;
7422 if (m3->mc_pg[csrc->mc_top] == psrc) {
7423 m3->mc_pg[csrc->mc_top] = pdst;
7424 m3->mc_ki[csrc->mc_top] += nkeys;
7429 unsigned int snum = cdst->mc_snum;
7430 uint16_t depth = cdst->mc_db->md_depth;
7431 mdb_cursor_pop(cdst);
7432 rc = mdb_rebalance(cdst);
7433 /* Did the tree shrink? */
7434 if (depth > cdst->mc_db->md_depth)
7436 cdst->mc_snum = snum;
7437 cdst->mc_top = snum-1;
7442 /** Copy the contents of a cursor.
7443 * @param[in] csrc The cursor to copy from.
7444 * @param[out] cdst The cursor to copy to.
7447 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7451 cdst->mc_txn = csrc->mc_txn;
7452 cdst->mc_dbi = csrc->mc_dbi;
7453 cdst->mc_db = csrc->mc_db;
7454 cdst->mc_dbx = csrc->mc_dbx;
7455 cdst->mc_snum = csrc->mc_snum;
7456 cdst->mc_top = csrc->mc_top;
7457 cdst->mc_flags = csrc->mc_flags;
7459 for (i=0; i<csrc->mc_snum; i++) {
7460 cdst->mc_pg[i] = csrc->mc_pg[i];
7461 cdst->mc_ki[i] = csrc->mc_ki[i];
7465 /** Rebalance the tree after a delete operation.
7466 * @param[in] mc Cursor pointing to the page where rebalancing
7468 * @return 0 on success, non-zero on failure.
7471 mdb_rebalance(MDB_cursor *mc)
7475 unsigned int ptop, minkeys;
7479 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7480 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7481 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7482 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7483 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7485 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7486 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7487 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7488 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7492 if (mc->mc_snum < 2) {
7493 MDB_page *mp = mc->mc_pg[0];
7495 DPUTS("Can't rebalance a subpage, ignoring");
7498 if (NUMKEYS(mp) == 0) {
7499 DPUTS("tree is completely empty");
7500 mc->mc_db->md_root = P_INVALID;
7501 mc->mc_db->md_depth = 0;
7502 mc->mc_db->md_leaf_pages = 0;
7503 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7506 /* Adjust cursors pointing to mp */
7509 mc->mc_flags &= ~C_INITIALIZED;
7511 MDB_cursor *m2, *m3;
7512 MDB_dbi dbi = mc->mc_dbi;
7514 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7515 if (mc->mc_flags & C_SUB)
7516 m3 = &m2->mc_xcursor->mx_cursor;
7519 if (m3->mc_snum < mc->mc_snum) continue;
7520 if (m3->mc_pg[0] == mp) {
7523 m3->mc_flags &= ~C_INITIALIZED;
7527 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7529 DPUTS("collapsing root page!");
7530 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7533 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7534 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7537 mc->mc_db->md_depth--;
7538 mc->mc_db->md_branch_pages--;
7539 mc->mc_ki[0] = mc->mc_ki[1];
7540 for (i = 1; i<mc->mc_db->md_depth; i++) {
7541 mc->mc_pg[i] = mc->mc_pg[i+1];
7542 mc->mc_ki[i] = mc->mc_ki[i+1];
7545 /* Adjust other cursors pointing to mp */
7546 MDB_cursor *m2, *m3;
7547 MDB_dbi dbi = mc->mc_dbi;
7549 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7550 if (mc->mc_flags & C_SUB)
7551 m3 = &m2->mc_xcursor->mx_cursor;
7554 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7555 if (m3->mc_pg[0] == mp) {
7558 for (i=0; i<m3->mc_snum; i++) {
7559 m3->mc_pg[i] = m3->mc_pg[i+1];
7560 m3->mc_ki[i] = m3->mc_ki[i+1];
7566 DPUTS("root page doesn't need rebalancing");
7570 /* The parent (branch page) must have at least 2 pointers,
7571 * otherwise the tree is invalid.
7573 ptop = mc->mc_top-1;
7574 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7576 /* Leaf page fill factor is below the threshold.
7577 * Try to move keys from left or right neighbor, or
7578 * merge with a neighbor page.
7583 mdb_cursor_copy(mc, &mn);
7584 mn.mc_xcursor = NULL;
7586 oldki = mc->mc_ki[mc->mc_top];
7587 if (mc->mc_ki[ptop] == 0) {
7588 /* We're the leftmost leaf in our parent.
7590 DPUTS("reading right neighbor");
7592 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7593 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7596 mn.mc_ki[mn.mc_top] = 0;
7597 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7599 /* There is at least one neighbor to the left.
7601 DPUTS("reading left neighbor");
7603 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7604 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7607 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7608 mc->mc_ki[mc->mc_top] = 0;
7611 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7612 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7613 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7615 /* If the neighbor page is above threshold and has enough keys,
7616 * move one key from it. Otherwise we should try to merge them.
7617 * (A branch page must never have less than 2 keys.)
7619 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7620 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7621 rc = mdb_node_move(&mn, mc);
7622 if (mc->mc_ki[ptop]) {
7626 if (mc->mc_ki[ptop] == 0) {
7627 rc = mdb_page_merge(&mn, mc);
7629 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7630 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7631 rc = mdb_page_merge(mc, &mn);
7632 mdb_cursor_copy(&mn, mc);
7634 mc->mc_flags &= ~C_EOF;
7636 mc->mc_ki[mc->mc_top] = oldki;
7640 /** Complete a delete operation started by #mdb_cursor_del(). */
7642 mdb_cursor_del0(MDB_cursor *mc)
7649 ki = mc->mc_ki[mc->mc_top];
7650 mdb_node_del(mc, mc->mc_db->md_pad);
7651 mc->mc_db->md_entries--;
7652 rc = mdb_rebalance(mc);
7654 if (rc == MDB_SUCCESS) {
7655 MDB_cursor *m2, *m3;
7656 MDB_dbi dbi = mc->mc_dbi;
7658 mp = mc->mc_pg[mc->mc_top];
7659 nkeys = NUMKEYS(mp);
7661 /* if mc points past last node in page, find next sibling */
7662 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7663 rc = mdb_cursor_sibling(mc, 1);
7664 if (rc == MDB_NOTFOUND) {
7665 mc->mc_flags |= C_EOF;
7670 /* Adjust other cursors pointing to mp */
7671 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7672 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7673 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7675 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7677 if (m3->mc_pg[mc->mc_top] == mp) {
7678 if (m3->mc_ki[mc->mc_top] >= ki) {
7679 m3->mc_flags |= C_DEL;
7680 if (m3->mc_ki[mc->mc_top] > ki)
7681 m3->mc_ki[mc->mc_top]--;
7682 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7683 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7685 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7686 rc = mdb_cursor_sibling(m3, 1);
7687 if (rc == MDB_NOTFOUND) {
7688 m3->mc_flags |= C_EOF;
7694 mc->mc_flags |= C_DEL;
7698 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7703 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7704 MDB_val *key, MDB_val *data)
7706 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7709 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7710 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7712 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7713 /* must ignore any data */
7717 return mdb_del0(txn, dbi, key, data, 0);
7721 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7722 MDB_val *key, MDB_val *data, unsigned flags)
7727 MDB_val rdata, *xdata;
7731 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7733 mdb_cursor_init(&mc, txn, dbi, &mx);
7742 flags |= MDB_NODUPDATA;
7744 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7746 /* let mdb_page_split know about this cursor if needed:
7747 * delete will trigger a rebalance; if it needs to move
7748 * a node from one page to another, it will have to
7749 * update the parent's separator key(s). If the new sepkey
7750 * is larger than the current one, the parent page may
7751 * run out of space, triggering a split. We need this
7752 * cursor to be consistent until the end of the rebalance.
7754 mc.mc_flags |= C_UNTRACK;
7755 mc.mc_next = txn->mt_cursors[dbi];
7756 txn->mt_cursors[dbi] = &mc;
7757 rc = mdb_cursor_del(&mc, flags);
7758 txn->mt_cursors[dbi] = mc.mc_next;
7763 /** Split a page and insert a new node.
7764 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7765 * The cursor will be updated to point to the actual page and index where
7766 * the node got inserted after the split.
7767 * @param[in] newkey The key for the newly inserted node.
7768 * @param[in] newdata The data for the newly inserted node.
7769 * @param[in] newpgno The page number, if the new node is a branch node.
7770 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7771 * @return 0 on success, non-zero on failure.
7774 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7775 unsigned int nflags)
7778 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7781 int i, j, split_indx, nkeys, pmax;
7782 MDB_env *env = mc->mc_txn->mt_env;
7784 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7785 MDB_page *copy = NULL;
7786 MDB_page *mp, *rp, *pp;
7791 mp = mc->mc_pg[mc->mc_top];
7792 newindx = mc->mc_ki[mc->mc_top];
7793 nkeys = NUMKEYS(mp);
7795 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7796 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7797 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7799 /* Create a right sibling. */
7800 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7802 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7804 if (mc->mc_snum < 2) {
7805 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7807 /* shift current top to make room for new parent */
7808 mc->mc_pg[1] = mc->mc_pg[0];
7809 mc->mc_ki[1] = mc->mc_ki[0];
7812 mc->mc_db->md_root = pp->mp_pgno;
7813 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7814 mc->mc_db->md_depth++;
7817 /* Add left (implicit) pointer. */
7818 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7819 /* undo the pre-push */
7820 mc->mc_pg[0] = mc->mc_pg[1];
7821 mc->mc_ki[0] = mc->mc_ki[1];
7822 mc->mc_db->md_root = mp->mp_pgno;
7823 mc->mc_db->md_depth--;
7830 ptop = mc->mc_top-1;
7831 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7834 mc->mc_flags |= C_SPLITTING;
7835 mdb_cursor_copy(mc, &mn);
7836 mn.mc_pg[mn.mc_top] = rp;
7837 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7839 if (nflags & MDB_APPEND) {
7840 mn.mc_ki[mn.mc_top] = 0;
7842 split_indx = newindx;
7846 split_indx = (nkeys+1) / 2;
7851 unsigned int lsize, rsize, ksize;
7852 /* Move half of the keys to the right sibling */
7853 x = mc->mc_ki[mc->mc_top] - split_indx;
7854 ksize = mc->mc_db->md_pad;
7855 split = LEAF2KEY(mp, split_indx, ksize);
7856 rsize = (nkeys - split_indx) * ksize;
7857 lsize = (nkeys - split_indx) * sizeof(indx_t);
7858 mp->mp_lower -= lsize;
7859 rp->mp_lower += lsize;
7860 mp->mp_upper += rsize - lsize;
7861 rp->mp_upper -= rsize - lsize;
7862 sepkey.mv_size = ksize;
7863 if (newindx == split_indx) {
7864 sepkey.mv_data = newkey->mv_data;
7866 sepkey.mv_data = split;
7869 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7870 memcpy(rp->mp_ptrs, split, rsize);
7871 sepkey.mv_data = rp->mp_ptrs;
7872 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7873 memcpy(ins, newkey->mv_data, ksize);
7874 mp->mp_lower += sizeof(indx_t);
7875 mp->mp_upper -= ksize - sizeof(indx_t);
7878 memcpy(rp->mp_ptrs, split, x * ksize);
7879 ins = LEAF2KEY(rp, x, ksize);
7880 memcpy(ins, newkey->mv_data, ksize);
7881 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7882 rp->mp_lower += sizeof(indx_t);
7883 rp->mp_upper -= ksize - sizeof(indx_t);
7884 mc->mc_ki[mc->mc_top] = x;
7885 mc->mc_pg[mc->mc_top] = rp;
7888 int psize, nsize, k;
7889 /* Maximum free space in an empty page */
7890 pmax = env->me_psize - PAGEHDRSZ;
7892 nsize = mdb_leaf_size(env, newkey, newdata);
7894 nsize = mdb_branch_size(env, newkey);
7895 nsize = EVEN(nsize);
7897 /* grab a page to hold a temporary copy */
7898 copy = mdb_page_malloc(mc->mc_txn, 1);
7903 copy->mp_pgno = mp->mp_pgno;
7904 copy->mp_flags = mp->mp_flags;
7905 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7906 copy->mp_upper = env->me_psize - PAGEBASE;
7908 /* prepare to insert */
7909 for (i=0, j=0; i<nkeys; i++) {
7911 copy->mp_ptrs[j++] = 0;
7913 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7916 /* When items are relatively large the split point needs
7917 * to be checked, because being off-by-one will make the
7918 * difference between success or failure in mdb_node_add.
7920 * It's also relevant if a page happens to be laid out
7921 * such that one half of its nodes are all "small" and
7922 * the other half of its nodes are "large." If the new
7923 * item is also "large" and falls on the half with
7924 * "large" nodes, it also may not fit.
7926 * As a final tweak, if the new item goes on the last
7927 * spot on the page (and thus, onto the new page), bias
7928 * the split so the new page is emptier than the old page.
7929 * This yields better packing during sequential inserts.
7931 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7932 /* Find split point */
7934 if (newindx <= split_indx || newindx >= nkeys) {
7936 k = newindx >= nkeys ? nkeys : split_indx+2;
7941 for (; i!=k; i+=j) {
7946 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7947 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7949 if (F_ISSET(node->mn_flags, F_BIGDATA))
7950 psize += sizeof(pgno_t);
7952 psize += NODEDSZ(node);
7954 psize = EVEN(psize);
7956 if (psize > pmax || i == k-j) {
7957 split_indx = i + (j<0);
7962 if (split_indx == newindx) {
7963 sepkey.mv_size = newkey->mv_size;
7964 sepkey.mv_data = newkey->mv_data;
7966 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
7967 sepkey.mv_size = node->mn_ksize;
7968 sepkey.mv_data = NODEKEY(node);
7973 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7975 /* Copy separator key to the parent.
7977 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7981 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7986 if (mn.mc_snum == mc->mc_snum) {
7987 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7988 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7989 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7990 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7995 /* Right page might now have changed parent.
7996 * Check if left page also changed parent.
7998 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7999 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8000 for (i=0; i<ptop; i++) {
8001 mc->mc_pg[i] = mn.mc_pg[i];
8002 mc->mc_ki[i] = mn.mc_ki[i];
8004 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8005 if (mn.mc_ki[ptop]) {
8006 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8008 /* find right page's left sibling */
8009 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8010 mdb_cursor_sibling(mc, 0);
8015 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8018 mc->mc_flags ^= C_SPLITTING;
8019 if (rc != MDB_SUCCESS) {
8022 if (nflags & MDB_APPEND) {
8023 mc->mc_pg[mc->mc_top] = rp;
8024 mc->mc_ki[mc->mc_top] = 0;
8025 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8028 for (i=0; i<mc->mc_top; i++)
8029 mc->mc_ki[i] = mn.mc_ki[i];
8030 } else if (!IS_LEAF2(mp)) {
8032 mc->mc_pg[mc->mc_top] = rp;
8037 rkey.mv_data = newkey->mv_data;
8038 rkey.mv_size = newkey->mv_size;
8044 /* Update index for the new key. */
8045 mc->mc_ki[mc->mc_top] = j;
8047 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8048 rkey.mv_data = NODEKEY(node);
8049 rkey.mv_size = node->mn_ksize;
8051 xdata.mv_data = NODEDATA(node);
8052 xdata.mv_size = NODEDSZ(node);
8055 pgno = NODEPGNO(node);
8056 flags = node->mn_flags;
8059 if (!IS_LEAF(mp) && j == 0) {
8060 /* First branch index doesn't need key data. */
8064 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8070 mc->mc_pg[mc->mc_top] = copy;
8075 } while (i != split_indx);
8077 nkeys = NUMKEYS(copy);
8078 for (i=0; i<nkeys; i++)
8079 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8080 mp->mp_lower = copy->mp_lower;
8081 mp->mp_upper = copy->mp_upper;
8082 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8083 env->me_psize - copy->mp_upper - PAGEBASE);
8085 /* reset back to original page */
8086 if (newindx < split_indx) {
8087 mc->mc_pg[mc->mc_top] = mp;
8088 if (nflags & MDB_RESERVE) {
8089 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8090 if (!(node->mn_flags & F_BIGDATA))
8091 newdata->mv_data = NODEDATA(node);
8094 mc->mc_pg[mc->mc_top] = rp;
8096 /* Make sure mc_ki is still valid.
8098 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8099 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8100 for (i=0; i<=ptop; i++) {
8101 mc->mc_pg[i] = mn.mc_pg[i];
8102 mc->mc_ki[i] = mn.mc_ki[i];
8109 /* Adjust other cursors pointing to mp */
8110 MDB_cursor *m2, *m3;
8111 MDB_dbi dbi = mc->mc_dbi;
8112 int fixup = NUMKEYS(mp);
8114 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8115 if (mc->mc_flags & C_SUB)
8116 m3 = &m2->mc_xcursor->mx_cursor;
8121 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8123 if (m3->mc_flags & C_SPLITTING)
8128 for (k=m3->mc_top; k>=0; k--) {
8129 m3->mc_ki[k+1] = m3->mc_ki[k];
8130 m3->mc_pg[k+1] = m3->mc_pg[k];
8132 if (m3->mc_ki[0] >= split_indx) {
8137 m3->mc_pg[0] = mc->mc_pg[0];
8141 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8142 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8143 m3->mc_ki[mc->mc_top]++;
8144 if (m3->mc_ki[mc->mc_top] >= fixup) {
8145 m3->mc_pg[mc->mc_top] = rp;
8146 m3->mc_ki[mc->mc_top] -= fixup;
8147 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8149 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8150 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8155 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8158 if (copy) /* tmp page */
8159 mdb_page_free(env, copy);
8161 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8166 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8167 MDB_val *key, MDB_val *data, unsigned int flags)
8172 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8175 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8178 mdb_cursor_init(&mc, txn, dbi, &mx);
8179 return mdb_cursor_put(&mc, key, data, flags);
8183 #define MDB_WBUF (1024*1024)
8186 /** State needed for a compacting copy. */
8187 typedef struct mdb_copy {
8188 pthread_mutex_t mc_mutex;
8189 pthread_cond_t mc_cond;
8196 pgno_t mc_next_pgno;
8199 volatile int mc_new;
8204 /** Dedicated writer thread for compacting copy. */
8205 static THREAD_RET ESECT
8206 mdb_env_copythr(void *arg)
8210 int toggle = 0, wsize, rc;
8213 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8216 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8219 pthread_mutex_lock(&my->mc_mutex);
8221 pthread_cond_signal(&my->mc_cond);
8224 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8225 if (my->mc_new < 0) {
8230 wsize = my->mc_wlen[toggle];
8231 ptr = my->mc_wbuf[toggle];
8234 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8238 } else if (len > 0) {
8252 /* If there's an overflow page tail, write it too */
8253 if (my->mc_olen[toggle]) {
8254 wsize = my->mc_olen[toggle];
8255 ptr = my->mc_over[toggle];
8256 my->mc_olen[toggle] = 0;
8259 my->mc_wlen[toggle] = 0;
8261 pthread_cond_signal(&my->mc_cond);
8263 pthread_cond_signal(&my->mc_cond);
8264 pthread_mutex_unlock(&my->mc_mutex);
8265 return (THREAD_RET)0;
8269 /** Tell the writer thread there's a buffer ready to write */
8271 mdb_env_cthr_toggle(mdb_copy *my, int st)
8273 int toggle = my->mc_toggle ^ 1;
8274 pthread_mutex_lock(&my->mc_mutex);
8275 if (my->mc_status) {
8276 pthread_mutex_unlock(&my->mc_mutex);
8277 return my->mc_status;
8279 while (my->mc_new == 1)
8280 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8282 my->mc_toggle = toggle;
8283 pthread_cond_signal(&my->mc_cond);
8284 pthread_mutex_unlock(&my->mc_mutex);
8288 /** Depth-first tree traversal for compacting copy. */
8290 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8293 MDB_txn *txn = my->mc_txn;
8295 MDB_page *mo, *mp, *leaf;
8300 /* Empty DB, nothing to do */
8301 if (*pg == P_INVALID)
8308 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8311 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8315 /* Make cursor pages writable */
8316 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8320 for (i=0; i<mc.mc_top; i++) {
8321 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8322 mc.mc_pg[i] = (MDB_page *)ptr;
8323 ptr += my->mc_env->me_psize;
8326 /* This is writable space for a leaf page. Usually not needed. */
8327 leaf = (MDB_page *)ptr;
8329 toggle = my->mc_toggle;
8330 while (mc.mc_snum > 0) {
8332 mp = mc.mc_pg[mc.mc_top];
8336 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8337 for (i=0; i<n; i++) {
8338 ni = NODEPTR(mp, i);
8339 if (ni->mn_flags & F_BIGDATA) {
8343 /* Need writable leaf */
8345 mc.mc_pg[mc.mc_top] = leaf;
8346 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8348 ni = NODEPTR(mp, i);
8351 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8352 rc = mdb_page_get(txn, pg, &omp, NULL);
8355 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8356 rc = mdb_env_cthr_toggle(my, 1);
8359 toggle = my->mc_toggle;
8361 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8362 memcpy(mo, omp, my->mc_env->me_psize);
8363 mo->mp_pgno = my->mc_next_pgno;
8364 my->mc_next_pgno += omp->mp_pages;
8365 my->mc_wlen[toggle] += my->mc_env->me_psize;
8366 if (omp->mp_pages > 1) {
8367 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8368 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8369 rc = mdb_env_cthr_toggle(my, 1);
8372 toggle = my->mc_toggle;
8374 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8375 } else if (ni->mn_flags & F_SUBDATA) {
8378 /* Need writable leaf */
8380 mc.mc_pg[mc.mc_top] = leaf;
8381 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8383 ni = NODEPTR(mp, i);
8386 memcpy(&db, NODEDATA(ni), sizeof(db));
8387 my->mc_toggle = toggle;
8388 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8391 toggle = my->mc_toggle;
8392 memcpy(NODEDATA(ni), &db, sizeof(db));
8397 mc.mc_ki[mc.mc_top]++;
8398 if (mc.mc_ki[mc.mc_top] < n) {
8401 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8403 rc = mdb_page_get(txn, pg, &mp, NULL);
8408 mc.mc_ki[mc.mc_top] = 0;
8409 if (IS_BRANCH(mp)) {
8410 /* Whenever we advance to a sibling branch page,
8411 * we must proceed all the way down to its first leaf.
8413 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8416 mc.mc_pg[mc.mc_top] = mp;
8420 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8421 rc = mdb_env_cthr_toggle(my, 1);
8424 toggle = my->mc_toggle;
8426 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8427 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8428 mo->mp_pgno = my->mc_next_pgno++;
8429 my->mc_wlen[toggle] += my->mc_env->me_psize;
8431 /* Update parent if there is one */
8432 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8433 SETPGNO(ni, mo->mp_pgno);
8434 mdb_cursor_pop(&mc);
8436 /* Otherwise we're done */
8446 /** Copy environment with compaction. */
8448 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8453 MDB_txn *txn = NULL;
8458 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8459 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8460 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8461 if (my.mc_wbuf[0] == NULL)
8464 pthread_mutex_init(&my.mc_mutex, NULL);
8465 pthread_cond_init(&my.mc_cond, NULL);
8466 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8470 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8471 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8476 my.mc_next_pgno = 2;
8482 THREAD_CREATE(thr, mdb_env_copythr, &my);
8484 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8488 mp = (MDB_page *)my.mc_wbuf[0];
8489 memset(mp, 0, 2*env->me_psize);
8491 mp->mp_flags = P_META;
8492 mm = (MDB_meta *)METADATA(mp);
8493 mdb_env_init_meta0(env, mm);
8494 mm->mm_address = env->me_metas[0]->mm_address;
8496 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8498 mp->mp_flags = P_META;
8499 *(MDB_meta *)METADATA(mp) = *mm;
8500 mm = (MDB_meta *)METADATA(mp);
8502 /* Count the number of free pages, subtract from lastpg to find
8503 * number of active pages
8506 MDB_ID freecount = 0;
8509 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8510 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8511 freecount += *(MDB_ID *)data.mv_data;
8512 freecount += txn->mt_dbs[0].md_branch_pages +
8513 txn->mt_dbs[0].md_leaf_pages +
8514 txn->mt_dbs[0].md_overflow_pages;
8516 /* Set metapage 1 */
8517 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8518 mm->mm_dbs[1] = txn->mt_dbs[1];
8519 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8522 my.mc_wlen[0] = env->me_psize * 2;
8524 pthread_mutex_lock(&my.mc_mutex);
8526 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8527 pthread_mutex_unlock(&my.mc_mutex);
8528 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8529 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8530 rc = mdb_env_cthr_toggle(&my, 1);
8531 mdb_env_cthr_toggle(&my, -1);
8532 pthread_mutex_lock(&my.mc_mutex);
8534 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8535 pthread_mutex_unlock(&my.mc_mutex);
8540 CloseHandle(my.mc_cond);
8541 CloseHandle(my.mc_mutex);
8542 _aligned_free(my.mc_wbuf[0]);
8544 pthread_cond_destroy(&my.mc_cond);
8545 pthread_mutex_destroy(&my.mc_mutex);
8546 free(my.mc_wbuf[0]);
8551 /** Copy environment as-is. */
8553 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8555 MDB_txn *txn = NULL;
8561 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8565 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8568 /* Do the lock/unlock of the reader mutex before starting the
8569 * write txn. Otherwise other read txns could block writers.
8571 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8576 /* We must start the actual read txn after blocking writers */
8577 mdb_txn_reset0(txn, "reset-stage1");
8579 /* Temporarily block writers until we snapshot the meta pages */
8582 rc = mdb_txn_renew0(txn);
8584 UNLOCK_MUTEX_W(env);
8589 wsize = env->me_psize * 2;
8593 DO_WRITE(rc, fd, ptr, w2, len);
8597 } else if (len > 0) {
8603 /* Non-blocking or async handles are not supported */
8609 UNLOCK_MUTEX_W(env);
8614 w2 = txn->mt_next_pgno * env->me_psize;
8617 LARGE_INTEGER fsize;
8618 GetFileSizeEx(env->me_fd, &fsize);
8619 if (w2 > fsize.QuadPart)
8620 w2 = fsize.QuadPart;
8625 fstat(env->me_fd, &st);
8626 if (w2 > (size_t)st.st_size)
8632 if (wsize > MAX_WRITE)
8636 DO_WRITE(rc, fd, ptr, w2, len);
8640 } else if (len > 0) {
8657 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8659 if (flags & MDB_CP_COMPACT)
8660 return mdb_env_copyfd1(env, fd);
8662 return mdb_env_copyfd0(env, fd);
8666 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8668 return mdb_env_copyfd2(env, fd, 0);
8672 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8676 HANDLE newfd = INVALID_HANDLE_VALUE;
8678 if (env->me_flags & MDB_NOSUBDIR) {
8679 lpath = (char *)path;
8682 len += sizeof(DATANAME);
8683 lpath = malloc(len);
8686 sprintf(lpath, "%s" DATANAME, path);
8689 /* The destination path must exist, but the destination file must not.
8690 * We don't want the OS to cache the writes, since the source data is
8691 * already in the OS cache.
8694 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8695 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8697 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8699 if (newfd == INVALID_HANDLE_VALUE) {
8704 if (env->me_psize >= env->me_os_psize) {
8706 /* Set O_DIRECT if the file system supports it */
8707 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8708 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8710 #ifdef F_NOCACHE /* __APPLE__ */
8711 rc = fcntl(newfd, F_NOCACHE, 1);
8719 rc = mdb_env_copyfd2(env, newfd, flags);
8722 if (!(env->me_flags & MDB_NOSUBDIR))
8724 if (newfd != INVALID_HANDLE_VALUE)
8725 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8732 mdb_env_copy(MDB_env *env, const char *path)
8734 return mdb_env_copy2(env, path, 0);
8738 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8740 if ((flag & CHANGEABLE) != flag)
8743 env->me_flags |= flag;
8745 env->me_flags &= ~flag;
8750 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8755 *arg = env->me_flags;
8760 mdb_env_set_userctx(MDB_env *env, void *ctx)
8764 env->me_userctx = ctx;
8769 mdb_env_get_userctx(MDB_env *env)
8771 return env ? env->me_userctx : NULL;
8775 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8780 env->me_assert_func = func;
8786 mdb_env_get_path(MDB_env *env, const char **arg)
8791 *arg = env->me_path;
8796 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8805 /** Common code for #mdb_stat() and #mdb_env_stat().
8806 * @param[in] env the environment to operate in.
8807 * @param[in] db the #MDB_db record containing the stats to return.
8808 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8809 * @return 0, this function always succeeds.
8812 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8814 arg->ms_psize = env->me_psize;
8815 arg->ms_depth = db->md_depth;
8816 arg->ms_branch_pages = db->md_branch_pages;
8817 arg->ms_leaf_pages = db->md_leaf_pages;
8818 arg->ms_overflow_pages = db->md_overflow_pages;
8819 arg->ms_entries = db->md_entries;
8825 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8829 if (env == NULL || arg == NULL)
8832 toggle = mdb_env_pick_meta(env);
8834 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8838 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8842 if (env == NULL || arg == NULL)
8845 toggle = mdb_env_pick_meta(env);
8846 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8847 arg->me_mapsize = env->me_mapsize;
8848 arg->me_maxreaders = env->me_maxreaders;
8850 /* me_numreaders may be zero if this process never used any readers. Use
8851 * the shared numreader count if it exists.
8853 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8855 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8856 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8860 /** Set the default comparison functions for a database.
8861 * Called immediately after a database is opened to set the defaults.
8862 * The user can then override them with #mdb_set_compare() or
8863 * #mdb_set_dupsort().
8864 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8865 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8868 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8870 uint16_t f = txn->mt_dbs[dbi].md_flags;
8872 txn->mt_dbxs[dbi].md_cmp =
8873 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8874 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8876 txn->mt_dbxs[dbi].md_dcmp =
8877 !(f & MDB_DUPSORT) ? 0 :
8878 ((f & MDB_INTEGERDUP)
8879 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8880 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8883 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8888 int rc, dbflag, exact;
8889 unsigned int unused = 0, seq;
8892 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8893 mdb_default_cmp(txn, FREE_DBI);
8896 if ((flags & VALID_FLAGS) != flags)
8898 if (txn->mt_flags & MDB_TXN_ERROR)
8904 if (flags & PERSISTENT_FLAGS) {
8905 uint16_t f2 = flags & PERSISTENT_FLAGS;
8906 /* make sure flag changes get committed */
8907 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8908 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8909 txn->mt_flags |= MDB_TXN_DIRTY;
8912 mdb_default_cmp(txn, MAIN_DBI);
8916 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8917 mdb_default_cmp(txn, MAIN_DBI);
8920 /* Is the DB already open? */
8922 for (i=2; i<txn->mt_numdbs; i++) {
8923 if (!txn->mt_dbxs[i].md_name.mv_size) {
8924 /* Remember this free slot */
8925 if (!unused) unused = i;
8928 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8929 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8935 /* If no free slot and max hit, fail */
8936 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8937 return MDB_DBS_FULL;
8939 /* Cannot mix named databases with some mainDB flags */
8940 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8941 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8943 /* Find the DB info */
8944 dbflag = DB_NEW|DB_VALID;
8947 key.mv_data = (void *)name;
8948 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8949 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8950 if (rc == MDB_SUCCESS) {
8951 /* make sure this is actually a DB */
8952 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8953 if (!(node->mn_flags & F_SUBDATA))
8954 return MDB_INCOMPATIBLE;
8955 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8956 /* Create if requested */
8958 data.mv_size = sizeof(MDB_db);
8959 data.mv_data = &dummy;
8960 memset(&dummy, 0, sizeof(dummy));
8961 dummy.md_root = P_INVALID;
8962 dummy.md_flags = flags & PERSISTENT_FLAGS;
8963 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8967 /* OK, got info, add to table */
8968 if (rc == MDB_SUCCESS) {
8969 unsigned int slot = unused ? unused : txn->mt_numdbs;
8970 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8971 txn->mt_dbxs[slot].md_name.mv_size = len;
8972 txn->mt_dbxs[slot].md_rel = NULL;
8973 txn->mt_dbflags[slot] = dbflag;
8974 /* txn-> and env-> are the same in read txns, use
8975 * tmp variable to avoid undefined assignment
8977 seq = ++txn->mt_env->me_dbiseqs[slot];
8978 txn->mt_dbiseqs[slot] = seq;
8980 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8982 mdb_default_cmp(txn, slot);
8991 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8993 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8996 if (txn->mt_flags & MDB_TXN_ERROR)
8999 if (txn->mt_dbflags[dbi] & DB_STALE) {
9002 /* Stale, must read the DB's root. cursor_init does it for us. */
9003 mdb_cursor_init(&mc, txn, dbi, &mx);
9005 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9008 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9011 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9013 ptr = env->me_dbxs[dbi].md_name.mv_data;
9014 /* If there was no name, this was already closed */
9016 env->me_dbxs[dbi].md_name.mv_data = NULL;
9017 env->me_dbxs[dbi].md_name.mv_size = 0;
9018 env->me_dbflags[dbi] = 0;
9019 env->me_dbiseqs[dbi]++;
9024 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9026 /* We could return the flags for the FREE_DBI too but what's the point? */
9027 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9029 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9033 /** Add all the DB's pages to the free list.
9034 * @param[in] mc Cursor on the DB to free.
9035 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9036 * @return 0 on success, non-zero on failure.
9039 mdb_drop0(MDB_cursor *mc, int subs)
9043 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9044 if (rc == MDB_SUCCESS) {
9045 MDB_txn *txn = mc->mc_txn;
9050 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9051 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9054 mdb_cursor_copy(mc, &mx);
9055 while (mc->mc_snum > 0) {
9056 MDB_page *mp = mc->mc_pg[mc->mc_top];
9057 unsigned n = NUMKEYS(mp);
9059 for (i=0; i<n; i++) {
9060 ni = NODEPTR(mp, i);
9061 if (ni->mn_flags & F_BIGDATA) {
9064 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9065 rc = mdb_page_get(txn, pg, &omp, NULL);
9068 mdb_cassert(mc, IS_OVERFLOW(omp));
9069 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9073 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9074 mdb_xcursor_init1(mc, ni);
9075 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9081 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9083 for (i=0; i<n; i++) {
9085 ni = NODEPTR(mp, i);
9088 mdb_midl_xappend(txn->mt_free_pgs, pg);
9093 mc->mc_ki[mc->mc_top] = i;
9094 rc = mdb_cursor_sibling(mc, 1);
9096 if (rc != MDB_NOTFOUND)
9098 /* no more siblings, go back to beginning
9099 * of previous level.
9103 for (i=1; i<mc->mc_snum; i++) {
9105 mc->mc_pg[i] = mx.mc_pg[i];
9110 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9113 txn->mt_flags |= MDB_TXN_ERROR;
9114 } else if (rc == MDB_NOTFOUND) {
9120 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9122 MDB_cursor *mc, *m2;
9125 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9128 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9131 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9134 rc = mdb_cursor_open(txn, dbi, &mc);
9138 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9139 /* Invalidate the dropped DB's cursors */
9140 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9141 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9145 /* Can't delete the main DB */
9146 if (del && dbi > MAIN_DBI) {
9147 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9149 txn->mt_dbflags[dbi] = DB_STALE;
9150 mdb_dbi_close(txn->mt_env, dbi);
9152 txn->mt_flags |= MDB_TXN_ERROR;
9155 /* reset the DB record, mark it dirty */
9156 txn->mt_dbflags[dbi] |= DB_DIRTY;
9157 txn->mt_dbs[dbi].md_depth = 0;
9158 txn->mt_dbs[dbi].md_branch_pages = 0;
9159 txn->mt_dbs[dbi].md_leaf_pages = 0;
9160 txn->mt_dbs[dbi].md_overflow_pages = 0;
9161 txn->mt_dbs[dbi].md_entries = 0;
9162 txn->mt_dbs[dbi].md_root = P_INVALID;
9164 txn->mt_flags |= MDB_TXN_DIRTY;
9167 mdb_cursor_close(mc);
9171 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9173 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9176 txn->mt_dbxs[dbi].md_cmp = cmp;
9180 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9182 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9185 txn->mt_dbxs[dbi].md_dcmp = cmp;
9189 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9191 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9194 txn->mt_dbxs[dbi].md_rel = rel;
9198 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9200 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9203 txn->mt_dbxs[dbi].md_relctx = ctx;
9208 mdb_env_get_maxkeysize(MDB_env *env)
9210 return ENV_MAXKEY(env);
9214 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9216 unsigned int i, rdrs;
9219 int rc = 0, first = 1;
9223 if (!env->me_txns) {
9224 return func("(no reader locks)\n", ctx);
9226 rdrs = env->me_txns->mti_numreaders;
9227 mr = env->me_txns->mti_readers;
9228 for (i=0; i<rdrs; i++) {
9230 txnid_t txnid = mr[i].mr_txnid;
9231 sprintf(buf, txnid == (txnid_t)-1 ?
9232 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9233 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9236 rc = func(" pid thread txnid\n", ctx);
9240 rc = func(buf, ctx);
9246 rc = func("(no active readers)\n", ctx);
9251 /** Insert pid into list if not already present.
9252 * return -1 if already present.
9255 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9257 /* binary search of pid in list */
9259 unsigned cursor = 1;
9261 unsigned n = ids[0];
9264 unsigned pivot = n >> 1;
9265 cursor = base + pivot + 1;
9266 val = pid - ids[cursor];
9271 } else if ( val > 0 ) {
9276 /* found, so it's a duplicate */
9285 for (n = ids[0]; n > cursor; n--)
9292 mdb_reader_check(MDB_env *env, int *dead)
9294 unsigned int i, j, rdrs;
9296 MDB_PID_T *pids, pid;
9305 rdrs = env->me_txns->mti_numreaders;
9306 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9310 mr = env->me_txns->mti_readers;
9311 for (i=0; i<rdrs; i++) {
9312 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9314 if (mdb_pid_insert(pids, pid) == 0) {
9315 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9317 /* Recheck, a new process may have reused pid */
9318 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9319 for (j=i; j<rdrs; j++)
9320 if (mr[j].mr_pid == pid) {
9321 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9322 (unsigned) pid, mr[j].mr_txnid));
9327 UNLOCK_MUTEX_R(env);
projects / openldap / blob