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
83 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
84 #include <netinet/in.h>
85 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
88 #if defined(__APPLE__) || defined (BSD)
89 # define MDB_USE_POSIX_SEM 1
90 # define MDB_FDATASYNC fsync
91 #elif defined(ANDROID)
92 # define MDB_FDATASYNC fsync
97 #ifdef MDB_USE_POSIX_SEM
98 # define MDB_USE_HASH 1
99 #include <semaphore.h>
104 #include <valgrind/memcheck.h>
105 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
106 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
107 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
108 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
109 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
111 #define VGMEMP_CREATE(h,r,z)
112 #define VGMEMP_ALLOC(h,a,s)
113 #define VGMEMP_FREE(h,a)
114 #define VGMEMP_DESTROY(h)
115 #define VGMEMP_DEFINED(a,s)
119 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
120 /* Solaris just defines one or the other */
121 # define LITTLE_ENDIAN 1234
122 # define BIG_ENDIAN 4321
123 # ifdef _LITTLE_ENDIAN
124 # define BYTE_ORDER LITTLE_ENDIAN
126 # define BYTE_ORDER BIG_ENDIAN
129 # define BYTE_ORDER __BYTE_ORDER
133 #ifndef LITTLE_ENDIAN
134 #define LITTLE_ENDIAN __LITTLE_ENDIAN
137 #define BIG_ENDIAN __BIG_ENDIAN
140 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
141 #define MISALIGNED_OK 1
147 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
148 # error "Unknown or unsupported endianness (BYTE_ORDER)"
149 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
150 # error "Two's complement, reasonably sized integer types, please"
154 /** Put infrequently used env functions in separate section */
156 # define ESECT __attribute__ ((section("__TEXT,text_env")))
158 # define ESECT __attribute__ ((section("text_env")))
164 /** @defgroup internal LMDB Internals
167 /** @defgroup compat Compatibility Macros
168 * A bunch of macros to minimize the amount of platform-specific ifdefs
169 * needed throughout the rest of the code. When the features this library
170 * needs are similar enough to POSIX to be hidden in a one-or-two line
171 * replacement, this macro approach is used.
175 /** Features under development */
180 /** Wrapper around __func__, which is a C99 feature */
181 #if __STDC_VERSION__ >= 199901L
182 # define mdb_func_ __func__
183 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
184 # define mdb_func_ __FUNCTION__
186 /* If a debug message says <mdb_unknown>(), update the #if statements above */
187 # define mdb_func_ "<mdb_unknown>"
191 #define MDB_USE_HASH 1
192 #define MDB_PIDLOCK 0
193 #define THREAD_RET DWORD
194 #define pthread_t HANDLE
195 #define pthread_mutex_t HANDLE
196 #define pthread_cond_t HANDLE
197 #define pthread_key_t DWORD
198 #define pthread_self() GetCurrentThreadId()
199 #define pthread_key_create(x,y) \
200 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
201 #define pthread_key_delete(x) TlsFree(x)
202 #define pthread_getspecific(x) TlsGetValue(x)
203 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
204 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
205 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
206 #define pthread_cond_signal(x) SetEvent(*x)
207 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
208 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
209 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
210 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
211 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
212 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
213 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
214 #define getpid() GetCurrentProcessId()
215 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
216 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
217 #define ErrCode() GetLastError()
218 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
219 #define close(fd) (CloseHandle(fd) ? 0 : -1)
220 #define munmap(ptr,len) UnmapViewOfFile(ptr)
221 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
222 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
224 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
228 #define THREAD_RET void *
229 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
230 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
231 #define Z "z" /**< printf format modifier for size_t */
233 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
234 #define MDB_PIDLOCK 1
236 #ifdef MDB_USE_POSIX_SEM
238 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
239 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
240 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
241 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
244 mdb_sem_wait(sem_t *sem)
247 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
252 /** Lock the reader mutex.
254 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
255 /** Unlock the reader mutex.
257 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
259 /** Lock the writer mutex.
260 * Only a single write transaction is allowed at a time. Other writers
261 * will block waiting for this mutex.
263 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
264 /** Unlock the writer mutex.
266 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
267 #endif /* MDB_USE_POSIX_SEM */
269 /** Get the error code for the last failed system function.
271 #define ErrCode() errno
273 /** An abstraction for a file handle.
274 * On POSIX systems file handles are small integers. On Windows
275 * they're opaque pointers.
279 /** A value for an invalid file handle.
280 * Mainly used to initialize file variables and signify that they are
283 #define INVALID_HANDLE_VALUE (-1)
285 /** Get the size of a memory page for the system.
286 * This is the basic size that the platform's memory manager uses, and is
287 * fundamental to the use of memory-mapped files.
289 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
292 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
295 #define MNAME_LEN (sizeof(pthread_mutex_t))
301 /** A flag for opening a file and requesting synchronous data writes.
302 * This is only used when writing a meta page. It's not strictly needed;
303 * we could just do a normal write and then immediately perform a flush.
304 * But if this flag is available it saves us an extra system call.
306 * @note If O_DSYNC is undefined but exists in /usr/include,
307 * preferably set some compiler flag to get the definition.
308 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
311 # define MDB_DSYNC O_DSYNC
315 /** Function for flushing the data of a file. Define this to fsync
316 * if fdatasync() is not supported.
318 #ifndef MDB_FDATASYNC
319 # define MDB_FDATASYNC fdatasync
323 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
334 /** A page number in the database.
335 * Note that 64 bit page numbers are overkill, since pages themselves
336 * already represent 12-13 bits of addressable memory, and the OS will
337 * always limit applications to a maximum of 63 bits of address space.
339 * @note In the #MDB_node structure, we only store 48 bits of this value,
340 * which thus limits us to only 60 bits of addressable data.
342 typedef MDB_ID pgno_t;
344 /** A transaction ID.
345 * See struct MDB_txn.mt_txnid for details.
347 typedef MDB_ID txnid_t;
349 /** @defgroup debug Debug Macros
353 /** Enable debug output. Needs variable argument macros (a C99 feature).
354 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
355 * read from and written to the database (used for free space management).
361 static int mdb_debug;
362 static txnid_t mdb_debug_start;
364 /** Print a debug message with printf formatting.
365 * Requires double parenthesis around 2 or more args.
367 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
368 # define DPRINTF0(fmt, ...) \
369 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
371 # define DPRINTF(args) ((void) 0)
373 /** Print a debug string.
374 * The string is printed literally, with no format processing.
376 #define DPUTS(arg) DPRINTF(("%s", arg))
377 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
379 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
382 /** @brief The maximum size of a database page.
384 * It is 32k or 64k, since value-PAGEBASE must fit in
385 * #MDB_page.%mp_upper.
387 * LMDB will use database pages < OS pages if needed.
388 * That causes more I/O in write transactions: The OS must
389 * know (read) the whole page before writing a partial page.
391 * Note that we don't currently support Huge pages. On Linux,
392 * regular data files cannot use Huge pages, and in general
393 * Huge pages aren't actually pageable. We rely on the OS
394 * demand-pager to read our data and page it out when memory
395 * pressure from other processes is high. So until OSs have
396 * actual paging support for Huge pages, they're not viable.
398 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
400 /** The minimum number of keys required in a database page.
401 * Setting this to a larger value will place a smaller bound on the
402 * maximum size of a data item. Data items larger than this size will
403 * be pushed into overflow pages instead of being stored directly in
404 * the B-tree node. This value used to default to 4. With a page size
405 * of 4096 bytes that meant that any item larger than 1024 bytes would
406 * go into an overflow page. That also meant that on average 2-3KB of
407 * each overflow page was wasted space. The value cannot be lower than
408 * 2 because then there would no longer be a tree structure. With this
409 * value, items larger than 2KB will go into overflow pages, and on
410 * average only 1KB will be wasted.
412 #define MDB_MINKEYS 2
414 /** A stamp that identifies a file as an LMDB file.
415 * There's nothing special about this value other than that it is easily
416 * recognizable, and it will reflect any byte order mismatches.
418 #define MDB_MAGIC 0xBEEFC0DE
420 /** The version number for a database's datafile format. */
421 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
422 /** The version number for a database's lockfile format. */
423 #define MDB_LOCK_VERSION 1
425 /** @brief The max size of a key we can write, or 0 for dynamic max.
427 * Define this as 0 to compute the max from the page size. 511
428 * is default for backwards compat: liblmdb <= 0.9.10 can break
429 * when modifying a DB with keys/dupsort data bigger than its max.
430 * #MDB_DEVEL sets the default to 0.
432 * Data items in an #MDB_DUPSORT database are also limited to
433 * this size, since they're actually keys of a sub-DB. Keys and
434 * #MDB_DUPSORT data items must fit on a node in a regular page.
436 #ifndef MDB_MAXKEYSIZE
437 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
440 /** The maximum size of a key we can write to the environment. */
442 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
444 #define ENV_MAXKEY(env) ((env)->me_maxkey)
447 /** @brief The maximum size of a data item.
449 * We only store a 32 bit value for node sizes.
451 #define MAXDATASIZE 0xffffffffUL
454 /** Key size which fits in a #DKBUF.
457 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
460 * This is used for printing a hex dump of a key's contents.
462 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
463 /** Display a key in hex.
465 * Invoke a function to display a key in hex.
467 #define DKEY(x) mdb_dkey(x, kbuf)
473 /** An invalid page number.
474 * Mainly used to denote an empty tree.
476 #define P_INVALID (~(pgno_t)0)
478 /** Test if the flags \b f are set in a flag word \b w. */
479 #define F_ISSET(w, f) (((w) & (f)) == (f))
481 /** Round \b n up to an even number. */
482 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
484 /** Used for offsets within a single page.
485 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
488 typedef uint16_t indx_t;
490 /** Default size of memory map.
491 * This is certainly too small for any actual applications. Apps should always set
492 * the size explicitly using #mdb_env_set_mapsize().
494 #define DEFAULT_MAPSIZE 1048576
496 /** @defgroup readers Reader Lock Table
497 * Readers don't acquire any locks for their data access. Instead, they
498 * simply record their transaction ID in the reader table. The reader
499 * mutex is needed just to find an empty slot in the reader table. The
500 * slot's address is saved in thread-specific data so that subsequent read
501 * transactions started by the same thread need no further locking to proceed.
503 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
505 * No reader table is used if the database is on a read-only filesystem, or
506 * if #MDB_NOLOCK is set.
508 * Since the database uses multi-version concurrency control, readers don't
509 * actually need any locking. This table is used to keep track of which
510 * readers are using data from which old transactions, so that we'll know
511 * when a particular old transaction is no longer in use. Old transactions
512 * that have discarded any data pages can then have those pages reclaimed
513 * for use by a later write transaction.
515 * The lock table is constructed such that reader slots are aligned with the
516 * processor's cache line size. Any slot is only ever used by one thread.
517 * This alignment guarantees that there will be no contention or cache
518 * thrashing as threads update their own slot info, and also eliminates
519 * any need for locking when accessing a slot.
521 * A writer thread will scan every slot in the table to determine the oldest
522 * outstanding reader transaction. Any freed pages older than this will be
523 * reclaimed by the writer. The writer doesn't use any locks when scanning
524 * this table. This means that there's no guarantee that the writer will
525 * see the most up-to-date reader info, but that's not required for correct
526 * operation - all we need is to know the upper bound on the oldest reader,
527 * we don't care at all about the newest reader. So the only consequence of
528 * reading stale information here is that old pages might hang around a
529 * while longer before being reclaimed. That's actually good anyway, because
530 * the longer we delay reclaiming old pages, the more likely it is that a
531 * string of contiguous pages can be found after coalescing old pages from
532 * many old transactions together.
535 /** Number of slots in the reader table.
536 * This value was chosen somewhat arbitrarily. 126 readers plus a
537 * couple mutexes fit exactly into 8KB on my development machine.
538 * Applications should set the table size using #mdb_env_set_maxreaders().
540 #define DEFAULT_READERS 126
542 /** The size of a CPU cache line in bytes. We want our lock structures
543 * aligned to this size to avoid false cache line sharing in the
545 * This value works for most CPUs. For Itanium this should be 128.
551 /** The information we store in a single slot of the reader table.
552 * In addition to a transaction ID, we also record the process and
553 * thread ID that owns a slot, so that we can detect stale information,
554 * e.g. threads or processes that went away without cleaning up.
555 * @note We currently don't check for stale records. We simply re-init
556 * the table when we know that we're the only process opening the
559 typedef struct MDB_rxbody {
560 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
561 * Multiple readers that start at the same time will probably have the
562 * same ID here. Again, it's not important to exclude them from
563 * anything; all we need to know is which version of the DB they
564 * started from so we can avoid overwriting any data used in that
565 * particular version.
568 /** The process ID of the process owning this reader txn. */
570 /** The thread ID of the thread owning this txn. */
574 /** The actual reader record, with cacheline padding. */
575 typedef struct MDB_reader {
578 /** shorthand for mrb_txnid */
579 #define mr_txnid mru.mrx.mrb_txnid
580 #define mr_pid mru.mrx.mrb_pid
581 #define mr_tid mru.mrx.mrb_tid
582 /** cache line alignment */
583 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
587 /** The header for the reader table.
588 * The table resides in a memory-mapped file. (This is a different file
589 * than is used for the main database.)
591 * For POSIX the actual mutexes reside in the shared memory of this
592 * mapped file. On Windows, mutexes are named objects allocated by the
593 * kernel; we store the mutex names in this mapped file so that other
594 * processes can grab them. This same approach is also used on
595 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
596 * process-shared POSIX mutexes. For these cases where a named object
597 * is used, the object name is derived from a 64 bit FNV hash of the
598 * environment pathname. As such, naming collisions are extremely
599 * unlikely. If a collision occurs, the results are unpredictable.
601 typedef struct MDB_txbody {
602 /** Stamp identifying this as an LMDB file. It must be set
605 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
607 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
608 char mtb_rmname[MNAME_LEN];
610 /** Mutex protecting access to this table.
611 * This is the reader lock that #LOCK_MUTEX_R acquires.
613 pthread_mutex_t mtb_mutex;
615 /** The ID of the last transaction committed to the database.
616 * This is recorded here only for convenience; the value can always
617 * be determined by reading the main database meta pages.
620 /** The number of slots that have been used in the reader table.
621 * This always records the maximum count, it is not decremented
622 * when readers release their slots.
624 unsigned mtb_numreaders;
627 /** The actual reader table definition. */
628 typedef struct MDB_txninfo {
631 #define mti_magic mt1.mtb.mtb_magic
632 #define mti_format mt1.mtb.mtb_format
633 #define mti_mutex mt1.mtb.mtb_mutex
634 #define mti_rmname mt1.mtb.mtb_rmname
635 #define mti_txnid mt1.mtb.mtb_txnid
636 #define mti_numreaders mt1.mtb.mtb_numreaders
637 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
640 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
641 char mt2_wmname[MNAME_LEN];
642 #define mti_wmname mt2.mt2_wmname
644 pthread_mutex_t mt2_wmutex;
645 #define mti_wmutex mt2.mt2_wmutex
647 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
649 MDB_reader mti_readers[1];
652 /** Lockfile format signature: version, features and field layout */
653 #define MDB_LOCK_FORMAT \
655 ((MDB_LOCK_VERSION) \
656 /* Flags which describe functionality */ \
657 + (((MDB_PIDLOCK) != 0) << 16)))
660 /** Common header for all page types.
661 * Overflow records occupy a number of contiguous pages with no
662 * headers on any page after the first.
664 typedef struct MDB_page {
665 #define mp_pgno mp_p.p_pgno
666 #define mp_next mp_p.p_next
668 pgno_t p_pgno; /**< page number */
669 struct MDB_page *p_next; /**< for in-memory list of freed pages */
672 /** @defgroup mdb_page Page Flags
674 * Flags for the page headers.
677 #define P_BRANCH 0x01 /**< branch page */
678 #define P_LEAF 0x02 /**< leaf page */
679 #define P_OVERFLOW 0x04 /**< overflow page */
680 #define P_META 0x08 /**< meta page */
681 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
682 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
683 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
684 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
685 #define P_KEEP 0x8000 /**< leave this page alone during spill */
687 uint16_t mp_flags; /**< @ref mdb_page */
688 #define mp_lower mp_pb.pb.pb_lower
689 #define mp_upper mp_pb.pb.pb_upper
690 #define mp_pages mp_pb.pb_pages
693 indx_t pb_lower; /**< lower bound of free space */
694 indx_t pb_upper; /**< upper bound of free space */
696 uint32_t pb_pages; /**< number of overflow pages */
698 indx_t mp_ptrs[1]; /**< dynamic size */
701 /** Size of the page header, excluding dynamic data at the end */
702 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
704 /** Address of first usable data byte in a page, after the header */
705 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
707 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
708 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
710 /** Number of nodes on a page */
711 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
713 /** The amount of space remaining in the page */
714 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
716 /** The percentage of space used in the page, in tenths of a percent. */
717 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
718 ((env)->me_psize - PAGEHDRSZ))
719 /** The minimum page fill factor, in tenths of a percent.
720 * Pages emptier than this are candidates for merging.
722 #define FILL_THRESHOLD 250
724 /** Test if a page is a leaf page */
725 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
726 /** Test if a page is a LEAF2 page */
727 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
728 /** Test if a page is a branch page */
729 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
730 /** Test if a page is an overflow page */
731 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
732 /** Test if a page is a sub page */
733 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
735 /** The number of overflow pages needed to store the given size. */
736 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
738 /** Link in #MDB_txn.%mt_loose_pages list */
739 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
741 /** Header for a single key/data pair within a page.
742 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
743 * We guarantee 2-byte alignment for 'MDB_node's.
745 typedef struct MDB_node {
746 /** lo and hi are used for data size on leaf nodes and for
747 * child pgno on branch nodes. On 64 bit platforms, flags
748 * is also used for pgno. (Branch nodes have no flags).
749 * They are in host byte order in case that lets some
750 * accesses be optimized into a 32-bit word access.
752 #if BYTE_ORDER == LITTLE_ENDIAN
753 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
755 unsigned short mn_hi, mn_lo;
757 /** @defgroup mdb_node Node Flags
759 * Flags for node headers.
762 #define F_BIGDATA 0x01 /**< data put on overflow page */
763 #define F_SUBDATA 0x02 /**< data is a sub-database */
764 #define F_DUPDATA 0x04 /**< data has duplicates */
766 /** valid flags for #mdb_node_add() */
767 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
770 unsigned short mn_flags; /**< @ref mdb_node */
771 unsigned short mn_ksize; /**< key size */
772 char mn_data[1]; /**< key and data are appended here */
775 /** Size of the node header, excluding dynamic data at the end */
776 #define NODESIZE offsetof(MDB_node, mn_data)
778 /** Bit position of top word in page number, for shifting mn_flags */
779 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
781 /** Size of a node in a branch page with a given key.
782 * This is just the node header plus the key, there is no data.
784 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
786 /** Size of a node in a leaf page with a given key and data.
787 * This is node header plus key plus data size.
789 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
791 /** Address of node \b i in page \b p */
792 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
794 /** Address of the key for the node */
795 #define NODEKEY(node) (void *)((node)->mn_data)
797 /** Address of the data for a node */
798 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
800 /** Get the page number pointed to by a branch node */
801 #define NODEPGNO(node) \
802 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
803 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
804 /** Set the page number in a branch node */
805 #define SETPGNO(node,pgno) do { \
806 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
807 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
809 /** Get the size of the data in a leaf node */
810 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
811 /** Set the size of the data for a leaf node */
812 #define SETDSZ(node,size) do { \
813 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
814 /** The size of a key in a node */
815 #define NODEKSZ(node) ((node)->mn_ksize)
817 /** Copy a page number from src to dst */
819 #define COPY_PGNO(dst,src) dst = src
821 #if SIZE_MAX > 4294967295UL
822 #define COPY_PGNO(dst,src) do { \
823 unsigned short *s, *d; \
824 s = (unsigned short *)&(src); \
825 d = (unsigned short *)&(dst); \
832 #define COPY_PGNO(dst,src) do { \
833 unsigned short *s, *d; \
834 s = (unsigned short *)&(src); \
835 d = (unsigned short *)&(dst); \
841 /** The address of a key in a LEAF2 page.
842 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
843 * There are no node headers, keys are stored contiguously.
845 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
847 /** Set the \b node's key into \b keyptr, if requested. */
848 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
849 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
851 /** Set the \b node's key into \b key. */
852 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
854 /** Information about a single database in the environment. */
855 typedef struct MDB_db {
856 uint32_t md_pad; /**< also ksize for LEAF2 pages */
857 uint16_t md_flags; /**< @ref mdb_dbi_open */
858 uint16_t md_depth; /**< depth of this tree */
859 pgno_t md_branch_pages; /**< number of internal pages */
860 pgno_t md_leaf_pages; /**< number of leaf pages */
861 pgno_t md_overflow_pages; /**< number of overflow pages */
862 size_t md_entries; /**< number of data items */
863 pgno_t md_root; /**< the root page of this tree */
866 /** mdb_dbi_open flags */
867 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
868 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
869 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
870 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
872 /** Handle for the DB used to track free pages. */
874 /** Handle for the default DB. */
877 /** Meta page content.
878 * A meta page is the start point for accessing a database snapshot.
879 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
881 typedef struct MDB_meta {
882 /** Stamp identifying this as an LMDB file. It must be set
885 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
887 void *mm_address; /**< address for fixed mapping */
888 size_t mm_mapsize; /**< size of mmap region */
889 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
890 /** The size of pages used in this DB */
891 #define mm_psize mm_dbs[0].md_pad
892 /** Any persistent environment flags. @ref mdb_env */
893 #define mm_flags mm_dbs[0].md_flags
894 pgno_t mm_last_pg; /**< last used page in file */
895 txnid_t mm_txnid; /**< txnid that committed this page */
898 /** Buffer for a stack-allocated meta page.
899 * The members define size and alignment, and silence type
900 * aliasing warnings. They are not used directly; that could
901 * mean incorrectly using several union members in parallel.
903 typedef union MDB_metabuf {
906 char mm_pad[PAGEHDRSZ];
911 /** Auxiliary DB info.
912 * The information here is mostly static/read-only. There is
913 * only a single copy of this record in the environment.
915 typedef struct MDB_dbx {
916 MDB_val md_name; /**< name of the database */
917 MDB_cmp_func *md_cmp; /**< function for comparing keys */
918 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
919 MDB_rel_func *md_rel; /**< user relocate function */
920 void *md_relctx; /**< user-provided context for md_rel */
923 /** A database transaction.
924 * Every operation requires a transaction handle.
927 MDB_txn *mt_parent; /**< parent of a nested txn */
928 MDB_txn *mt_child; /**< nested txn under this txn */
929 pgno_t mt_next_pgno; /**< next unallocated page */
930 /** The ID of this transaction. IDs are integers incrementing from 1.
931 * Only committed write transactions increment the ID. If a transaction
932 * aborts, the ID may be re-used by the next writer.
935 MDB_env *mt_env; /**< the DB environment */
936 /** The list of pages that became unused during this transaction.
939 /** The list of loose pages that became unused and may be reused
940 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
942 MDB_page *mt_loose_pgs;
943 /** The sorted list of dirty pages we temporarily wrote to disk
944 * because the dirty list was full. page numbers in here are
945 * shifted left by 1, deleted slots have the LSB set.
947 MDB_IDL mt_spill_pgs;
949 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
951 /** For read txns: This thread/txn's reader table slot, or NULL. */
954 /** Array of records for each DB known in the environment. */
956 /** Array of MDB_db records for each known DB */
958 /** Array of sequence numbers for each DB handle */
959 unsigned int *mt_dbiseqs;
960 /** @defgroup mt_dbflag Transaction DB Flags
964 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
965 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
966 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
967 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
969 /** In write txns, array of cursors for each DB */
970 MDB_cursor **mt_cursors;
971 /** Array of flags for each DB */
972 unsigned char *mt_dbflags;
973 /** Number of DB records in use. This number only ever increments;
974 * we don't decrement it when individual DB handles are closed.
978 /** @defgroup mdb_txn Transaction Flags
982 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
983 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
984 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
985 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
987 unsigned int mt_flags; /**< @ref mdb_txn */
988 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
989 * Includes ancestor txns' dirty pages not hidden by other txns'
990 * dirty/spilled pages. Thus commit(nested txn) has room to merge
991 * dirty_list into mt_parent after freeing hidden mt_parent pages.
993 unsigned int mt_dirty_room;
996 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
997 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
998 * raise this on a 64 bit machine.
1000 #define CURSOR_STACK 32
1004 /** Cursors are used for all DB operations.
1005 * A cursor holds a path of (page pointer, key index) from the DB
1006 * root to a position in the DB, plus other state. #MDB_DUPSORT
1007 * cursors include an xcursor to the current data item. Write txns
1008 * track their cursors and keep them up to date when data moves.
1009 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1010 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1013 /** Next cursor on this DB in this txn */
1014 MDB_cursor *mc_next;
1015 /** Backup of the original cursor if this cursor is a shadow */
1016 MDB_cursor *mc_backup;
1017 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1018 struct MDB_xcursor *mc_xcursor;
1019 /** The transaction that owns this cursor */
1021 /** The database handle this cursor operates on */
1023 /** The database record for this cursor */
1025 /** The database auxiliary record for this cursor */
1027 /** The @ref mt_dbflag for this database */
1028 unsigned char *mc_dbflag;
1029 unsigned short mc_snum; /**< number of pushed pages */
1030 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1031 /** @defgroup mdb_cursor Cursor Flags
1033 * Cursor state flags.
1036 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1037 #define C_EOF 0x02 /**< No more data */
1038 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1039 #define C_DEL 0x08 /**< last op was a cursor_del */
1040 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1041 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1043 unsigned int mc_flags; /**< @ref mdb_cursor */
1044 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1045 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1048 /** Context for sorted-dup records.
1049 * We could have gone to a fully recursive design, with arbitrarily
1050 * deep nesting of sub-databases. But for now we only handle these
1051 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1053 typedef struct MDB_xcursor {
1054 /** A sub-cursor for traversing the Dup DB */
1055 MDB_cursor mx_cursor;
1056 /** The database record for this Dup DB */
1058 /** The auxiliary DB record for this Dup DB */
1060 /** The @ref mt_dbflag for this Dup DB */
1061 unsigned char mx_dbflag;
1064 /** State of FreeDB old pages, stored in the MDB_env */
1065 typedef struct MDB_pgstate {
1066 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1067 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1070 /** The database environment. */
1072 HANDLE me_fd; /**< The main data file */
1073 HANDLE me_lfd; /**< The lock file */
1074 HANDLE me_mfd; /**< just for writing the meta pages */
1075 /** Failed to update the meta page. Probably an I/O error. */
1076 #define MDB_FATAL_ERROR 0x80000000U
1077 /** Some fields are initialized. */
1078 #define MDB_ENV_ACTIVE 0x20000000U
1079 /** me_txkey is set */
1080 #define MDB_ENV_TXKEY 0x10000000U
1081 uint32_t me_flags; /**< @ref mdb_env */
1082 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1083 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1084 unsigned int me_maxreaders; /**< size of the reader table */
1085 unsigned int me_numreaders; /**< max numreaders set by this env */
1086 MDB_dbi me_numdbs; /**< number of DBs opened */
1087 MDB_dbi me_maxdbs; /**< size of the DB table */
1088 MDB_PID_T me_pid; /**< process ID of this env */
1089 char *me_path; /**< path to the DB files */
1090 char *me_map; /**< the memory map of the data file */
1091 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1092 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1093 void *me_pbuf; /**< scratch area for DUPSORT put() */
1094 MDB_txn *me_txn; /**< current write transaction */
1095 size_t me_mapsize; /**< size of the data memory map */
1096 off_t me_size; /**< current file size */
1097 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1098 MDB_dbx *me_dbxs; /**< array of static DB info */
1099 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1100 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1101 pthread_key_t me_txkey; /**< thread-key for readers */
1102 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1103 # define me_pglast me_pgstate.mf_pglast
1104 # define me_pghead me_pgstate.mf_pghead
1105 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1106 /** IDL of pages that became unused in a write txn */
1107 MDB_IDL me_free_pgs;
1108 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1109 MDB_ID2L me_dirty_list;
1110 /** Max number of freelist items that can fit in a single overflow page */
1112 /** Max size of a node on a page */
1113 unsigned int me_nodemax;
1114 #if !(MDB_MAXKEYSIZE)
1115 unsigned int me_maxkey; /**< max size of a key */
1117 int me_live_reader; /**< have liveness lock in reader table */
1119 int me_pidquery; /**< Used in OpenProcess */
1120 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1122 #elif defined(MDB_USE_POSIX_SEM)
1123 sem_t *me_rmutex; /* Shared mutexes are not supported */
1126 void *me_userctx; /**< User-settable context */
1127 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1130 /** Nested transaction */
1131 typedef struct MDB_ntxn {
1132 MDB_txn mnt_txn; /**< the transaction */
1133 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1136 /** max number of pages to commit in one writev() call */
1137 #define MDB_COMMIT_PAGES 64
1138 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1139 #undef MDB_COMMIT_PAGES
1140 #define MDB_COMMIT_PAGES IOV_MAX
1143 /** max bytes to write in one call */
1144 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1146 /** Check \b txn and \b dbi arguments to a function */
1147 #define TXN_DBI_EXIST(txn, dbi) \
1148 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1150 /** Check for misused \b dbi handles */
1151 #define TXN_DBI_CHANGED(txn, dbi) \
1152 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1154 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1155 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1156 static int mdb_page_touch(MDB_cursor *mc);
1158 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1159 static int mdb_page_search_root(MDB_cursor *mc,
1160 MDB_val *key, int modify);
1161 #define MDB_PS_MODIFY 1
1162 #define MDB_PS_ROOTONLY 2
1163 #define MDB_PS_FIRST 4
1164 #define MDB_PS_LAST 8
1165 static int mdb_page_search(MDB_cursor *mc,
1166 MDB_val *key, int flags);
1167 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1169 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1170 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1171 pgno_t newpgno, unsigned int nflags);
1173 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1174 static int mdb_env_pick_meta(const MDB_env *env);
1175 static int mdb_env_write_meta(MDB_txn *txn);
1176 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1177 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1179 static void mdb_env_close0(MDB_env *env, int excl);
1181 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1182 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1183 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1184 static void mdb_node_del(MDB_cursor *mc, int ksize);
1185 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1186 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1187 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1188 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1189 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1191 static int mdb_rebalance(MDB_cursor *mc);
1192 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1194 static void mdb_cursor_pop(MDB_cursor *mc);
1195 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1197 static int mdb_cursor_del0(MDB_cursor *mc);
1198 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1199 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1200 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1201 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1202 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1204 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1205 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1207 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1208 static void mdb_xcursor_init0(MDB_cursor *mc);
1209 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1211 static int mdb_drop0(MDB_cursor *mc, int subs);
1212 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1215 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1219 static SECURITY_DESCRIPTOR mdb_null_sd;
1220 static SECURITY_ATTRIBUTES mdb_all_sa;
1221 static int mdb_sec_inited;
1224 /** Return the library version info. */
1226 mdb_version(int *major, int *minor, int *patch)
1228 if (major) *major = MDB_VERSION_MAJOR;
1229 if (minor) *minor = MDB_VERSION_MINOR;
1230 if (patch) *patch = MDB_VERSION_PATCH;
1231 return MDB_VERSION_STRING;
1234 /** Table of descriptions for LMDB @ref errors */
1235 static char *const mdb_errstr[] = {
1236 "MDB_KEYEXIST: Key/data pair already exists",
1237 "MDB_NOTFOUND: No matching key/data pair found",
1238 "MDB_PAGE_NOTFOUND: Requested page not found",
1239 "MDB_CORRUPTED: Located page was wrong type",
1240 "MDB_PANIC: Update of meta page failed",
1241 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1242 "MDB_INVALID: File is not an LMDB file",
1243 "MDB_MAP_FULL: Environment mapsize limit reached",
1244 "MDB_DBS_FULL: Environment maxdbs limit reached",
1245 "MDB_READERS_FULL: Environment maxreaders limit reached",
1246 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1247 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1248 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1249 "MDB_PAGE_FULL: Internal error - page has no more space",
1250 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1251 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1252 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1253 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1254 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1255 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1259 mdb_strerror(int err)
1262 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1263 * This works as long as no function between the call to mdb_strerror
1264 * and the actual use of the message uses more than 4K of stack.
1267 char buf[1024], *ptr = buf;
1271 return ("Successful return: 0");
1273 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1274 i = err - MDB_KEYEXIST;
1275 return mdb_errstr[i];
1279 /* These are the C-runtime error codes we use. The comment indicates
1280 * their numeric value, and the Win32 error they would correspond to
1281 * if the error actually came from a Win32 API. A major mess, we should
1282 * have used LMDB-specific error codes for everything.
1285 case ENOENT: /* 2, FILE_NOT_FOUND */
1286 case EIO: /* 5, ACCESS_DENIED */
1287 case ENOMEM: /* 12, INVALID_ACCESS */
1288 case EACCES: /* 13, INVALID_DATA */
1289 case EBUSY: /* 16, CURRENT_DIRECTORY */
1290 case EINVAL: /* 22, BAD_COMMAND */
1291 case ENOSPC: /* 28, OUT_OF_PAPER */
1292 return strerror(err);
1297 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1298 FORMAT_MESSAGE_IGNORE_INSERTS,
1299 NULL, err, 0, ptr, sizeof(buf), pad);
1302 return strerror(err);
1306 /** assert(3) variant in cursor context */
1307 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1308 /** assert(3) variant in transaction context */
1309 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1310 /** assert(3) variant in environment context */
1311 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1314 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1315 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1318 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1319 const char *func, const char *file, int line)
1322 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1323 file, line, expr_txt, func);
1324 if (env->me_assert_func)
1325 env->me_assert_func(env, buf);
1326 fprintf(stderr, "%s\n", buf);
1330 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1334 /** Return the page number of \b mp which may be sub-page, for debug output */
1336 mdb_dbg_pgno(MDB_page *mp)
1339 COPY_PGNO(ret, mp->mp_pgno);
1343 /** Display a key in hexadecimal and return the address of the result.
1344 * @param[in] key the key to display
1345 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1346 * @return The key in hexadecimal form.
1349 mdb_dkey(MDB_val *key, char *buf)
1352 unsigned char *c = key->mv_data;
1358 if (key->mv_size > DKBUF_MAXKEYSIZE)
1359 return "MDB_MAXKEYSIZE";
1360 /* may want to make this a dynamic check: if the key is mostly
1361 * printable characters, print it as-is instead of converting to hex.
1365 for (i=0; i<key->mv_size; i++)
1366 ptr += sprintf(ptr, "%02x", *c++);
1368 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1374 mdb_leafnode_type(MDB_node *n)
1376 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1377 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1378 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1381 /** Display all the keys in the page. */
1383 mdb_page_list(MDB_page *mp)
1385 pgno_t pgno = mdb_dbg_pgno(mp);
1386 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1388 unsigned int i, nkeys, nsize, total = 0;
1392 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1393 case P_BRANCH: type = "Branch page"; break;
1394 case P_LEAF: type = "Leaf page"; break;
1395 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1396 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1397 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1399 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1400 pgno, mp->mp_pages, state);
1403 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1404 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1407 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1411 nkeys = NUMKEYS(mp);
1412 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1414 for (i=0; i<nkeys; i++) {
1415 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1416 key.mv_size = nsize = mp->mp_pad;
1417 key.mv_data = LEAF2KEY(mp, i, nsize);
1419 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1422 node = NODEPTR(mp, i);
1423 key.mv_size = node->mn_ksize;
1424 key.mv_data = node->mn_data;
1425 nsize = NODESIZE + key.mv_size;
1426 if (IS_BRANCH(mp)) {
1427 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1431 if (F_ISSET(node->mn_flags, F_BIGDATA))
1432 nsize += sizeof(pgno_t);
1434 nsize += NODEDSZ(node);
1436 nsize += sizeof(indx_t);
1437 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1438 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1440 total = EVEN(total);
1442 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1443 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1447 mdb_cursor_chk(MDB_cursor *mc)
1453 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1454 for (i=0; i<mc->mc_top; i++) {
1456 node = NODEPTR(mp, mc->mc_ki[i]);
1457 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1460 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1466 /** Count all the pages in each DB and in the freelist
1467 * and make sure it matches the actual number of pages
1469 * All named DBs must be open for a correct count.
1471 static void mdb_audit(MDB_txn *txn)
1475 MDB_ID freecount, count;
1480 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1481 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1482 freecount += *(MDB_ID *)data.mv_data;
1483 mdb_tassert(txn, rc == MDB_NOTFOUND);
1486 for (i = 0; i<txn->mt_numdbs; i++) {
1488 if (!(txn->mt_dbflags[i] & DB_VALID))
1490 mdb_cursor_init(&mc, txn, i, &mx);
1491 if (txn->mt_dbs[i].md_root == P_INVALID)
1493 count += txn->mt_dbs[i].md_branch_pages +
1494 txn->mt_dbs[i].md_leaf_pages +
1495 txn->mt_dbs[i].md_overflow_pages;
1496 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1497 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1498 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1501 mp = mc.mc_pg[mc.mc_top];
1502 for (j=0; j<NUMKEYS(mp); j++) {
1503 MDB_node *leaf = NODEPTR(mp, j);
1504 if (leaf->mn_flags & F_SUBDATA) {
1506 memcpy(&db, NODEDATA(leaf), sizeof(db));
1507 count += db.md_branch_pages + db.md_leaf_pages +
1508 db.md_overflow_pages;
1512 mdb_tassert(txn, rc == MDB_NOTFOUND);
1515 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1516 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1517 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1523 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1525 return txn->mt_dbxs[dbi].md_cmp(a, b);
1529 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1531 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1534 /** Allocate memory for a page.
1535 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1538 mdb_page_malloc(MDB_txn *txn, unsigned num)
1540 MDB_env *env = txn->mt_env;
1541 MDB_page *ret = env->me_dpages;
1542 size_t psize = env->me_psize, sz = psize, off;
1543 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1544 * For a single page alloc, we init everything after the page header.
1545 * For multi-page, we init the final page; if the caller needed that
1546 * many pages they will be filling in at least up to the last page.
1550 VGMEMP_ALLOC(env, ret, sz);
1551 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1552 env->me_dpages = ret->mp_next;
1555 psize -= off = PAGEHDRSZ;
1560 if ((ret = malloc(sz)) != NULL) {
1561 VGMEMP_ALLOC(env, ret, sz);
1562 if (!(env->me_flags & MDB_NOMEMINIT)) {
1563 memset((char *)ret + off, 0, psize);
1567 txn->mt_flags |= MDB_TXN_ERROR;
1571 /** Free a single page.
1572 * Saves single pages to a list, for future reuse.
1573 * (This is not used for multi-page overflow pages.)
1576 mdb_page_free(MDB_env *env, MDB_page *mp)
1578 mp->mp_next = env->me_dpages;
1579 VGMEMP_FREE(env, mp);
1580 env->me_dpages = mp;
1583 /** Free a dirty page */
1585 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1587 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1588 mdb_page_free(env, dp);
1590 /* large pages just get freed directly */
1591 VGMEMP_FREE(env, dp);
1596 /** Return all dirty pages to dpage list */
1598 mdb_dlist_free(MDB_txn *txn)
1600 MDB_env *env = txn->mt_env;
1601 MDB_ID2L dl = txn->mt_u.dirty_list;
1602 unsigned i, n = dl[0].mid;
1604 for (i = 1; i <= n; i++) {
1605 mdb_dpage_free(env, dl[i].mptr);
1610 /** Loosen or free a single page.
1611 * Saves single pages to a list for future reuse
1612 * in this same txn. It has been pulled from the freeDB
1613 * and already resides on the dirty list, but has been
1614 * deleted. Use these pages first before pulling again
1617 * If the page wasn't dirtied in this txn, just add it
1618 * to this txn's free list.
1621 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1624 pgno_t pgno = mp->mp_pgno;
1626 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1627 if (mc->mc_txn->mt_parent) {
1628 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1629 /* If txn has a parent, make sure the page is in our
1633 unsigned x = mdb_mid2l_search(dl, pgno);
1634 if (x <= dl[0].mid && dl[x].mid == pgno) {
1635 if (mp != dl[x].mptr) { /* bad cursor? */
1636 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1637 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1638 return MDB_CORRUPTED;
1645 /* no parent txn, so it's just ours */
1650 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1652 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1653 mc->mc_txn->mt_loose_pgs = mp;
1654 mp->mp_flags |= P_LOOSE;
1656 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1664 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1665 * @param[in] mc A cursor handle for the current operation.
1666 * @param[in] pflags Flags of the pages to update:
1667 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1668 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1669 * @return 0 on success, non-zero on failure.
1672 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1674 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1675 MDB_txn *txn = mc->mc_txn;
1681 int rc = MDB_SUCCESS, level;
1683 /* Mark pages seen by cursors */
1684 if (mc->mc_flags & C_UNTRACK)
1685 mc = NULL; /* will find mc in mt_cursors */
1686 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1687 for (; mc; mc=mc->mc_next) {
1688 if (!(mc->mc_flags & C_INITIALIZED))
1690 for (m3 = mc;; m3 = &mx->mx_cursor) {
1692 for (j=0; j<m3->mc_snum; j++) {
1694 if ((mp->mp_flags & Mask) == pflags)
1695 mp->mp_flags ^= P_KEEP;
1697 mx = m3->mc_xcursor;
1698 /* Proceed to mx if it is at a sub-database */
1699 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1701 if (! (mp && (mp->mp_flags & P_LEAF)))
1703 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1704 if (!(leaf->mn_flags & F_SUBDATA))
1713 /* Mark dirty root pages */
1714 for (i=0; i<txn->mt_numdbs; i++) {
1715 if (txn->mt_dbflags[i] & DB_DIRTY) {
1716 pgno_t pgno = txn->mt_dbs[i].md_root;
1717 if (pgno == P_INVALID)
1719 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1721 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1722 dp->mp_flags ^= P_KEEP;
1730 static int mdb_page_flush(MDB_txn *txn, int keep);
1732 /** Spill pages from the dirty list back to disk.
1733 * This is intended to prevent running into #MDB_TXN_FULL situations,
1734 * but note that they may still occur in a few cases:
1735 * 1) our estimate of the txn size could be too small. Currently this
1736 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1737 * 2) child txns may run out of space if their parents dirtied a
1738 * lot of pages and never spilled them. TODO: we probably should do
1739 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1740 * the parent's dirty_room is below a given threshold.
1742 * Otherwise, if not using nested txns, it is expected that apps will
1743 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1744 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1745 * If the txn never references them again, they can be left alone.
1746 * If the txn only reads them, they can be used without any fuss.
1747 * If the txn writes them again, they can be dirtied immediately without
1748 * going thru all of the work of #mdb_page_touch(). Such references are
1749 * handled by #mdb_page_unspill().
1751 * Also note, we never spill DB root pages, nor pages of active cursors,
1752 * because we'll need these back again soon anyway. And in nested txns,
1753 * we can't spill a page in a child txn if it was already spilled in a
1754 * parent txn. That would alter the parent txns' data even though
1755 * the child hasn't committed yet, and we'd have no way to undo it if
1756 * the child aborted.
1758 * @param[in] m0 cursor A cursor handle identifying the transaction and
1759 * database for which we are checking space.
1760 * @param[in] key For a put operation, the key being stored.
1761 * @param[in] data For a put operation, the data being stored.
1762 * @return 0 on success, non-zero on failure.
1765 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1767 MDB_txn *txn = m0->mc_txn;
1769 MDB_ID2L dl = txn->mt_u.dirty_list;
1770 unsigned int i, j, need;
1773 if (m0->mc_flags & C_SUB)
1776 /* Estimate how much space this op will take */
1777 i = m0->mc_db->md_depth;
1778 /* Named DBs also dirty the main DB */
1779 if (m0->mc_dbi > MAIN_DBI)
1780 i += txn->mt_dbs[MAIN_DBI].md_depth;
1781 /* For puts, roughly factor in the key+data size */
1783 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1784 i += i; /* double it for good measure */
1787 if (txn->mt_dirty_room > i)
1790 if (!txn->mt_spill_pgs) {
1791 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1792 if (!txn->mt_spill_pgs)
1795 /* purge deleted slots */
1796 MDB_IDL sl = txn->mt_spill_pgs;
1797 unsigned int num = sl[0];
1799 for (i=1; i<=num; i++) {
1806 /* Preserve pages which may soon be dirtied again */
1807 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1810 /* Less aggressive spill - we originally spilled the entire dirty list,
1811 * with a few exceptions for cursor pages and DB root pages. But this
1812 * turns out to be a lot of wasted effort because in a large txn many
1813 * of those pages will need to be used again. So now we spill only 1/8th
1814 * of the dirty pages. Testing revealed this to be a good tradeoff,
1815 * better than 1/2, 1/4, or 1/10.
1817 if (need < MDB_IDL_UM_MAX / 8)
1818 need = MDB_IDL_UM_MAX / 8;
1820 /* Save the page IDs of all the pages we're flushing */
1821 /* flush from the tail forward, this saves a lot of shifting later on. */
1822 for (i=dl[0].mid; i && need; i--) {
1823 MDB_ID pn = dl[i].mid << 1;
1825 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1827 /* Can't spill twice, make sure it's not already in a parent's
1830 if (txn->mt_parent) {
1832 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1833 if (tx2->mt_spill_pgs) {
1834 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1835 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1836 dp->mp_flags |= P_KEEP;
1844 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1848 mdb_midl_sort(txn->mt_spill_pgs);
1850 /* Flush the spilled part of dirty list */
1851 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1854 /* Reset any dirty pages we kept that page_flush didn't see */
1855 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1858 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1862 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1864 mdb_find_oldest(MDB_txn *txn)
1867 txnid_t mr, oldest = txn->mt_txnid - 1;
1868 if (txn->mt_env->me_txns) {
1869 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1870 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1881 /** Add a page to the txn's dirty list */
1883 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1886 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1888 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1889 insert = mdb_mid2l_append;
1891 insert = mdb_mid2l_insert;
1893 mid.mid = mp->mp_pgno;
1895 rc = insert(txn->mt_u.dirty_list, &mid);
1896 mdb_tassert(txn, rc == 0);
1897 txn->mt_dirty_room--;
1900 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1901 * me_pghead and mt_next_pgno.
1903 * If there are free pages available from older transactions, they
1904 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1905 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1906 * and move me_pglast to say which records were consumed. Only this
1907 * function can create me_pghead and move me_pglast/mt_next_pgno.
1908 * @param[in] mc cursor A cursor handle identifying the transaction and
1909 * database for which we are allocating.
1910 * @param[in] num the number of pages to allocate.
1911 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1912 * will always be satisfied by a single contiguous chunk of memory.
1913 * @return 0 on success, non-zero on failure.
1916 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1918 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1919 /* Get at most <Max_retries> more freeDB records once me_pghead
1920 * has enough pages. If not enough, use new pages from the map.
1921 * If <Paranoid> and mc is updating the freeDB, only get new
1922 * records if me_pghead is empty. Then the freelist cannot play
1923 * catch-up with itself by growing while trying to save it.
1925 enum { Paranoid = 1, Max_retries = 500 };
1927 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1929 int rc, retry = num * 20;
1930 MDB_txn *txn = mc->mc_txn;
1931 MDB_env *env = txn->mt_env;
1932 pgno_t pgno, *mop = env->me_pghead;
1933 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1935 txnid_t oldest = 0, last;
1939 /* If there are any loose pages, just use them */
1940 if (num == 1 && txn->mt_loose_pgs) {
1941 np = txn->mt_loose_pgs;
1942 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1943 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1951 /* If our dirty list is already full, we can't do anything */
1952 if (txn->mt_dirty_room == 0) {
1957 for (op = MDB_FIRST;; op = MDB_NEXT) {
1960 pgno_t *idl, old_id, new_id;
1962 /* Seek a big enough contiguous page range. Prefer
1963 * pages at the tail, just truncating the list.
1969 if (mop[i-n2] == pgno+n2)
1976 if (op == MDB_FIRST) { /* 1st iteration */
1977 /* Prepare to fetch more and coalesce */
1978 oldest = mdb_find_oldest(txn);
1979 last = env->me_pglast;
1980 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1983 key.mv_data = &last; /* will look up last+1 */
1984 key.mv_size = sizeof(last);
1986 if (Paranoid && mc->mc_dbi == FREE_DBI)
1989 if (Paranoid && retry < 0 && mop_len)
1993 /* Do not fetch more if the record will be too recent */
1996 rc = mdb_cursor_get(&m2, &key, NULL, op);
1998 if (rc == MDB_NOTFOUND)
2002 last = *(txnid_t*)key.mv_data;
2005 np = m2.mc_pg[m2.mc_top];
2006 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2007 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2010 idl = (MDB_ID *) data.mv_data;
2013 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2018 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2020 mop = env->me_pghead;
2022 env->me_pglast = last;
2024 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2025 last, txn->mt_dbs[FREE_DBI].md_root, i));
2027 DPRINTF(("IDL %"Z"u", idl[k]));
2029 /* Merge in descending sorted order */
2032 mop[0] = (pgno_t)-1;
2036 for (; old_id < new_id; old_id = mop[--j])
2043 /* Use new pages from the map when nothing suitable in the freeDB */
2045 pgno = txn->mt_next_pgno;
2046 if (pgno + num >= env->me_maxpg) {
2047 DPUTS("DB size maxed out");
2053 if (env->me_flags & MDB_WRITEMAP) {
2054 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2056 if (!(np = mdb_page_malloc(txn, num))) {
2062 mop[0] = mop_len -= num;
2063 /* Move any stragglers down */
2064 for (j = i-num; j < mop_len; )
2065 mop[++j] = mop[++i];
2067 txn->mt_next_pgno = pgno + num;
2070 mdb_page_dirty(txn, np);
2076 txn->mt_flags |= MDB_TXN_ERROR;
2080 /** Copy the used portions of a non-overflow page.
2081 * @param[in] dst page to copy into
2082 * @param[in] src page to copy from
2083 * @param[in] psize size of a page
2086 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2088 enum { Align = sizeof(pgno_t) };
2089 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2091 /* If page isn't full, just copy the used portion. Adjust
2092 * alignment so memcpy may copy words instead of bytes.
2094 if ((unused &= -Align) && !IS_LEAF2(src)) {
2095 upper = (upper + PAGEBASE) & -Align;
2096 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2097 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2100 memcpy(dst, src, psize - unused);
2104 /** Pull a page off the txn's spill list, if present.
2105 * If a page being referenced was spilled to disk in this txn, bring
2106 * it back and make it dirty/writable again.
2107 * @param[in] txn the transaction handle.
2108 * @param[in] mp the page being referenced. It must not be dirty.
2109 * @param[out] ret the writable page, if any. ret is unchanged if
2110 * mp wasn't spilled.
2113 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2115 MDB_env *env = txn->mt_env;
2118 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2120 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2121 if (!tx2->mt_spill_pgs)
2123 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2124 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2127 if (txn->mt_dirty_room == 0)
2128 return MDB_TXN_FULL;
2129 if (IS_OVERFLOW(mp))
2133 if (env->me_flags & MDB_WRITEMAP) {
2136 np = mdb_page_malloc(txn, num);
2140 memcpy(np, mp, num * env->me_psize);
2142 mdb_page_copy(np, mp, env->me_psize);
2145 /* If in current txn, this page is no longer spilled.
2146 * If it happens to be the last page, truncate the spill list.
2147 * Otherwise mark it as deleted by setting the LSB.
2149 if (x == txn->mt_spill_pgs[0])
2150 txn->mt_spill_pgs[0]--;
2152 txn->mt_spill_pgs[x] |= 1;
2153 } /* otherwise, if belonging to a parent txn, the
2154 * page remains spilled until child commits
2157 mdb_page_dirty(txn, np);
2158 np->mp_flags |= P_DIRTY;
2166 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2167 * @param[in] mc cursor pointing to the page to be touched
2168 * @return 0 on success, non-zero on failure.
2171 mdb_page_touch(MDB_cursor *mc)
2173 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2174 MDB_txn *txn = mc->mc_txn;
2175 MDB_cursor *m2, *m3;
2179 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2180 if (txn->mt_flags & MDB_TXN_SPILLS) {
2182 rc = mdb_page_unspill(txn, mp, &np);
2188 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2189 (rc = mdb_page_alloc(mc, 1, &np)))
2192 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2193 mp->mp_pgno, pgno));
2194 mdb_cassert(mc, mp->mp_pgno != pgno);
2195 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2196 /* Update the parent page, if any, to point to the new page */
2198 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2199 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2200 SETPGNO(node, pgno);
2202 mc->mc_db->md_root = pgno;
2204 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2205 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2207 /* If txn has a parent, make sure the page is in our
2211 unsigned x = mdb_mid2l_search(dl, pgno);
2212 if (x <= dl[0].mid && dl[x].mid == pgno) {
2213 if (mp != dl[x].mptr) { /* bad cursor? */
2214 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2215 txn->mt_flags |= MDB_TXN_ERROR;
2216 return MDB_CORRUPTED;
2221 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2223 np = mdb_page_malloc(txn, 1);
2228 rc = mdb_mid2l_insert(dl, &mid);
2229 mdb_cassert(mc, rc == 0);
2234 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2236 np->mp_flags |= P_DIRTY;
2239 /* Adjust cursors pointing to mp */
2240 mc->mc_pg[mc->mc_top] = np;
2241 m2 = txn->mt_cursors[mc->mc_dbi];
2242 if (mc->mc_flags & C_SUB) {
2243 for (; m2; m2=m2->mc_next) {
2244 m3 = &m2->mc_xcursor->mx_cursor;
2245 if (m3->mc_snum < mc->mc_snum) continue;
2246 if (m3->mc_pg[mc->mc_top] == mp)
2247 m3->mc_pg[mc->mc_top] = np;
2250 for (; m2; m2=m2->mc_next) {
2251 if (m2->mc_snum < mc->mc_snum) continue;
2252 if (m2->mc_pg[mc->mc_top] == mp) {
2253 m2->mc_pg[mc->mc_top] = np;
2254 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2256 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2258 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2259 if (!(leaf->mn_flags & F_SUBDATA))
2260 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2268 txn->mt_flags |= MDB_TXN_ERROR;
2273 mdb_env_sync(MDB_env *env, int force)
2276 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2277 if (env->me_flags & MDB_WRITEMAP) {
2278 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2279 ? MS_ASYNC : MS_SYNC;
2280 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2283 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2287 if (MDB_FDATASYNC(env->me_fd))
2294 /** Back up parent txn's cursors, then grab the originals for tracking */
2296 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2298 MDB_cursor *mc, *bk;
2303 for (i = src->mt_numdbs; --i >= 0; ) {
2304 if ((mc = src->mt_cursors[i]) != NULL) {
2305 size = sizeof(MDB_cursor);
2307 size += sizeof(MDB_xcursor);
2308 for (; mc; mc = bk->mc_next) {
2314 mc->mc_db = &dst->mt_dbs[i];
2315 /* Kill pointers into src - and dst to reduce abuse: The
2316 * user may not use mc until dst ends. Otherwise we'd...
2318 mc->mc_txn = NULL; /* ...set this to dst */
2319 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2320 if ((mx = mc->mc_xcursor) != NULL) {
2321 *(MDB_xcursor *)(bk+1) = *mx;
2322 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2324 mc->mc_next = dst->mt_cursors[i];
2325 dst->mt_cursors[i] = mc;
2332 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2333 * @param[in] txn the transaction handle.
2334 * @param[in] merge true to keep changes to parent cursors, false to revert.
2335 * @return 0 on success, non-zero on failure.
2338 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2340 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2344 for (i = txn->mt_numdbs; --i >= 0; ) {
2345 for (mc = cursors[i]; mc; mc = next) {
2347 if ((bk = mc->mc_backup) != NULL) {
2349 /* Commit changes to parent txn */
2350 mc->mc_next = bk->mc_next;
2351 mc->mc_backup = bk->mc_backup;
2352 mc->mc_txn = bk->mc_txn;
2353 mc->mc_db = bk->mc_db;
2354 mc->mc_dbflag = bk->mc_dbflag;
2355 if ((mx = mc->mc_xcursor) != NULL)
2356 mx->mx_cursor.mc_txn = bk->mc_txn;
2358 /* Abort nested txn */
2360 if ((mx = mc->mc_xcursor) != NULL)
2361 *mx = *(MDB_xcursor *)(bk+1);
2365 /* Only malloced cursors are permanently tracked. */
2373 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2376 mdb_txn_reset0(MDB_txn *txn, const char *act);
2378 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2384 Pidset = F_SETLK, Pidcheck = F_GETLK
2388 /** Set or check a pid lock. Set returns 0 on success.
2389 * Check returns 0 if the process is certainly dead, nonzero if it may
2390 * be alive (the lock exists or an error happened so we do not know).
2392 * On Windows Pidset is a no-op, we merely check for the existence
2393 * of the process with the given pid. On POSIX we use a single byte
2394 * lock on the lockfile, set at an offset equal to the pid.
2397 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2399 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2402 if (op == Pidcheck) {
2403 h = OpenProcess(env->me_pidquery, FALSE, pid);
2404 /* No documented "no such process" code, but other program use this: */
2406 return ErrCode() != ERROR_INVALID_PARAMETER;
2407 /* A process exists until all handles to it close. Has it exited? */
2408 ret = WaitForSingleObject(h, 0) != 0;
2415 struct flock lock_info;
2416 memset(&lock_info, 0, sizeof(lock_info));
2417 lock_info.l_type = F_WRLCK;
2418 lock_info.l_whence = SEEK_SET;
2419 lock_info.l_start = pid;
2420 lock_info.l_len = 1;
2421 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2422 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2424 } else if ((rc = ErrCode()) == EINTR) {
2432 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2433 * @param[in] txn the transaction handle to initialize
2434 * @return 0 on success, non-zero on failure.
2437 mdb_txn_renew0(MDB_txn *txn)
2439 MDB_env *env = txn->mt_env;
2440 MDB_txninfo *ti = env->me_txns;
2444 int rc, new_notls = 0;
2447 txn->mt_numdbs = env->me_numdbs;
2448 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2450 if (txn->mt_flags & MDB_TXN_RDONLY) {
2452 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2453 txn->mt_txnid = meta->mm_txnid;
2454 txn->mt_u.reader = NULL;
2456 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2457 pthread_getspecific(env->me_txkey);
2459 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2460 return MDB_BAD_RSLOT;
2462 MDB_PID_T pid = env->me_pid;
2463 MDB_THR_T tid = pthread_self();
2465 if (!env->me_live_reader) {
2466 rc = mdb_reader_pid(env, Pidset, pid);
2469 env->me_live_reader = 1;
2473 nr = ti->mti_numreaders;
2474 for (i=0; i<nr; i++)
2475 if (ti->mti_readers[i].mr_pid == 0)
2477 if (i == env->me_maxreaders) {
2478 UNLOCK_MUTEX_R(env);
2479 return MDB_READERS_FULL;
2481 ti->mti_readers[i].mr_pid = pid;
2482 ti->mti_readers[i].mr_tid = tid;
2484 ti->mti_numreaders = ++nr;
2485 /* Save numreaders for un-mutexed mdb_env_close() */
2486 env->me_numreaders = nr;
2487 UNLOCK_MUTEX_R(env);
2489 r = &ti->mti_readers[i];
2490 new_notls = (env->me_flags & MDB_NOTLS);
2491 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2496 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2497 txn->mt_u.reader = r;
2498 meta = env->me_metas[txn->mt_txnid & 1];
2504 txn->mt_txnid = ti->mti_txnid;
2505 meta = env->me_metas[txn->mt_txnid & 1];
2507 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2508 txn->mt_txnid = meta->mm_txnid;
2512 if (txn->mt_txnid == mdb_debug_start)
2515 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2516 txn->mt_u.dirty_list = env->me_dirty_list;
2517 txn->mt_u.dirty_list[0].mid = 0;
2518 txn->mt_free_pgs = env->me_free_pgs;
2519 txn->mt_free_pgs[0] = 0;
2520 txn->mt_spill_pgs = NULL;
2522 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2525 /* Copy the DB info and flags */
2526 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2528 /* Moved to here to avoid a data race in read TXNs */
2529 txn->mt_next_pgno = meta->mm_last_pg+1;
2531 for (i=2; i<txn->mt_numdbs; i++) {
2532 x = env->me_dbflags[i];
2533 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2534 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2536 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2538 if (env->me_maxpg < txn->mt_next_pgno) {
2539 mdb_txn_reset0(txn, "renew0-mapfail");
2541 txn->mt_u.reader->mr_pid = 0;
2542 txn->mt_u.reader = NULL;
2544 return MDB_MAP_RESIZED;
2551 mdb_txn_renew(MDB_txn *txn)
2555 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2558 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2559 DPUTS("environment had fatal error, must shutdown!");
2563 rc = mdb_txn_renew0(txn);
2564 if (rc == MDB_SUCCESS) {
2565 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2566 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2567 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2573 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2577 int rc, size, tsize = sizeof(MDB_txn);
2579 if (env->me_flags & MDB_FATAL_ERROR) {
2580 DPUTS("environment had fatal error, must shutdown!");
2583 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2586 /* Nested transactions: Max 1 child, write txns only, no writemap */
2587 if (parent->mt_child ||
2588 (flags & MDB_RDONLY) ||
2589 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2590 (env->me_flags & MDB_WRITEMAP))
2592 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2594 tsize = sizeof(MDB_ntxn);
2596 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2597 if (!(flags & MDB_RDONLY)) {
2598 size += env->me_maxdbs * sizeof(MDB_cursor *);
2599 /* child txns use parent's dbiseqs */
2601 size += env->me_maxdbs * sizeof(unsigned int);
2604 if ((txn = calloc(1, size)) == NULL) {
2605 DPRINTF(("calloc: %s", strerror(errno)));
2608 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2609 if (flags & MDB_RDONLY) {
2610 txn->mt_flags |= MDB_TXN_RDONLY;
2611 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2612 txn->mt_dbiseqs = env->me_dbiseqs;
2614 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2616 txn->mt_dbiseqs = parent->mt_dbiseqs;
2617 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2619 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2620 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2627 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2628 if (!txn->mt_u.dirty_list ||
2629 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2631 free(txn->mt_u.dirty_list);
2635 txn->mt_txnid = parent->mt_txnid;
2636 txn->mt_dirty_room = parent->mt_dirty_room;
2637 txn->mt_u.dirty_list[0].mid = 0;
2638 txn->mt_spill_pgs = NULL;
2639 txn->mt_next_pgno = parent->mt_next_pgno;
2640 parent->mt_child = txn;
2641 txn->mt_parent = parent;
2642 txn->mt_numdbs = parent->mt_numdbs;
2643 txn->mt_flags = parent->mt_flags;
2644 txn->mt_dbxs = parent->mt_dbxs;
2645 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2646 /* Copy parent's mt_dbflags, but clear DB_NEW */
2647 for (i=0; i<txn->mt_numdbs; i++)
2648 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2650 ntxn = (MDB_ntxn *)txn;
2651 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2652 if (env->me_pghead) {
2653 size = MDB_IDL_SIZEOF(env->me_pghead);
2654 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2656 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2661 rc = mdb_cursor_shadow(parent, txn);
2663 mdb_txn_reset0(txn, "beginchild-fail");
2665 rc = mdb_txn_renew0(txn);
2671 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2672 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2673 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2680 mdb_txn_env(MDB_txn *txn)
2682 if(!txn) return NULL;
2686 /** Export or close DBI handles opened in this txn. */
2688 mdb_dbis_update(MDB_txn *txn, int keep)
2691 MDB_dbi n = txn->mt_numdbs;
2692 MDB_env *env = txn->mt_env;
2693 unsigned char *tdbflags = txn->mt_dbflags;
2695 for (i = n; --i >= 2;) {
2696 if (tdbflags[i] & DB_NEW) {
2698 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2700 char *ptr = env->me_dbxs[i].md_name.mv_data;
2702 env->me_dbxs[i].md_name.mv_data = NULL;
2703 env->me_dbxs[i].md_name.mv_size = 0;
2704 env->me_dbflags[i] = 0;
2705 env->me_dbiseqs[i]++;
2711 if (keep && env->me_numdbs < n)
2715 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2716 * May be called twice for readonly txns: First reset it, then abort.
2717 * @param[in] txn the transaction handle to reset
2718 * @param[in] act why the transaction is being reset
2721 mdb_txn_reset0(MDB_txn *txn, const char *act)
2723 MDB_env *env = txn->mt_env;
2725 /* Close any DBI handles opened in this txn */
2726 mdb_dbis_update(txn, 0);
2728 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2729 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2730 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2732 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2733 if (txn->mt_u.reader) {
2734 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2735 if (!(env->me_flags & MDB_NOTLS))
2736 txn->mt_u.reader = NULL; /* txn does not own reader */
2738 txn->mt_numdbs = 0; /* close nothing if called again */
2739 txn->mt_dbxs = NULL; /* mark txn as reset */
2741 mdb_cursors_close(txn, 0);
2743 if (!(env->me_flags & MDB_WRITEMAP)) {
2744 mdb_dlist_free(txn);
2746 mdb_midl_free(env->me_pghead);
2748 if (txn->mt_parent) {
2749 txn->mt_parent->mt_child = NULL;
2750 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2751 mdb_midl_free(txn->mt_free_pgs);
2752 mdb_midl_free(txn->mt_spill_pgs);
2753 free(txn->mt_u.dirty_list);
2757 if (mdb_midl_shrink(&txn->mt_free_pgs))
2758 env->me_free_pgs = txn->mt_free_pgs;
2759 env->me_pghead = NULL;
2763 /* The writer mutex was locked in mdb_txn_begin. */
2765 UNLOCK_MUTEX_W(env);
2770 mdb_txn_reset(MDB_txn *txn)
2775 /* This call is only valid for read-only txns */
2776 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2779 mdb_txn_reset0(txn, "reset");
2783 mdb_txn_abort(MDB_txn *txn)
2789 mdb_txn_abort(txn->mt_child);
2791 mdb_txn_reset0(txn, "abort");
2792 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2793 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2794 txn->mt_u.reader->mr_pid = 0;
2799 /** Save the freelist as of this transaction to the freeDB.
2800 * This changes the freelist. Keep trying until it stabilizes.
2803 mdb_freelist_save(MDB_txn *txn)
2805 /* env->me_pghead[] can grow and shrink during this call.
2806 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2807 * Page numbers cannot disappear from txn->mt_free_pgs[].
2810 MDB_env *env = txn->mt_env;
2811 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2812 txnid_t pglast = 0, head_id = 0;
2813 pgno_t freecnt = 0, *free_pgs, *mop;
2814 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2816 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2818 if (env->me_pghead) {
2819 /* Make sure first page of freeDB is touched and on freelist */
2820 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2821 if (rc && rc != MDB_NOTFOUND)
2825 /* Dispose of loose pages. Usually they will have all
2826 * been used up by the time we get here.
2828 if (txn->mt_loose_pgs) {
2829 MDB_page *mp = txn->mt_loose_pgs;
2830 /* Just return them to freeDB */
2831 if (env->me_pghead) {
2833 mop = env->me_pghead;
2834 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2835 pgno_t pg = mp->mp_pgno;
2837 for (i = mop[0]; i && mop[i] < pg; i--)
2843 /* Oh well, they were wasted. Put on freelist */
2844 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2845 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
2848 txn->mt_loose_pgs = NULL;
2851 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2852 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2853 ? SSIZE_MAX : maxfree_1pg;
2856 /* Come back here after each Put() in case freelist changed */
2861 /* If using records from freeDB which we have not yet
2862 * deleted, delete them and any we reserved for me_pghead.
2864 while (pglast < env->me_pglast) {
2865 rc = mdb_cursor_first(&mc, &key, NULL);
2868 pglast = head_id = *(txnid_t *)key.mv_data;
2869 total_room = head_room = 0;
2870 mdb_tassert(txn, pglast <= env->me_pglast);
2871 rc = mdb_cursor_del(&mc, 0);
2876 /* Save the IDL of pages freed by this txn, to a single record */
2877 if (freecnt < txn->mt_free_pgs[0]) {
2879 /* Make sure last page of freeDB is touched and on freelist */
2880 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2881 if (rc && rc != MDB_NOTFOUND)
2884 free_pgs = txn->mt_free_pgs;
2885 /* Write to last page of freeDB */
2886 key.mv_size = sizeof(txn->mt_txnid);
2887 key.mv_data = &txn->mt_txnid;
2889 freecnt = free_pgs[0];
2890 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2891 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2894 /* Retry if mt_free_pgs[] grew during the Put() */
2895 free_pgs = txn->mt_free_pgs;
2896 } while (freecnt < free_pgs[0]);
2897 mdb_midl_sort(free_pgs);
2898 memcpy(data.mv_data, free_pgs, data.mv_size);
2901 unsigned int i = free_pgs[0];
2902 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2903 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2905 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2911 mop = env->me_pghead;
2912 mop_len = mop ? mop[0] : 0;
2914 /* Reserve records for me_pghead[]. Split it if multi-page,
2915 * to avoid searching freeDB for a page range. Use keys in
2916 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2918 if (total_room >= mop_len) {
2919 if (total_room == mop_len || --more < 0)
2921 } else if (head_room >= maxfree_1pg && head_id > 1) {
2922 /* Keep current record (overflow page), add a new one */
2926 /* (Re)write {key = head_id, IDL length = head_room} */
2927 total_room -= head_room;
2928 head_room = mop_len - total_room;
2929 if (head_room > maxfree_1pg && head_id > 1) {
2930 /* Overflow multi-page for part of me_pghead */
2931 head_room /= head_id; /* amortize page sizes */
2932 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2933 } else if (head_room < 0) {
2934 /* Rare case, not bothering to delete this record */
2937 key.mv_size = sizeof(head_id);
2938 key.mv_data = &head_id;
2939 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2940 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2943 /* IDL is initially empty, zero out at least the length */
2944 pgs = (pgno_t *)data.mv_data;
2945 j = head_room > clean_limit ? head_room : 0;
2949 total_room += head_room;
2952 /* Fill in the reserved me_pghead records */
2958 rc = mdb_cursor_first(&mc, &key, &data);
2959 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2960 txnid_t id = *(txnid_t *)key.mv_data;
2961 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2964 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2966 if (len > mop_len) {
2968 data.mv_size = (len + 1) * sizeof(MDB_ID);
2970 data.mv_data = mop -= len;
2973 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2975 if (rc || !(mop_len -= len))
2982 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2983 * @param[in] txn the transaction that's being committed
2984 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2985 * @return 0 on success, non-zero on failure.
2988 mdb_page_flush(MDB_txn *txn, int keep)
2990 MDB_env *env = txn->mt_env;
2991 MDB_ID2L dl = txn->mt_u.dirty_list;
2992 unsigned psize = env->me_psize, j;
2993 int i, pagecount = dl[0].mid, rc;
2994 size_t size = 0, pos = 0;
2996 MDB_page *dp = NULL;
3000 struct iovec iov[MDB_COMMIT_PAGES];
3001 ssize_t wpos = 0, wsize = 0, wres;
3002 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3008 if (env->me_flags & MDB_WRITEMAP) {
3009 /* Clear dirty flags */
3010 while (++i <= pagecount) {
3012 /* Don't flush this page yet */
3013 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3014 dp->mp_flags &= ~P_KEEP;
3018 dp->mp_flags &= ~P_DIRTY;
3023 /* Write the pages */
3025 if (++i <= pagecount) {
3027 /* Don't flush this page yet */
3028 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3029 dp->mp_flags &= ~P_KEEP;
3034 /* clear dirty flag */
3035 dp->mp_flags &= ~P_DIRTY;
3038 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3043 /* Windows actually supports scatter/gather I/O, but only on
3044 * unbuffered file handles. Since we're relying on the OS page
3045 * cache for all our data, that's self-defeating. So we just
3046 * write pages one at a time. We use the ov structure to set
3047 * the write offset, to at least save the overhead of a Seek
3050 DPRINTF(("committing page %"Z"u", pgno));
3051 memset(&ov, 0, sizeof(ov));
3052 ov.Offset = pos & 0xffffffff;
3053 ov.OffsetHigh = pos >> 16 >> 16;
3054 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3056 DPRINTF(("WriteFile: %d", rc));
3060 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3061 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3063 /* Write previous page(s) */
3064 #ifdef MDB_USE_PWRITEV
3065 wres = pwritev(env->me_fd, iov, n, wpos);
3068 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3070 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3072 DPRINTF(("lseek: %s", strerror(rc)));
3075 wres = writev(env->me_fd, iov, n);
3078 if (wres != wsize) {
3081 DPRINTF(("Write error: %s", strerror(rc)));
3083 rc = EIO; /* TODO: Use which error code? */
3084 DPUTS("short write, filesystem full?");
3095 DPRINTF(("committing page %"Z"u", pgno));
3096 next_pos = pos + size;
3097 iov[n].iov_len = size;
3098 iov[n].iov_base = (char *)dp;
3104 for (i = keep; ++i <= pagecount; ) {
3106 /* This is a page we skipped above */
3109 dl[j].mid = dp->mp_pgno;
3112 mdb_dpage_free(env, dp);
3117 txn->mt_dirty_room += i - j;
3123 mdb_txn_commit(MDB_txn *txn)
3129 if (txn == NULL || txn->mt_env == NULL)
3132 if (txn->mt_child) {
3133 rc = mdb_txn_commit(txn->mt_child);
3134 txn->mt_child = NULL;
3141 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3142 mdb_dbis_update(txn, 1);
3143 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3148 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3149 DPUTS("error flag is set, can't commit");
3151 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3156 if (txn->mt_parent) {
3157 MDB_txn *parent = txn->mt_parent;
3161 unsigned x, y, len, ps_len;
3163 /* Append our free list to parent's */
3164 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3167 mdb_midl_free(txn->mt_free_pgs);
3168 /* Failures after this must either undo the changes
3169 * to the parent or set MDB_TXN_ERROR in the parent.
3172 parent->mt_next_pgno = txn->mt_next_pgno;
3173 parent->mt_flags = txn->mt_flags;
3175 /* Merge our cursors into parent's and close them */
3176 mdb_cursors_close(txn, 1);
3178 /* Update parent's DB table. */
3179 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3180 parent->mt_numdbs = txn->mt_numdbs;
3181 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3182 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3183 for (i=2; i<txn->mt_numdbs; i++) {
3184 /* preserve parent's DB_NEW status */
3185 x = parent->mt_dbflags[i] & DB_NEW;
3186 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3189 dst = parent->mt_u.dirty_list;
3190 src = txn->mt_u.dirty_list;
3191 /* Remove anything in our dirty list from parent's spill list */
3192 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3194 pspill[0] = (pgno_t)-1;
3195 /* Mark our dirty pages as deleted in parent spill list */
3196 for (i=0, len=src[0].mid; ++i <= len; ) {
3197 MDB_ID pn = src[i].mid << 1;
3198 while (pn > pspill[x])
3200 if (pn == pspill[x]) {
3205 /* Squash deleted pagenums if we deleted any */
3206 for (x=y; ++x <= ps_len; )
3207 if (!(pspill[x] & 1))
3208 pspill[++y] = pspill[x];
3212 /* Find len = length of merging our dirty list with parent's */
3214 dst[0].mid = 0; /* simplify loops */
3215 if (parent->mt_parent) {
3216 len = x + src[0].mid;
3217 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3218 for (i = x; y && i; y--) {
3219 pgno_t yp = src[y].mid;
3220 while (yp < dst[i].mid)
3222 if (yp == dst[i].mid) {
3227 } else { /* Simplify the above for single-ancestor case */
3228 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3230 /* Merge our dirty list with parent's */
3232 for (i = len; y; dst[i--] = src[y--]) {
3233 pgno_t yp = src[y].mid;
3234 while (yp < dst[x].mid)
3235 dst[i--] = dst[x--];
3236 if (yp == dst[x].mid)
3237 free(dst[x--].mptr);
3239 mdb_tassert(txn, i == x);
3241 free(txn->mt_u.dirty_list);
3242 parent->mt_dirty_room = txn->mt_dirty_room;
3243 if (txn->mt_spill_pgs) {
3244 if (parent->mt_spill_pgs) {
3245 /* TODO: Prevent failure here, so parent does not fail */
3246 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3248 parent->mt_flags |= MDB_TXN_ERROR;
3249 mdb_midl_free(txn->mt_spill_pgs);
3250 mdb_midl_sort(parent->mt_spill_pgs);
3252 parent->mt_spill_pgs = txn->mt_spill_pgs;
3256 /* Append our loose page list to parent's */
3257 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3259 *lp = txn->mt_loose_pgs;
3261 parent->mt_child = NULL;
3262 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3267 if (txn != env->me_txn) {
3268 DPUTS("attempt to commit unknown transaction");
3273 mdb_cursors_close(txn, 0);
3275 if (!txn->mt_u.dirty_list[0].mid &&
3276 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3279 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3280 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3282 /* Update DB root pointers */
3283 if (txn->mt_numdbs > 2) {
3287 data.mv_size = sizeof(MDB_db);
3289 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3290 for (i = 2; i < txn->mt_numdbs; i++) {
3291 if (txn->mt_dbflags[i] & DB_DIRTY) {
3292 if (TXN_DBI_CHANGED(txn, i)) {
3296 data.mv_data = &txn->mt_dbs[i];
3297 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3304 rc = mdb_freelist_save(txn);
3308 mdb_midl_free(env->me_pghead);
3309 env->me_pghead = NULL;
3310 if (mdb_midl_shrink(&txn->mt_free_pgs))
3311 env->me_free_pgs = txn->mt_free_pgs;
3317 if ((rc = mdb_page_flush(txn, 0)) ||
3318 (rc = mdb_env_sync(env, 0)) ||
3319 (rc = mdb_env_write_meta(txn)))
3325 mdb_dbis_update(txn, 1);
3328 UNLOCK_MUTEX_W(env);
3338 /** Read the environment parameters of a DB environment before
3339 * mapping it into memory.
3340 * @param[in] env the environment handle
3341 * @param[out] meta address of where to store the meta information
3342 * @return 0 on success, non-zero on failure.
3345 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3351 enum { Size = sizeof(pbuf) };
3353 /* We don't know the page size yet, so use a minimum value.
3354 * Read both meta pages so we can use the latest one.
3357 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3361 memset(&ov, 0, sizeof(ov));
3363 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3364 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3367 rc = pread(env->me_fd, &pbuf, Size, off);
3370 if (rc == 0 && off == 0)
3372 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3373 DPRINTF(("read: %s", mdb_strerror(rc)));
3377 p = (MDB_page *)&pbuf;
3379 if (!F_ISSET(p->mp_flags, P_META)) {
3380 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3385 if (m->mm_magic != MDB_MAGIC) {
3386 DPUTS("meta has invalid magic");
3390 if (m->mm_version != MDB_DATA_VERSION) {
3391 DPRINTF(("database is version %u, expected version %u",
3392 m->mm_version, MDB_DATA_VERSION));
3393 return MDB_VERSION_MISMATCH;
3396 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3403 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3405 meta->mm_magic = MDB_MAGIC;
3406 meta->mm_version = MDB_DATA_VERSION;
3407 meta->mm_mapsize = env->me_mapsize;
3408 meta->mm_psize = env->me_psize;
3409 meta->mm_last_pg = 1;
3410 meta->mm_flags = env->me_flags & 0xffff;
3411 meta->mm_flags |= MDB_INTEGERKEY;
3412 meta->mm_dbs[0].md_root = P_INVALID;
3413 meta->mm_dbs[1].md_root = P_INVALID;
3416 /** Write the environment parameters of a freshly created DB environment.
3417 * @param[in] env the environment handle
3418 * @param[out] meta address of where to store the meta information
3419 * @return 0 on success, non-zero on failure.
3422 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3430 memset(&ov, 0, sizeof(ov));
3431 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3433 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3436 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3437 len = pwrite(fd, ptr, size, pos); \
3438 rc = (len >= 0); } while(0)
3441 DPUTS("writing new meta page");
3443 psize = env->me_psize;
3445 mdb_env_init_meta0(env, meta);
3447 p = calloc(2, psize);
3449 p->mp_flags = P_META;
3450 *(MDB_meta *)METADATA(p) = *meta;
3452 q = (MDB_page *)((char *)p + psize);
3454 q->mp_flags = P_META;
3455 *(MDB_meta *)METADATA(q) = *meta;
3457 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3460 else if ((unsigned) len == psize * 2)
3468 /** Update the environment info to commit a transaction.
3469 * @param[in] txn the transaction that's being committed
3470 * @return 0 on success, non-zero on failure.
3473 mdb_env_write_meta(MDB_txn *txn)
3476 MDB_meta meta, metab, *mp;
3478 int rc, len, toggle;
3487 toggle = txn->mt_txnid & 1;
3488 DPRINTF(("writing meta page %d for root page %"Z"u",
3489 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3492 mp = env->me_metas[toggle];
3494 if (env->me_flags & MDB_WRITEMAP) {
3495 /* Persist any increases of mapsize config */
3496 if (env->me_mapsize > mp->mm_mapsize)
3497 mp->mm_mapsize = env->me_mapsize;
3498 mp->mm_dbs[0] = txn->mt_dbs[0];
3499 mp->mm_dbs[1] = txn->mt_dbs[1];
3500 mp->mm_last_pg = txn->mt_next_pgno - 1;
3501 mp->mm_txnid = txn->mt_txnid;
3502 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3503 unsigned meta_size = env->me_psize;
3504 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3507 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3508 if (meta_size < env->me_os_psize)
3509 meta_size += meta_size;
3514 if (MDB_MSYNC(ptr, meta_size, rc)) {
3521 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3522 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3524 ptr = (char *)&meta;
3525 if (env->me_mapsize > mp->mm_mapsize) {
3526 /* Persist any increases of mapsize config */
3527 meta.mm_mapsize = env->me_mapsize;
3528 off = offsetof(MDB_meta, mm_mapsize);
3530 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3532 len = sizeof(MDB_meta) - off;
3535 meta.mm_dbs[0] = txn->mt_dbs[0];
3536 meta.mm_dbs[1] = txn->mt_dbs[1];
3537 meta.mm_last_pg = txn->mt_next_pgno - 1;
3538 meta.mm_txnid = txn->mt_txnid;
3541 off += env->me_psize;
3544 /* Write to the SYNC fd */
3545 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3546 env->me_fd : env->me_mfd;
3549 memset(&ov, 0, sizeof(ov));
3551 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3555 rc = pwrite(mfd, ptr, len, off);
3558 rc = rc < 0 ? ErrCode() : EIO;
3559 DPUTS("write failed, disk error?");
3560 /* On a failure, the pagecache still contains the new data.
3561 * Write some old data back, to prevent it from being used.
3562 * Use the non-SYNC fd; we know it will fail anyway.
3564 meta.mm_last_pg = metab.mm_last_pg;
3565 meta.mm_txnid = metab.mm_txnid;
3567 memset(&ov, 0, sizeof(ov));
3569 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3571 r2 = pwrite(env->me_fd, ptr, len, off);
3572 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3575 env->me_flags |= MDB_FATAL_ERROR;
3579 /* Memory ordering issues are irrelevant; since the entire writer
3580 * is wrapped by wmutex, all of these changes will become visible
3581 * after the wmutex is unlocked. Since the DB is multi-version,
3582 * readers will get consistent data regardless of how fresh or
3583 * how stale their view of these values is.
3586 env->me_txns->mti_txnid = txn->mt_txnid;
3591 /** Check both meta pages to see which one is newer.
3592 * @param[in] env the environment handle
3593 * @return meta toggle (0 or 1).
3596 mdb_env_pick_meta(const MDB_env *env)
3598 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3602 mdb_env_create(MDB_env **env)
3606 e = calloc(1, sizeof(MDB_env));
3610 e->me_maxreaders = DEFAULT_READERS;
3611 e->me_maxdbs = e->me_numdbs = 2;
3612 e->me_fd = INVALID_HANDLE_VALUE;
3613 e->me_lfd = INVALID_HANDLE_VALUE;
3614 e->me_mfd = INVALID_HANDLE_VALUE;
3615 #ifdef MDB_USE_POSIX_SEM
3616 e->me_rmutex = SEM_FAILED;
3617 e->me_wmutex = SEM_FAILED;
3619 e->me_pid = getpid();
3620 GET_PAGESIZE(e->me_os_psize);
3621 VGMEMP_CREATE(e,0,0);
3627 mdb_env_map(MDB_env *env, void *addr)
3630 unsigned int flags = env->me_flags;
3634 LONG sizelo, sizehi;
3637 if (flags & MDB_RDONLY) {
3638 /* Don't set explicit map size, use whatever exists */
3643 msize = env->me_mapsize;
3644 sizelo = msize & 0xffffffff;
3645 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3647 /* Windows won't create mappings for zero length files.
3648 * and won't map more than the file size.
3649 * Just set the maxsize right now.
3651 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3652 || !SetEndOfFile(env->me_fd)
3653 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3657 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3658 PAGE_READWRITE : PAGE_READONLY,
3659 sizehi, sizelo, NULL);
3662 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3663 FILE_MAP_WRITE : FILE_MAP_READ,
3665 rc = env->me_map ? 0 : ErrCode();
3670 int prot = PROT_READ;
3671 if (flags & MDB_WRITEMAP) {
3673 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3676 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3678 if (env->me_map == MAP_FAILED) {
3683 if (flags & MDB_NORDAHEAD) {
3684 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3686 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3688 #ifdef POSIX_MADV_RANDOM
3689 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3690 #endif /* POSIX_MADV_RANDOM */
3691 #endif /* MADV_RANDOM */
3695 /* Can happen because the address argument to mmap() is just a
3696 * hint. mmap() can pick another, e.g. if the range is in use.
3697 * The MAP_FIXED flag would prevent that, but then mmap could
3698 * instead unmap existing pages to make room for the new map.
3700 if (addr && env->me_map != addr)
3701 return EBUSY; /* TODO: Make a new MDB_* error code? */
3703 p = (MDB_page *)env->me_map;
3704 env->me_metas[0] = METADATA(p);
3705 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3711 mdb_env_set_mapsize(MDB_env *env, size_t size)
3713 /* If env is already open, caller is responsible for making
3714 * sure there are no active txns.
3722 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3723 else if (size < env->me_mapsize) {
3724 /* If the configured size is smaller, make sure it's
3725 * still big enough. Silently round up to minimum if not.
3727 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3731 munmap(env->me_map, env->me_mapsize);
3732 env->me_mapsize = size;
3733 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3734 rc = mdb_env_map(env, old);
3738 env->me_mapsize = size;
3740 env->me_maxpg = env->me_mapsize / env->me_psize;
3745 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3749 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3754 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3756 if (env->me_map || readers < 1)
3758 env->me_maxreaders = readers;
3763 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3765 if (!env || !readers)
3767 *readers = env->me_maxreaders;
3771 /** Further setup required for opening an LMDB environment
3774 mdb_env_open2(MDB_env *env)
3776 unsigned int flags = env->me_flags;
3777 int i, newenv = 0, rc;
3781 /* See if we should use QueryLimited */
3783 if ((rc & 0xff) > 5)
3784 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3786 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3789 memset(&meta, 0, sizeof(meta));
3791 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3794 DPUTS("new mdbenv");
3796 env->me_psize = env->me_os_psize;
3797 if (env->me_psize > MAX_PAGESIZE)
3798 env->me_psize = MAX_PAGESIZE;
3800 env->me_psize = meta.mm_psize;
3803 /* Was a mapsize configured? */
3804 if (!env->me_mapsize) {
3805 /* If this is a new environment, take the default,
3806 * else use the size recorded in the existing env.
3808 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3809 } else if (env->me_mapsize < meta.mm_mapsize) {
3810 /* If the configured size is smaller, make sure it's
3811 * still big enough. Silently round up to minimum if not.
3813 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3814 if (env->me_mapsize < minsize)
3815 env->me_mapsize = minsize;
3818 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3823 if (flags & MDB_FIXEDMAP)
3824 meta.mm_address = env->me_map;
3825 i = mdb_env_init_meta(env, &meta);
3826 if (i != MDB_SUCCESS) {
3831 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3832 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3834 #if !(MDB_MAXKEYSIZE)
3835 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3837 env->me_maxpg = env->me_mapsize / env->me_psize;
3841 int toggle = mdb_env_pick_meta(env);
3842 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3844 DPRINTF(("opened database version %u, pagesize %u",
3845 env->me_metas[0]->mm_version, env->me_psize));
3846 DPRINTF(("using meta page %d", toggle));
3847 DPRINTF(("depth: %u", db->md_depth));
3848 DPRINTF(("entries: %"Z"u", db->md_entries));
3849 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3850 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3851 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3852 DPRINTF(("root: %"Z"u", db->md_root));
3860 /** Release a reader thread's slot in the reader lock table.
3861 * This function is called automatically when a thread exits.
3862 * @param[in] ptr This points to the slot in the reader lock table.
3865 mdb_env_reader_dest(void *ptr)
3867 MDB_reader *reader = ptr;
3873 /** Junk for arranging thread-specific callbacks on Windows. This is
3874 * necessarily platform and compiler-specific. Windows supports up
3875 * to 1088 keys. Let's assume nobody opens more than 64 environments
3876 * in a single process, for now. They can override this if needed.
3878 #ifndef MAX_TLS_KEYS
3879 #define MAX_TLS_KEYS 64
3881 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3882 static int mdb_tls_nkeys;
3884 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3888 case DLL_PROCESS_ATTACH: break;
3889 case DLL_THREAD_ATTACH: break;
3890 case DLL_THREAD_DETACH:
3891 for (i=0; i<mdb_tls_nkeys; i++) {
3892 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3894 mdb_env_reader_dest(r);
3898 case DLL_PROCESS_DETACH: break;
3903 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3905 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3909 /* Force some symbol references.
3910 * _tls_used forces the linker to create the TLS directory if not already done
3911 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3913 #pragma comment(linker, "/INCLUDE:_tls_used")
3914 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3915 #pragma const_seg(".CRT$XLB")
3916 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3917 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3920 #pragma comment(linker, "/INCLUDE:__tls_used")
3921 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3922 #pragma data_seg(".CRT$XLB")
3923 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3925 #endif /* WIN 32/64 */
3926 #endif /* !__GNUC__ */
3929 /** Downgrade the exclusive lock on the region back to shared */
3931 mdb_env_share_locks(MDB_env *env, int *excl)
3933 int rc = 0, toggle = mdb_env_pick_meta(env);
3935 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3940 /* First acquire a shared lock. The Unlock will
3941 * then release the existing exclusive lock.
3943 memset(&ov, 0, sizeof(ov));
3944 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3947 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3953 struct flock lock_info;
3954 /* The shared lock replaces the existing lock */
3955 memset((void *)&lock_info, 0, sizeof(lock_info));
3956 lock_info.l_type = F_RDLCK;
3957 lock_info.l_whence = SEEK_SET;
3958 lock_info.l_start = 0;
3959 lock_info.l_len = 1;
3960 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3961 (rc = ErrCode()) == EINTR) ;
3962 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3969 /** Try to get exlusive lock, otherwise shared.
3970 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3973 mdb_env_excl_lock(MDB_env *env, int *excl)
3977 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3981 memset(&ov, 0, sizeof(ov));
3982 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3989 struct flock lock_info;
3990 memset((void *)&lock_info, 0, sizeof(lock_info));
3991 lock_info.l_type = F_WRLCK;
3992 lock_info.l_whence = SEEK_SET;
3993 lock_info.l_start = 0;
3994 lock_info.l_len = 1;
3995 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3996 (rc = ErrCode()) == EINTR) ;
4000 # ifdef MDB_USE_POSIX_SEM
4001 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4004 lock_info.l_type = F_RDLCK;
4005 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4006 (rc = ErrCode()) == EINTR) ;
4016 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4018 * @(#) $Revision: 5.1 $
4019 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4020 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4022 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4026 * Please do not copyright this code. This code is in the public domain.
4028 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4029 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4030 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4031 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4032 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4033 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4034 * PERFORMANCE OF THIS SOFTWARE.
4037 * chongo <Landon Curt Noll> /\oo/\
4038 * http://www.isthe.com/chongo/
4040 * Share and Enjoy! :-)
4043 typedef unsigned long long mdb_hash_t;
4044 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4046 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4047 * @param[in] val value to hash
4048 * @param[in] hval initial value for hash
4049 * @return 64 bit hash
4051 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4052 * hval arg on the first call.
4055 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4057 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4058 unsigned char *end = s + val->mv_size;
4060 * FNV-1a hash each octet of the string
4063 /* xor the bottom with the current octet */
4064 hval ^= (mdb_hash_t)*s++;
4066 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4067 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4068 (hval << 7) + (hval << 8) + (hval << 40);
4070 /* return our new hash value */
4074 /** Hash the string and output the encoded hash.
4075 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4076 * very short name limits. We don't care about the encoding being reversible,
4077 * we just want to preserve as many bits of the input as possible in a
4078 * small printable string.
4079 * @param[in] str string to hash
4080 * @param[out] encbuf an array of 11 chars to hold the hash
4082 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4085 mdb_pack85(unsigned long l, char *out)
4089 for (i=0; i<5; i++) {
4090 *out++ = mdb_a85[l % 85];
4096 mdb_hash_enc(MDB_val *val, char *encbuf)
4098 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4100 mdb_pack85(h, encbuf);
4101 mdb_pack85(h>>32, encbuf+5);
4106 /** Open and/or initialize the lock region for the environment.
4107 * @param[in] env The LMDB environment.
4108 * @param[in] lpath The pathname of the file used for the lock region.
4109 * @param[in] mode The Unix permissions for the file, if we create it.
4110 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4111 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4112 * @return 0 on success, non-zero on failure.
4115 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4118 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4120 # define MDB_ERRCODE_ROFS EROFS
4121 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4122 # define MDB_CLOEXEC O_CLOEXEC
4125 # define MDB_CLOEXEC 0
4132 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4133 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4134 FILE_ATTRIBUTE_NORMAL, NULL);
4136 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4138 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4140 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4145 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4146 /* Lose record locks when exec*() */
4147 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4148 fcntl(env->me_lfd, F_SETFD, fdflags);
4151 if (!(env->me_flags & MDB_NOTLS)) {
4152 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4155 env->me_flags |= MDB_ENV_TXKEY;
4157 /* Windows TLS callbacks need help finding their TLS info. */
4158 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4162 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4166 /* Try to get exclusive lock. If we succeed, then
4167 * nobody is using the lock region and we should initialize it.
4169 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4172 size = GetFileSize(env->me_lfd, NULL);
4174 size = lseek(env->me_lfd, 0, SEEK_END);
4175 if (size == -1) goto fail_errno;
4177 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4178 if (size < rsize && *excl > 0) {
4180 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4181 || !SetEndOfFile(env->me_lfd))
4184 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4188 size = rsize - sizeof(MDB_txninfo);
4189 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4194 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4196 if (!mh) goto fail_errno;
4197 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4199 if (!env->me_txns) goto fail_errno;
4201 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4203 if (m == MAP_FAILED) goto fail_errno;
4209 BY_HANDLE_FILE_INFORMATION stbuf;
4218 if (!mdb_sec_inited) {
4219 InitializeSecurityDescriptor(&mdb_null_sd,
4220 SECURITY_DESCRIPTOR_REVISION);
4221 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4222 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4223 mdb_all_sa.bInheritHandle = FALSE;
4224 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4227 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4228 idbuf.volume = stbuf.dwVolumeSerialNumber;
4229 idbuf.nhigh = stbuf.nFileIndexHigh;
4230 idbuf.nlow = stbuf.nFileIndexLow;
4231 val.mv_data = &idbuf;
4232 val.mv_size = sizeof(idbuf);
4233 mdb_hash_enc(&val, encbuf);
4234 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4235 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4236 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4237 if (!env->me_rmutex) goto fail_errno;
4238 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4239 if (!env->me_wmutex) goto fail_errno;
4240 #elif defined(MDB_USE_POSIX_SEM)
4249 #if defined(__NetBSD__)
4250 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4252 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4253 idbuf.dev = stbuf.st_dev;
4254 idbuf.ino = stbuf.st_ino;
4255 val.mv_data = &idbuf;
4256 val.mv_size = sizeof(idbuf);
4257 mdb_hash_enc(&val, encbuf);
4258 #ifdef MDB_SHORT_SEMNAMES
4259 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4261 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4262 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4263 /* Clean up after a previous run, if needed: Try to
4264 * remove both semaphores before doing anything else.
4266 sem_unlink(env->me_txns->mti_rmname);
4267 sem_unlink(env->me_txns->mti_wmname);
4268 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4269 O_CREAT|O_EXCL, mode, 1);
4270 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4271 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4272 O_CREAT|O_EXCL, mode, 1);
4273 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4274 #else /* MDB_USE_POSIX_SEM */
4275 pthread_mutexattr_t mattr;
4277 if ((rc = pthread_mutexattr_init(&mattr))
4278 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4279 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4280 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4282 pthread_mutexattr_destroy(&mattr);
4283 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4285 env->me_txns->mti_magic = MDB_MAGIC;
4286 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4287 env->me_txns->mti_txnid = 0;
4288 env->me_txns->mti_numreaders = 0;
4291 if (env->me_txns->mti_magic != MDB_MAGIC) {
4292 DPUTS("lock region has invalid magic");
4296 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4297 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4298 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4299 rc = MDB_VERSION_MISMATCH;
4303 if (rc && rc != EACCES && rc != EAGAIN) {
4307 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4308 if (!env->me_rmutex) goto fail_errno;
4309 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4310 if (!env->me_wmutex) goto fail_errno;
4311 #elif defined(MDB_USE_POSIX_SEM)
4312 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4313 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4314 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4315 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4326 /** The name of the lock file in the DB environment */
4327 #define LOCKNAME "/lock.mdb"
4328 /** The name of the data file in the DB environment */
4329 #define DATANAME "/data.mdb"
4330 /** The suffix of the lock file when no subdir is used */
4331 #define LOCKSUFF "-lock"
4332 /** Only a subset of the @ref mdb_env flags can be changed
4333 * at runtime. Changing other flags requires closing the
4334 * environment and re-opening it with the new flags.
4336 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4337 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4338 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4340 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4341 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4345 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4347 int oflags, rc, len, excl = -1;
4348 char *lpath, *dpath;
4350 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4354 if (flags & MDB_NOSUBDIR) {
4355 rc = len + sizeof(LOCKSUFF) + len + 1;
4357 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4362 if (flags & MDB_NOSUBDIR) {
4363 dpath = lpath + len + sizeof(LOCKSUFF);
4364 sprintf(lpath, "%s" LOCKSUFF, path);
4365 strcpy(dpath, path);
4367 dpath = lpath + len + sizeof(LOCKNAME);
4368 sprintf(lpath, "%s" LOCKNAME, path);
4369 sprintf(dpath, "%s" DATANAME, path);
4373 flags |= env->me_flags;
4374 if (flags & MDB_RDONLY) {
4375 /* silently ignore WRITEMAP when we're only getting read access */
4376 flags &= ~MDB_WRITEMAP;
4378 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4379 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4382 env->me_flags = flags |= MDB_ENV_ACTIVE;
4386 env->me_path = strdup(path);
4387 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4388 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4389 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4390 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4395 /* For RDONLY, get lockfile after we know datafile exists */
4396 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4397 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4403 if (F_ISSET(flags, MDB_RDONLY)) {
4404 oflags = GENERIC_READ;
4405 len = OPEN_EXISTING;
4407 oflags = GENERIC_READ|GENERIC_WRITE;
4410 mode = FILE_ATTRIBUTE_NORMAL;
4411 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4412 NULL, len, mode, NULL);
4414 if (F_ISSET(flags, MDB_RDONLY))
4417 oflags = O_RDWR | O_CREAT;
4419 env->me_fd = open(dpath, oflags, mode);
4421 if (env->me_fd == INVALID_HANDLE_VALUE) {
4426 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4427 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4432 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4433 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4434 env->me_mfd = env->me_fd;
4436 /* Synchronous fd for meta writes. Needed even with
4437 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4440 len = OPEN_EXISTING;
4441 env->me_mfd = CreateFile(dpath, oflags,
4442 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4443 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4446 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4448 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4453 DPRINTF(("opened dbenv %p", (void *) env));
4455 rc = mdb_env_share_locks(env, &excl);
4459 if (!((flags & MDB_RDONLY) ||
4460 (env->me_pbuf = calloc(1, env->me_psize))))
4466 mdb_env_close0(env, excl);
4472 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4474 mdb_env_close0(MDB_env *env, int excl)
4478 if (!(env->me_flags & MDB_ENV_ACTIVE))
4481 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4482 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4483 free(env->me_dbxs[i].md_name.mv_data);
4486 free(env->me_dbiseqs);
4487 free(env->me_dbflags);
4490 free(env->me_dirty_list);
4491 mdb_midl_free(env->me_free_pgs);
4493 if (env->me_flags & MDB_ENV_TXKEY) {
4494 pthread_key_delete(env->me_txkey);
4496 /* Delete our key from the global list */
4497 for (i=0; i<mdb_tls_nkeys; i++)
4498 if (mdb_tls_keys[i] == env->me_txkey) {
4499 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4507 munmap(env->me_map, env->me_mapsize);
4509 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4510 (void) close(env->me_mfd);
4511 if (env->me_fd != INVALID_HANDLE_VALUE)
4512 (void) close(env->me_fd);
4514 MDB_PID_T pid = env->me_pid;
4515 /* Clearing readers is done in this function because
4516 * me_txkey with its destructor must be disabled first.
4518 for (i = env->me_numreaders; --i >= 0; )
4519 if (env->me_txns->mti_readers[i].mr_pid == pid)
4520 env->me_txns->mti_readers[i].mr_pid = 0;
4522 if (env->me_rmutex) {
4523 CloseHandle(env->me_rmutex);
4524 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4526 /* Windows automatically destroys the mutexes when
4527 * the last handle closes.
4529 #elif defined(MDB_USE_POSIX_SEM)
4530 if (env->me_rmutex != SEM_FAILED) {
4531 sem_close(env->me_rmutex);
4532 if (env->me_wmutex != SEM_FAILED)
4533 sem_close(env->me_wmutex);
4534 /* If we have the filelock: If we are the
4535 * only remaining user, clean up semaphores.
4538 mdb_env_excl_lock(env, &excl);
4540 sem_unlink(env->me_txns->mti_rmname);
4541 sem_unlink(env->me_txns->mti_wmname);
4545 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4547 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4550 /* Unlock the lockfile. Windows would have unlocked it
4551 * after closing anyway, but not necessarily at once.
4553 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4556 (void) close(env->me_lfd);
4559 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4564 mdb_env_close(MDB_env *env)
4571 VGMEMP_DESTROY(env);
4572 while ((dp = env->me_dpages) != NULL) {
4573 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4574 env->me_dpages = dp->mp_next;
4578 mdb_env_close0(env, 0);
4582 /** Compare two items pointing at aligned size_t's */
4584 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4586 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4587 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4590 /** Compare two items pointing at aligned unsigned int's */
4592 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4594 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4595 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4598 /** Compare two items pointing at unsigned ints of unknown alignment.
4599 * Nodes and keys are guaranteed to be 2-byte aligned.
4602 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4604 #if BYTE_ORDER == LITTLE_ENDIAN
4605 unsigned short *u, *c;
4608 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4609 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4612 } while(!x && u > (unsigned short *)a->mv_data);
4615 unsigned short *u, *c, *end;
4618 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4619 u = (unsigned short *)a->mv_data;
4620 c = (unsigned short *)b->mv_data;
4623 } while(!x && u < end);
4628 /** Compare two items pointing at size_t's of unknown alignment. */
4629 #ifdef MISALIGNED_OK
4630 # define mdb_cmp_clong mdb_cmp_long
4632 # define mdb_cmp_clong mdb_cmp_cint
4635 /** Compare two items lexically */
4637 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4644 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4650 diff = memcmp(a->mv_data, b->mv_data, len);
4651 return diff ? diff : len_diff<0 ? -1 : len_diff;
4654 /** Compare two items in reverse byte order */
4656 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4658 const unsigned char *p1, *p2, *p1_lim;
4662 p1_lim = (const unsigned char *)a->mv_data;
4663 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4664 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4666 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4672 while (p1 > p1_lim) {
4673 diff = *--p1 - *--p2;
4677 return len_diff<0 ? -1 : len_diff;
4680 /** Search for key within a page, using binary search.
4681 * Returns the smallest entry larger or equal to the key.
4682 * If exactp is non-null, stores whether the found entry was an exact match
4683 * in *exactp (1 or 0).
4684 * Updates the cursor index with the index of the found entry.
4685 * If no entry larger or equal to the key is found, returns NULL.
4688 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4690 unsigned int i = 0, nkeys;
4693 MDB_page *mp = mc->mc_pg[mc->mc_top];
4694 MDB_node *node = NULL;
4699 nkeys = NUMKEYS(mp);
4701 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4702 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4705 low = IS_LEAF(mp) ? 0 : 1;
4707 cmp = mc->mc_dbx->md_cmp;
4709 /* Branch pages have no data, so if using integer keys,
4710 * alignment is guaranteed. Use faster mdb_cmp_int.
4712 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4713 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4720 nodekey.mv_size = mc->mc_db->md_pad;
4721 node = NODEPTR(mp, 0); /* fake */
4722 while (low <= high) {
4723 i = (low + high) >> 1;
4724 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4725 rc = cmp(key, &nodekey);
4726 DPRINTF(("found leaf index %u [%s], rc = %i",
4727 i, DKEY(&nodekey), rc));
4736 while (low <= high) {
4737 i = (low + high) >> 1;
4739 node = NODEPTR(mp, i);
4740 nodekey.mv_size = NODEKSZ(node);
4741 nodekey.mv_data = NODEKEY(node);
4743 rc = cmp(key, &nodekey);
4746 DPRINTF(("found leaf index %u [%s], rc = %i",
4747 i, DKEY(&nodekey), rc));
4749 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4750 i, DKEY(&nodekey), NODEPGNO(node), rc));
4761 if (rc > 0) { /* Found entry is less than the key. */
4762 i++; /* Skip to get the smallest entry larger than key. */
4764 node = NODEPTR(mp, i);
4767 *exactp = (rc == 0 && nkeys > 0);
4768 /* store the key index */
4769 mc->mc_ki[mc->mc_top] = i;
4771 /* There is no entry larger or equal to the key. */
4774 /* nodeptr is fake for LEAF2 */
4780 mdb_cursor_adjust(MDB_cursor *mc, func)
4784 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4785 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4792 /** Pop a page off the top of the cursor's stack. */
4794 mdb_cursor_pop(MDB_cursor *mc)
4798 MDB_page *top = mc->mc_pg[mc->mc_top];
4804 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4805 DDBI(mc), (void *) mc));
4809 /** Push a page onto the top of the cursor's stack. */
4811 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4813 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4814 DDBI(mc), (void *) mc));
4816 if (mc->mc_snum >= CURSOR_STACK) {
4817 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4818 return MDB_CURSOR_FULL;
4821 mc->mc_top = mc->mc_snum++;
4822 mc->mc_pg[mc->mc_top] = mp;
4823 mc->mc_ki[mc->mc_top] = 0;
4828 /** Find the address of the page corresponding to a given page number.
4829 * @param[in] txn the transaction for this access.
4830 * @param[in] pgno the page number for the page to retrieve.
4831 * @param[out] ret address of a pointer where the page's address will be stored.
4832 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4833 * @return 0 on success, non-zero on failure.
4836 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4838 MDB_env *env = txn->mt_env;
4842 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4846 MDB_ID2L dl = tx2->mt_u.dirty_list;
4848 /* Spilled pages were dirtied in this txn and flushed
4849 * because the dirty list got full. Bring this page
4850 * back in from the map (but don't unspill it here,
4851 * leave that unless page_touch happens again).
4853 if (tx2->mt_spill_pgs) {
4854 MDB_ID pn = pgno << 1;
4855 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4856 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4857 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4862 unsigned x = mdb_mid2l_search(dl, pgno);
4863 if (x <= dl[0].mid && dl[x].mid == pgno) {
4869 } while ((tx2 = tx2->mt_parent) != NULL);
4872 if (pgno < txn->mt_next_pgno) {
4874 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4876 DPRINTF(("page %"Z"u not found", pgno));
4877 txn->mt_flags |= MDB_TXN_ERROR;
4878 return MDB_PAGE_NOTFOUND;
4888 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4889 * The cursor is at the root page, set up the rest of it.
4892 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4894 MDB_page *mp = mc->mc_pg[mc->mc_top];
4898 while (IS_BRANCH(mp)) {
4902 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4903 mdb_cassert(mc, NUMKEYS(mp) > 1);
4904 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4906 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4908 if (flags & MDB_PS_LAST)
4909 i = NUMKEYS(mp) - 1;
4912 node = mdb_node_search(mc, key, &exact);
4914 i = NUMKEYS(mp) - 1;
4916 i = mc->mc_ki[mc->mc_top];
4918 mdb_cassert(mc, i > 0);
4922 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4925 mdb_cassert(mc, i < NUMKEYS(mp));
4926 node = NODEPTR(mp, i);
4928 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4931 mc->mc_ki[mc->mc_top] = i;
4932 if ((rc = mdb_cursor_push(mc, mp)))
4935 if (flags & MDB_PS_MODIFY) {
4936 if ((rc = mdb_page_touch(mc)) != 0)
4938 mp = mc->mc_pg[mc->mc_top];
4943 DPRINTF(("internal error, index points to a %02X page!?",
4945 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4946 return MDB_CORRUPTED;
4949 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4950 key ? DKEY(key) : "null"));
4951 mc->mc_flags |= C_INITIALIZED;
4952 mc->mc_flags &= ~C_EOF;
4957 /** Search for the lowest key under the current branch page.
4958 * This just bypasses a NUMKEYS check in the current page
4959 * before calling mdb_page_search_root(), because the callers
4960 * are all in situations where the current page is known to
4964 mdb_page_search_lowest(MDB_cursor *mc)
4966 MDB_page *mp = mc->mc_pg[mc->mc_top];
4967 MDB_node *node = NODEPTR(mp, 0);
4970 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4973 mc->mc_ki[mc->mc_top] = 0;
4974 if ((rc = mdb_cursor_push(mc, mp)))
4976 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4979 /** Search for the page a given key should be in.
4980 * Push it and its parent pages on the cursor stack.
4981 * @param[in,out] mc the cursor for this operation.
4982 * @param[in] key the key to search for, or NULL for first/last page.
4983 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4984 * are touched (updated with new page numbers).
4985 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4986 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4987 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4988 * @return 0 on success, non-zero on failure.
4991 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4996 /* Make sure the txn is still viable, then find the root from
4997 * the txn's db table and set it as the root of the cursor's stack.
4999 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5000 DPUTS("transaction has failed, must abort");
5003 /* Make sure we're using an up-to-date root */
5004 if (*mc->mc_dbflag & DB_STALE) {
5006 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5008 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5009 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5016 MDB_node *leaf = mdb_node_search(&mc2,
5017 &mc->mc_dbx->md_name, &exact);
5019 return MDB_NOTFOUND;
5020 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5023 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5025 /* The txn may not know this DBI, or another process may
5026 * have dropped and recreated the DB with other flags.
5028 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5029 return MDB_INCOMPATIBLE;
5030 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5032 *mc->mc_dbflag &= ~DB_STALE;
5034 root = mc->mc_db->md_root;
5036 if (root == P_INVALID) { /* Tree is empty. */
5037 DPUTS("tree is empty");
5038 return MDB_NOTFOUND;
5042 mdb_cassert(mc, root > 1);
5043 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5044 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5050 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5051 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5053 if (flags & MDB_PS_MODIFY) {
5054 if ((rc = mdb_page_touch(mc)))
5058 if (flags & MDB_PS_ROOTONLY)
5061 return mdb_page_search_root(mc, key, flags);
5065 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5067 MDB_txn *txn = mc->mc_txn;
5068 pgno_t pg = mp->mp_pgno;
5069 unsigned x = 0, ovpages = mp->mp_pages;
5070 MDB_env *env = txn->mt_env;
5071 MDB_IDL sl = txn->mt_spill_pgs;
5072 MDB_ID pn = pg << 1;
5075 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5076 /* If the page is dirty or on the spill list we just acquired it,
5077 * so we should give it back to our current free list, if any.
5078 * Otherwise put it onto the list of pages we freed in this txn.
5080 * Won't create me_pghead: me_pglast must be inited along with it.
5081 * Unsupported in nested txns: They would need to hide the page
5082 * range in ancestor txns' dirty and spilled lists.
5084 if (env->me_pghead &&
5086 ((mp->mp_flags & P_DIRTY) ||
5087 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5091 MDB_ID2 *dl, ix, iy;
5092 rc = mdb_midl_need(&env->me_pghead, ovpages);
5095 if (!(mp->mp_flags & P_DIRTY)) {
5096 /* This page is no longer spilled */
5103 /* Remove from dirty list */
5104 dl = txn->mt_u.dirty_list;
5106 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5112 mdb_cassert(mc, x > 1);
5114 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5115 txn->mt_flags |= MDB_TXN_ERROR;
5116 return MDB_CORRUPTED;
5119 if (!(env->me_flags & MDB_WRITEMAP))
5120 mdb_dpage_free(env, mp);
5122 /* Insert in me_pghead */
5123 mop = env->me_pghead;
5124 j = mop[0] + ovpages;
5125 for (i = mop[0]; i && mop[i] < pg; i--)
5131 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5135 mc->mc_db->md_overflow_pages -= ovpages;
5139 /** Return the data associated with a given node.
5140 * @param[in] txn The transaction for this operation.
5141 * @param[in] leaf The node being read.
5142 * @param[out] data Updated to point to the node's data.
5143 * @return 0 on success, non-zero on failure.
5146 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5148 MDB_page *omp; /* overflow page */
5152 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5153 data->mv_size = NODEDSZ(leaf);
5154 data->mv_data = NODEDATA(leaf);
5158 /* Read overflow data.
5160 data->mv_size = NODEDSZ(leaf);
5161 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5162 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5163 DPRINTF(("read overflow page %"Z"u failed", pgno));
5166 data->mv_data = METADATA(omp);
5172 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5173 MDB_val *key, MDB_val *data)
5180 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5182 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5185 if (txn->mt_flags & MDB_TXN_ERROR)
5188 mdb_cursor_init(&mc, txn, dbi, &mx);
5189 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5192 /** Find a sibling for a page.
5193 * Replaces the page at the top of the cursor's stack with the
5194 * specified sibling, if one exists.
5195 * @param[in] mc The cursor for this operation.
5196 * @param[in] move_right Non-zero if the right sibling is requested,
5197 * otherwise the left sibling.
5198 * @return 0 on success, non-zero on failure.
5201 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5207 if (mc->mc_snum < 2) {
5208 return MDB_NOTFOUND; /* root has no siblings */
5212 DPRINTF(("parent page is page %"Z"u, index %u",
5213 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5215 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5216 : (mc->mc_ki[mc->mc_top] == 0)) {
5217 DPRINTF(("no more keys left, moving to %s sibling",
5218 move_right ? "right" : "left"));
5219 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5220 /* undo cursor_pop before returning */
5227 mc->mc_ki[mc->mc_top]++;
5229 mc->mc_ki[mc->mc_top]--;
5230 DPRINTF(("just moving to %s index key %u",
5231 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5233 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5235 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5236 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5237 /* mc will be inconsistent if caller does mc_snum++ as above */
5238 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5242 mdb_cursor_push(mc, mp);
5244 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5249 /** Move the cursor to the next data item. */
5251 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5257 if (mc->mc_flags & C_EOF) {
5258 return MDB_NOTFOUND;
5261 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5263 mp = mc->mc_pg[mc->mc_top];
5265 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5266 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5267 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5268 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5269 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5270 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5271 if (rc == MDB_SUCCESS)
5272 MDB_GET_KEY(leaf, key);
5277 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5278 if (op == MDB_NEXT_DUP)
5279 return MDB_NOTFOUND;
5283 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5284 mdb_dbg_pgno(mp), (void *) mc));
5285 if (mc->mc_flags & C_DEL)
5288 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5289 DPUTS("=====> move to next sibling page");
5290 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5291 mc->mc_flags |= C_EOF;
5294 mp = mc->mc_pg[mc->mc_top];
5295 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5297 mc->mc_ki[mc->mc_top]++;
5300 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5301 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5304 key->mv_size = mc->mc_db->md_pad;
5305 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5309 mdb_cassert(mc, IS_LEAF(mp));
5310 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5312 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5313 mdb_xcursor_init1(mc, leaf);
5316 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5319 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5320 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5321 if (rc != MDB_SUCCESS)
5326 MDB_GET_KEY(leaf, key);
5330 /** Move the cursor to the previous data item. */
5332 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5338 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5340 mp = mc->mc_pg[mc->mc_top];
5342 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5343 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5344 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5345 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5346 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5347 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5348 if (rc == MDB_SUCCESS)
5349 MDB_GET_KEY(leaf, key);
5353 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5354 if (op == MDB_PREV_DUP)
5355 return MDB_NOTFOUND;
5360 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5361 mdb_dbg_pgno(mp), (void *) mc));
5363 if (mc->mc_ki[mc->mc_top] == 0) {
5364 DPUTS("=====> move to prev sibling page");
5365 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5368 mp = mc->mc_pg[mc->mc_top];
5369 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5370 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5372 mc->mc_ki[mc->mc_top]--;
5374 mc->mc_flags &= ~C_EOF;
5376 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5377 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5380 key->mv_size = mc->mc_db->md_pad;
5381 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5385 mdb_cassert(mc, IS_LEAF(mp));
5386 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5388 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5389 mdb_xcursor_init1(mc, leaf);
5392 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5395 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5396 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5397 if (rc != MDB_SUCCESS)
5402 MDB_GET_KEY(leaf, key);
5406 /** Set the cursor on a specific data item. */
5408 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5409 MDB_cursor_op op, int *exactp)
5413 MDB_node *leaf = NULL;
5416 if (key->mv_size == 0)
5417 return MDB_BAD_VALSIZE;
5420 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5422 /* See if we're already on the right page */
5423 if (mc->mc_flags & C_INITIALIZED) {
5426 mp = mc->mc_pg[mc->mc_top];
5428 mc->mc_ki[mc->mc_top] = 0;
5429 return MDB_NOTFOUND;
5431 if (mp->mp_flags & P_LEAF2) {
5432 nodekey.mv_size = mc->mc_db->md_pad;
5433 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5435 leaf = NODEPTR(mp, 0);
5436 MDB_GET_KEY2(leaf, nodekey);
5438 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5440 /* Probably happens rarely, but first node on the page
5441 * was the one we wanted.
5443 mc->mc_ki[mc->mc_top] = 0;
5450 unsigned int nkeys = NUMKEYS(mp);
5452 if (mp->mp_flags & P_LEAF2) {
5453 nodekey.mv_data = LEAF2KEY(mp,
5454 nkeys-1, nodekey.mv_size);
5456 leaf = NODEPTR(mp, nkeys-1);
5457 MDB_GET_KEY2(leaf, nodekey);
5459 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5461 /* last node was the one we wanted */
5462 mc->mc_ki[mc->mc_top] = nkeys-1;
5468 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5469 /* This is definitely the right page, skip search_page */
5470 if (mp->mp_flags & P_LEAF2) {
5471 nodekey.mv_data = LEAF2KEY(mp,
5472 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5474 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5475 MDB_GET_KEY2(leaf, nodekey);
5477 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5479 /* current node was the one we wanted */
5489 /* If any parents have right-sibs, search.
5490 * Otherwise, there's nothing further.
5492 for (i=0; i<mc->mc_top; i++)
5494 NUMKEYS(mc->mc_pg[i])-1)
5496 if (i == mc->mc_top) {
5497 /* There are no other pages */
5498 mc->mc_ki[mc->mc_top] = nkeys;
5499 return MDB_NOTFOUND;
5503 /* There are no other pages */
5504 mc->mc_ki[mc->mc_top] = 0;
5505 if (op == MDB_SET_RANGE && !exactp) {
5509 return MDB_NOTFOUND;
5513 rc = mdb_page_search(mc, key, 0);
5514 if (rc != MDB_SUCCESS)
5517 mp = mc->mc_pg[mc->mc_top];
5518 mdb_cassert(mc, IS_LEAF(mp));
5521 leaf = mdb_node_search(mc, key, exactp);
5522 if (exactp != NULL && !*exactp) {
5523 /* MDB_SET specified and not an exact match. */
5524 return MDB_NOTFOUND;
5528 DPUTS("===> inexact leaf not found, goto sibling");
5529 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5530 return rc; /* no entries matched */
5531 mp = mc->mc_pg[mc->mc_top];
5532 mdb_cassert(mc, IS_LEAF(mp));
5533 leaf = NODEPTR(mp, 0);
5537 mc->mc_flags |= C_INITIALIZED;
5538 mc->mc_flags &= ~C_EOF;
5541 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5542 key->mv_size = mc->mc_db->md_pad;
5543 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5548 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5549 mdb_xcursor_init1(mc, leaf);
5552 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5553 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5554 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5557 if (op == MDB_GET_BOTH) {
5563 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5564 if (rc != MDB_SUCCESS)
5567 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5569 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5571 rc = mc->mc_dbx->md_dcmp(data, &d2);
5573 if (op == MDB_GET_BOTH || rc > 0)
5574 return MDB_NOTFOUND;
5581 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5582 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5587 /* The key already matches in all other cases */
5588 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5589 MDB_GET_KEY(leaf, key);
5590 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5595 /** Move the cursor to the first item in the database. */
5597 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5603 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5605 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5606 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5607 if (rc != MDB_SUCCESS)
5610 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5612 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5613 mc->mc_flags |= C_INITIALIZED;
5614 mc->mc_flags &= ~C_EOF;
5616 mc->mc_ki[mc->mc_top] = 0;
5618 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5619 key->mv_size = mc->mc_db->md_pad;
5620 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5625 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5626 mdb_xcursor_init1(mc, leaf);
5627 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5631 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5635 MDB_GET_KEY(leaf, key);
5639 /** Move the cursor to the last item in the database. */
5641 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5647 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5649 if (!(mc->mc_flags & C_EOF)) {
5651 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5652 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5653 if (rc != MDB_SUCCESS)
5656 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5659 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5660 mc->mc_flags |= C_INITIALIZED|C_EOF;
5661 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5663 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5664 key->mv_size = mc->mc_db->md_pad;
5665 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5670 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5671 mdb_xcursor_init1(mc, leaf);
5672 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5676 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5681 MDB_GET_KEY(leaf, key);
5686 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5691 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5696 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5700 case MDB_GET_CURRENT:
5701 if (!(mc->mc_flags & C_INITIALIZED)) {
5704 MDB_page *mp = mc->mc_pg[mc->mc_top];
5705 int nkeys = NUMKEYS(mp);
5706 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5707 mc->mc_ki[mc->mc_top] = nkeys;
5713 key->mv_size = mc->mc_db->md_pad;
5714 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5716 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5717 MDB_GET_KEY(leaf, key);
5719 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5720 if (mc->mc_flags & C_DEL)
5721 mdb_xcursor_init1(mc, leaf);
5722 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5724 rc = mdb_node_read(mc->mc_txn, leaf, data);
5731 case MDB_GET_BOTH_RANGE:
5736 if (mc->mc_xcursor == NULL) {
5737 rc = MDB_INCOMPATIBLE;
5747 rc = mdb_cursor_set(mc, key, data, op,
5748 op == MDB_SET_RANGE ? NULL : &exact);
5751 case MDB_GET_MULTIPLE:
5752 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5756 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5757 rc = MDB_INCOMPATIBLE;
5761 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5762 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5765 case MDB_NEXT_MULTIPLE:
5770 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5771 rc = MDB_INCOMPATIBLE;
5774 if (!(mc->mc_flags & C_INITIALIZED))
5775 rc = mdb_cursor_first(mc, key, data);
5777 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5778 if (rc == MDB_SUCCESS) {
5779 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5782 mx = &mc->mc_xcursor->mx_cursor;
5783 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5785 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5786 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5794 case MDB_NEXT_NODUP:
5795 if (!(mc->mc_flags & C_INITIALIZED))
5796 rc = mdb_cursor_first(mc, key, data);
5798 rc = mdb_cursor_next(mc, key, data, op);
5802 case MDB_PREV_NODUP:
5803 if (!(mc->mc_flags & C_INITIALIZED)) {
5804 rc = mdb_cursor_last(mc, key, data);
5807 mc->mc_flags |= C_INITIALIZED;
5808 mc->mc_ki[mc->mc_top]++;
5810 rc = mdb_cursor_prev(mc, key, data, op);
5813 rc = mdb_cursor_first(mc, key, data);
5816 mfunc = mdb_cursor_first;
5818 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5822 if (mc->mc_xcursor == NULL) {
5823 rc = MDB_INCOMPATIBLE;
5826 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5830 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5833 rc = mdb_cursor_last(mc, key, data);
5836 mfunc = mdb_cursor_last;
5839 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5844 if (mc->mc_flags & C_DEL)
5845 mc->mc_flags ^= C_DEL;
5850 /** Touch all the pages in the cursor stack. Set mc_top.
5851 * Makes sure all the pages are writable, before attempting a write operation.
5852 * @param[in] mc The cursor to operate on.
5855 mdb_cursor_touch(MDB_cursor *mc)
5857 int rc = MDB_SUCCESS;
5859 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5862 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5864 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5865 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5868 *mc->mc_dbflag |= DB_DIRTY;
5873 rc = mdb_page_touch(mc);
5874 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5875 mc->mc_top = mc->mc_snum-1;
5880 /** Do not spill pages to disk if txn is getting full, may fail instead */
5881 #define MDB_NOSPILL 0x8000
5884 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5887 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5889 MDB_node *leaf = NULL;
5892 MDB_val xdata, *rdata, dkey, olddata;
5894 int do_sub = 0, insert_key, insert_data;
5895 unsigned int mcount = 0, dcount = 0, nospill;
5898 unsigned int nflags;
5901 if (mc == NULL || key == NULL)
5904 env = mc->mc_txn->mt_env;
5906 /* Check this first so counter will always be zero on any
5909 if (flags & MDB_MULTIPLE) {
5910 dcount = data[1].mv_size;
5911 data[1].mv_size = 0;
5912 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5913 return MDB_INCOMPATIBLE;
5916 nospill = flags & MDB_NOSPILL;
5917 flags &= ~MDB_NOSPILL;
5919 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5920 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5922 if (key->mv_size-1 >= ENV_MAXKEY(env))
5923 return MDB_BAD_VALSIZE;
5925 #if SIZE_MAX > MAXDATASIZE
5926 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5927 return MDB_BAD_VALSIZE;
5929 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5930 return MDB_BAD_VALSIZE;
5933 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5934 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5938 if (flags == MDB_CURRENT) {
5939 if (!(mc->mc_flags & C_INITIALIZED))
5942 } else if (mc->mc_db->md_root == P_INVALID) {
5943 /* new database, cursor has nothing to point to */
5946 mc->mc_flags &= ~C_INITIALIZED;
5951 if (flags & MDB_APPEND) {
5953 rc = mdb_cursor_last(mc, &k2, &d2);
5955 rc = mc->mc_dbx->md_cmp(key, &k2);
5958 mc->mc_ki[mc->mc_top]++;
5960 /* new key is <= last key */
5965 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5967 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5968 DPRINTF(("duplicate key [%s]", DKEY(key)));
5970 return MDB_KEYEXIST;
5972 if (rc && rc != MDB_NOTFOUND)
5976 if (mc->mc_flags & C_DEL)
5977 mc->mc_flags ^= C_DEL;
5979 /* Cursor is positioned, check for room in the dirty list */
5981 if (flags & MDB_MULTIPLE) {
5983 xdata.mv_size = data->mv_size * dcount;
5987 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5991 if (rc == MDB_NO_ROOT) {
5993 /* new database, write a root leaf page */
5994 DPUTS("allocating new root leaf page");
5995 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5998 mdb_cursor_push(mc, np);
5999 mc->mc_db->md_root = np->mp_pgno;
6000 mc->mc_db->md_depth++;
6001 *mc->mc_dbflag |= DB_DIRTY;
6002 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6004 np->mp_flags |= P_LEAF2;
6005 mc->mc_flags |= C_INITIALIZED;
6007 /* make sure all cursor pages are writable */
6008 rc2 = mdb_cursor_touch(mc);
6013 insert_key = insert_data = rc;
6015 /* The key does not exist */
6016 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6017 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6018 LEAFSIZE(key, data) > env->me_nodemax)
6020 /* Too big for a node, insert in sub-DB. Set up an empty
6021 * "old sub-page" for prep_subDB to expand to a full page.
6023 fp_flags = P_LEAF|P_DIRTY;
6025 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6026 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6027 olddata.mv_size = PAGEHDRSZ;
6031 /* there's only a key anyway, so this is a no-op */
6032 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6034 unsigned int ksize = mc->mc_db->md_pad;
6035 if (key->mv_size != ksize)
6036 return MDB_BAD_VALSIZE;
6037 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6038 memcpy(ptr, key->mv_data, ksize);
6043 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6044 olddata.mv_size = NODEDSZ(leaf);
6045 olddata.mv_data = NODEDATA(leaf);
6048 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6049 /* Prepare (sub-)page/sub-DB to accept the new item,
6050 * if needed. fp: old sub-page or a header faking
6051 * it. mp: new (sub-)page. offset: growth in page
6052 * size. xdata: node data with new page or DB.
6054 unsigned i, offset = 0;
6055 mp = fp = xdata.mv_data = env->me_pbuf;
6056 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6058 /* Was a single item before, must convert now */
6059 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6060 /* Just overwrite the current item */
6061 if (flags == MDB_CURRENT)
6064 #if UINT_MAX < SIZE_MAX
6065 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6066 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6068 /* does data match? */
6069 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6070 if (flags & MDB_NODUPDATA)
6071 return MDB_KEYEXIST;
6076 /* Back up original data item */
6077 dkey.mv_size = olddata.mv_size;
6078 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6080 /* Make sub-page header for the dup items, with dummy body */
6081 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6082 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6083 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6084 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6085 fp->mp_flags |= P_LEAF2;
6086 fp->mp_pad = data->mv_size;
6087 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6089 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6090 (dkey.mv_size & 1) + (data->mv_size & 1);
6092 fp->mp_upper = xdata.mv_size - PAGEBASE;
6093 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6094 } else if (leaf->mn_flags & F_SUBDATA) {
6095 /* Data is on sub-DB, just store it */
6096 flags |= F_DUPDATA|F_SUBDATA;
6099 /* Data is on sub-page */
6100 fp = olddata.mv_data;
6103 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6104 offset = EVEN(NODESIZE + sizeof(indx_t) +
6108 offset = fp->mp_pad;
6109 if (SIZELEFT(fp) < offset) {
6110 offset *= 4; /* space for 4 more */
6113 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6115 fp->mp_flags |= P_DIRTY;
6116 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6117 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6121 xdata.mv_size = olddata.mv_size + offset;
6124 fp_flags = fp->mp_flags;
6125 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6126 /* Too big for a sub-page, convert to sub-DB */
6127 fp_flags &= ~P_SUBP;
6129 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6130 fp_flags |= P_LEAF2;
6131 dummy.md_pad = fp->mp_pad;
6132 dummy.md_flags = MDB_DUPFIXED;
6133 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6134 dummy.md_flags |= MDB_INTEGERKEY;
6140 dummy.md_branch_pages = 0;
6141 dummy.md_leaf_pages = 1;
6142 dummy.md_overflow_pages = 0;
6143 dummy.md_entries = NUMKEYS(fp);
6144 xdata.mv_size = sizeof(MDB_db);
6145 xdata.mv_data = &dummy;
6146 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6148 offset = env->me_psize - olddata.mv_size;
6149 flags |= F_DUPDATA|F_SUBDATA;
6150 dummy.md_root = mp->mp_pgno;
6153 mp->mp_flags = fp_flags | P_DIRTY;
6154 mp->mp_pad = fp->mp_pad;
6155 mp->mp_lower = fp->mp_lower;
6156 mp->mp_upper = fp->mp_upper + offset;
6157 if (fp_flags & P_LEAF2) {
6158 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6160 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6161 olddata.mv_size - fp->mp_upper - PAGEBASE);
6162 for (i=0; i<NUMKEYS(fp); i++)
6163 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6171 mdb_node_del(mc, 0);
6175 /* overflow page overwrites need special handling */
6176 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6179 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6181 memcpy(&pg, olddata.mv_data, sizeof(pg));
6182 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6184 ovpages = omp->mp_pages;
6186 /* Is the ov page large enough? */
6187 if (ovpages >= dpages) {
6188 if (!(omp->mp_flags & P_DIRTY) &&
6189 (level || (env->me_flags & MDB_WRITEMAP)))
6191 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6194 level = 0; /* dirty in this txn or clean */
6197 if (omp->mp_flags & P_DIRTY) {
6198 /* yes, overwrite it. Note in this case we don't
6199 * bother to try shrinking the page if the new data
6200 * is smaller than the overflow threshold.
6203 /* It is writable only in a parent txn */
6204 size_t sz = (size_t) env->me_psize * ovpages, off;
6205 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6211 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6212 mdb_cassert(mc, rc2 == 0);
6213 if (!(flags & MDB_RESERVE)) {
6214 /* Copy end of page, adjusting alignment so
6215 * compiler may copy words instead of bytes.
6217 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6218 memcpy((size_t *)((char *)np + off),
6219 (size_t *)((char *)omp + off), sz - off);
6222 memcpy(np, omp, sz); /* Copy beginning of page */
6225 SETDSZ(leaf, data->mv_size);
6226 if (F_ISSET(flags, MDB_RESERVE))
6227 data->mv_data = METADATA(omp);
6229 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6233 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6235 } else if (data->mv_size == olddata.mv_size) {
6236 /* same size, just replace it. Note that we could
6237 * also reuse this node if the new data is smaller,
6238 * but instead we opt to shrink the node in that case.
6240 if (F_ISSET(flags, MDB_RESERVE))
6241 data->mv_data = olddata.mv_data;
6242 else if (!(mc->mc_flags & C_SUB))
6243 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6245 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6248 mdb_node_del(mc, 0);
6254 nflags = flags & NODE_ADD_FLAGS;
6255 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6256 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6257 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6258 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6260 nflags |= MDB_SPLIT_REPLACE;
6261 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6263 /* There is room already in this leaf page. */
6264 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6265 if (rc == 0 && insert_key) {
6266 /* Adjust other cursors pointing to mp */
6267 MDB_cursor *m2, *m3;
6268 MDB_dbi dbi = mc->mc_dbi;
6269 unsigned i = mc->mc_top;
6270 MDB_page *mp = mc->mc_pg[i];
6272 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6273 if (mc->mc_flags & C_SUB)
6274 m3 = &m2->mc_xcursor->mx_cursor;
6277 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6278 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6285 if (rc == MDB_SUCCESS) {
6286 /* Now store the actual data in the child DB. Note that we're
6287 * storing the user data in the keys field, so there are strict
6288 * size limits on dupdata. The actual data fields of the child
6289 * DB are all zero size.
6297 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6298 if (flags & MDB_CURRENT) {
6299 xflags = MDB_CURRENT|MDB_NOSPILL;
6301 mdb_xcursor_init1(mc, leaf);
6302 xflags = (flags & MDB_NODUPDATA) ?
6303 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6305 /* converted, write the original data first */
6307 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6311 /* Adjust other cursors pointing to mp */
6313 unsigned i = mc->mc_top;
6314 MDB_page *mp = mc->mc_pg[i];
6316 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6317 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6318 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6319 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6320 mdb_xcursor_init1(m2, leaf);
6324 /* we've done our job */
6327 ecount = mc->mc_xcursor->mx_db.md_entries;
6328 if (flags & MDB_APPENDDUP)
6329 xflags |= MDB_APPEND;
6330 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6331 if (flags & F_SUBDATA) {
6332 void *db = NODEDATA(leaf);
6333 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6335 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6337 /* Increment count unless we just replaced an existing item. */
6339 mc->mc_db->md_entries++;
6341 /* Invalidate txn if we created an empty sub-DB */
6344 /* If we succeeded and the key didn't exist before,
6345 * make sure the cursor is marked valid.
6347 mc->mc_flags |= C_INITIALIZED;
6349 if (flags & MDB_MULTIPLE) {
6352 /* let caller know how many succeeded, if any */
6353 data[1].mv_size = mcount;
6354 if (mcount < dcount) {
6355 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6356 insert_key = insert_data = 0;
6363 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6366 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6371 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6377 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6378 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6380 if (!(mc->mc_flags & C_INITIALIZED))
6383 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6384 return MDB_NOTFOUND;
6386 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6389 rc = mdb_cursor_touch(mc);
6393 mp = mc->mc_pg[mc->mc_top];
6396 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6398 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6399 if (flags & MDB_NODUPDATA) {
6400 /* mdb_cursor_del0() will subtract the final entry */
6401 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6403 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6404 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6406 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6409 /* If sub-DB still has entries, we're done */
6410 if (mc->mc_xcursor->mx_db.md_entries) {
6411 if (leaf->mn_flags & F_SUBDATA) {
6412 /* update subDB info */
6413 void *db = NODEDATA(leaf);
6414 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6417 /* shrink fake page */
6418 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6419 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6420 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6421 /* fix other sub-DB cursors pointed at this fake page */
6422 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6423 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6424 if (m2->mc_pg[mc->mc_top] == mp &&
6425 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6426 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6429 mc->mc_db->md_entries--;
6430 mc->mc_flags |= C_DEL;
6433 /* otherwise fall thru and delete the sub-DB */
6436 if (leaf->mn_flags & F_SUBDATA) {
6437 /* add all the child DB's pages to the free list */
6438 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6444 /* add overflow pages to free list */
6445 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6449 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6450 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6451 (rc = mdb_ovpage_free(mc, omp)))
6456 return mdb_cursor_del0(mc);
6459 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6463 /** Allocate and initialize new pages for a database.
6464 * @param[in] mc a cursor on the database being added to.
6465 * @param[in] flags flags defining what type of page is being allocated.
6466 * @param[in] num the number of pages to allocate. This is usually 1,
6467 * unless allocating overflow pages for a large record.
6468 * @param[out] mp Address of a page, or NULL on failure.
6469 * @return 0 on success, non-zero on failure.
6472 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6477 if ((rc = mdb_page_alloc(mc, num, &np)))
6479 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6480 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6481 np->mp_flags = flags | P_DIRTY;
6482 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6483 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6486 mc->mc_db->md_branch_pages++;
6487 else if (IS_LEAF(np))
6488 mc->mc_db->md_leaf_pages++;
6489 else if (IS_OVERFLOW(np)) {
6490 mc->mc_db->md_overflow_pages += num;
6498 /** Calculate the size of a leaf node.
6499 * The size depends on the environment's page size; if a data item
6500 * is too large it will be put onto an overflow page and the node
6501 * size will only include the key and not the data. Sizes are always
6502 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6503 * of the #MDB_node headers.
6504 * @param[in] env The environment handle.
6505 * @param[in] key The key for the node.
6506 * @param[in] data The data for the node.
6507 * @return The number of bytes needed to store the node.
6510 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6514 sz = LEAFSIZE(key, data);
6515 if (sz > env->me_nodemax) {
6516 /* put on overflow page */
6517 sz -= data->mv_size - sizeof(pgno_t);
6520 return EVEN(sz + sizeof(indx_t));
6523 /** Calculate the size of a branch node.
6524 * The size should depend on the environment's page size but since
6525 * we currently don't support spilling large keys onto overflow
6526 * pages, it's simply the size of the #MDB_node header plus the
6527 * size of the key. Sizes are always rounded up to an even number
6528 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6529 * @param[in] env The environment handle.
6530 * @param[in] key The key for the node.
6531 * @return The number of bytes needed to store the node.
6534 mdb_branch_size(MDB_env *env, MDB_val *key)
6539 if (sz > env->me_nodemax) {
6540 /* put on overflow page */
6541 /* not implemented */
6542 /* sz -= key->size - sizeof(pgno_t); */
6545 return sz + sizeof(indx_t);
6548 /** Add a node to the page pointed to by the cursor.
6549 * @param[in] mc The cursor for this operation.
6550 * @param[in] indx The index on the page where the new node should be added.
6551 * @param[in] key The key for the new node.
6552 * @param[in] data The data for the new node, if any.
6553 * @param[in] pgno The page number, if adding a branch node.
6554 * @param[in] flags Flags for the node.
6555 * @return 0 on success, non-zero on failure. Possible errors are:
6557 * <li>ENOMEM - failed to allocate overflow pages for the node.
6558 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6559 * should never happen since all callers already calculate the
6560 * page's free space before calling this function.
6564 mdb_node_add(MDB_cursor *mc, indx_t indx,
6565 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6568 size_t node_size = NODESIZE;
6572 MDB_page *mp = mc->mc_pg[mc->mc_top];
6573 MDB_page *ofp = NULL; /* overflow page */
6576 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6578 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6579 IS_LEAF(mp) ? "leaf" : "branch",
6580 IS_SUBP(mp) ? "sub-" : "",
6581 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6582 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6585 /* Move higher keys up one slot. */
6586 int ksize = mc->mc_db->md_pad, dif;
6587 char *ptr = LEAF2KEY(mp, indx, ksize);
6588 dif = NUMKEYS(mp) - indx;
6590 memmove(ptr+ksize, ptr, dif*ksize);
6591 /* insert new key */
6592 memcpy(ptr, key->mv_data, ksize);
6594 /* Just using these for counting */
6595 mp->mp_lower += sizeof(indx_t);
6596 mp->mp_upper -= ksize - sizeof(indx_t);
6600 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6602 node_size += key->mv_size;
6604 mdb_cassert(mc, data);
6605 if (F_ISSET(flags, F_BIGDATA)) {
6606 /* Data already on overflow page. */
6607 node_size += sizeof(pgno_t);
6608 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6609 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6611 /* Put data on overflow page. */
6612 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6613 data->mv_size, node_size+data->mv_size));
6614 node_size = EVEN(node_size + sizeof(pgno_t));
6615 if ((ssize_t)node_size > room)
6617 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6619 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6623 node_size += data->mv_size;
6626 node_size = EVEN(node_size);
6627 if ((ssize_t)node_size > room)
6631 /* Move higher pointers up one slot. */
6632 for (i = NUMKEYS(mp); i > indx; i--)
6633 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6635 /* Adjust free space offsets. */
6636 ofs = mp->mp_upper - node_size;
6637 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6638 mp->mp_ptrs[indx] = ofs;
6640 mp->mp_lower += sizeof(indx_t);
6642 /* Write the node data. */
6643 node = NODEPTR(mp, indx);
6644 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6645 node->mn_flags = flags;
6647 SETDSZ(node,data->mv_size);
6652 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6655 mdb_cassert(mc, key);
6657 if (F_ISSET(flags, F_BIGDATA))
6658 memcpy(node->mn_data + key->mv_size, data->mv_data,
6660 else if (F_ISSET(flags, MDB_RESERVE))
6661 data->mv_data = node->mn_data + key->mv_size;
6663 memcpy(node->mn_data + key->mv_size, data->mv_data,
6666 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6668 if (F_ISSET(flags, MDB_RESERVE))
6669 data->mv_data = METADATA(ofp);
6671 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6678 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6679 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6680 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6681 DPRINTF(("node size = %"Z"u", node_size));
6682 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6683 return MDB_PAGE_FULL;
6686 /** Delete the specified node from a page.
6687 * @param[in] mc Cursor pointing to the node to delete.
6688 * @param[in] ksize The size of a node. Only used if the page is
6689 * part of a #MDB_DUPFIXED database.
6692 mdb_node_del(MDB_cursor *mc, int ksize)
6694 MDB_page *mp = mc->mc_pg[mc->mc_top];
6695 indx_t indx = mc->mc_ki[mc->mc_top];
6697 indx_t i, j, numkeys, ptr;
6701 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6702 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6703 numkeys = NUMKEYS(mp);
6704 mdb_cassert(mc, indx < numkeys);
6707 int x = numkeys - 1 - indx;
6708 base = LEAF2KEY(mp, indx, ksize);
6710 memmove(base, base + ksize, x * ksize);
6711 mp->mp_lower -= sizeof(indx_t);
6712 mp->mp_upper += ksize - sizeof(indx_t);
6716 node = NODEPTR(mp, indx);
6717 sz = NODESIZE + node->mn_ksize;
6719 if (F_ISSET(node->mn_flags, F_BIGDATA))
6720 sz += sizeof(pgno_t);
6722 sz += NODEDSZ(node);
6726 ptr = mp->mp_ptrs[indx];
6727 for (i = j = 0; i < numkeys; i++) {
6729 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6730 if (mp->mp_ptrs[i] < ptr)
6731 mp->mp_ptrs[j] += sz;
6736 base = (char *)mp + mp->mp_upper + PAGEBASE;
6737 memmove(base + sz, base, ptr - mp->mp_upper);
6739 mp->mp_lower -= sizeof(indx_t);
6743 /** Compact the main page after deleting a node on a subpage.
6744 * @param[in] mp The main page to operate on.
6745 * @param[in] indx The index of the subpage on the main page.
6748 mdb_node_shrink(MDB_page *mp, indx_t indx)
6754 indx_t i, numkeys, ptr;
6756 node = NODEPTR(mp, indx);
6757 sp = (MDB_page *)NODEDATA(node);
6758 delta = SIZELEFT(sp);
6759 xp = (MDB_page *)((char *)sp + delta);
6761 /* shift subpage upward */
6763 nsize = NUMKEYS(sp) * sp->mp_pad;
6765 return; /* do not make the node uneven-sized */
6766 memmove(METADATA(xp), METADATA(sp), nsize);
6769 numkeys = NUMKEYS(sp);
6770 for (i=numkeys-1; i>=0; i--)
6771 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6773 xp->mp_upper = sp->mp_lower;
6774 xp->mp_lower = sp->mp_lower;
6775 xp->mp_flags = sp->mp_flags;
6776 xp->mp_pad = sp->mp_pad;
6777 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6779 nsize = NODEDSZ(node) - delta;
6780 SETDSZ(node, nsize);
6782 /* shift lower nodes upward */
6783 ptr = mp->mp_ptrs[indx];
6784 numkeys = NUMKEYS(mp);
6785 for (i = 0; i < numkeys; i++) {
6786 if (mp->mp_ptrs[i] <= ptr)
6787 mp->mp_ptrs[i] += delta;
6790 base = (char *)mp + mp->mp_upper + PAGEBASE;
6791 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6792 mp->mp_upper += delta;
6795 /** Initial setup of a sorted-dups cursor.
6796 * Sorted duplicates are implemented as a sub-database for the given key.
6797 * The duplicate data items are actually keys of the sub-database.
6798 * Operations on the duplicate data items are performed using a sub-cursor
6799 * initialized when the sub-database is first accessed. This function does
6800 * the preliminary setup of the sub-cursor, filling in the fields that
6801 * depend only on the parent DB.
6802 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6805 mdb_xcursor_init0(MDB_cursor *mc)
6807 MDB_xcursor *mx = mc->mc_xcursor;
6809 mx->mx_cursor.mc_xcursor = NULL;
6810 mx->mx_cursor.mc_txn = mc->mc_txn;
6811 mx->mx_cursor.mc_db = &mx->mx_db;
6812 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6813 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6814 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6815 mx->mx_cursor.mc_snum = 0;
6816 mx->mx_cursor.mc_top = 0;
6817 mx->mx_cursor.mc_flags = C_SUB;
6818 mx->mx_dbx.md_name.mv_size = 0;
6819 mx->mx_dbx.md_name.mv_data = NULL;
6820 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6821 mx->mx_dbx.md_dcmp = NULL;
6822 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6825 /** Final setup of a sorted-dups cursor.
6826 * Sets up the fields that depend on the data from the main cursor.
6827 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6828 * @param[in] node The data containing the #MDB_db record for the
6829 * sorted-dup database.
6832 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6834 MDB_xcursor *mx = mc->mc_xcursor;
6836 if (node->mn_flags & F_SUBDATA) {
6837 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6838 mx->mx_cursor.mc_pg[0] = 0;
6839 mx->mx_cursor.mc_snum = 0;
6840 mx->mx_cursor.mc_top = 0;
6841 mx->mx_cursor.mc_flags = C_SUB;
6843 MDB_page *fp = NODEDATA(node);
6844 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6845 mx->mx_db.md_flags = 0;
6846 mx->mx_db.md_depth = 1;
6847 mx->mx_db.md_branch_pages = 0;
6848 mx->mx_db.md_leaf_pages = 1;
6849 mx->mx_db.md_overflow_pages = 0;
6850 mx->mx_db.md_entries = NUMKEYS(fp);
6851 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6852 mx->mx_cursor.mc_snum = 1;
6853 mx->mx_cursor.mc_top = 0;
6854 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6855 mx->mx_cursor.mc_pg[0] = fp;
6856 mx->mx_cursor.mc_ki[0] = 0;
6857 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6858 mx->mx_db.md_flags = MDB_DUPFIXED;
6859 mx->mx_db.md_pad = fp->mp_pad;
6860 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6861 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6864 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6865 mx->mx_db.md_root));
6866 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6867 #if UINT_MAX < SIZE_MAX
6868 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6869 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6873 /** Initialize a cursor for a given transaction and database. */
6875 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6878 mc->mc_backup = NULL;
6881 mc->mc_db = &txn->mt_dbs[dbi];
6882 mc->mc_dbx = &txn->mt_dbxs[dbi];
6883 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6888 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6889 mdb_tassert(txn, mx != NULL);
6890 mc->mc_xcursor = mx;
6891 mdb_xcursor_init0(mc);
6893 mc->mc_xcursor = NULL;
6895 if (*mc->mc_dbflag & DB_STALE) {
6896 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6901 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6904 size_t size = sizeof(MDB_cursor);
6906 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6909 if (txn->mt_flags & MDB_TXN_ERROR)
6912 /* Allow read access to the freelist */
6913 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6916 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6917 size += sizeof(MDB_xcursor);
6919 if ((mc = malloc(size)) != NULL) {
6920 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6921 if (txn->mt_cursors) {
6922 mc->mc_next = txn->mt_cursors[dbi];
6923 txn->mt_cursors[dbi] = mc;
6924 mc->mc_flags |= C_UNTRACK;
6936 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6938 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6941 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6944 if (txn->mt_flags & MDB_TXN_ERROR)
6947 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6951 /* Return the count of duplicate data items for the current key */
6953 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6957 if (mc == NULL || countp == NULL)
6960 if (mc->mc_xcursor == NULL)
6961 return MDB_INCOMPATIBLE;
6963 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6966 if (!(mc->mc_flags & C_INITIALIZED))
6969 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6970 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6973 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6976 *countp = mc->mc_xcursor->mx_db.md_entries;
6982 mdb_cursor_close(MDB_cursor *mc)
6984 if (mc && !mc->mc_backup) {
6985 /* remove from txn, if tracked */
6986 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6987 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6988 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6990 *prev = mc->mc_next;
6997 mdb_cursor_txn(MDB_cursor *mc)
6999 if (!mc) return NULL;
7004 mdb_cursor_dbi(MDB_cursor *mc)
7009 /** Replace the key for a branch node with a new key.
7010 * @param[in] mc Cursor pointing to the node to operate on.
7011 * @param[in] key The new key to use.
7012 * @return 0 on success, non-zero on failure.
7015 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7021 int delta, ksize, oksize;
7022 indx_t ptr, i, numkeys, indx;
7025 indx = mc->mc_ki[mc->mc_top];
7026 mp = mc->mc_pg[mc->mc_top];
7027 node = NODEPTR(mp, indx);
7028 ptr = mp->mp_ptrs[indx];
7032 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7033 k2.mv_data = NODEKEY(node);
7034 k2.mv_size = node->mn_ksize;
7035 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7037 mdb_dkey(&k2, kbuf2),
7043 /* Sizes must be 2-byte aligned. */
7044 ksize = EVEN(key->mv_size);
7045 oksize = EVEN(node->mn_ksize);
7046 delta = ksize - oksize;
7048 /* Shift node contents if EVEN(key length) changed. */
7050 if (delta > 0 && SIZELEFT(mp) < delta) {
7052 /* not enough space left, do a delete and split */
7053 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7054 pgno = NODEPGNO(node);
7055 mdb_node_del(mc, 0);
7056 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7059 numkeys = NUMKEYS(mp);
7060 for (i = 0; i < numkeys; i++) {
7061 if (mp->mp_ptrs[i] <= ptr)
7062 mp->mp_ptrs[i] -= delta;
7065 base = (char *)mp + mp->mp_upper + PAGEBASE;
7066 len = ptr - mp->mp_upper + NODESIZE;
7067 memmove(base - delta, base, len);
7068 mp->mp_upper -= delta;
7070 node = NODEPTR(mp, indx);
7073 /* But even if no shift was needed, update ksize */
7074 if (node->mn_ksize != key->mv_size)
7075 node->mn_ksize = key->mv_size;
7078 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7084 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7086 /** Move a node from csrc to cdst.
7089 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7096 unsigned short flags;
7100 /* Mark src and dst as dirty. */
7101 if ((rc = mdb_page_touch(csrc)) ||
7102 (rc = mdb_page_touch(cdst)))
7105 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7106 key.mv_size = csrc->mc_db->md_pad;
7107 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7109 data.mv_data = NULL;
7113 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7114 mdb_cassert(csrc, !((size_t)srcnode & 1));
7115 srcpg = NODEPGNO(srcnode);
7116 flags = srcnode->mn_flags;
7117 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7118 unsigned int snum = csrc->mc_snum;
7120 /* must find the lowest key below src */
7121 rc = mdb_page_search_lowest(csrc);
7124 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7125 key.mv_size = csrc->mc_db->md_pad;
7126 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7128 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7129 key.mv_size = NODEKSZ(s2);
7130 key.mv_data = NODEKEY(s2);
7132 csrc->mc_snum = snum--;
7133 csrc->mc_top = snum;
7135 key.mv_size = NODEKSZ(srcnode);
7136 key.mv_data = NODEKEY(srcnode);
7138 data.mv_size = NODEDSZ(srcnode);
7139 data.mv_data = NODEDATA(srcnode);
7141 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7142 unsigned int snum = cdst->mc_snum;
7145 /* must find the lowest key below dst */
7146 mdb_cursor_copy(cdst, &mn);
7147 rc = mdb_page_search_lowest(&mn);
7150 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7151 bkey.mv_size = mn.mc_db->md_pad;
7152 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7154 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7155 bkey.mv_size = NODEKSZ(s2);
7156 bkey.mv_data = NODEKEY(s2);
7158 mn.mc_snum = snum--;
7161 rc = mdb_update_key(&mn, &bkey);
7166 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7167 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7168 csrc->mc_ki[csrc->mc_top],
7170 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7171 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7173 /* Add the node to the destination page.
7175 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7176 if (rc != MDB_SUCCESS)
7179 /* Delete the node from the source page.
7181 mdb_node_del(csrc, key.mv_size);
7184 /* Adjust other cursors pointing to mp */
7185 MDB_cursor *m2, *m3;
7186 MDB_dbi dbi = csrc->mc_dbi;
7187 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7189 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7190 if (csrc->mc_flags & C_SUB)
7191 m3 = &m2->mc_xcursor->mx_cursor;
7194 if (m3 == csrc) continue;
7195 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7196 csrc->mc_ki[csrc->mc_top]) {
7197 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7198 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7203 /* Update the parent separators.
7205 if (csrc->mc_ki[csrc->mc_top] == 0) {
7206 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7207 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7208 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7210 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7211 key.mv_size = NODEKSZ(srcnode);
7212 key.mv_data = NODEKEY(srcnode);
7214 DPRINTF(("update separator for source page %"Z"u to [%s]",
7215 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7216 mdb_cursor_copy(csrc, &mn);
7219 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7222 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7224 indx_t ix = csrc->mc_ki[csrc->mc_top];
7225 nullkey.mv_size = 0;
7226 csrc->mc_ki[csrc->mc_top] = 0;
7227 rc = mdb_update_key(csrc, &nullkey);
7228 csrc->mc_ki[csrc->mc_top] = ix;
7229 mdb_cassert(csrc, rc == MDB_SUCCESS);
7233 if (cdst->mc_ki[cdst->mc_top] == 0) {
7234 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7235 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7236 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7238 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7239 key.mv_size = NODEKSZ(srcnode);
7240 key.mv_data = NODEKEY(srcnode);
7242 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7243 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7244 mdb_cursor_copy(cdst, &mn);
7247 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7250 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7252 indx_t ix = cdst->mc_ki[cdst->mc_top];
7253 nullkey.mv_size = 0;
7254 cdst->mc_ki[cdst->mc_top] = 0;
7255 rc = mdb_update_key(cdst, &nullkey);
7256 cdst->mc_ki[cdst->mc_top] = ix;
7257 mdb_cassert(csrc, rc == MDB_SUCCESS);
7264 /** Merge one page into another.
7265 * The nodes from the page pointed to by \b csrc will
7266 * be copied to the page pointed to by \b cdst and then
7267 * the \b csrc page will be freed.
7268 * @param[in] csrc Cursor pointing to the source page.
7269 * @param[in] cdst Cursor pointing to the destination page.
7270 * @return 0 on success, non-zero on failure.
7273 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7275 MDB_page *psrc, *pdst;
7282 psrc = csrc->mc_pg[csrc->mc_top];
7283 pdst = cdst->mc_pg[cdst->mc_top];
7285 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7287 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7288 mdb_cassert(csrc, cdst->mc_snum > 1);
7290 /* Mark dst as dirty. */
7291 if ((rc = mdb_page_touch(cdst)))
7294 /* Move all nodes from src to dst.
7296 j = nkeys = NUMKEYS(pdst);
7297 if (IS_LEAF2(psrc)) {
7298 key.mv_size = csrc->mc_db->md_pad;
7299 key.mv_data = METADATA(psrc);
7300 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7301 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7302 if (rc != MDB_SUCCESS)
7304 key.mv_data = (char *)key.mv_data + key.mv_size;
7307 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7308 srcnode = NODEPTR(psrc, i);
7309 if (i == 0 && IS_BRANCH(psrc)) {
7312 mdb_cursor_copy(csrc, &mn);
7313 /* must find the lowest key below src */
7314 rc = mdb_page_search_lowest(&mn);
7317 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7318 key.mv_size = mn.mc_db->md_pad;
7319 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7321 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7322 key.mv_size = NODEKSZ(s2);
7323 key.mv_data = NODEKEY(s2);
7326 key.mv_size = srcnode->mn_ksize;
7327 key.mv_data = NODEKEY(srcnode);
7330 data.mv_size = NODEDSZ(srcnode);
7331 data.mv_data = NODEDATA(srcnode);
7332 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7333 if (rc != MDB_SUCCESS)
7338 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7339 pdst->mp_pgno, NUMKEYS(pdst),
7340 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7342 /* Unlink the src page from parent and add to free list.
7345 mdb_node_del(csrc, 0);
7346 if (csrc->mc_ki[csrc->mc_top] == 0) {
7348 rc = mdb_update_key(csrc, &key);
7356 psrc = csrc->mc_pg[csrc->mc_top];
7357 /* If not operating on FreeDB, allow this page to be reused
7358 * in this txn. Otherwise just add to free list.
7360 rc = mdb_page_loose(csrc, psrc);
7364 csrc->mc_db->md_leaf_pages--;
7366 csrc->mc_db->md_branch_pages--;
7368 /* Adjust other cursors pointing to mp */
7369 MDB_cursor *m2, *m3;
7370 MDB_dbi dbi = csrc->mc_dbi;
7372 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7373 if (csrc->mc_flags & C_SUB)
7374 m3 = &m2->mc_xcursor->mx_cursor;
7377 if (m3 == csrc) continue;
7378 if (m3->mc_snum < csrc->mc_snum) continue;
7379 if (m3->mc_pg[csrc->mc_top] == psrc) {
7380 m3->mc_pg[csrc->mc_top] = pdst;
7381 m3->mc_ki[csrc->mc_top] += nkeys;
7386 unsigned int snum = cdst->mc_snum;
7387 uint16_t depth = cdst->mc_db->md_depth;
7388 mdb_cursor_pop(cdst);
7389 rc = mdb_rebalance(cdst);
7390 /* Did the tree shrink? */
7391 if (depth > cdst->mc_db->md_depth)
7393 cdst->mc_snum = snum;
7394 cdst->mc_top = snum-1;
7399 /** Copy the contents of a cursor.
7400 * @param[in] csrc The cursor to copy from.
7401 * @param[out] cdst The cursor to copy to.
7404 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7408 cdst->mc_txn = csrc->mc_txn;
7409 cdst->mc_dbi = csrc->mc_dbi;
7410 cdst->mc_db = csrc->mc_db;
7411 cdst->mc_dbx = csrc->mc_dbx;
7412 cdst->mc_snum = csrc->mc_snum;
7413 cdst->mc_top = csrc->mc_top;
7414 cdst->mc_flags = csrc->mc_flags;
7416 for (i=0; i<csrc->mc_snum; i++) {
7417 cdst->mc_pg[i] = csrc->mc_pg[i];
7418 cdst->mc_ki[i] = csrc->mc_ki[i];
7422 /** Rebalance the tree after a delete operation.
7423 * @param[in] mc Cursor pointing to the page where rebalancing
7425 * @return 0 on success, non-zero on failure.
7428 mdb_rebalance(MDB_cursor *mc)
7432 unsigned int ptop, minkeys;
7436 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7437 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7438 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7439 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7440 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7442 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7443 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7444 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7445 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7449 if (mc->mc_snum < 2) {
7450 MDB_page *mp = mc->mc_pg[0];
7452 DPUTS("Can't rebalance a subpage, ignoring");
7455 if (NUMKEYS(mp) == 0) {
7456 DPUTS("tree is completely empty");
7457 mc->mc_db->md_root = P_INVALID;
7458 mc->mc_db->md_depth = 0;
7459 mc->mc_db->md_leaf_pages = 0;
7460 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7463 /* Adjust cursors pointing to mp */
7466 mc->mc_flags &= ~C_INITIALIZED;
7468 MDB_cursor *m2, *m3;
7469 MDB_dbi dbi = mc->mc_dbi;
7471 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7472 if (mc->mc_flags & C_SUB)
7473 m3 = &m2->mc_xcursor->mx_cursor;
7476 if (m3->mc_snum < mc->mc_snum) continue;
7477 if (m3->mc_pg[0] == mp) {
7480 m3->mc_flags &= ~C_INITIALIZED;
7484 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7486 DPUTS("collapsing root page!");
7487 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7490 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7491 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7494 mc->mc_db->md_depth--;
7495 mc->mc_db->md_branch_pages--;
7496 mc->mc_ki[0] = mc->mc_ki[1];
7497 for (i = 1; i<mc->mc_db->md_depth; i++) {
7498 mc->mc_pg[i] = mc->mc_pg[i+1];
7499 mc->mc_ki[i] = mc->mc_ki[i+1];
7502 /* Adjust other cursors pointing to mp */
7503 MDB_cursor *m2, *m3;
7504 MDB_dbi dbi = mc->mc_dbi;
7506 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7507 if (mc->mc_flags & C_SUB)
7508 m3 = &m2->mc_xcursor->mx_cursor;
7511 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7512 if (m3->mc_pg[0] == mp) {
7515 for (i=0; i<m3->mc_snum; i++) {
7516 m3->mc_pg[i] = m3->mc_pg[i+1];
7517 m3->mc_ki[i] = m3->mc_ki[i+1];
7523 DPUTS("root page doesn't need rebalancing");
7527 /* The parent (branch page) must have at least 2 pointers,
7528 * otherwise the tree is invalid.
7530 ptop = mc->mc_top-1;
7531 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7533 /* Leaf page fill factor is below the threshold.
7534 * Try to move keys from left or right neighbor, or
7535 * merge with a neighbor page.
7540 mdb_cursor_copy(mc, &mn);
7541 mn.mc_xcursor = NULL;
7543 oldki = mc->mc_ki[mc->mc_top];
7544 if (mc->mc_ki[ptop] == 0) {
7545 /* We're the leftmost leaf in our parent.
7547 DPUTS("reading right neighbor");
7549 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7550 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7553 mn.mc_ki[mn.mc_top] = 0;
7554 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7556 /* There is at least one neighbor to the left.
7558 DPUTS("reading left neighbor");
7560 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7561 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7564 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7565 mc->mc_ki[mc->mc_top] = 0;
7568 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7569 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7570 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7572 /* If the neighbor page is above threshold and has enough keys,
7573 * move one key from it. Otherwise we should try to merge them.
7574 * (A branch page must never have less than 2 keys.)
7576 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7577 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7578 rc = mdb_node_move(&mn, mc);
7579 if (mc->mc_ki[ptop]) {
7583 if (mc->mc_ki[ptop] == 0) {
7584 rc = mdb_page_merge(&mn, mc);
7586 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7587 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7588 rc = mdb_page_merge(mc, &mn);
7589 mdb_cursor_copy(&mn, mc);
7591 mc->mc_flags &= ~C_EOF;
7593 mc->mc_ki[mc->mc_top] = oldki;
7597 /** Complete a delete operation started by #mdb_cursor_del(). */
7599 mdb_cursor_del0(MDB_cursor *mc)
7606 ki = mc->mc_ki[mc->mc_top];
7607 mdb_node_del(mc, mc->mc_db->md_pad);
7608 mc->mc_db->md_entries--;
7609 rc = mdb_rebalance(mc);
7611 if (rc == MDB_SUCCESS) {
7612 MDB_cursor *m2, *m3;
7613 MDB_dbi dbi = mc->mc_dbi;
7615 mp = mc->mc_pg[mc->mc_top];
7616 nkeys = NUMKEYS(mp);
7618 /* if mc points past last node in page, find next sibling */
7619 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7620 rc = mdb_cursor_sibling(mc, 1);
7621 if (rc == MDB_NOTFOUND) {
7622 mc->mc_flags |= C_EOF;
7627 /* Adjust other cursors pointing to mp */
7628 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7629 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7630 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7632 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7634 if (m3->mc_pg[mc->mc_top] == mp) {
7635 if (m3->mc_ki[mc->mc_top] >= ki) {
7636 m3->mc_flags |= C_DEL;
7637 if (m3->mc_ki[mc->mc_top] > ki)
7638 m3->mc_ki[mc->mc_top]--;
7639 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7640 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7642 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7643 rc = mdb_cursor_sibling(m3, 1);
7644 if (rc == MDB_NOTFOUND) {
7645 m3->mc_flags |= C_EOF;
7651 mc->mc_flags |= C_DEL;
7655 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7660 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7661 MDB_val *key, MDB_val *data)
7663 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7666 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7667 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7669 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7670 /* must ignore any data */
7674 return mdb_del0(txn, dbi, key, data, 0);
7678 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7679 MDB_val *key, MDB_val *data, unsigned flags)
7684 MDB_val rdata, *xdata;
7688 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7690 mdb_cursor_init(&mc, txn, dbi, &mx);
7699 flags |= MDB_NODUPDATA;
7701 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7703 /* let mdb_page_split know about this cursor if needed:
7704 * delete will trigger a rebalance; if it needs to move
7705 * a node from one page to another, it will have to
7706 * update the parent's separator key(s). If the new sepkey
7707 * is larger than the current one, the parent page may
7708 * run out of space, triggering a split. We need this
7709 * cursor to be consistent until the end of the rebalance.
7711 mc.mc_flags |= C_UNTRACK;
7712 mc.mc_next = txn->mt_cursors[dbi];
7713 txn->mt_cursors[dbi] = &mc;
7714 rc = mdb_cursor_del(&mc, flags);
7715 txn->mt_cursors[dbi] = mc.mc_next;
7720 /** Split a page and insert a new node.
7721 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7722 * The cursor will be updated to point to the actual page and index where
7723 * the node got inserted after the split.
7724 * @param[in] newkey The key for the newly inserted node.
7725 * @param[in] newdata The data for the newly inserted node.
7726 * @param[in] newpgno The page number, if the new node is a branch node.
7727 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7728 * @return 0 on success, non-zero on failure.
7731 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7732 unsigned int nflags)
7735 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7738 int i, j, split_indx, nkeys, pmax;
7739 MDB_env *env = mc->mc_txn->mt_env;
7741 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7742 MDB_page *copy = NULL;
7743 MDB_page *mp, *rp, *pp;
7748 mp = mc->mc_pg[mc->mc_top];
7749 newindx = mc->mc_ki[mc->mc_top];
7750 nkeys = NUMKEYS(mp);
7752 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7753 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7754 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7756 /* Create a right sibling. */
7757 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7759 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7761 if (mc->mc_snum < 2) {
7762 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7764 /* shift current top to make room for new parent */
7765 mc->mc_pg[1] = mc->mc_pg[0];
7766 mc->mc_ki[1] = mc->mc_ki[0];
7769 mc->mc_db->md_root = pp->mp_pgno;
7770 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7771 mc->mc_db->md_depth++;
7774 /* Add left (implicit) pointer. */
7775 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7776 /* undo the pre-push */
7777 mc->mc_pg[0] = mc->mc_pg[1];
7778 mc->mc_ki[0] = mc->mc_ki[1];
7779 mc->mc_db->md_root = mp->mp_pgno;
7780 mc->mc_db->md_depth--;
7787 ptop = mc->mc_top-1;
7788 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7791 mc->mc_flags |= C_SPLITTING;
7792 mdb_cursor_copy(mc, &mn);
7793 mn.mc_pg[mn.mc_top] = rp;
7794 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7796 if (nflags & MDB_APPEND) {
7797 mn.mc_ki[mn.mc_top] = 0;
7799 split_indx = newindx;
7803 split_indx = (nkeys+1) / 2;
7808 unsigned int lsize, rsize, ksize;
7809 /* Move half of the keys to the right sibling */
7810 x = mc->mc_ki[mc->mc_top] - split_indx;
7811 ksize = mc->mc_db->md_pad;
7812 split = LEAF2KEY(mp, split_indx, ksize);
7813 rsize = (nkeys - split_indx) * ksize;
7814 lsize = (nkeys - split_indx) * sizeof(indx_t);
7815 mp->mp_lower -= lsize;
7816 rp->mp_lower += lsize;
7817 mp->mp_upper += rsize - lsize;
7818 rp->mp_upper -= rsize - lsize;
7819 sepkey.mv_size = ksize;
7820 if (newindx == split_indx) {
7821 sepkey.mv_data = newkey->mv_data;
7823 sepkey.mv_data = split;
7826 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7827 memcpy(rp->mp_ptrs, split, rsize);
7828 sepkey.mv_data = rp->mp_ptrs;
7829 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7830 memcpy(ins, newkey->mv_data, ksize);
7831 mp->mp_lower += sizeof(indx_t);
7832 mp->mp_upper -= ksize - sizeof(indx_t);
7835 memcpy(rp->mp_ptrs, split, x * ksize);
7836 ins = LEAF2KEY(rp, x, ksize);
7837 memcpy(ins, newkey->mv_data, ksize);
7838 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7839 rp->mp_lower += sizeof(indx_t);
7840 rp->mp_upper -= ksize - sizeof(indx_t);
7841 mc->mc_ki[mc->mc_top] = x;
7842 mc->mc_pg[mc->mc_top] = rp;
7845 int psize, nsize, k;
7846 /* Maximum free space in an empty page */
7847 pmax = env->me_psize - PAGEHDRSZ;
7849 nsize = mdb_leaf_size(env, newkey, newdata);
7851 nsize = mdb_branch_size(env, newkey);
7852 nsize = EVEN(nsize);
7854 /* grab a page to hold a temporary copy */
7855 copy = mdb_page_malloc(mc->mc_txn, 1);
7860 copy->mp_pgno = mp->mp_pgno;
7861 copy->mp_flags = mp->mp_flags;
7862 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7863 copy->mp_upper = env->me_psize - PAGEBASE;
7865 /* prepare to insert */
7866 for (i=0, j=0; i<nkeys; i++) {
7868 copy->mp_ptrs[j++] = 0;
7870 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7873 /* When items are relatively large the split point needs
7874 * to be checked, because being off-by-one will make the
7875 * difference between success or failure in mdb_node_add.
7877 * It's also relevant if a page happens to be laid out
7878 * such that one half of its nodes are all "small" and
7879 * the other half of its nodes are "large." If the new
7880 * item is also "large" and falls on the half with
7881 * "large" nodes, it also may not fit.
7883 * As a final tweak, if the new item goes on the last
7884 * spot on the page (and thus, onto the new page), bias
7885 * the split so the new page is emptier than the old page.
7886 * This yields better packing during sequential inserts.
7888 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7889 /* Find split point */
7891 if (newindx <= split_indx || newindx >= nkeys) {
7893 k = newindx >= nkeys ? nkeys : split_indx+2;
7898 for (; i!=k; i+=j) {
7903 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7904 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7906 if (F_ISSET(node->mn_flags, F_BIGDATA))
7907 psize += sizeof(pgno_t);
7909 psize += NODEDSZ(node);
7911 psize = EVEN(psize);
7913 if (psize > pmax || i == k-j) {
7914 split_indx = i + (j<0);
7919 if (split_indx == newindx) {
7920 sepkey.mv_size = newkey->mv_size;
7921 sepkey.mv_data = newkey->mv_data;
7923 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
7924 sepkey.mv_size = node->mn_ksize;
7925 sepkey.mv_data = NODEKEY(node);
7930 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7932 /* Copy separator key to the parent.
7934 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7938 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7943 if (mn.mc_snum == mc->mc_snum) {
7944 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7945 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7946 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7947 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7952 /* Right page might now have changed parent.
7953 * Check if left page also changed parent.
7955 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7956 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7957 for (i=0; i<ptop; i++) {
7958 mc->mc_pg[i] = mn.mc_pg[i];
7959 mc->mc_ki[i] = mn.mc_ki[i];
7961 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7962 if (mn.mc_ki[ptop]) {
7963 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7965 /* find right page's left sibling */
7966 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7967 mdb_cursor_sibling(mc, 0);
7972 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7975 mc->mc_flags ^= C_SPLITTING;
7976 if (rc != MDB_SUCCESS) {
7979 if (nflags & MDB_APPEND) {
7980 mc->mc_pg[mc->mc_top] = rp;
7981 mc->mc_ki[mc->mc_top] = 0;
7982 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7985 for (i=0; i<mc->mc_top; i++)
7986 mc->mc_ki[i] = mn.mc_ki[i];
7987 } else if (!IS_LEAF2(mp)) {
7989 mc->mc_pg[mc->mc_top] = rp;
7994 rkey.mv_data = newkey->mv_data;
7995 rkey.mv_size = newkey->mv_size;
8001 /* Update index for the new key. */
8002 mc->mc_ki[mc->mc_top] = j;
8004 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8005 rkey.mv_data = NODEKEY(node);
8006 rkey.mv_size = node->mn_ksize;
8008 xdata.mv_data = NODEDATA(node);
8009 xdata.mv_size = NODEDSZ(node);
8012 pgno = NODEPGNO(node);
8013 flags = node->mn_flags;
8016 if (!IS_LEAF(mp) && j == 0) {
8017 /* First branch index doesn't need key data. */
8021 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8027 mc->mc_pg[mc->mc_top] = copy;
8032 } while (i != split_indx);
8034 nkeys = NUMKEYS(copy);
8035 for (i=0; i<nkeys; i++)
8036 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8037 mp->mp_lower = copy->mp_lower;
8038 mp->mp_upper = copy->mp_upper;
8039 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8040 env->me_psize - copy->mp_upper - PAGEBASE);
8042 /* reset back to original page */
8043 if (newindx < split_indx) {
8044 mc->mc_pg[mc->mc_top] = mp;
8045 if (nflags & MDB_RESERVE) {
8046 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8047 if (!(node->mn_flags & F_BIGDATA))
8048 newdata->mv_data = NODEDATA(node);
8051 mc->mc_pg[mc->mc_top] = rp;
8053 /* Make sure mc_ki is still valid.
8055 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8056 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8057 for (i=0; i<=ptop; i++) {
8058 mc->mc_pg[i] = mn.mc_pg[i];
8059 mc->mc_ki[i] = mn.mc_ki[i];
8066 /* Adjust other cursors pointing to mp */
8067 MDB_cursor *m2, *m3;
8068 MDB_dbi dbi = mc->mc_dbi;
8069 int fixup = NUMKEYS(mp);
8071 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8072 if (mc->mc_flags & C_SUB)
8073 m3 = &m2->mc_xcursor->mx_cursor;
8078 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8080 if (m3->mc_flags & C_SPLITTING)
8085 for (k=m3->mc_top; k>=0; k--) {
8086 m3->mc_ki[k+1] = m3->mc_ki[k];
8087 m3->mc_pg[k+1] = m3->mc_pg[k];
8089 if (m3->mc_ki[0] >= split_indx) {
8094 m3->mc_pg[0] = mc->mc_pg[0];
8098 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8099 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8100 m3->mc_ki[mc->mc_top]++;
8101 if (m3->mc_ki[mc->mc_top] >= fixup) {
8102 m3->mc_pg[mc->mc_top] = rp;
8103 m3->mc_ki[mc->mc_top] -= fixup;
8104 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8106 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8107 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8112 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8115 if (copy) /* tmp page */
8116 mdb_page_free(env, copy);
8118 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8123 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8124 MDB_val *key, MDB_val *data, unsigned int flags)
8129 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8132 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8135 mdb_cursor_init(&mc, txn, dbi, &mx);
8136 return mdb_cursor_put(&mc, key, data, flags);
8140 #define MDB_WBUF (1024*1024)
8143 /** State needed for a compacting copy. */
8144 typedef struct mdb_copy {
8145 pthread_mutex_t mc_mutex;
8146 pthread_cond_t mc_cond;
8153 pgno_t mc_next_pgno;
8156 volatile int mc_new;
8161 /** Dedicated writer thread for compacting copy. */
8162 static THREAD_RET ESECT
8163 mdb_env_copythr(void *arg)
8167 int toggle = 0, wsize, rc;
8170 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8173 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8176 pthread_mutex_lock(&my->mc_mutex);
8178 pthread_cond_signal(&my->mc_cond);
8181 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8182 if (my->mc_new < 0) {
8187 wsize = my->mc_wlen[toggle];
8188 ptr = my->mc_wbuf[toggle];
8191 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8195 } else if (len > 0) {
8209 /* If there's an overflow page tail, write it too */
8210 if (my->mc_olen[toggle]) {
8211 wsize = my->mc_olen[toggle];
8212 ptr = my->mc_over[toggle];
8213 my->mc_olen[toggle] = 0;
8216 my->mc_wlen[toggle] = 0;
8218 pthread_cond_signal(&my->mc_cond);
8220 pthread_cond_signal(&my->mc_cond);
8221 pthread_mutex_unlock(&my->mc_mutex);
8222 return (THREAD_RET)0;
8226 /** Tell the writer thread there's a buffer ready to write */
8228 mdb_env_cthr_toggle(mdb_copy *my, int st)
8230 int toggle = my->mc_toggle ^ 1;
8231 pthread_mutex_lock(&my->mc_mutex);
8232 if (my->mc_status) {
8233 pthread_mutex_unlock(&my->mc_mutex);
8234 return my->mc_status;
8236 while (my->mc_new == 1)
8237 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8239 my->mc_toggle = toggle;
8240 pthread_cond_signal(&my->mc_cond);
8241 pthread_mutex_unlock(&my->mc_mutex);
8245 /** Depth-first tree traversal for compacting copy. */
8247 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8250 MDB_txn *txn = my->mc_txn;
8252 MDB_page *mo, *mp, *leaf;
8257 /* Empty DB, nothing to do */
8258 if (*pg == P_INVALID)
8265 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8268 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8272 /* Make cursor pages writable */
8273 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8277 for (i=0; i<mc.mc_top; i++) {
8278 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8279 mc.mc_pg[i] = (MDB_page *)ptr;
8280 ptr += my->mc_env->me_psize;
8283 /* This is writable space for a leaf page. Usually not needed. */
8284 leaf = (MDB_page *)ptr;
8286 toggle = my->mc_toggle;
8287 while (mc.mc_snum > 0) {
8289 mp = mc.mc_pg[mc.mc_top];
8293 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8294 for (i=0; i<n; i++) {
8295 ni = NODEPTR(mp, i);
8296 if (ni->mn_flags & F_BIGDATA) {
8300 /* Need writable leaf */
8302 mc.mc_pg[mc.mc_top] = leaf;
8303 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8305 ni = NODEPTR(mp, i);
8308 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8309 rc = mdb_page_get(txn, pg, &omp, NULL);
8312 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8313 rc = mdb_env_cthr_toggle(my, 1);
8316 toggle = my->mc_toggle;
8318 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8319 memcpy(mo, omp, my->mc_env->me_psize);
8320 mo->mp_pgno = my->mc_next_pgno;
8321 my->mc_next_pgno += omp->mp_pages;
8322 my->mc_wlen[toggle] += my->mc_env->me_psize;
8323 if (omp->mp_pages > 1) {
8324 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8325 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8326 rc = mdb_env_cthr_toggle(my, 1);
8329 toggle = my->mc_toggle;
8331 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8332 } else if (ni->mn_flags & F_SUBDATA) {
8335 /* Need writable leaf */
8337 mc.mc_pg[mc.mc_top] = leaf;
8338 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8340 ni = NODEPTR(mp, i);
8343 memcpy(&db, NODEDATA(ni), sizeof(db));
8344 my->mc_toggle = toggle;
8345 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8348 toggle = my->mc_toggle;
8349 memcpy(NODEDATA(ni), &db, sizeof(db));
8354 mc.mc_ki[mc.mc_top]++;
8355 if (mc.mc_ki[mc.mc_top] < n) {
8358 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8360 rc = mdb_page_get(txn, pg, &mp, NULL);
8365 mc.mc_ki[mc.mc_top] = 0;
8366 if (IS_BRANCH(mp)) {
8367 /* Whenever we advance to a sibling branch page,
8368 * we must proceed all the way down to its first leaf.
8370 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8373 mc.mc_pg[mc.mc_top] = mp;
8377 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8378 rc = mdb_env_cthr_toggle(my, 1);
8381 toggle = my->mc_toggle;
8383 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8384 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8385 mo->mp_pgno = my->mc_next_pgno++;
8386 my->mc_wlen[toggle] += my->mc_env->me_psize;
8388 /* Update parent if there is one */
8389 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8390 SETPGNO(ni, mo->mp_pgno);
8391 mdb_cursor_pop(&mc);
8393 /* Otherwise we're done */
8403 /** Copy environment with compaction. */
8405 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8410 MDB_txn *txn = NULL;
8415 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8416 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8417 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize);
8418 if (my.mc_wbuf[0] == NULL)
8421 pthread_mutex_init(&my.mc_mutex, NULL);
8422 pthread_cond_init(&my.mc_cond, NULL);
8423 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2);
8427 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8428 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8433 my.mc_next_pgno = 2;
8439 THREAD_CREATE(thr, mdb_env_copythr, &my);
8441 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8445 mp = (MDB_page *)my.mc_wbuf[0];
8446 memset(mp, 0, 2*env->me_psize);
8448 mp->mp_flags = P_META;
8449 mm = (MDB_meta *)METADATA(mp);
8450 mdb_env_init_meta0(env, mm);
8451 mm->mm_address = env->me_metas[0]->mm_address;
8453 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8455 mp->mp_flags = P_META;
8456 *(MDB_meta *)METADATA(mp) = *mm;
8457 mm = (MDB_meta *)METADATA(mp);
8459 /* Count the number of free pages, subtract from lastpg to find
8460 * number of active pages
8463 MDB_ID freecount = 0;
8466 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8467 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8468 freecount += *(MDB_ID *)data.mv_data;
8469 freecount += txn->mt_dbs[0].md_branch_pages +
8470 txn->mt_dbs[0].md_leaf_pages +
8471 txn->mt_dbs[0].md_overflow_pages;
8473 /* Set metapage 1 */
8474 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8475 mm->mm_dbs[1] = txn->mt_dbs[1];
8476 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8479 my.mc_wlen[0] = env->me_psize * 2;
8481 pthread_mutex_lock(&my.mc_mutex);
8483 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8484 pthread_mutex_unlock(&my.mc_mutex);
8485 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8486 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8487 rc = mdb_env_cthr_toggle(&my, 1);
8488 mdb_env_cthr_toggle(&my, -1);
8489 pthread_mutex_lock(&my.mc_mutex);
8491 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8492 pthread_mutex_unlock(&my.mc_mutex);
8497 CloseHandle(my.mc_cond);
8498 CloseHandle(my.mc_mutex);
8499 _aligned_free(my.mc_wbuf[0]);
8501 pthread_cond_destroy(&my.mc_cond);
8502 pthread_mutex_destroy(&my.mc_mutex);
8503 free(my.mc_wbuf[0]);
8508 /** Copy environment as-is. */
8510 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8512 MDB_txn *txn = NULL;
8518 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8522 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8525 /* Do the lock/unlock of the reader mutex before starting the
8526 * write txn. Otherwise other read txns could block writers.
8528 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8533 /* We must start the actual read txn after blocking writers */
8534 mdb_txn_reset0(txn, "reset-stage1");
8536 /* Temporarily block writers until we snapshot the meta pages */
8539 rc = mdb_txn_renew0(txn);
8541 UNLOCK_MUTEX_W(env);
8546 wsize = env->me_psize * 2;
8550 DO_WRITE(rc, fd, ptr, w2, len);
8554 } else if (len > 0) {
8560 /* Non-blocking or async handles are not supported */
8566 UNLOCK_MUTEX_W(env);
8571 w2 = txn->mt_next_pgno * env->me_psize;
8574 LARGE_INTEGER fsize;
8575 GetFileSizeEx(env->me_fd, &fsize);
8576 if (w2 > fsize.QuadPart)
8577 w2 = fsize.QuadPart;
8582 fstat(env->me_fd, &st);
8583 if (w2 > (size_t)st.st_size)
8589 if (wsize > MAX_WRITE)
8593 DO_WRITE(rc, fd, ptr, w2, len);
8597 } else if (len > 0) {
8614 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8616 if (flags & MDB_CP_COMPACT)
8617 return mdb_env_copyfd1(env, fd);
8619 return mdb_env_copyfd0(env, fd);
8623 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8625 return mdb_env_copyfd2(env, fd, 0);
8629 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8633 HANDLE newfd = INVALID_HANDLE_VALUE;
8635 if (env->me_flags & MDB_NOSUBDIR) {
8636 lpath = (char *)path;
8639 len += sizeof(DATANAME);
8640 lpath = malloc(len);
8643 sprintf(lpath, "%s" DATANAME, path);
8646 /* The destination path must exist, but the destination file must not.
8647 * We don't want the OS to cache the writes, since the source data is
8648 * already in the OS cache.
8651 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8652 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8654 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8656 if (newfd == INVALID_HANDLE_VALUE) {
8662 /* Set O_DIRECT if the file system supports it */
8663 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8664 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8666 #ifdef F_NOCACHE /* __APPLE__ */
8667 rc = fcntl(newfd, F_NOCACHE, 1);
8674 rc = mdb_env_copyfd2(env, newfd, flags);
8677 if (!(env->me_flags & MDB_NOSUBDIR))
8679 if (newfd != INVALID_HANDLE_VALUE)
8680 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8687 mdb_env_copy(MDB_env *env, const char *path)
8689 return mdb_env_copy2(env, path, 0);
8693 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8695 if ((flag & CHANGEABLE) != flag)
8698 env->me_flags |= flag;
8700 env->me_flags &= ~flag;
8705 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8710 *arg = env->me_flags;
8715 mdb_env_set_userctx(MDB_env *env, void *ctx)
8719 env->me_userctx = ctx;
8724 mdb_env_get_userctx(MDB_env *env)
8726 return env ? env->me_userctx : NULL;
8730 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8735 env->me_assert_func = func;
8741 mdb_env_get_path(MDB_env *env, const char **arg)
8746 *arg = env->me_path;
8751 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8760 /** Common code for #mdb_stat() and #mdb_env_stat().
8761 * @param[in] env the environment to operate in.
8762 * @param[in] db the #MDB_db record containing the stats to return.
8763 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8764 * @return 0, this function always succeeds.
8767 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8769 arg->ms_psize = env->me_psize;
8770 arg->ms_depth = db->md_depth;
8771 arg->ms_branch_pages = db->md_branch_pages;
8772 arg->ms_leaf_pages = db->md_leaf_pages;
8773 arg->ms_overflow_pages = db->md_overflow_pages;
8774 arg->ms_entries = db->md_entries;
8780 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8784 if (env == NULL || arg == NULL)
8787 toggle = mdb_env_pick_meta(env);
8789 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8793 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8797 if (env == NULL || arg == NULL)
8800 toggle = mdb_env_pick_meta(env);
8801 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8802 arg->me_mapsize = env->me_mapsize;
8803 arg->me_maxreaders = env->me_maxreaders;
8805 /* me_numreaders may be zero if this process never used any readers. Use
8806 * the shared numreader count if it exists.
8808 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8810 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8811 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8815 /** Set the default comparison functions for a database.
8816 * Called immediately after a database is opened to set the defaults.
8817 * The user can then override them with #mdb_set_compare() or
8818 * #mdb_set_dupsort().
8819 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8820 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8823 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8825 uint16_t f = txn->mt_dbs[dbi].md_flags;
8827 txn->mt_dbxs[dbi].md_cmp =
8828 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8829 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8831 txn->mt_dbxs[dbi].md_dcmp =
8832 !(f & MDB_DUPSORT) ? 0 :
8833 ((f & MDB_INTEGERDUP)
8834 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8835 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8838 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8843 int rc, dbflag, exact;
8844 unsigned int unused = 0, seq;
8847 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8848 mdb_default_cmp(txn, FREE_DBI);
8851 if ((flags & VALID_FLAGS) != flags)
8853 if (txn->mt_flags & MDB_TXN_ERROR)
8859 if (flags & PERSISTENT_FLAGS) {
8860 uint16_t f2 = flags & PERSISTENT_FLAGS;
8861 /* make sure flag changes get committed */
8862 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8863 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8864 txn->mt_flags |= MDB_TXN_DIRTY;
8867 mdb_default_cmp(txn, MAIN_DBI);
8871 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8872 mdb_default_cmp(txn, MAIN_DBI);
8875 /* Is the DB already open? */
8877 for (i=2; i<txn->mt_numdbs; i++) {
8878 if (!txn->mt_dbxs[i].md_name.mv_size) {
8879 /* Remember this free slot */
8880 if (!unused) unused = i;
8883 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8884 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8890 /* If no free slot and max hit, fail */
8891 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8892 return MDB_DBS_FULL;
8894 /* Cannot mix named databases with some mainDB flags */
8895 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8896 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8898 /* Find the DB info */
8899 dbflag = DB_NEW|DB_VALID;
8902 key.mv_data = (void *)name;
8903 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8904 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8905 if (rc == MDB_SUCCESS) {
8906 /* make sure this is actually a DB */
8907 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8908 if (!(node->mn_flags & F_SUBDATA))
8909 return MDB_INCOMPATIBLE;
8910 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8911 /* Create if requested */
8913 data.mv_size = sizeof(MDB_db);
8914 data.mv_data = &dummy;
8915 memset(&dummy, 0, sizeof(dummy));
8916 dummy.md_root = P_INVALID;
8917 dummy.md_flags = flags & PERSISTENT_FLAGS;
8918 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8922 /* OK, got info, add to table */
8923 if (rc == MDB_SUCCESS) {
8924 unsigned int slot = unused ? unused : txn->mt_numdbs;
8925 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8926 txn->mt_dbxs[slot].md_name.mv_size = len;
8927 txn->mt_dbxs[slot].md_rel = NULL;
8928 txn->mt_dbflags[slot] = dbflag;
8929 /* txn-> and env-> are the same in read txns, use
8930 * tmp variable to avoid undefined assignment
8932 seq = ++txn->mt_env->me_dbiseqs[slot];
8933 txn->mt_dbiseqs[slot] = seq;
8935 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8937 mdb_default_cmp(txn, slot);
8946 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8948 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8951 if (txn->mt_flags & MDB_TXN_ERROR)
8954 if (txn->mt_dbflags[dbi] & DB_STALE) {
8957 /* Stale, must read the DB's root. cursor_init does it for us. */
8958 mdb_cursor_init(&mc, txn, dbi, &mx);
8960 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8963 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8966 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8968 ptr = env->me_dbxs[dbi].md_name.mv_data;
8969 /* If there was no name, this was already closed */
8971 env->me_dbxs[dbi].md_name.mv_data = NULL;
8972 env->me_dbxs[dbi].md_name.mv_size = 0;
8973 env->me_dbflags[dbi] = 0;
8974 env->me_dbiseqs[dbi]++;
8979 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8981 /* We could return the flags for the FREE_DBI too but what's the point? */
8982 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8984 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8988 /** Add all the DB's pages to the free list.
8989 * @param[in] mc Cursor on the DB to free.
8990 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8991 * @return 0 on success, non-zero on failure.
8994 mdb_drop0(MDB_cursor *mc, int subs)
8998 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8999 if (rc == MDB_SUCCESS) {
9000 MDB_txn *txn = mc->mc_txn;
9005 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9006 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9009 mdb_cursor_copy(mc, &mx);
9010 while (mc->mc_snum > 0) {
9011 MDB_page *mp = mc->mc_pg[mc->mc_top];
9012 unsigned n = NUMKEYS(mp);
9014 for (i=0; i<n; i++) {
9015 ni = NODEPTR(mp, i);
9016 if (ni->mn_flags & F_BIGDATA) {
9019 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9020 rc = mdb_page_get(txn, pg, &omp, NULL);
9023 mdb_cassert(mc, IS_OVERFLOW(omp));
9024 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9028 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9029 mdb_xcursor_init1(mc, ni);
9030 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9036 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9038 for (i=0; i<n; i++) {
9040 ni = NODEPTR(mp, i);
9043 mdb_midl_xappend(txn->mt_free_pgs, pg);
9048 mc->mc_ki[mc->mc_top] = i;
9049 rc = mdb_cursor_sibling(mc, 1);
9051 if (rc != MDB_NOTFOUND)
9053 /* no more siblings, go back to beginning
9054 * of previous level.
9058 for (i=1; i<mc->mc_snum; i++) {
9060 mc->mc_pg[i] = mx.mc_pg[i];
9065 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9068 txn->mt_flags |= MDB_TXN_ERROR;
9069 } else if (rc == MDB_NOTFOUND) {
9075 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9077 MDB_cursor *mc, *m2;
9080 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9083 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9086 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9089 rc = mdb_cursor_open(txn, dbi, &mc);
9093 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9094 /* Invalidate the dropped DB's cursors */
9095 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9096 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9100 /* Can't delete the main DB */
9101 if (del && dbi > MAIN_DBI) {
9102 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9104 txn->mt_dbflags[dbi] = DB_STALE;
9105 mdb_dbi_close(txn->mt_env, dbi);
9107 txn->mt_flags |= MDB_TXN_ERROR;
9110 /* reset the DB record, mark it dirty */
9111 txn->mt_dbflags[dbi] |= DB_DIRTY;
9112 txn->mt_dbs[dbi].md_depth = 0;
9113 txn->mt_dbs[dbi].md_branch_pages = 0;
9114 txn->mt_dbs[dbi].md_leaf_pages = 0;
9115 txn->mt_dbs[dbi].md_overflow_pages = 0;
9116 txn->mt_dbs[dbi].md_entries = 0;
9117 txn->mt_dbs[dbi].md_root = P_INVALID;
9119 txn->mt_flags |= MDB_TXN_DIRTY;
9122 mdb_cursor_close(mc);
9126 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9128 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9131 txn->mt_dbxs[dbi].md_cmp = cmp;
9135 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9137 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9140 txn->mt_dbxs[dbi].md_dcmp = cmp;
9144 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9146 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9149 txn->mt_dbxs[dbi].md_rel = rel;
9153 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9155 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9158 txn->mt_dbxs[dbi].md_relctx = ctx;
9163 mdb_env_get_maxkeysize(MDB_env *env)
9165 return ENV_MAXKEY(env);
9169 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9171 unsigned int i, rdrs;
9174 int rc = 0, first = 1;
9178 if (!env->me_txns) {
9179 return func("(no reader locks)\n", ctx);
9181 rdrs = env->me_txns->mti_numreaders;
9182 mr = env->me_txns->mti_readers;
9183 for (i=0; i<rdrs; i++) {
9185 txnid_t txnid = mr[i].mr_txnid;
9186 sprintf(buf, txnid == (txnid_t)-1 ?
9187 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9188 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9191 rc = func(" pid thread txnid\n", ctx);
9195 rc = func(buf, ctx);
9201 rc = func("(no active readers)\n", ctx);
9206 /** Insert pid into list if not already present.
9207 * return -1 if already present.
9210 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9212 /* binary search of pid in list */
9214 unsigned cursor = 1;
9216 unsigned n = ids[0];
9219 unsigned pivot = n >> 1;
9220 cursor = base + pivot + 1;
9221 val = pid - ids[cursor];
9226 } else if ( val > 0 ) {
9231 /* found, so it's a duplicate */
9240 for (n = ids[0]; n > cursor; n--)
9247 mdb_reader_check(MDB_env *env, int *dead)
9249 unsigned int i, j, rdrs;
9251 MDB_PID_T *pids, pid;
9260 rdrs = env->me_txns->mti_numreaders;
9261 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9265 mr = env->me_txns->mti_readers;
9266 for (i=0; i<rdrs; i++) {
9267 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9269 if (mdb_pid_insert(pids, pid) == 0) {
9270 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9272 /* Recheck, a new process may have reused pid */
9273 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9274 for (j=i; j<rdrs; j++)
9275 if (mr[j].mr_pid == pid) {
9276 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9277 (unsigned) pid, mr[j].mr_txnid));
9282 UNLOCK_MUTEX_R(env);