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 */
155 #define ESECT __attribute__ ((section("text_env")))
160 /** @defgroup internal LMDB Internals
163 /** @defgroup compat Compatibility Macros
164 * A bunch of macros to minimize the amount of platform-specific ifdefs
165 * needed throughout the rest of the code. When the features this library
166 * needs are similar enough to POSIX to be hidden in a one-or-two line
167 * replacement, this macro approach is used.
171 /** Wrapper around __func__, which is a C99 feature */
172 #if __STDC_VERSION__ >= 199901L
173 # define mdb_func_ __func__
174 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
175 # define mdb_func_ __FUNCTION__
177 /* If a debug message says <mdb_unknown>(), update the #if statements above */
178 # define mdb_func_ "<mdb_unknown>"
182 #define MDB_USE_HASH 1
183 #define MDB_PIDLOCK 0
184 #define THREAD_RET DWORD
185 #define pthread_t HANDLE
186 #define pthread_mutex_t HANDLE
187 #define pthread_cond_t HANDLE
188 #define pthread_key_t DWORD
189 #define pthread_self() GetCurrentThreadId()
190 #define pthread_key_create(x,y) \
191 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
192 #define pthread_key_delete(x) TlsFree(x)
193 #define pthread_getspecific(x) TlsGetValue(x)
194 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
195 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
196 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
197 #define pthread_cond_signal(x) SetEvent(*x)
198 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
199 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
200 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
201 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
202 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
203 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
204 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
205 #define getpid() GetCurrentProcessId()
206 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
207 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
208 #define ErrCode() GetLastError()
209 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
210 #define close(fd) (CloseHandle(fd) ? 0 : -1)
211 #define munmap(ptr,len) UnmapViewOfFile(ptr)
212 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
213 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
215 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
219 #define THREAD_RET void *
220 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
221 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
222 #define Z "z" /**< printf format modifier for size_t */
224 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
225 #define MDB_PIDLOCK 1
227 #ifdef MDB_USE_POSIX_SEM
229 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
230 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
231 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
232 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
235 mdb_sem_wait(sem_t *sem)
238 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
243 /** Lock the reader mutex.
245 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
246 /** Unlock the reader mutex.
248 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
250 /** Lock the writer mutex.
251 * Only a single write transaction is allowed at a time. Other writers
252 * will block waiting for this mutex.
254 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
255 /** Unlock the writer mutex.
257 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
258 #endif /* MDB_USE_POSIX_SEM */
260 /** Get the error code for the last failed system function.
262 #define ErrCode() errno
264 /** An abstraction for a file handle.
265 * On POSIX systems file handles are small integers. On Windows
266 * they're opaque pointers.
270 /** A value for an invalid file handle.
271 * Mainly used to initialize file variables and signify that they are
274 #define INVALID_HANDLE_VALUE (-1)
276 /** Get the size of a memory page for the system.
277 * This is the basic size that the platform's memory manager uses, and is
278 * fundamental to the use of memory-mapped files.
280 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
283 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
286 #define MNAME_LEN (sizeof(pthread_mutex_t))
292 /** A flag for opening a file and requesting synchronous data writes.
293 * This is only used when writing a meta page. It's not strictly needed;
294 * we could just do a normal write and then immediately perform a flush.
295 * But if this flag is available it saves us an extra system call.
297 * @note If O_DSYNC is undefined but exists in /usr/include,
298 * preferably set some compiler flag to get the definition.
299 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
302 # define MDB_DSYNC O_DSYNC
306 /** Function for flushing the data of a file. Define this to fsync
307 * if fdatasync() is not supported.
309 #ifndef MDB_FDATASYNC
310 # define MDB_FDATASYNC fdatasync
314 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
325 /** A page number in the database.
326 * Note that 64 bit page numbers are overkill, since pages themselves
327 * already represent 12-13 bits of addressable memory, and the OS will
328 * always limit applications to a maximum of 63 bits of address space.
330 * @note In the #MDB_node structure, we only store 48 bits of this value,
331 * which thus limits us to only 60 bits of addressable data.
333 typedef MDB_ID pgno_t;
335 /** A transaction ID.
336 * See struct MDB_txn.mt_txnid for details.
338 typedef MDB_ID txnid_t;
340 /** @defgroup debug Debug Macros
344 /** Enable debug output. Needs variable argument macros (a C99 feature).
345 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
346 * read from and written to the database (used for free space management).
352 static int mdb_debug;
353 static txnid_t mdb_debug_start;
355 /** Print a debug message with printf formatting.
356 * Requires double parenthesis around 2 or more args.
358 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
359 # define DPRINTF0(fmt, ...) \
360 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
362 # define DPRINTF(args) ((void) 0)
364 /** Print a debug string.
365 * The string is printed literally, with no format processing.
367 #define DPUTS(arg) DPRINTF(("%s", arg))
368 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
370 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
373 /** @brief The maximum size of a database page.
375 * This is 32k, since it must fit in #MDB_page.%mp_upper.
377 * LMDB will use database pages < OS pages if needed.
378 * That causes more I/O in write transactions: The OS must
379 * know (read) the whole page before writing a partial page.
381 * Note that we don't currently support Huge pages. On Linux,
382 * regular data files cannot use Huge pages, and in general
383 * Huge pages aren't actually pageable. We rely on the OS
384 * demand-pager to read our data and page it out when memory
385 * pressure from other processes is high. So until OSs have
386 * actual paging support for Huge pages, they're not viable.
388 #define MAX_PAGESIZE 0x8000
390 /** The minimum number of keys required in a database page.
391 * Setting this to a larger value will place a smaller bound on the
392 * maximum size of a data item. Data items larger than this size will
393 * be pushed into overflow pages instead of being stored directly in
394 * the B-tree node. This value used to default to 4. With a page size
395 * of 4096 bytes that meant that any item larger than 1024 bytes would
396 * go into an overflow page. That also meant that on average 2-3KB of
397 * each overflow page was wasted space. The value cannot be lower than
398 * 2 because then there would no longer be a tree structure. With this
399 * value, items larger than 2KB will go into overflow pages, and on
400 * average only 1KB will be wasted.
402 #define MDB_MINKEYS 2
404 /** A stamp that identifies a file as an LMDB file.
405 * There's nothing special about this value other than that it is easily
406 * recognizable, and it will reflect any byte order mismatches.
408 #define MDB_MAGIC 0xBEEFC0DE
410 /** The version number for a database's datafile format. */
411 #define MDB_DATA_VERSION 1
412 /** The version number for a database's lockfile format. */
413 #define MDB_LOCK_VERSION 1
415 /** @brief The max size of a key we can write, or 0 for dynamic max.
417 * Define this as 0 to compute the max from the page size. 511
418 * is default for backwards compat: liblmdb <= 0.9.10 can break
419 * when modifying a DB with keys/dupsort data bigger than its max.
421 * Data items in an #MDB_DUPSORT database are also limited to
422 * this size, since they're actually keys of a sub-DB. Keys and
423 * #MDB_DUPSORT data items must fit on a node in a regular page.
425 #ifndef MDB_MAXKEYSIZE
426 #define MDB_MAXKEYSIZE 511
429 /** The maximum size of a key we can write to the environment. */
431 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
433 #define ENV_MAXKEY(env) ((env)->me_maxkey)
436 /** @brief The maximum size of a data item.
438 * We only store a 32 bit value for node sizes.
440 #define MAXDATASIZE 0xffffffffUL
443 /** Key size which fits in a #DKBUF.
446 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
449 * This is used for printing a hex dump of a key's contents.
451 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
452 /** Display a key in hex.
454 * Invoke a function to display a key in hex.
456 #define DKEY(x) mdb_dkey(x, kbuf)
462 /** An invalid page number.
463 * Mainly used to denote an empty tree.
465 #define P_INVALID (~(pgno_t)0)
467 /** Test if the flags \b f are set in a flag word \b w. */
468 #define F_ISSET(w, f) (((w) & (f)) == (f))
470 /** Round \b n up to an even number. */
471 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
473 /** Used for offsets within a single page.
474 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
477 typedef uint16_t indx_t;
479 /** Default size of memory map.
480 * This is certainly too small for any actual applications. Apps should always set
481 * the size explicitly using #mdb_env_set_mapsize().
483 #define DEFAULT_MAPSIZE 1048576
485 /** @defgroup readers Reader Lock Table
486 * Readers don't acquire any locks for their data access. Instead, they
487 * simply record their transaction ID in the reader table. The reader
488 * mutex is needed just to find an empty slot in the reader table. The
489 * slot's address is saved in thread-specific data so that subsequent read
490 * transactions started by the same thread need no further locking to proceed.
492 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
494 * No reader table is used if the database is on a read-only filesystem, or
495 * if #MDB_NOLOCK is set.
497 * Since the database uses multi-version concurrency control, readers don't
498 * actually need any locking. This table is used to keep track of which
499 * readers are using data from which old transactions, so that we'll know
500 * when a particular old transaction is no longer in use. Old transactions
501 * that have discarded any data pages can then have those pages reclaimed
502 * for use by a later write transaction.
504 * The lock table is constructed such that reader slots are aligned with the
505 * processor's cache line size. Any slot is only ever used by one thread.
506 * This alignment guarantees that there will be no contention or cache
507 * thrashing as threads update their own slot info, and also eliminates
508 * any need for locking when accessing a slot.
510 * A writer thread will scan every slot in the table to determine the oldest
511 * outstanding reader transaction. Any freed pages older than this will be
512 * reclaimed by the writer. The writer doesn't use any locks when scanning
513 * this table. This means that there's no guarantee that the writer will
514 * see the most up-to-date reader info, but that's not required for correct
515 * operation - all we need is to know the upper bound on the oldest reader,
516 * we don't care at all about the newest reader. So the only consequence of
517 * reading stale information here is that old pages might hang around a
518 * while longer before being reclaimed. That's actually good anyway, because
519 * the longer we delay reclaiming old pages, the more likely it is that a
520 * string of contiguous pages can be found after coalescing old pages from
521 * many old transactions together.
524 /** Number of slots in the reader table.
525 * This value was chosen somewhat arbitrarily. 126 readers plus a
526 * couple mutexes fit exactly into 8KB on my development machine.
527 * Applications should set the table size using #mdb_env_set_maxreaders().
529 #define DEFAULT_READERS 126
531 /** The size of a CPU cache line in bytes. We want our lock structures
532 * aligned to this size to avoid false cache line sharing in the
534 * This value works for most CPUs. For Itanium this should be 128.
540 /** The information we store in a single slot of the reader table.
541 * In addition to a transaction ID, we also record the process and
542 * thread ID that owns a slot, so that we can detect stale information,
543 * e.g. threads or processes that went away without cleaning up.
544 * @note We currently don't check for stale records. We simply re-init
545 * the table when we know that we're the only process opening the
548 typedef struct MDB_rxbody {
549 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
550 * Multiple readers that start at the same time will probably have the
551 * same ID here. Again, it's not important to exclude them from
552 * anything; all we need to know is which version of the DB they
553 * started from so we can avoid overwriting any data used in that
554 * particular version.
557 /** The process ID of the process owning this reader txn. */
559 /** The thread ID of the thread owning this txn. */
563 /** The actual reader record, with cacheline padding. */
564 typedef struct MDB_reader {
567 /** shorthand for mrb_txnid */
568 #define mr_txnid mru.mrx.mrb_txnid
569 #define mr_pid mru.mrx.mrb_pid
570 #define mr_tid mru.mrx.mrb_tid
571 /** cache line alignment */
572 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
576 /** The header for the reader table.
577 * The table resides in a memory-mapped file. (This is a different file
578 * than is used for the main database.)
580 * For POSIX the actual mutexes reside in the shared memory of this
581 * mapped file. On Windows, mutexes are named objects allocated by the
582 * kernel; we store the mutex names in this mapped file so that other
583 * processes can grab them. This same approach is also used on
584 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
585 * process-shared POSIX mutexes. For these cases where a named object
586 * is used, the object name is derived from a 64 bit FNV hash of the
587 * environment pathname. As such, naming collisions are extremely
588 * unlikely. If a collision occurs, the results are unpredictable.
590 typedef struct MDB_txbody {
591 /** Stamp identifying this as an LMDB file. It must be set
594 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
596 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
597 char mtb_rmname[MNAME_LEN];
599 /** Mutex protecting access to this table.
600 * This is the reader lock that #LOCK_MUTEX_R acquires.
602 pthread_mutex_t mtb_mutex;
604 /** The ID of the last transaction committed to the database.
605 * This is recorded here only for convenience; the value can always
606 * be determined by reading the main database meta pages.
609 /** The number of slots that have been used in the reader table.
610 * This always records the maximum count, it is not decremented
611 * when readers release their slots.
613 unsigned mtb_numreaders;
616 /** The actual reader table definition. */
617 typedef struct MDB_txninfo {
620 #define mti_magic mt1.mtb.mtb_magic
621 #define mti_format mt1.mtb.mtb_format
622 #define mti_mutex mt1.mtb.mtb_mutex
623 #define mti_rmname mt1.mtb.mtb_rmname
624 #define mti_txnid mt1.mtb.mtb_txnid
625 #define mti_numreaders mt1.mtb.mtb_numreaders
626 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
629 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
630 char mt2_wmname[MNAME_LEN];
631 #define mti_wmname mt2.mt2_wmname
633 pthread_mutex_t mt2_wmutex;
634 #define mti_wmutex mt2.mt2_wmutex
636 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
638 MDB_reader mti_readers[1];
641 /** Lockfile format signature: version, features and field layout */
642 #define MDB_LOCK_FORMAT \
644 ((MDB_LOCK_VERSION) \
645 /* Flags which describe functionality */ \
646 + (((MDB_PIDLOCK) != 0) << 16)))
649 /** Common header for all page types.
650 * Overflow records occupy a number of contiguous pages with no
651 * headers on any page after the first.
653 typedef struct MDB_page {
654 #define mp_pgno mp_p.p_pgno
655 #define mp_next mp_p.p_next
657 pgno_t p_pgno; /**< page number */
658 void * p_next; /**< for in-memory list of freed structs */
661 /** @defgroup mdb_page Page Flags
663 * Flags for the page headers.
666 #define P_BRANCH 0x01 /**< branch page */
667 #define P_LEAF 0x02 /**< leaf page */
668 #define P_OVERFLOW 0x04 /**< overflow page */
669 #define P_META 0x08 /**< meta page */
670 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
671 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
672 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
673 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
674 #define P_KEEP 0x8000 /**< leave this page alone during spill */
676 uint16_t mp_flags; /**< @ref mdb_page */
677 #define mp_lower mp_pb.pb.pb_lower
678 #define mp_upper mp_pb.pb.pb_upper
679 #define mp_pages mp_pb.pb_pages
682 indx_t pb_lower; /**< lower bound of free space */
683 indx_t pb_upper; /**< upper bound of free space */
685 uint32_t pb_pages; /**< number of overflow pages */
687 indx_t mp_ptrs[1]; /**< dynamic size */
690 /** Size of the page header, excluding dynamic data at the end */
691 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
693 /** Address of first usable data byte in a page, after the header */
694 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
696 /** Number of nodes on a page */
697 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
699 /** The amount of space remaining in the page */
700 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
702 /** The percentage of space used in the page, in tenths of a percent. */
703 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
704 ((env)->me_psize - PAGEHDRSZ))
705 /** The minimum page fill factor, in tenths of a percent.
706 * Pages emptier than this are candidates for merging.
708 #define FILL_THRESHOLD 250
710 /** Test if a page is a leaf page */
711 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
712 /** Test if a page is a LEAF2 page */
713 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
714 /** Test if a page is a branch page */
715 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
716 /** Test if a page is an overflow page */
717 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
718 /** Test if a page is a sub page */
719 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
721 /** The number of overflow pages needed to store the given size. */
722 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
724 /** Link in #MDB_txn.%mt_loose_pages list */
725 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)METADATA(p))
727 /** Header for a single key/data pair within a page.
728 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
729 * We guarantee 2-byte alignment for 'MDB_node's.
731 typedef struct MDB_node {
732 /** lo and hi are used for data size on leaf nodes and for
733 * child pgno on branch nodes. On 64 bit platforms, flags
734 * is also used for pgno. (Branch nodes have no flags).
735 * They are in host byte order in case that lets some
736 * accesses be optimized into a 32-bit word access.
738 #if BYTE_ORDER == LITTLE_ENDIAN
739 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
741 unsigned short mn_hi, mn_lo;
743 /** @defgroup mdb_node Node Flags
745 * Flags for node headers.
748 #define F_BIGDATA 0x01 /**< data put on overflow page */
749 #define F_SUBDATA 0x02 /**< data is a sub-database */
750 #define F_DUPDATA 0x04 /**< data has duplicates */
752 /** valid flags for #mdb_node_add() */
753 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
756 unsigned short mn_flags; /**< @ref mdb_node */
757 unsigned short mn_ksize; /**< key size */
758 char mn_data[1]; /**< key and data are appended here */
761 /** Size of the node header, excluding dynamic data at the end */
762 #define NODESIZE offsetof(MDB_node, mn_data)
764 /** Bit position of top word in page number, for shifting mn_flags */
765 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
767 /** Size of a node in a branch page with a given key.
768 * This is just the node header plus the key, there is no data.
770 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
772 /** Size of a node in a leaf page with a given key and data.
773 * This is node header plus key plus data size.
775 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
777 /** Address of node \b i in page \b p */
778 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
780 /** Address of the key for the node */
781 #define NODEKEY(node) (void *)((node)->mn_data)
783 /** Address of the data for a node */
784 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
786 /** Get the page number pointed to by a branch node */
787 #define NODEPGNO(node) \
788 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
789 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
790 /** Set the page number in a branch node */
791 #define SETPGNO(node,pgno) do { \
792 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
793 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
795 /** Get the size of the data in a leaf node */
796 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
797 /** Set the size of the data for a leaf node */
798 #define SETDSZ(node,size) do { \
799 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
800 /** The size of a key in a node */
801 #define NODEKSZ(node) ((node)->mn_ksize)
803 /** Copy a page number from src to dst */
805 #define COPY_PGNO(dst,src) dst = src
807 #if SIZE_MAX > 4294967295UL
808 #define COPY_PGNO(dst,src) do { \
809 unsigned short *s, *d; \
810 s = (unsigned short *)&(src); \
811 d = (unsigned short *)&(dst); \
818 #define COPY_PGNO(dst,src) do { \
819 unsigned short *s, *d; \
820 s = (unsigned short *)&(src); \
821 d = (unsigned short *)&(dst); \
827 /** The address of a key in a LEAF2 page.
828 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
829 * There are no node headers, keys are stored contiguously.
831 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
833 /** Set the \b node's key into \b keyptr, if requested. */
834 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
835 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
837 /** Set the \b node's key into \b key. */
838 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
840 /** Information about a single database in the environment. */
841 typedef struct MDB_db {
842 uint32_t md_pad; /**< also ksize for LEAF2 pages */
843 uint16_t md_flags; /**< @ref mdb_dbi_open */
844 uint16_t md_depth; /**< depth of this tree */
845 pgno_t md_branch_pages; /**< number of internal pages */
846 pgno_t md_leaf_pages; /**< number of leaf pages */
847 pgno_t md_overflow_pages; /**< number of overflow pages */
848 size_t md_entries; /**< number of data items */
849 pgno_t md_root; /**< the root page of this tree */
852 /** mdb_dbi_open flags */
853 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
854 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
855 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
856 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
858 /** Handle for the DB used to track free pages. */
860 /** Handle for the default DB. */
863 /** Meta page content.
864 * A meta page is the start point for accessing a database snapshot.
865 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
867 typedef struct MDB_meta {
868 /** Stamp identifying this as an LMDB file. It must be set
871 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
873 void *mm_address; /**< address for fixed mapping */
874 size_t mm_mapsize; /**< size of mmap region */
875 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
876 /** The size of pages used in this DB */
877 #define mm_psize mm_dbs[0].md_pad
878 /** Any persistent environment flags. @ref mdb_env */
879 #define mm_flags mm_dbs[0].md_flags
880 pgno_t mm_last_pg; /**< last used page in file */
881 txnid_t mm_txnid; /**< txnid that committed this page */
884 /** Buffer for a stack-allocated meta page.
885 * The members define size and alignment, and silence type
886 * aliasing warnings. They are not used directly; that could
887 * mean incorrectly using several union members in parallel.
889 typedef union MDB_metabuf {
892 char mm_pad[PAGEHDRSZ];
897 /** Auxiliary DB info.
898 * The information here is mostly static/read-only. There is
899 * only a single copy of this record in the environment.
901 typedef struct MDB_dbx {
902 MDB_val md_name; /**< name of the database */
903 MDB_cmp_func *md_cmp; /**< function for comparing keys */
904 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
905 MDB_rel_func *md_rel; /**< user relocate function */
906 void *md_relctx; /**< user-provided context for md_rel */
909 /** A database transaction.
910 * Every operation requires a transaction handle.
913 MDB_txn *mt_parent; /**< parent of a nested txn */
914 MDB_txn *mt_child; /**< nested txn under this txn */
915 pgno_t mt_next_pgno; /**< next unallocated page */
916 /** The ID of this transaction. IDs are integers incrementing from 1.
917 * Only committed write transactions increment the ID. If a transaction
918 * aborts, the ID may be re-used by the next writer.
921 MDB_env *mt_env; /**< the DB environment */
922 /** The list of pages that became unused during this transaction.
925 /** The list of loose pages that became unused and may be reused
926 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
928 MDB_page *mt_loose_pgs;
929 /** The sorted list of dirty pages we temporarily wrote to disk
930 * because the dirty list was full. page numbers in here are
931 * shifted left by 1, deleted slots have the LSB set.
933 MDB_IDL mt_spill_pgs;
935 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
937 /** For read txns: This thread/txn's reader table slot, or NULL. */
940 /** Array of records for each DB known in the environment. */
942 /** Array of MDB_db records for each known DB */
944 /** @defgroup mt_dbflag Transaction DB Flags
948 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
949 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
950 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
951 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
953 /** In write txns, array of cursors for each DB */
954 MDB_cursor **mt_cursors;
955 /** Array of flags for each DB */
956 unsigned char *mt_dbflags;
957 /** Number of DB records in use. This number only ever increments;
958 * we don't decrement it when individual DB handles are closed.
962 /** @defgroup mdb_txn Transaction Flags
966 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
967 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
968 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
969 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
971 unsigned int mt_flags; /**< @ref mdb_txn */
972 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
973 * Includes ancestor txns' dirty pages not hidden by other txns'
974 * dirty/spilled pages. Thus commit(nested txn) has room to merge
975 * dirty_list into mt_parent after freeing hidden mt_parent pages.
977 unsigned int mt_dirty_room;
980 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
981 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
982 * raise this on a 64 bit machine.
984 #define CURSOR_STACK 32
988 /** Cursors are used for all DB operations.
989 * A cursor holds a path of (page pointer, key index) from the DB
990 * root to a position in the DB, plus other state. #MDB_DUPSORT
991 * cursors include an xcursor to the current data item. Write txns
992 * track their cursors and keep them up to date when data moves.
993 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
994 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
997 /** Next cursor on this DB in this txn */
999 /** Backup of the original cursor if this cursor is a shadow */
1000 MDB_cursor *mc_backup;
1001 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1002 struct MDB_xcursor *mc_xcursor;
1003 /** The transaction that owns this cursor */
1005 /** The database handle this cursor operates on */
1007 /** The database record for this cursor */
1009 /** The database auxiliary record for this cursor */
1011 /** The @ref mt_dbflag for this database */
1012 unsigned char *mc_dbflag;
1013 unsigned short mc_snum; /**< number of pushed pages */
1014 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1015 /** @defgroup mdb_cursor Cursor Flags
1017 * Cursor state flags.
1020 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1021 #define C_EOF 0x02 /**< No more data */
1022 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1023 #define C_DEL 0x08 /**< last op was a cursor_del */
1024 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1025 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1027 unsigned int mc_flags; /**< @ref mdb_cursor */
1028 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1029 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1032 /** Context for sorted-dup records.
1033 * We could have gone to a fully recursive design, with arbitrarily
1034 * deep nesting of sub-databases. But for now we only handle these
1035 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1037 typedef struct MDB_xcursor {
1038 /** A sub-cursor for traversing the Dup DB */
1039 MDB_cursor mx_cursor;
1040 /** The database record for this Dup DB */
1042 /** The auxiliary DB record for this Dup DB */
1044 /** The @ref mt_dbflag for this Dup DB */
1045 unsigned char mx_dbflag;
1048 /** State of FreeDB old pages, stored in the MDB_env */
1049 typedef struct MDB_pgstate {
1050 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1051 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1054 /** The database environment. */
1056 HANDLE me_fd; /**< The main data file */
1057 HANDLE me_lfd; /**< The lock file */
1058 HANDLE me_mfd; /**< just for writing the meta pages */
1059 /** Failed to update the meta page. Probably an I/O error. */
1060 #define MDB_FATAL_ERROR 0x80000000U
1061 /** Some fields are initialized. */
1062 #define MDB_ENV_ACTIVE 0x20000000U
1063 /** me_txkey is set */
1064 #define MDB_ENV_TXKEY 0x10000000U
1065 uint32_t me_flags; /**< @ref mdb_env */
1066 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1067 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1068 unsigned int me_maxreaders; /**< size of the reader table */
1069 unsigned int me_numreaders; /**< max numreaders set by this env */
1070 MDB_dbi me_numdbs; /**< number of DBs opened */
1071 MDB_dbi me_maxdbs; /**< size of the DB table */
1072 MDB_PID_T me_pid; /**< process ID of this env */
1073 char *me_path; /**< path to the DB files */
1074 char *me_map; /**< the memory map of the data file */
1075 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1076 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1077 void *me_pbuf; /**< scratch area for DUPSORT put() */
1078 MDB_txn *me_txn; /**< current write transaction */
1079 size_t me_mapsize; /**< size of the data memory map */
1080 off_t me_size; /**< current file size */
1081 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1082 MDB_dbx *me_dbxs; /**< array of static DB info */
1083 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1084 pthread_key_t me_txkey; /**< thread-key for readers */
1085 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1086 # define me_pglast me_pgstate.mf_pglast
1087 # define me_pghead me_pgstate.mf_pghead
1088 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1089 /** IDL of pages that became unused in a write txn */
1090 MDB_IDL me_free_pgs;
1091 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1092 MDB_ID2L me_dirty_list;
1093 /** Max number of freelist items that can fit in a single overflow page */
1095 /** Max size of a node on a page */
1096 unsigned int me_nodemax;
1097 #if !(MDB_MAXKEYSIZE)
1098 unsigned int me_maxkey; /**< max size of a key */
1100 int me_live_reader; /**< have liveness lock in reader table */
1102 int me_pidquery; /**< Used in OpenProcess */
1103 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1105 #elif defined(MDB_USE_POSIX_SEM)
1106 sem_t *me_rmutex; /* Shared mutexes are not supported */
1109 void *me_userctx; /**< User-settable context */
1110 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1113 /** Nested transaction */
1114 typedef struct MDB_ntxn {
1115 MDB_txn mnt_txn; /**< the transaction */
1116 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1119 /** max number of pages to commit in one writev() call */
1120 #define MDB_COMMIT_PAGES 64
1121 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1122 #undef MDB_COMMIT_PAGES
1123 #define MDB_COMMIT_PAGES IOV_MAX
1126 /** max bytes to write in one call */
1127 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1129 /** Check \b txn and \b dbi arguments to a function */
1130 #define TXN_DBI_EXIST(txn, dbi) \
1131 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1133 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1134 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1135 static int mdb_page_touch(MDB_cursor *mc);
1137 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1138 static int mdb_page_search_root(MDB_cursor *mc,
1139 MDB_val *key, int modify);
1140 #define MDB_PS_MODIFY 1
1141 #define MDB_PS_ROOTONLY 2
1142 #define MDB_PS_FIRST 4
1143 #define MDB_PS_LAST 8
1144 static int mdb_page_search(MDB_cursor *mc,
1145 MDB_val *key, int flags);
1146 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1148 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1149 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1150 pgno_t newpgno, unsigned int nflags);
1152 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1153 static int mdb_env_pick_meta(const MDB_env *env);
1154 static int mdb_env_write_meta(MDB_txn *txn);
1155 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1156 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1158 static void mdb_env_close0(MDB_env *env, int excl);
1160 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1161 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1162 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1163 static void mdb_node_del(MDB_cursor *mc, int ksize);
1164 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1165 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1166 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1167 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1168 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1170 static int mdb_rebalance(MDB_cursor *mc);
1171 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1173 static void mdb_cursor_pop(MDB_cursor *mc);
1174 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1176 static int mdb_cursor_del0(MDB_cursor *mc);
1177 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1178 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1179 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1180 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1181 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1183 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1184 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1186 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1187 static void mdb_xcursor_init0(MDB_cursor *mc);
1188 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1190 static int mdb_drop0(MDB_cursor *mc, int subs);
1191 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1194 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1198 static SECURITY_DESCRIPTOR mdb_null_sd;
1199 static SECURITY_ATTRIBUTES mdb_all_sa;
1200 static int mdb_sec_inited;
1203 /** Return the library version info. */
1205 mdb_version(int *major, int *minor, int *patch)
1207 if (major) *major = MDB_VERSION_MAJOR;
1208 if (minor) *minor = MDB_VERSION_MINOR;
1209 if (patch) *patch = MDB_VERSION_PATCH;
1210 return MDB_VERSION_STRING;
1213 /** Table of descriptions for LMDB @ref errors */
1214 static char *const mdb_errstr[] = {
1215 "MDB_KEYEXIST: Key/data pair already exists",
1216 "MDB_NOTFOUND: No matching key/data pair found",
1217 "MDB_PAGE_NOTFOUND: Requested page not found",
1218 "MDB_CORRUPTED: Located page was wrong type",
1219 "MDB_PANIC: Update of meta page failed",
1220 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1221 "MDB_INVALID: File is not an LMDB file",
1222 "MDB_MAP_FULL: Environment mapsize limit reached",
1223 "MDB_DBS_FULL: Environment maxdbs limit reached",
1224 "MDB_READERS_FULL: Environment maxreaders limit reached",
1225 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1226 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1227 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1228 "MDB_PAGE_FULL: Internal error - page has no more space",
1229 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1230 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1231 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1232 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1233 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1237 mdb_strerror(int err)
1241 return ("Successful return: 0");
1243 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1244 i = err - MDB_KEYEXIST;
1245 return mdb_errstr[i];
1248 return strerror(err);
1251 /** assert(3) variant in cursor context */
1252 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1253 /** assert(3) variant in transaction context */
1254 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1255 /** assert(3) variant in environment context */
1256 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1259 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1260 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1263 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1264 const char *func, const char *file, int line)
1267 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1268 file, line, expr_txt, func);
1269 if (env->me_assert_func)
1270 env->me_assert_func(env, buf);
1271 fprintf(stderr, "%s\n", buf);
1275 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1279 /** Return the page number of \b mp which may be sub-page, for debug output */
1281 mdb_dbg_pgno(MDB_page *mp)
1284 COPY_PGNO(ret, mp->mp_pgno);
1288 /** Display a key in hexadecimal and return the address of the result.
1289 * @param[in] key the key to display
1290 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1291 * @return The key in hexadecimal form.
1294 mdb_dkey(MDB_val *key, char *buf)
1297 unsigned char *c = key->mv_data;
1303 if (key->mv_size > DKBUF_MAXKEYSIZE)
1304 return "MDB_MAXKEYSIZE";
1305 /* may want to make this a dynamic check: if the key is mostly
1306 * printable characters, print it as-is instead of converting to hex.
1310 for (i=0; i<key->mv_size; i++)
1311 ptr += sprintf(ptr, "%02x", *c++);
1313 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1319 mdb_leafnode_type(MDB_node *n)
1321 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1322 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1323 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1326 /** Display all the keys in the page. */
1328 mdb_page_list(MDB_page *mp)
1330 pgno_t pgno = mdb_dbg_pgno(mp);
1331 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1333 unsigned int i, nkeys, nsize, total = 0;
1337 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1338 case P_BRANCH: type = "Branch page"; break;
1339 case P_LEAF: type = "Leaf page"; break;
1340 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1341 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1342 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1344 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1345 pgno, mp->mp_pages, state);
1348 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1349 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1352 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1356 nkeys = NUMKEYS(mp);
1357 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1359 for (i=0; i<nkeys; i++) {
1360 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1361 key.mv_size = nsize = mp->mp_pad;
1362 key.mv_data = LEAF2KEY(mp, i, nsize);
1364 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1367 node = NODEPTR(mp, i);
1368 key.mv_size = node->mn_ksize;
1369 key.mv_data = node->mn_data;
1370 nsize = NODESIZE + key.mv_size;
1371 if (IS_BRANCH(mp)) {
1372 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1376 if (F_ISSET(node->mn_flags, F_BIGDATA))
1377 nsize += sizeof(pgno_t);
1379 nsize += NODEDSZ(node);
1381 nsize += sizeof(indx_t);
1382 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1383 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1385 total = EVEN(total);
1387 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1388 IS_LEAF2(mp) ? PAGEHDRSZ : mp->mp_lower, total, SIZELEFT(mp));
1392 mdb_cursor_chk(MDB_cursor *mc)
1398 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1399 for (i=0; i<mc->mc_top; i++) {
1401 node = NODEPTR(mp, mc->mc_ki[i]);
1402 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1405 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1411 /** Count all the pages in each DB and in the freelist
1412 * and make sure it matches the actual number of pages
1414 * All named DBs must be open for a correct count.
1416 static void mdb_audit(MDB_txn *txn)
1420 MDB_ID freecount, count;
1425 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1426 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1427 freecount += *(MDB_ID *)data.mv_data;
1428 mdb_tassert(txn, rc == MDB_NOTFOUND);
1431 for (i = 0; i<txn->mt_numdbs; i++) {
1433 if (!(txn->mt_dbflags[i] & DB_VALID))
1435 mdb_cursor_init(&mc, txn, i, &mx);
1436 if (txn->mt_dbs[i].md_root == P_INVALID)
1438 count += txn->mt_dbs[i].md_branch_pages +
1439 txn->mt_dbs[i].md_leaf_pages +
1440 txn->mt_dbs[i].md_overflow_pages;
1441 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1442 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1443 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1446 mp = mc.mc_pg[mc.mc_top];
1447 for (j=0; j<NUMKEYS(mp); j++) {
1448 MDB_node *leaf = NODEPTR(mp, j);
1449 if (leaf->mn_flags & F_SUBDATA) {
1451 memcpy(&db, NODEDATA(leaf), sizeof(db));
1452 count += db.md_branch_pages + db.md_leaf_pages +
1453 db.md_overflow_pages;
1457 mdb_tassert(txn, rc == MDB_NOTFOUND);
1460 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1461 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1462 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1468 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1470 return txn->mt_dbxs[dbi].md_cmp(a, b);
1474 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1476 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1479 /** Allocate memory for a page.
1480 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1483 mdb_page_malloc(MDB_txn *txn, unsigned num)
1485 MDB_env *env = txn->mt_env;
1486 MDB_page *ret = env->me_dpages;
1487 size_t psize = env->me_psize, sz = psize, off;
1488 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1489 * For a single page alloc, we init everything after the page header.
1490 * For multi-page, we init the final page; if the caller needed that
1491 * many pages they will be filling in at least up to the last page.
1495 VGMEMP_ALLOC(env, ret, sz);
1496 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1497 env->me_dpages = ret->mp_next;
1500 psize -= off = PAGEHDRSZ;
1505 if ((ret = malloc(sz)) != NULL) {
1506 VGMEMP_ALLOC(env, ret, sz);
1507 if (!(env->me_flags & MDB_NOMEMINIT)) {
1508 memset((char *)ret + off, 0, psize);
1512 txn->mt_flags |= MDB_TXN_ERROR;
1516 /** Free a single page.
1517 * Saves single pages to a list, for future reuse.
1518 * (This is not used for multi-page overflow pages.)
1521 mdb_page_free(MDB_env *env, MDB_page *mp)
1523 mp->mp_next = env->me_dpages;
1524 VGMEMP_FREE(env, mp);
1525 env->me_dpages = mp;
1528 /** Free a dirty page */
1530 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1532 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1533 mdb_page_free(env, dp);
1535 /* large pages just get freed directly */
1536 VGMEMP_FREE(env, dp);
1541 /** Return all dirty pages to dpage list */
1543 mdb_dlist_free(MDB_txn *txn)
1545 MDB_env *env = txn->mt_env;
1546 MDB_ID2L dl = txn->mt_u.dirty_list;
1547 unsigned i, n = dl[0].mid;
1549 for (i = 1; i <= n; i++) {
1550 mdb_dpage_free(env, dl[i].mptr);
1555 /** Loosen or free a single page.
1556 * Saves single pages to a list for future reuse
1557 * in this same txn. It has been pulled from the freeDB
1558 * and already resides on the dirty list, but has been
1559 * deleted. Use these pages first before pulling again
1562 * If the page wasn't dirtied in this txn, just add it
1563 * to this txn's free list.
1566 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1569 pgno_t pgno = mp->mp_pgno;
1571 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1572 if (mc->mc_txn->mt_parent) {
1573 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1574 /* If txn has a parent, make sure the page is in our
1578 unsigned x = mdb_mid2l_search(dl, pgno);
1579 if (x <= dl[0].mid && dl[x].mid == pgno) {
1580 if (mp != dl[x].mptr) { /* bad cursor? */
1581 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1582 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1583 return MDB_CORRUPTED;
1590 /* no parent txn, so it's just ours */
1595 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1596 mc->mc_txn->mt_loose_pgs = mp;
1597 mp->mp_flags |= P_LOOSE;
1599 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1607 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1608 * @param[in] mc A cursor handle for the current operation.
1609 * @param[in] pflags Flags of the pages to update:
1610 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1611 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1612 * @return 0 on success, non-zero on failure.
1615 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1617 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1618 MDB_txn *txn = mc->mc_txn;
1624 int rc = MDB_SUCCESS, level;
1626 /* Mark pages seen by cursors */
1627 if (mc->mc_flags & C_UNTRACK)
1628 mc = NULL; /* will find mc in mt_cursors */
1629 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1630 for (; mc; mc=mc->mc_next) {
1631 if (!(mc->mc_flags & C_INITIALIZED))
1633 for (m3 = mc;; m3 = &mx->mx_cursor) {
1635 for (j=0; j<m3->mc_snum; j++) {
1637 if ((mp->mp_flags & Mask) == pflags)
1638 mp->mp_flags ^= P_KEEP;
1640 mx = m3->mc_xcursor;
1641 /* Proceed to mx if it is at a sub-database */
1642 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1644 if (! (mp && (mp->mp_flags & P_LEAF)))
1646 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1647 if (!(leaf->mn_flags & F_SUBDATA))
1655 /* Loose pages shouldn't be spilled */
1656 for (dp = txn->mt_loose_pgs; dp; dp = NEXT_LOOSE_PAGE(dp)) {
1657 if ((dp->mp_flags & Mask) == pflags)
1658 dp->mp_flags ^= P_KEEP;
1662 /* Mark dirty root pages */
1663 for (i=0; i<txn->mt_numdbs; i++) {
1664 if (txn->mt_dbflags[i] & DB_DIRTY) {
1665 pgno_t pgno = txn->mt_dbs[i].md_root;
1666 if (pgno == P_INVALID)
1668 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1670 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1671 dp->mp_flags ^= P_KEEP;
1679 static int mdb_page_flush(MDB_txn *txn, int keep);
1681 /** Spill pages from the dirty list back to disk.
1682 * This is intended to prevent running into #MDB_TXN_FULL situations,
1683 * but note that they may still occur in a few cases:
1684 * 1) our estimate of the txn size could be too small. Currently this
1685 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1686 * 2) child txns may run out of space if their parents dirtied a
1687 * lot of pages and never spilled them. TODO: we probably should do
1688 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1689 * the parent's dirty_room is below a given threshold.
1691 * Otherwise, if not using nested txns, it is expected that apps will
1692 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1693 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1694 * If the txn never references them again, they can be left alone.
1695 * If the txn only reads them, they can be used without any fuss.
1696 * If the txn writes them again, they can be dirtied immediately without
1697 * going thru all of the work of #mdb_page_touch(). Such references are
1698 * handled by #mdb_page_unspill().
1700 * Also note, we never spill DB root pages, nor pages of active cursors,
1701 * because we'll need these back again soon anyway. And in nested txns,
1702 * we can't spill a page in a child txn if it was already spilled in a
1703 * parent txn. That would alter the parent txns' data even though
1704 * the child hasn't committed yet, and we'd have no way to undo it if
1705 * the child aborted.
1707 * @param[in] m0 cursor A cursor handle identifying the transaction and
1708 * database for which we are checking space.
1709 * @param[in] key For a put operation, the key being stored.
1710 * @param[in] data For a put operation, the data being stored.
1711 * @return 0 on success, non-zero on failure.
1714 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1716 MDB_txn *txn = m0->mc_txn;
1718 MDB_ID2L dl = txn->mt_u.dirty_list;
1719 unsigned int i, j, need;
1722 if (m0->mc_flags & C_SUB)
1725 /* Estimate how much space this op will take */
1726 i = m0->mc_db->md_depth;
1727 /* Named DBs also dirty the main DB */
1728 if (m0->mc_dbi > MAIN_DBI)
1729 i += txn->mt_dbs[MAIN_DBI].md_depth;
1730 /* For puts, roughly factor in the key+data size */
1732 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1733 i += i; /* double it for good measure */
1736 if (txn->mt_dirty_room > i)
1739 if (!txn->mt_spill_pgs) {
1740 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1741 if (!txn->mt_spill_pgs)
1744 /* purge deleted slots */
1745 MDB_IDL sl = txn->mt_spill_pgs;
1746 unsigned int num = sl[0];
1748 for (i=1; i<=num; i++) {
1755 /* Preserve pages which may soon be dirtied again */
1756 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1759 /* Less aggressive spill - we originally spilled the entire dirty list,
1760 * with a few exceptions for cursor pages and DB root pages. But this
1761 * turns out to be a lot of wasted effort because in a large txn many
1762 * of those pages will need to be used again. So now we spill only 1/8th
1763 * of the dirty pages. Testing revealed this to be a good tradeoff,
1764 * better than 1/2, 1/4, or 1/10.
1766 if (need < MDB_IDL_UM_MAX / 8)
1767 need = MDB_IDL_UM_MAX / 8;
1769 /* Save the page IDs of all the pages we're flushing */
1770 /* flush from the tail forward, this saves a lot of shifting later on. */
1771 for (i=dl[0].mid; i && need; i--) {
1772 MDB_ID pn = dl[i].mid << 1;
1774 if (dp->mp_flags & P_KEEP)
1776 /* Can't spill twice, make sure it's not already in a parent's
1779 if (txn->mt_parent) {
1781 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1782 if (tx2->mt_spill_pgs) {
1783 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1784 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1785 dp->mp_flags |= P_KEEP;
1793 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1797 mdb_midl_sort(txn->mt_spill_pgs);
1799 /* Flush the spilled part of dirty list */
1800 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1803 /* Reset any dirty pages we kept that page_flush didn't see */
1804 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1807 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1811 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1813 mdb_find_oldest(MDB_txn *txn)
1816 txnid_t mr, oldest = txn->mt_txnid - 1;
1817 if (txn->mt_env->me_txns) {
1818 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1819 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1830 /** Add a page to the txn's dirty list */
1832 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1835 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1837 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1838 insert = mdb_mid2l_append;
1840 insert = mdb_mid2l_insert;
1842 mid.mid = mp->mp_pgno;
1844 rc = insert(txn->mt_u.dirty_list, &mid);
1845 mdb_tassert(txn, rc == 0);
1846 txn->mt_dirty_room--;
1849 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1850 * me_pghead and mt_next_pgno.
1852 * If there are free pages available from older transactions, they
1853 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1854 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1855 * and move me_pglast to say which records were consumed. Only this
1856 * function can create me_pghead and move me_pglast/mt_next_pgno.
1857 * @param[in] mc cursor A cursor handle identifying the transaction and
1858 * database for which we are allocating.
1859 * @param[in] num the number of pages to allocate.
1860 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1861 * will always be satisfied by a single contiguous chunk of memory.
1862 * @return 0 on success, non-zero on failure.
1865 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1867 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1868 /* Get at most <Max_retries> more freeDB records once me_pghead
1869 * has enough pages. If not enough, use new pages from the map.
1870 * If <Paranoid> and mc is updating the freeDB, only get new
1871 * records if me_pghead is empty. Then the freelist cannot play
1872 * catch-up with itself by growing while trying to save it.
1874 enum { Paranoid = 1, Max_retries = 500 };
1876 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1878 int rc, retry = num * 20;
1879 MDB_txn *txn = mc->mc_txn;
1880 MDB_env *env = txn->mt_env;
1881 pgno_t pgno, *mop = env->me_pghead;
1882 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1884 txnid_t oldest = 0, last;
1888 /* If there are any loose pages, just use them */
1889 if (num == 1 && txn->mt_loose_pgs) {
1890 np = txn->mt_loose_pgs;
1891 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1898 /* If our dirty list is already full, we can't do anything */
1899 if (txn->mt_dirty_room == 0) {
1904 for (op = MDB_FIRST;; op = MDB_NEXT) {
1907 pgno_t *idl, old_id, new_id;
1909 /* Seek a big enough contiguous page range. Prefer
1910 * pages at the tail, just truncating the list.
1916 if (mop[i-n2] == pgno+n2)
1923 if (op == MDB_FIRST) { /* 1st iteration */
1924 /* Prepare to fetch more and coalesce */
1925 oldest = mdb_find_oldest(txn);
1926 last = env->me_pglast;
1927 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1930 key.mv_data = &last; /* will look up last+1 */
1931 key.mv_size = sizeof(last);
1933 if (Paranoid && mc->mc_dbi == FREE_DBI)
1936 if (Paranoid && retry < 0 && mop_len)
1940 /* Do not fetch more if the record will be too recent */
1943 rc = mdb_cursor_get(&m2, &key, NULL, op);
1945 if (rc == MDB_NOTFOUND)
1949 last = *(txnid_t*)key.mv_data;
1952 np = m2.mc_pg[m2.mc_top];
1953 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1954 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1957 idl = (MDB_ID *) data.mv_data;
1960 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1965 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1967 mop = env->me_pghead;
1969 env->me_pglast = last;
1971 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1972 last, txn->mt_dbs[FREE_DBI].md_root, i));
1974 DPRINTF(("IDL %"Z"u", idl[k]));
1976 /* Merge in descending sorted order */
1979 mop[0] = (pgno_t)-1;
1983 for (; old_id < new_id; old_id = mop[--j])
1990 /* Use new pages from the map when nothing suitable in the freeDB */
1992 pgno = txn->mt_next_pgno;
1993 if (pgno + num >= env->me_maxpg) {
1994 DPUTS("DB size maxed out");
2000 if (env->me_flags & MDB_WRITEMAP) {
2001 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2003 if (!(np = mdb_page_malloc(txn, num))) {
2009 mop[0] = mop_len -= num;
2010 /* Move any stragglers down */
2011 for (j = i-num; j < mop_len; )
2012 mop[++j] = mop[++i];
2014 txn->mt_next_pgno = pgno + num;
2017 mdb_page_dirty(txn, np);
2023 txn->mt_flags |= MDB_TXN_ERROR;
2027 /** Copy the used portions of a non-overflow page.
2028 * @param[in] dst page to copy into
2029 * @param[in] src page to copy from
2030 * @param[in] psize size of a page
2033 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2035 enum { Align = sizeof(pgno_t) };
2036 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2038 /* If page isn't full, just copy the used portion. Adjust
2039 * alignment so memcpy may copy words instead of bytes.
2041 if ((unused &= -Align) && !IS_LEAF2(src)) {
2043 memcpy(dst, src, (lower + (Align-1)) & -Align);
2044 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2047 memcpy(dst, src, psize - unused);
2051 /** Pull a page off the txn's spill list, if present.
2052 * If a page being referenced was spilled to disk in this txn, bring
2053 * it back and make it dirty/writable again.
2054 * @param[in] txn the transaction handle.
2055 * @param[in] mp the page being referenced. It must not be dirty.
2056 * @param[out] ret the writable page, if any. ret is unchanged if
2057 * mp wasn't spilled.
2060 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2062 MDB_env *env = txn->mt_env;
2065 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2067 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2068 if (!tx2->mt_spill_pgs)
2070 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2071 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2074 if (txn->mt_dirty_room == 0)
2075 return MDB_TXN_FULL;
2076 if (IS_OVERFLOW(mp))
2080 if (env->me_flags & MDB_WRITEMAP) {
2083 np = mdb_page_malloc(txn, num);
2087 memcpy(np, mp, num * env->me_psize);
2089 mdb_page_copy(np, mp, env->me_psize);
2092 /* If in current txn, this page is no longer spilled.
2093 * If it happens to be the last page, truncate the spill list.
2094 * Otherwise mark it as deleted by setting the LSB.
2096 if (x == txn->mt_spill_pgs[0])
2097 txn->mt_spill_pgs[0]--;
2099 txn->mt_spill_pgs[x] |= 1;
2100 } /* otherwise, if belonging to a parent txn, the
2101 * page remains spilled until child commits
2104 mdb_page_dirty(txn, np);
2105 np->mp_flags |= P_DIRTY;
2113 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2114 * @param[in] mc cursor pointing to the page to be touched
2115 * @return 0 on success, non-zero on failure.
2118 mdb_page_touch(MDB_cursor *mc)
2120 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2121 MDB_txn *txn = mc->mc_txn;
2122 MDB_cursor *m2, *m3;
2126 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2127 if (txn->mt_flags & MDB_TXN_SPILLS) {
2129 rc = mdb_page_unspill(txn, mp, &np);
2135 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2136 (rc = mdb_page_alloc(mc, 1, &np)))
2139 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2140 mp->mp_pgno, pgno));
2141 mdb_cassert(mc, mp->mp_pgno != pgno);
2142 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2143 /* Update the parent page, if any, to point to the new page */
2145 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2146 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2147 SETPGNO(node, pgno);
2149 mc->mc_db->md_root = pgno;
2151 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2152 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2154 /* If txn has a parent, make sure the page is in our
2158 unsigned x = mdb_mid2l_search(dl, pgno);
2159 if (x <= dl[0].mid && dl[x].mid == pgno) {
2160 if (mp != dl[x].mptr) { /* bad cursor? */
2161 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2162 txn->mt_flags |= MDB_TXN_ERROR;
2163 return MDB_CORRUPTED;
2168 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2170 np = mdb_page_malloc(txn, 1);
2175 rc = mdb_mid2l_insert(dl, &mid);
2176 mdb_cassert(mc, rc == 0);
2181 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2183 np->mp_flags |= P_DIRTY;
2186 /* Adjust cursors pointing to mp */
2187 mc->mc_pg[mc->mc_top] = np;
2188 m2 = txn->mt_cursors[mc->mc_dbi];
2189 if (mc->mc_flags & C_SUB) {
2190 for (; m2; m2=m2->mc_next) {
2191 m3 = &m2->mc_xcursor->mx_cursor;
2192 if (m3->mc_snum < mc->mc_snum) continue;
2193 if (m3->mc_pg[mc->mc_top] == mp)
2194 m3->mc_pg[mc->mc_top] = np;
2197 for (; m2; m2=m2->mc_next) {
2198 if (m2->mc_snum < mc->mc_snum) continue;
2199 if (m2->mc_pg[mc->mc_top] == mp) {
2200 m2->mc_pg[mc->mc_top] = np;
2201 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2203 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2205 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2206 if (!(leaf->mn_flags & F_SUBDATA))
2207 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2215 txn->mt_flags |= MDB_TXN_ERROR;
2220 mdb_env_sync(MDB_env *env, int force)
2223 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2224 if (env->me_flags & MDB_WRITEMAP) {
2225 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2226 ? MS_ASYNC : MS_SYNC;
2227 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2230 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2234 if (MDB_FDATASYNC(env->me_fd))
2241 /** Back up parent txn's cursors, then grab the originals for tracking */
2243 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2245 MDB_cursor *mc, *bk;
2250 for (i = src->mt_numdbs; --i >= 0; ) {
2251 if ((mc = src->mt_cursors[i]) != NULL) {
2252 size = sizeof(MDB_cursor);
2254 size += sizeof(MDB_xcursor);
2255 for (; mc; mc = bk->mc_next) {
2261 mc->mc_db = &dst->mt_dbs[i];
2262 /* Kill pointers into src - and dst to reduce abuse: The
2263 * user may not use mc until dst ends. Otherwise we'd...
2265 mc->mc_txn = NULL; /* ...set this to dst */
2266 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2267 if ((mx = mc->mc_xcursor) != NULL) {
2268 *(MDB_xcursor *)(bk+1) = *mx;
2269 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2271 mc->mc_next = dst->mt_cursors[i];
2272 dst->mt_cursors[i] = mc;
2279 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2280 * @param[in] txn the transaction handle.
2281 * @param[in] merge true to keep changes to parent cursors, false to revert.
2282 * @return 0 on success, non-zero on failure.
2285 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2287 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2291 for (i = txn->mt_numdbs; --i >= 0; ) {
2292 for (mc = cursors[i]; mc; mc = next) {
2294 if ((bk = mc->mc_backup) != NULL) {
2296 /* Commit changes to parent txn */
2297 mc->mc_next = bk->mc_next;
2298 mc->mc_backup = bk->mc_backup;
2299 mc->mc_txn = bk->mc_txn;
2300 mc->mc_db = bk->mc_db;
2301 mc->mc_dbflag = bk->mc_dbflag;
2302 if ((mx = mc->mc_xcursor) != NULL)
2303 mx->mx_cursor.mc_txn = bk->mc_txn;
2305 /* Abort nested txn */
2307 if ((mx = mc->mc_xcursor) != NULL)
2308 *mx = *(MDB_xcursor *)(bk+1);
2312 /* Only malloced cursors are permanently tracked. */
2320 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2323 mdb_txn_reset0(MDB_txn *txn, const char *act);
2325 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2331 Pidset = F_SETLK, Pidcheck = F_GETLK
2335 /** Set or check a pid lock. Set returns 0 on success.
2336 * Check returns 0 if the process is certainly dead, nonzero if it may
2337 * be alive (the lock exists or an error happened so we do not know).
2339 * On Windows Pidset is a no-op, we merely check for the existence
2340 * of the process with the given pid. On POSIX we use a single byte
2341 * lock on the lockfile, set at an offset equal to the pid.
2344 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2346 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2349 if (op == Pidcheck) {
2350 h = OpenProcess(env->me_pidquery, FALSE, pid);
2351 /* No documented "no such process" code, but other program use this: */
2353 return ErrCode() != ERROR_INVALID_PARAMETER;
2354 /* A process exists until all handles to it close. Has it exited? */
2355 ret = WaitForSingleObject(h, 0) != 0;
2362 struct flock lock_info;
2363 memset(&lock_info, 0, sizeof(lock_info));
2364 lock_info.l_type = F_WRLCK;
2365 lock_info.l_whence = SEEK_SET;
2366 lock_info.l_start = pid;
2367 lock_info.l_len = 1;
2368 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2369 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2371 } else if ((rc = ErrCode()) == EINTR) {
2379 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2380 * @param[in] txn the transaction handle to initialize
2381 * @return 0 on success, non-zero on failure.
2384 mdb_txn_renew0(MDB_txn *txn)
2386 MDB_env *env = txn->mt_env;
2387 MDB_txninfo *ti = env->me_txns;
2391 int rc, new_notls = 0;
2394 txn->mt_numdbs = env->me_numdbs;
2395 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2397 if (txn->mt_flags & MDB_TXN_RDONLY) {
2399 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2400 txn->mt_txnid = meta->mm_txnid;
2401 txn->mt_u.reader = NULL;
2403 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2404 pthread_getspecific(env->me_txkey);
2406 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2407 return MDB_BAD_RSLOT;
2409 MDB_PID_T pid = env->me_pid;
2410 MDB_THR_T tid = pthread_self();
2412 if (!env->me_live_reader) {
2413 rc = mdb_reader_pid(env, Pidset, pid);
2416 env->me_live_reader = 1;
2420 nr = ti->mti_numreaders;
2421 for (i=0; i<nr; i++)
2422 if (ti->mti_readers[i].mr_pid == 0)
2424 if (i == env->me_maxreaders) {
2425 UNLOCK_MUTEX_R(env);
2426 return MDB_READERS_FULL;
2428 ti->mti_readers[i].mr_pid = pid;
2429 ti->mti_readers[i].mr_tid = tid;
2431 ti->mti_numreaders = ++nr;
2432 /* Save numreaders for un-mutexed mdb_env_close() */
2433 env->me_numreaders = nr;
2434 UNLOCK_MUTEX_R(env);
2436 r = &ti->mti_readers[i];
2437 new_notls = (env->me_flags & MDB_NOTLS);
2438 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2443 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2444 txn->mt_u.reader = r;
2445 meta = env->me_metas[txn->mt_txnid & 1];
2451 txn->mt_txnid = ti->mti_txnid;
2452 meta = env->me_metas[txn->mt_txnid & 1];
2454 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2455 txn->mt_txnid = meta->mm_txnid;
2459 if (txn->mt_txnid == mdb_debug_start)
2462 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2463 txn->mt_u.dirty_list = env->me_dirty_list;
2464 txn->mt_u.dirty_list[0].mid = 0;
2465 txn->mt_free_pgs = env->me_free_pgs;
2466 txn->mt_free_pgs[0] = 0;
2467 txn->mt_spill_pgs = NULL;
2471 /* Copy the DB info and flags */
2472 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2474 /* Moved to here to avoid a data race in read TXNs */
2475 txn->mt_next_pgno = meta->mm_last_pg+1;
2477 for (i=2; i<txn->mt_numdbs; i++) {
2478 x = env->me_dbflags[i];
2479 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2480 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2482 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2484 if (env->me_maxpg < txn->mt_next_pgno) {
2485 mdb_txn_reset0(txn, "renew0-mapfail");
2487 txn->mt_u.reader->mr_pid = 0;
2488 txn->mt_u.reader = NULL;
2490 return MDB_MAP_RESIZED;
2497 mdb_txn_renew(MDB_txn *txn)
2501 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2504 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2505 DPUTS("environment had fatal error, must shutdown!");
2509 rc = mdb_txn_renew0(txn);
2510 if (rc == MDB_SUCCESS) {
2511 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2512 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2513 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2519 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2523 int rc, size, tsize = sizeof(MDB_txn);
2525 if (env->me_flags & MDB_FATAL_ERROR) {
2526 DPUTS("environment had fatal error, must shutdown!");
2529 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2532 /* Nested transactions: Max 1 child, write txns only, no writemap */
2533 if (parent->mt_child ||
2534 (flags & MDB_RDONLY) ||
2535 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2536 (env->me_flags & MDB_WRITEMAP))
2538 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2540 tsize = sizeof(MDB_ntxn);
2542 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2543 if (!(flags & MDB_RDONLY))
2544 size += env->me_maxdbs * sizeof(MDB_cursor *);
2546 if ((txn = calloc(1, size)) == NULL) {
2547 DPRINTF(("calloc: %s", strerror(ErrCode())));
2550 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2551 if (flags & MDB_RDONLY) {
2552 txn->mt_flags |= MDB_TXN_RDONLY;
2553 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2555 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2556 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2562 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2563 if (!txn->mt_u.dirty_list ||
2564 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2566 free(txn->mt_u.dirty_list);
2570 txn->mt_txnid = parent->mt_txnid;
2571 txn->mt_dirty_room = parent->mt_dirty_room;
2572 txn->mt_u.dirty_list[0].mid = 0;
2573 txn->mt_spill_pgs = NULL;
2574 txn->mt_next_pgno = parent->mt_next_pgno;
2575 parent->mt_child = txn;
2576 txn->mt_parent = parent;
2577 txn->mt_numdbs = parent->mt_numdbs;
2578 txn->mt_flags = parent->mt_flags;
2579 txn->mt_dbxs = parent->mt_dbxs;
2580 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2581 /* Copy parent's mt_dbflags, but clear DB_NEW */
2582 for (i=0; i<txn->mt_numdbs; i++)
2583 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2585 ntxn = (MDB_ntxn *)txn;
2586 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2587 if (env->me_pghead) {
2588 size = MDB_IDL_SIZEOF(env->me_pghead);
2589 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2591 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2596 rc = mdb_cursor_shadow(parent, txn);
2598 mdb_txn_reset0(txn, "beginchild-fail");
2600 rc = mdb_txn_renew0(txn);
2606 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2607 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2608 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2615 mdb_txn_env(MDB_txn *txn)
2617 if(!txn) return NULL;
2621 /** Export or close DBI handles opened in this txn. */
2623 mdb_dbis_update(MDB_txn *txn, int keep)
2626 MDB_dbi n = txn->mt_numdbs;
2627 MDB_env *env = txn->mt_env;
2628 unsigned char *tdbflags = txn->mt_dbflags;
2630 for (i = n; --i >= 2;) {
2631 if (tdbflags[i] & DB_NEW) {
2633 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2635 char *ptr = env->me_dbxs[i].md_name.mv_data;
2636 env->me_dbxs[i].md_name.mv_data = NULL;
2637 env->me_dbxs[i].md_name.mv_size = 0;
2638 env->me_dbflags[i] = 0;
2643 if (keep && env->me_numdbs < n)
2647 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2648 * May be called twice for readonly txns: First reset it, then abort.
2649 * @param[in] txn the transaction handle to reset
2650 * @param[in] act why the transaction is being reset
2653 mdb_txn_reset0(MDB_txn *txn, const char *act)
2655 MDB_env *env = txn->mt_env;
2657 /* Close any DBI handles opened in this txn */
2658 mdb_dbis_update(txn, 0);
2660 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2661 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2662 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2664 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2665 if (txn->mt_u.reader) {
2666 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2667 if (!(env->me_flags & MDB_NOTLS))
2668 txn->mt_u.reader = NULL; /* txn does not own reader */
2670 txn->mt_numdbs = 0; /* close nothing if called again */
2671 txn->mt_dbxs = NULL; /* mark txn as reset */
2673 mdb_cursors_close(txn, 0);
2675 if (!(env->me_flags & MDB_WRITEMAP)) {
2676 mdb_dlist_free(txn);
2678 mdb_midl_free(env->me_pghead);
2680 if (txn->mt_parent) {
2681 txn->mt_parent->mt_child = NULL;
2682 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2683 mdb_midl_free(txn->mt_free_pgs);
2684 mdb_midl_free(txn->mt_spill_pgs);
2685 free(txn->mt_u.dirty_list);
2689 if (mdb_midl_shrink(&txn->mt_free_pgs))
2690 env->me_free_pgs = txn->mt_free_pgs;
2691 env->me_pghead = NULL;
2695 /* The writer mutex was locked in mdb_txn_begin. */
2697 UNLOCK_MUTEX_W(env);
2702 mdb_txn_reset(MDB_txn *txn)
2707 /* This call is only valid for read-only txns */
2708 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2711 mdb_txn_reset0(txn, "reset");
2715 mdb_txn_abort(MDB_txn *txn)
2721 mdb_txn_abort(txn->mt_child);
2723 mdb_txn_reset0(txn, "abort");
2724 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2725 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2726 txn->mt_u.reader->mr_pid = 0;
2731 /** Save the freelist as of this transaction to the freeDB.
2732 * This changes the freelist. Keep trying until it stabilizes.
2735 mdb_freelist_save(MDB_txn *txn)
2737 /* env->me_pghead[] can grow and shrink during this call.
2738 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2739 * Page numbers cannot disappear from txn->mt_free_pgs[].
2742 MDB_env *env = txn->mt_env;
2743 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2744 txnid_t pglast = 0, head_id = 0;
2745 pgno_t freecnt = 0, *free_pgs, *mop;
2746 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2748 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2750 if (env->me_pghead) {
2751 /* Make sure first page of freeDB is touched and on freelist */
2752 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2753 if (rc && rc != MDB_NOTFOUND)
2757 /* Dispose of loose pages. Usually they will have all
2758 * been used up by the time we get here.
2760 if (txn->mt_loose_pgs) {
2761 MDB_page *mp = txn->mt_loose_pgs;
2762 /* Just return them to freeDB */
2763 if (env->me_pghead) {
2765 mop = env->me_pghead;
2766 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2767 pgno_t pg = mp->mp_pgno;
2769 for (i = mop[0]; i && mop[i] < pg; i--)
2775 /* Oh well, they were wasted. Put on freelist */
2776 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2777 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
2780 txn->mt_loose_pgs = NULL;
2783 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2784 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2785 ? SSIZE_MAX : maxfree_1pg;
2788 /* Come back here after each Put() in case freelist changed */
2793 /* If using records from freeDB which we have not yet
2794 * deleted, delete them and any we reserved for me_pghead.
2796 while (pglast < env->me_pglast) {
2797 rc = mdb_cursor_first(&mc, &key, NULL);
2800 pglast = head_id = *(txnid_t *)key.mv_data;
2801 total_room = head_room = 0;
2802 mdb_tassert(txn, pglast <= env->me_pglast);
2803 rc = mdb_cursor_del(&mc, 0);
2808 /* Save the IDL of pages freed by this txn, to a single record */
2809 if (freecnt < txn->mt_free_pgs[0]) {
2811 /* Make sure last page of freeDB is touched and on freelist */
2812 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2813 if (rc && rc != MDB_NOTFOUND)
2816 free_pgs = txn->mt_free_pgs;
2817 /* Write to last page of freeDB */
2818 key.mv_size = sizeof(txn->mt_txnid);
2819 key.mv_data = &txn->mt_txnid;
2821 freecnt = free_pgs[0];
2822 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2823 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2826 /* Retry if mt_free_pgs[] grew during the Put() */
2827 free_pgs = txn->mt_free_pgs;
2828 } while (freecnt < free_pgs[0]);
2829 mdb_midl_sort(free_pgs);
2830 memcpy(data.mv_data, free_pgs, data.mv_size);
2833 unsigned int i = free_pgs[0];
2834 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2835 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2837 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2843 mop = env->me_pghead;
2844 mop_len = mop ? mop[0] : 0;
2846 /* Reserve records for me_pghead[]. Split it if multi-page,
2847 * to avoid searching freeDB for a page range. Use keys in
2848 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2850 if (total_room >= mop_len) {
2851 if (total_room == mop_len || --more < 0)
2853 } else if (head_room >= maxfree_1pg && head_id > 1) {
2854 /* Keep current record (overflow page), add a new one */
2858 /* (Re)write {key = head_id, IDL length = head_room} */
2859 total_room -= head_room;
2860 head_room = mop_len - total_room;
2861 if (head_room > maxfree_1pg && head_id > 1) {
2862 /* Overflow multi-page for part of me_pghead */
2863 head_room /= head_id; /* amortize page sizes */
2864 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2865 } else if (head_room < 0) {
2866 /* Rare case, not bothering to delete this record */
2869 key.mv_size = sizeof(head_id);
2870 key.mv_data = &head_id;
2871 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2872 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2875 /* IDL is initially empty, zero out at least the length */
2876 pgs = (pgno_t *)data.mv_data;
2877 j = head_room > clean_limit ? head_room : 0;
2881 total_room += head_room;
2884 /* Fill in the reserved me_pghead records */
2890 rc = mdb_cursor_first(&mc, &key, &data);
2891 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2892 txnid_t id = *(txnid_t *)key.mv_data;
2893 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2896 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2898 if (len > mop_len) {
2900 data.mv_size = (len + 1) * sizeof(MDB_ID);
2902 data.mv_data = mop -= len;
2905 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2907 if (rc || !(mop_len -= len))
2914 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2915 * @param[in] txn the transaction that's being committed
2916 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2917 * @return 0 on success, non-zero on failure.
2920 mdb_page_flush(MDB_txn *txn, int keep)
2922 MDB_env *env = txn->mt_env;
2923 MDB_ID2L dl = txn->mt_u.dirty_list;
2924 unsigned psize = env->me_psize, j;
2925 int i, pagecount = dl[0].mid, rc;
2926 size_t size = 0, pos = 0;
2928 MDB_page *dp = NULL;
2932 struct iovec iov[MDB_COMMIT_PAGES];
2933 ssize_t wpos = 0, wsize = 0, wres;
2934 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2940 if (env->me_flags & MDB_WRITEMAP) {
2941 /* Clear dirty flags */
2942 while (++i <= pagecount) {
2944 /* Don't flush this page yet */
2945 if (dp->mp_flags & P_KEEP) {
2946 dp->mp_flags ^= P_KEEP;
2950 dp->mp_flags &= ~P_DIRTY;
2955 /* Write the pages */
2957 if (++i <= pagecount) {
2959 /* Don't flush this page yet */
2960 if (dp->mp_flags & P_KEEP) {
2961 dp->mp_flags ^= P_KEEP;
2966 /* clear dirty flag */
2967 dp->mp_flags &= ~P_DIRTY;
2970 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2975 /* Windows actually supports scatter/gather I/O, but only on
2976 * unbuffered file handles. Since we're relying on the OS page
2977 * cache for all our data, that's self-defeating. So we just
2978 * write pages one at a time. We use the ov structure to set
2979 * the write offset, to at least save the overhead of a Seek
2982 DPRINTF(("committing page %"Z"u", pgno));
2983 memset(&ov, 0, sizeof(ov));
2984 ov.Offset = pos & 0xffffffff;
2985 ov.OffsetHigh = pos >> 16 >> 16;
2986 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2988 DPRINTF(("WriteFile: %d", rc));
2992 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2993 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2995 /* Write previous page(s) */
2996 #ifdef MDB_USE_PWRITEV
2997 wres = pwritev(env->me_fd, iov, n, wpos);
3000 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3002 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3004 DPRINTF(("lseek: %s", strerror(rc)));
3007 wres = writev(env->me_fd, iov, n);
3010 if (wres != wsize) {
3013 DPRINTF(("Write error: %s", strerror(rc)));
3015 rc = EIO; /* TODO: Use which error code? */
3016 DPUTS("short write, filesystem full?");
3027 DPRINTF(("committing page %"Z"u", pgno));
3028 next_pos = pos + size;
3029 iov[n].iov_len = size;
3030 iov[n].iov_base = (char *)dp;
3036 for (i = keep; ++i <= pagecount; ) {
3038 /* This is a page we skipped above */
3041 dl[j].mid = dp->mp_pgno;
3044 mdb_dpage_free(env, dp);
3049 txn->mt_dirty_room += i - j;
3055 mdb_txn_commit(MDB_txn *txn)
3061 if (txn == NULL || txn->mt_env == NULL)
3064 if (txn->mt_child) {
3065 rc = mdb_txn_commit(txn->mt_child);
3066 txn->mt_child = NULL;
3073 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3074 mdb_dbis_update(txn, 1);
3075 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3080 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3081 DPUTS("error flag is set, can't commit");
3083 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3088 if (txn->mt_parent) {
3089 MDB_txn *parent = txn->mt_parent;
3092 unsigned x, y, len, ps_len;
3094 /* Append our free list to parent's */
3095 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3098 mdb_midl_free(txn->mt_free_pgs);
3099 /* Failures after this must either undo the changes
3100 * to the parent or set MDB_TXN_ERROR in the parent.
3103 parent->mt_next_pgno = txn->mt_next_pgno;
3104 parent->mt_flags = txn->mt_flags;
3106 /* Merge our cursors into parent's and close them */
3107 mdb_cursors_close(txn, 1);
3109 /* Update parent's DB table. */
3110 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3111 parent->mt_numdbs = txn->mt_numdbs;
3112 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3113 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3114 for (i=2; i<txn->mt_numdbs; i++) {
3115 /* preserve parent's DB_NEW status */
3116 x = parent->mt_dbflags[i] & DB_NEW;
3117 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3120 dst = parent->mt_u.dirty_list;
3121 src = txn->mt_u.dirty_list;
3122 /* Remove anything in our dirty list from parent's spill list */
3123 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3125 pspill[0] = (pgno_t)-1;
3126 /* Mark our dirty pages as deleted in parent spill list */
3127 for (i=0, len=src[0].mid; ++i <= len; ) {
3128 MDB_ID pn = src[i].mid << 1;
3129 while (pn > pspill[x])
3131 if (pn == pspill[x]) {
3136 /* Squash deleted pagenums if we deleted any */
3137 for (x=y; ++x <= ps_len; )
3138 if (!(pspill[x] & 1))
3139 pspill[++y] = pspill[x];
3143 /* Find len = length of merging our dirty list with parent's */
3145 dst[0].mid = 0; /* simplify loops */
3146 if (parent->mt_parent) {
3147 len = x + src[0].mid;
3148 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3149 for (i = x; y && i; y--) {
3150 pgno_t yp = src[y].mid;
3151 while (yp < dst[i].mid)
3153 if (yp == dst[i].mid) {
3158 } else { /* Simplify the above for single-ancestor case */
3159 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3161 /* Merge our dirty list with parent's */
3163 for (i = len; y; dst[i--] = src[y--]) {
3164 pgno_t yp = src[y].mid;
3165 while (yp < dst[x].mid)
3166 dst[i--] = dst[x--];
3167 if (yp == dst[x].mid)
3168 free(dst[x--].mptr);
3170 mdb_tassert(txn, i == x);
3172 free(txn->mt_u.dirty_list);
3173 parent->mt_dirty_room = txn->mt_dirty_room;
3174 if (txn->mt_spill_pgs) {
3175 if (parent->mt_spill_pgs) {
3176 /* TODO: Prevent failure here, so parent does not fail */
3177 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3179 parent->mt_flags |= MDB_TXN_ERROR;
3180 mdb_midl_free(txn->mt_spill_pgs);
3181 mdb_midl_sort(parent->mt_spill_pgs);
3183 parent->mt_spill_pgs = txn->mt_spill_pgs;
3187 parent->mt_child = NULL;
3188 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3193 if (txn != env->me_txn) {
3194 DPUTS("attempt to commit unknown transaction");
3199 mdb_cursors_close(txn, 0);
3201 if (!txn->mt_u.dirty_list[0].mid &&
3202 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3205 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3206 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3208 /* Update DB root pointers */
3209 if (txn->mt_numdbs > 2) {
3213 data.mv_size = sizeof(MDB_db);
3215 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3216 for (i = 2; i < txn->mt_numdbs; i++) {
3217 if (txn->mt_dbflags[i] & DB_DIRTY) {
3218 data.mv_data = &txn->mt_dbs[i];
3219 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3226 rc = mdb_freelist_save(txn);
3230 mdb_midl_free(env->me_pghead);
3231 env->me_pghead = NULL;
3232 if (mdb_midl_shrink(&txn->mt_free_pgs))
3233 env->me_free_pgs = txn->mt_free_pgs;
3239 if ((rc = mdb_page_flush(txn, 0)) ||
3240 (rc = mdb_env_sync(env, 0)) ||
3241 (rc = mdb_env_write_meta(txn)))
3247 mdb_dbis_update(txn, 1);
3250 UNLOCK_MUTEX_W(env);
3260 /** Read the environment parameters of a DB environment before
3261 * mapping it into memory.
3262 * @param[in] env the environment handle
3263 * @param[out] meta address of where to store the meta information
3264 * @return 0 on success, non-zero on failure.
3267 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3273 enum { Size = sizeof(pbuf) };
3275 /* We don't know the page size yet, so use a minimum value.
3276 * Read both meta pages so we can use the latest one.
3279 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3283 memset(&ov, 0, sizeof(ov));
3285 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3286 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3289 rc = pread(env->me_fd, &pbuf, Size, off);
3292 if (rc == 0 && off == 0)
3294 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3295 DPRINTF(("read: %s", mdb_strerror(rc)));
3299 p = (MDB_page *)&pbuf;
3301 if (!F_ISSET(p->mp_flags, P_META)) {
3302 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3307 if (m->mm_magic != MDB_MAGIC) {
3308 DPUTS("meta has invalid magic");
3312 if (m->mm_version != MDB_DATA_VERSION) {
3313 DPRINTF(("database is version %u, expected version %u",
3314 m->mm_version, MDB_DATA_VERSION));
3315 return MDB_VERSION_MISMATCH;
3318 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3325 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3327 meta->mm_magic = MDB_MAGIC;
3328 meta->mm_version = MDB_DATA_VERSION;
3329 meta->mm_mapsize = env->me_mapsize;
3330 meta->mm_psize = env->me_psize;
3331 meta->mm_last_pg = 1;
3332 meta->mm_flags = env->me_flags & 0xffff;
3333 meta->mm_flags |= MDB_INTEGERKEY;
3334 meta->mm_dbs[0].md_root = P_INVALID;
3335 meta->mm_dbs[1].md_root = P_INVALID;
3338 /** Write the environment parameters of a freshly created DB environment.
3339 * @param[in] env the environment handle
3340 * @param[out] meta address of where to store the meta information
3341 * @return 0 on success, non-zero on failure.
3344 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3352 memset(&ov, 0, sizeof(ov));
3353 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3355 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3358 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3359 len = pwrite(fd, ptr, size, pos); \
3360 rc = (len >= 0); } while(0)
3363 DPUTS("writing new meta page");
3365 psize = env->me_psize;
3367 mdb_env_init_meta0(env, meta);
3369 p = calloc(2, psize);
3371 p->mp_flags = P_META;
3372 *(MDB_meta *)METADATA(p) = *meta;
3374 q = (MDB_page *)((char *)p + psize);
3376 q->mp_flags = P_META;
3377 *(MDB_meta *)METADATA(q) = *meta;
3379 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3382 else if ((unsigned) len == psize * 2)
3390 /** Update the environment info to commit a transaction.
3391 * @param[in] txn the transaction that's being committed
3392 * @return 0 on success, non-zero on failure.
3395 mdb_env_write_meta(MDB_txn *txn)
3398 MDB_meta meta, metab, *mp;
3400 int rc, len, toggle;
3409 toggle = txn->mt_txnid & 1;
3410 DPRINTF(("writing meta page %d for root page %"Z"u",
3411 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3414 mp = env->me_metas[toggle];
3416 if (env->me_flags & MDB_WRITEMAP) {
3417 /* Persist any increases of mapsize config */
3418 if (env->me_mapsize > mp->mm_mapsize)
3419 mp->mm_mapsize = env->me_mapsize;
3420 mp->mm_dbs[0] = txn->mt_dbs[0];
3421 mp->mm_dbs[1] = txn->mt_dbs[1];
3422 mp->mm_last_pg = txn->mt_next_pgno - 1;
3423 mp->mm_txnid = txn->mt_txnid;
3424 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3425 unsigned meta_size = env->me_psize;
3426 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3429 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3430 if (meta_size < env->me_os_psize)
3431 meta_size += meta_size;
3436 if (MDB_MSYNC(ptr, meta_size, rc)) {
3443 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3444 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3446 ptr = (char *)&meta;
3447 if (env->me_mapsize > mp->mm_mapsize) {
3448 /* Persist any increases of mapsize config */
3449 meta.mm_mapsize = env->me_mapsize;
3450 off = offsetof(MDB_meta, mm_mapsize);
3452 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3454 len = sizeof(MDB_meta) - off;
3457 meta.mm_dbs[0] = txn->mt_dbs[0];
3458 meta.mm_dbs[1] = txn->mt_dbs[1];
3459 meta.mm_last_pg = txn->mt_next_pgno - 1;
3460 meta.mm_txnid = txn->mt_txnid;
3463 off += env->me_psize;
3466 /* Write to the SYNC fd */
3467 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3468 env->me_fd : env->me_mfd;
3471 memset(&ov, 0, sizeof(ov));
3473 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3477 rc = pwrite(mfd, ptr, len, off);
3480 rc = rc < 0 ? ErrCode() : EIO;
3481 DPUTS("write failed, disk error?");
3482 /* On a failure, the pagecache still contains the new data.
3483 * Write some old data back, to prevent it from being used.
3484 * Use the non-SYNC fd; we know it will fail anyway.
3486 meta.mm_last_pg = metab.mm_last_pg;
3487 meta.mm_txnid = metab.mm_txnid;
3489 memset(&ov, 0, sizeof(ov));
3491 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3493 r2 = pwrite(env->me_fd, ptr, len, off);
3494 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3497 env->me_flags |= MDB_FATAL_ERROR;
3501 /* Memory ordering issues are irrelevant; since the entire writer
3502 * is wrapped by wmutex, all of these changes will become visible
3503 * after the wmutex is unlocked. Since the DB is multi-version,
3504 * readers will get consistent data regardless of how fresh or
3505 * how stale their view of these values is.
3508 env->me_txns->mti_txnid = txn->mt_txnid;
3513 /** Check both meta pages to see which one is newer.
3514 * @param[in] env the environment handle
3515 * @return meta toggle (0 or 1).
3518 mdb_env_pick_meta(const MDB_env *env)
3520 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3524 mdb_env_create(MDB_env **env)
3528 e = calloc(1, sizeof(MDB_env));
3532 e->me_maxreaders = DEFAULT_READERS;
3533 e->me_maxdbs = e->me_numdbs = 2;
3534 e->me_fd = INVALID_HANDLE_VALUE;
3535 e->me_lfd = INVALID_HANDLE_VALUE;
3536 e->me_mfd = INVALID_HANDLE_VALUE;
3537 #ifdef MDB_USE_POSIX_SEM
3538 e->me_rmutex = SEM_FAILED;
3539 e->me_wmutex = SEM_FAILED;
3541 e->me_pid = getpid();
3542 GET_PAGESIZE(e->me_os_psize);
3543 VGMEMP_CREATE(e,0,0);
3549 mdb_env_map(MDB_env *env, void *addr, int newsize)
3552 unsigned int flags = env->me_flags;
3556 LONG sizelo, sizehi;
3559 if (flags & MDB_RDONLY) {
3564 msize = env->me_mapsize;
3565 sizelo = msize & 0xffffffff;
3566 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3569 /* Windows won't create mappings for zero length files.
3570 * Just allocate the maxsize right now.
3573 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3574 || !SetEndOfFile(env->me_fd)
3575 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3578 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3579 PAGE_READWRITE : PAGE_READONLY,
3580 sizehi, sizelo, NULL);
3583 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3584 FILE_MAP_WRITE : FILE_MAP_READ,
3586 rc = env->me_map ? 0 : ErrCode();
3591 int prot = PROT_READ;
3592 if (flags & MDB_WRITEMAP) {
3594 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3597 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3599 if (env->me_map == MAP_FAILED) {
3604 if (flags & MDB_NORDAHEAD) {
3605 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3607 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3609 #ifdef POSIX_MADV_RANDOM
3610 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3611 #endif /* POSIX_MADV_RANDOM */
3612 #endif /* MADV_RANDOM */
3616 /* Can happen because the address argument to mmap() is just a
3617 * hint. mmap() can pick another, e.g. if the range is in use.
3618 * The MAP_FIXED flag would prevent that, but then mmap could
3619 * instead unmap existing pages to make room for the new map.
3621 if (addr && env->me_map != addr)
3622 return EBUSY; /* TODO: Make a new MDB_* error code? */
3624 p = (MDB_page *)env->me_map;
3625 env->me_metas[0] = METADATA(p);
3626 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3632 mdb_env_set_mapsize(MDB_env *env, size_t size)
3634 /* If env is already open, caller is responsible for making
3635 * sure there are no active txns.
3643 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3644 else if (size < env->me_mapsize) {
3645 /* If the configured size is smaller, make sure it's
3646 * still big enough. Silently round up to minimum if not.
3648 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3652 munmap(env->me_map, env->me_mapsize);
3653 env->me_mapsize = size;
3654 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3655 rc = mdb_env_map(env, old, 1);
3659 env->me_mapsize = size;
3661 env->me_maxpg = env->me_mapsize / env->me_psize;
3666 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3670 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3675 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3677 if (env->me_map || readers < 1)
3679 env->me_maxreaders = readers;
3684 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3686 if (!env || !readers)
3688 *readers = env->me_maxreaders;
3692 /** Further setup required for opening an LMDB environment
3695 mdb_env_open2(MDB_env *env)
3697 unsigned int flags = env->me_flags;
3698 int i, newenv = 0, rc;
3702 /* See if we should use QueryLimited */
3704 if ((rc & 0xff) > 5)
3705 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3707 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3710 memset(&meta, 0, sizeof(meta));
3712 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3715 DPUTS("new mdbenv");
3717 env->me_psize = env->me_os_psize;
3718 if (env->me_psize > MAX_PAGESIZE)
3719 env->me_psize = MAX_PAGESIZE;
3721 env->me_psize = meta.mm_psize;
3724 /* Was a mapsize configured? */
3725 if (!env->me_mapsize) {
3726 /* If this is a new environment, take the default,
3727 * else use the size recorded in the existing env.
3729 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3730 } else if (env->me_mapsize < meta.mm_mapsize) {
3731 /* If the configured size is smaller, make sure it's
3732 * still big enough. Silently round up to minimum if not.
3734 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3735 if (env->me_mapsize < minsize)
3736 env->me_mapsize = minsize;
3739 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3744 if (flags & MDB_FIXEDMAP)
3745 meta.mm_address = env->me_map;
3746 i = mdb_env_init_meta(env, &meta);
3747 if (i != MDB_SUCCESS) {
3752 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3753 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3755 #if !(MDB_MAXKEYSIZE)
3756 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3758 env->me_maxpg = env->me_mapsize / env->me_psize;
3762 int toggle = mdb_env_pick_meta(env);
3763 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3765 DPRINTF(("opened database version %u, pagesize %u",
3766 env->me_metas[0]->mm_version, env->me_psize));
3767 DPRINTF(("using meta page %d", toggle));
3768 DPRINTF(("depth: %u", db->md_depth));
3769 DPRINTF(("entries: %"Z"u", db->md_entries));
3770 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3771 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3772 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3773 DPRINTF(("root: %"Z"u", db->md_root));
3781 /** Release a reader thread's slot in the reader lock table.
3782 * This function is called automatically when a thread exits.
3783 * @param[in] ptr This points to the slot in the reader lock table.
3786 mdb_env_reader_dest(void *ptr)
3788 MDB_reader *reader = ptr;
3794 /** Junk for arranging thread-specific callbacks on Windows. This is
3795 * necessarily platform and compiler-specific. Windows supports up
3796 * to 1088 keys. Let's assume nobody opens more than 64 environments
3797 * in a single process, for now. They can override this if needed.
3799 #ifndef MAX_TLS_KEYS
3800 #define MAX_TLS_KEYS 64
3802 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3803 static int mdb_tls_nkeys;
3805 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3809 case DLL_PROCESS_ATTACH: break;
3810 case DLL_THREAD_ATTACH: break;
3811 case DLL_THREAD_DETACH:
3812 for (i=0; i<mdb_tls_nkeys; i++) {
3813 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3815 mdb_env_reader_dest(r);
3819 case DLL_PROCESS_DETACH: break;
3824 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3826 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3830 /* Force some symbol references.
3831 * _tls_used forces the linker to create the TLS directory if not already done
3832 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3834 #pragma comment(linker, "/INCLUDE:_tls_used")
3835 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3836 #pragma const_seg(".CRT$XLB")
3837 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3838 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3841 #pragma comment(linker, "/INCLUDE:__tls_used")
3842 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3843 #pragma data_seg(".CRT$XLB")
3844 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3846 #endif /* WIN 32/64 */
3847 #endif /* !__GNUC__ */
3850 /** Downgrade the exclusive lock on the region back to shared */
3852 mdb_env_share_locks(MDB_env *env, int *excl)
3854 int rc = 0, toggle = mdb_env_pick_meta(env);
3856 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3861 /* First acquire a shared lock. The Unlock will
3862 * then release the existing exclusive lock.
3864 memset(&ov, 0, sizeof(ov));
3865 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3868 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3874 struct flock lock_info;
3875 /* The shared lock replaces the existing lock */
3876 memset((void *)&lock_info, 0, sizeof(lock_info));
3877 lock_info.l_type = F_RDLCK;
3878 lock_info.l_whence = SEEK_SET;
3879 lock_info.l_start = 0;
3880 lock_info.l_len = 1;
3881 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3882 (rc = ErrCode()) == EINTR) ;
3883 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3890 /** Try to get exlusive lock, otherwise shared.
3891 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3894 mdb_env_excl_lock(MDB_env *env, int *excl)
3898 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3902 memset(&ov, 0, sizeof(ov));
3903 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3910 struct flock lock_info;
3911 memset((void *)&lock_info, 0, sizeof(lock_info));
3912 lock_info.l_type = F_WRLCK;
3913 lock_info.l_whence = SEEK_SET;
3914 lock_info.l_start = 0;
3915 lock_info.l_len = 1;
3916 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3917 (rc = ErrCode()) == EINTR) ;
3921 # ifdef MDB_USE_POSIX_SEM
3922 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3925 lock_info.l_type = F_RDLCK;
3926 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3927 (rc = ErrCode()) == EINTR) ;
3937 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3939 * @(#) $Revision: 5.1 $
3940 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3941 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3943 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3947 * Please do not copyright this code. This code is in the public domain.
3949 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3950 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3951 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3952 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3953 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3954 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3955 * PERFORMANCE OF THIS SOFTWARE.
3958 * chongo <Landon Curt Noll> /\oo/\
3959 * http://www.isthe.com/chongo/
3961 * Share and Enjoy! :-)
3964 typedef unsigned long long mdb_hash_t;
3965 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3967 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3968 * @param[in] val value to hash
3969 * @param[in] hval initial value for hash
3970 * @return 64 bit hash
3972 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3973 * hval arg on the first call.
3976 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3978 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3979 unsigned char *end = s + val->mv_size;
3981 * FNV-1a hash each octet of the string
3984 /* xor the bottom with the current octet */
3985 hval ^= (mdb_hash_t)*s++;
3987 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3988 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3989 (hval << 7) + (hval << 8) + (hval << 40);
3991 /* return our new hash value */
3995 /** Hash the string and output the encoded hash.
3996 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3997 * very short name limits. We don't care about the encoding being reversible,
3998 * we just want to preserve as many bits of the input as possible in a
3999 * small printable string.
4000 * @param[in] str string to hash
4001 * @param[out] encbuf an array of 11 chars to hold the hash
4003 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4006 mdb_pack85(unsigned long l, char *out)
4010 for (i=0; i<5; i++) {
4011 *out++ = mdb_a85[l % 85];
4017 mdb_hash_enc(MDB_val *val, char *encbuf)
4019 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4021 mdb_pack85(h, encbuf);
4022 mdb_pack85(h>>32, encbuf+5);
4027 /** Open and/or initialize the lock region for the environment.
4028 * @param[in] env The LMDB environment.
4029 * @param[in] lpath The pathname of the file used for the lock region.
4030 * @param[in] mode The Unix permissions for the file, if we create it.
4031 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4032 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4033 * @return 0 on success, non-zero on failure.
4036 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4039 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4041 # define MDB_ERRCODE_ROFS EROFS
4042 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4043 # define MDB_CLOEXEC O_CLOEXEC
4046 # define MDB_CLOEXEC 0
4053 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4054 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4055 FILE_ATTRIBUTE_NORMAL, NULL);
4057 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4059 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4061 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4066 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4067 /* Lose record locks when exec*() */
4068 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4069 fcntl(env->me_lfd, F_SETFD, fdflags);
4072 if (!(env->me_flags & MDB_NOTLS)) {
4073 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4076 env->me_flags |= MDB_ENV_TXKEY;
4078 /* Windows TLS callbacks need help finding their TLS info. */
4079 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4083 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4087 /* Try to get exclusive lock. If we succeed, then
4088 * nobody is using the lock region and we should initialize it.
4090 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4093 size = GetFileSize(env->me_lfd, NULL);
4095 size = lseek(env->me_lfd, 0, SEEK_END);
4096 if (size == -1) goto fail_errno;
4098 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4099 if (size < rsize && *excl > 0) {
4101 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4102 || !SetEndOfFile(env->me_lfd))
4105 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4109 size = rsize - sizeof(MDB_txninfo);
4110 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4115 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4117 if (!mh) goto fail_errno;
4118 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4120 if (!env->me_txns) goto fail_errno;
4122 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4124 if (m == MAP_FAILED) goto fail_errno;
4130 BY_HANDLE_FILE_INFORMATION stbuf;
4139 if (!mdb_sec_inited) {
4140 InitializeSecurityDescriptor(&mdb_null_sd,
4141 SECURITY_DESCRIPTOR_REVISION);
4142 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4143 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4144 mdb_all_sa.bInheritHandle = FALSE;
4145 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4148 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4149 idbuf.volume = stbuf.dwVolumeSerialNumber;
4150 idbuf.nhigh = stbuf.nFileIndexHigh;
4151 idbuf.nlow = stbuf.nFileIndexLow;
4152 val.mv_data = &idbuf;
4153 val.mv_size = sizeof(idbuf);
4154 mdb_hash_enc(&val, encbuf);
4155 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4156 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4157 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4158 if (!env->me_rmutex) goto fail_errno;
4159 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4160 if (!env->me_wmutex) goto fail_errno;
4161 #elif defined(MDB_USE_POSIX_SEM)
4170 #if defined(__NetBSD__)
4171 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4173 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4174 idbuf.dev = stbuf.st_dev;
4175 idbuf.ino = stbuf.st_ino;
4176 val.mv_data = &idbuf;
4177 val.mv_size = sizeof(idbuf);
4178 mdb_hash_enc(&val, encbuf);
4179 #ifdef MDB_SHORT_SEMNAMES
4180 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4182 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4183 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4184 /* Clean up after a previous run, if needed: Try to
4185 * remove both semaphores before doing anything else.
4187 sem_unlink(env->me_txns->mti_rmname);
4188 sem_unlink(env->me_txns->mti_wmname);
4189 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4190 O_CREAT|O_EXCL, mode, 1);
4191 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4192 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4193 O_CREAT|O_EXCL, mode, 1);
4194 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4195 #else /* MDB_USE_POSIX_SEM */
4196 pthread_mutexattr_t mattr;
4198 if ((rc = pthread_mutexattr_init(&mattr))
4199 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4200 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4201 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4203 pthread_mutexattr_destroy(&mattr);
4204 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4206 env->me_txns->mti_magic = MDB_MAGIC;
4207 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4208 env->me_txns->mti_txnid = 0;
4209 env->me_txns->mti_numreaders = 0;
4212 if (env->me_txns->mti_magic != MDB_MAGIC) {
4213 DPUTS("lock region has invalid magic");
4217 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4218 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4219 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4220 rc = MDB_VERSION_MISMATCH;
4224 if (rc && rc != EACCES && rc != EAGAIN) {
4228 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4229 if (!env->me_rmutex) goto fail_errno;
4230 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4231 if (!env->me_wmutex) goto fail_errno;
4232 #elif defined(MDB_USE_POSIX_SEM)
4233 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4234 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4235 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4236 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4247 /** The name of the lock file in the DB environment */
4248 #define LOCKNAME "/lock.mdb"
4249 /** The name of the data file in the DB environment */
4250 #define DATANAME "/data.mdb"
4251 /** The suffix of the lock file when no subdir is used */
4252 #define LOCKSUFF "-lock"
4253 /** Only a subset of the @ref mdb_env flags can be changed
4254 * at runtime. Changing other flags requires closing the
4255 * environment and re-opening it with the new flags.
4257 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4258 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4259 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4261 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4262 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4266 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4268 int oflags, rc, len, excl = -1;
4269 char *lpath, *dpath;
4271 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4275 if (flags & MDB_NOSUBDIR) {
4276 rc = len + sizeof(LOCKSUFF) + len + 1;
4278 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4283 if (flags & MDB_NOSUBDIR) {
4284 dpath = lpath + len + sizeof(LOCKSUFF);
4285 sprintf(lpath, "%s" LOCKSUFF, path);
4286 strcpy(dpath, path);
4288 dpath = lpath + len + sizeof(LOCKNAME);
4289 sprintf(lpath, "%s" LOCKNAME, path);
4290 sprintf(dpath, "%s" DATANAME, path);
4294 flags |= env->me_flags;
4295 if (flags & MDB_RDONLY) {
4296 /* silently ignore WRITEMAP when we're only getting read access */
4297 flags &= ~MDB_WRITEMAP;
4299 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4300 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4303 env->me_flags = flags |= MDB_ENV_ACTIVE;
4307 env->me_path = strdup(path);
4308 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4309 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4310 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4315 /* For RDONLY, get lockfile after we know datafile exists */
4316 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4317 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4323 if (F_ISSET(flags, MDB_RDONLY)) {
4324 oflags = GENERIC_READ;
4325 len = OPEN_EXISTING;
4327 oflags = GENERIC_READ|GENERIC_WRITE;
4330 mode = FILE_ATTRIBUTE_NORMAL;
4331 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4332 NULL, len, mode, NULL);
4334 if (F_ISSET(flags, MDB_RDONLY))
4337 oflags = O_RDWR | O_CREAT;
4339 env->me_fd = open(dpath, oflags, mode);
4341 if (env->me_fd == INVALID_HANDLE_VALUE) {
4346 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4347 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4352 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4353 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4354 env->me_mfd = env->me_fd;
4356 /* Synchronous fd for meta writes. Needed even with
4357 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4360 len = OPEN_EXISTING;
4361 env->me_mfd = CreateFile(dpath, oflags,
4362 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4363 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4366 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4368 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4373 DPRINTF(("opened dbenv %p", (void *) env));
4375 rc = mdb_env_share_locks(env, &excl);
4379 if (!((flags & MDB_RDONLY) ||
4380 (env->me_pbuf = calloc(1, env->me_psize))))
4386 mdb_env_close0(env, excl);
4392 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4394 mdb_env_close0(MDB_env *env, int excl)
4398 if (!(env->me_flags & MDB_ENV_ACTIVE))
4401 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4402 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4403 free(env->me_dbxs[i].md_name.mv_data);
4406 free(env->me_dbflags);
4409 free(env->me_dirty_list);
4410 mdb_midl_free(env->me_free_pgs);
4412 if (env->me_flags & MDB_ENV_TXKEY) {
4413 pthread_key_delete(env->me_txkey);
4415 /* Delete our key from the global list */
4416 for (i=0; i<mdb_tls_nkeys; i++)
4417 if (mdb_tls_keys[i] == env->me_txkey) {
4418 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4426 munmap(env->me_map, env->me_mapsize);
4428 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4429 (void) close(env->me_mfd);
4430 if (env->me_fd != INVALID_HANDLE_VALUE)
4431 (void) close(env->me_fd);
4433 MDB_PID_T pid = env->me_pid;
4434 /* Clearing readers is done in this function because
4435 * me_txkey with its destructor must be disabled first.
4437 for (i = env->me_numreaders; --i >= 0; )
4438 if (env->me_txns->mti_readers[i].mr_pid == pid)
4439 env->me_txns->mti_readers[i].mr_pid = 0;
4441 if (env->me_rmutex) {
4442 CloseHandle(env->me_rmutex);
4443 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4445 /* Windows automatically destroys the mutexes when
4446 * the last handle closes.
4448 #elif defined(MDB_USE_POSIX_SEM)
4449 if (env->me_rmutex != SEM_FAILED) {
4450 sem_close(env->me_rmutex);
4451 if (env->me_wmutex != SEM_FAILED)
4452 sem_close(env->me_wmutex);
4453 /* If we have the filelock: If we are the
4454 * only remaining user, clean up semaphores.
4457 mdb_env_excl_lock(env, &excl);
4459 sem_unlink(env->me_txns->mti_rmname);
4460 sem_unlink(env->me_txns->mti_wmname);
4464 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4466 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4469 /* Unlock the lockfile. Windows would have unlocked it
4470 * after closing anyway, but not necessarily at once.
4472 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4475 (void) close(env->me_lfd);
4478 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4483 mdb_env_close(MDB_env *env)
4490 VGMEMP_DESTROY(env);
4491 while ((dp = env->me_dpages) != NULL) {
4492 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4493 env->me_dpages = dp->mp_next;
4497 mdb_env_close0(env, 0);
4501 /** Compare two items pointing at aligned size_t's */
4503 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4505 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4506 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4509 /** Compare two items pointing at aligned unsigned int's */
4511 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4513 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4514 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4517 /** Compare two items pointing at unsigned ints of unknown alignment.
4518 * Nodes and keys are guaranteed to be 2-byte aligned.
4521 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4523 #if BYTE_ORDER == LITTLE_ENDIAN
4524 unsigned short *u, *c;
4527 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4528 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4531 } while(!x && u > (unsigned short *)a->mv_data);
4534 unsigned short *u, *c, *end;
4537 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4538 u = (unsigned short *)a->mv_data;
4539 c = (unsigned short *)b->mv_data;
4542 } while(!x && u < end);
4547 /** Compare two items pointing at size_t's of unknown alignment. */
4548 #ifdef MISALIGNED_OK
4549 # define mdb_cmp_clong mdb_cmp_long
4551 # define mdb_cmp_clong mdb_cmp_cint
4554 /** Compare two items lexically */
4556 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4563 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4569 diff = memcmp(a->mv_data, b->mv_data, len);
4570 return diff ? diff : len_diff<0 ? -1 : len_diff;
4573 /** Compare two items in reverse byte order */
4575 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4577 const unsigned char *p1, *p2, *p1_lim;
4581 p1_lim = (const unsigned char *)a->mv_data;
4582 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4583 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4585 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4591 while (p1 > p1_lim) {
4592 diff = *--p1 - *--p2;
4596 return len_diff<0 ? -1 : len_diff;
4599 /** Search for key within a page, using binary search.
4600 * Returns the smallest entry larger or equal to the key.
4601 * If exactp is non-null, stores whether the found entry was an exact match
4602 * in *exactp (1 or 0).
4603 * Updates the cursor index with the index of the found entry.
4604 * If no entry larger or equal to the key is found, returns NULL.
4607 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4609 unsigned int i = 0, nkeys;
4612 MDB_page *mp = mc->mc_pg[mc->mc_top];
4613 MDB_node *node = NULL;
4618 nkeys = NUMKEYS(mp);
4620 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4621 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4624 low = IS_LEAF(mp) ? 0 : 1;
4626 cmp = mc->mc_dbx->md_cmp;
4628 /* Branch pages have no data, so if using integer keys,
4629 * alignment is guaranteed. Use faster mdb_cmp_int.
4631 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4632 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4639 nodekey.mv_size = mc->mc_db->md_pad;
4640 node = NODEPTR(mp, 0); /* fake */
4641 while (low <= high) {
4642 i = (low + high) >> 1;
4643 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4644 rc = cmp(key, &nodekey);
4645 DPRINTF(("found leaf index %u [%s], rc = %i",
4646 i, DKEY(&nodekey), rc));
4655 while (low <= high) {
4656 i = (low + high) >> 1;
4658 node = NODEPTR(mp, i);
4659 nodekey.mv_size = NODEKSZ(node);
4660 nodekey.mv_data = NODEKEY(node);
4662 rc = cmp(key, &nodekey);
4665 DPRINTF(("found leaf index %u [%s], rc = %i",
4666 i, DKEY(&nodekey), rc));
4668 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4669 i, DKEY(&nodekey), NODEPGNO(node), rc));
4680 if (rc > 0) { /* Found entry is less than the key. */
4681 i++; /* Skip to get the smallest entry larger than key. */
4683 node = NODEPTR(mp, i);
4686 *exactp = (rc == 0 && nkeys > 0);
4687 /* store the key index */
4688 mc->mc_ki[mc->mc_top] = i;
4690 /* There is no entry larger or equal to the key. */
4693 /* nodeptr is fake for LEAF2 */
4699 mdb_cursor_adjust(MDB_cursor *mc, func)
4703 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4704 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4711 /** Pop a page off the top of the cursor's stack. */
4713 mdb_cursor_pop(MDB_cursor *mc)
4717 MDB_page *top = mc->mc_pg[mc->mc_top];
4723 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4724 DDBI(mc), (void *) mc));
4728 /** Push a page onto the top of the cursor's stack. */
4730 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4732 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4733 DDBI(mc), (void *) mc));
4735 if (mc->mc_snum >= CURSOR_STACK) {
4736 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4737 return MDB_CURSOR_FULL;
4740 mc->mc_top = mc->mc_snum++;
4741 mc->mc_pg[mc->mc_top] = mp;
4742 mc->mc_ki[mc->mc_top] = 0;
4747 /** Find the address of the page corresponding to a given page number.
4748 * @param[in] txn the transaction for this access.
4749 * @param[in] pgno the page number for the page to retrieve.
4750 * @param[out] ret address of a pointer where the page's address will be stored.
4751 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4752 * @return 0 on success, non-zero on failure.
4755 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4757 MDB_env *env = txn->mt_env;
4761 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4765 MDB_ID2L dl = tx2->mt_u.dirty_list;
4767 /* Spilled pages were dirtied in this txn and flushed
4768 * because the dirty list got full. Bring this page
4769 * back in from the map (but don't unspill it here,
4770 * leave that unless page_touch happens again).
4772 if (tx2->mt_spill_pgs) {
4773 MDB_ID pn = pgno << 1;
4774 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4775 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4776 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4781 unsigned x = mdb_mid2l_search(dl, pgno);
4782 if (x <= dl[0].mid && dl[x].mid == pgno) {
4788 } while ((tx2 = tx2->mt_parent) != NULL);
4791 if (pgno < txn->mt_next_pgno) {
4793 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4795 DPRINTF(("page %"Z"u not found", pgno));
4796 txn->mt_flags |= MDB_TXN_ERROR;
4797 return MDB_PAGE_NOTFOUND;
4807 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4808 * The cursor is at the root page, set up the rest of it.
4811 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4813 MDB_page *mp = mc->mc_pg[mc->mc_top];
4817 while (IS_BRANCH(mp)) {
4821 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4822 mdb_cassert(mc, NUMKEYS(mp) > 1);
4823 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4825 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4827 if (flags & MDB_PS_LAST)
4828 i = NUMKEYS(mp) - 1;
4831 node = mdb_node_search(mc, key, &exact);
4833 i = NUMKEYS(mp) - 1;
4835 i = mc->mc_ki[mc->mc_top];
4837 mdb_cassert(mc, i > 0);
4841 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4844 mdb_cassert(mc, i < NUMKEYS(mp));
4845 node = NODEPTR(mp, i);
4847 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4850 mc->mc_ki[mc->mc_top] = i;
4851 if ((rc = mdb_cursor_push(mc, mp)))
4854 if (flags & MDB_PS_MODIFY) {
4855 if ((rc = mdb_page_touch(mc)) != 0)
4857 mp = mc->mc_pg[mc->mc_top];
4862 DPRINTF(("internal error, index points to a %02X page!?",
4864 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4865 return MDB_CORRUPTED;
4868 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4869 key ? DKEY(key) : "null"));
4870 mc->mc_flags |= C_INITIALIZED;
4871 mc->mc_flags &= ~C_EOF;
4876 /** Search for the lowest key under the current branch page.
4877 * This just bypasses a NUMKEYS check in the current page
4878 * before calling mdb_page_search_root(), because the callers
4879 * are all in situations where the current page is known to
4883 mdb_page_search_lowest(MDB_cursor *mc)
4885 MDB_page *mp = mc->mc_pg[mc->mc_top];
4886 MDB_node *node = NODEPTR(mp, 0);
4889 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4892 mc->mc_ki[mc->mc_top] = 0;
4893 if ((rc = mdb_cursor_push(mc, mp)))
4895 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4898 /** Search for the page a given key should be in.
4899 * Push it and its parent pages on the cursor stack.
4900 * @param[in,out] mc the cursor for this operation.
4901 * @param[in] key the key to search for, or NULL for first/last page.
4902 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4903 * are touched (updated with new page numbers).
4904 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4905 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4906 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4907 * @return 0 on success, non-zero on failure.
4910 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4915 /* Make sure the txn is still viable, then find the root from
4916 * the txn's db table and set it as the root of the cursor's stack.
4918 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4919 DPUTS("transaction has failed, must abort");
4922 /* Make sure we're using an up-to-date root */
4923 if (*mc->mc_dbflag & DB_STALE) {
4925 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4926 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4933 MDB_node *leaf = mdb_node_search(&mc2,
4934 &mc->mc_dbx->md_name, &exact);
4936 return MDB_NOTFOUND;
4937 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4940 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4942 /* The txn may not know this DBI, or another process may
4943 * have dropped and recreated the DB with other flags.
4945 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4946 return MDB_INCOMPATIBLE;
4947 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4949 *mc->mc_dbflag &= ~DB_STALE;
4951 root = mc->mc_db->md_root;
4953 if (root == P_INVALID) { /* Tree is empty. */
4954 DPUTS("tree is empty");
4955 return MDB_NOTFOUND;
4959 mdb_cassert(mc, root > 1);
4960 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4961 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4967 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4968 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4970 if (flags & MDB_PS_MODIFY) {
4971 if ((rc = mdb_page_touch(mc)))
4975 if (flags & MDB_PS_ROOTONLY)
4978 return mdb_page_search_root(mc, key, flags);
4982 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4984 MDB_txn *txn = mc->mc_txn;
4985 pgno_t pg = mp->mp_pgno;
4986 unsigned x = 0, ovpages = mp->mp_pages;
4987 MDB_env *env = txn->mt_env;
4988 MDB_IDL sl = txn->mt_spill_pgs;
4989 MDB_ID pn = pg << 1;
4992 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4993 /* If the page is dirty or on the spill list we just acquired it,
4994 * so we should give it back to our current free list, if any.
4995 * Otherwise put it onto the list of pages we freed in this txn.
4997 * Won't create me_pghead: me_pglast must be inited along with it.
4998 * Unsupported in nested txns: They would need to hide the page
4999 * range in ancestor txns' dirty and spilled lists.
5001 if (env->me_pghead &&
5003 ((mp->mp_flags & P_DIRTY) ||
5004 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5008 MDB_ID2 *dl, ix, iy;
5009 rc = mdb_midl_need(&env->me_pghead, ovpages);
5012 if (!(mp->mp_flags & P_DIRTY)) {
5013 /* This page is no longer spilled */
5020 /* Remove from dirty list */
5021 dl = txn->mt_u.dirty_list;
5023 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5029 mdb_cassert(mc, x > 1);
5031 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5032 txn->mt_flags |= MDB_TXN_ERROR;
5033 return MDB_CORRUPTED;
5036 if (!(env->me_flags & MDB_WRITEMAP))
5037 mdb_dpage_free(env, mp);
5039 /* Insert in me_pghead */
5040 mop = env->me_pghead;
5041 j = mop[0] + ovpages;
5042 for (i = mop[0]; i && mop[i] < pg; i--)
5048 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5052 mc->mc_db->md_overflow_pages -= ovpages;
5056 /** Return the data associated with a given node.
5057 * @param[in] txn The transaction for this operation.
5058 * @param[in] leaf The node being read.
5059 * @param[out] data Updated to point to the node's data.
5060 * @return 0 on success, non-zero on failure.
5063 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5065 MDB_page *omp; /* overflow page */
5069 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5070 data->mv_size = NODEDSZ(leaf);
5071 data->mv_data = NODEDATA(leaf);
5075 /* Read overflow data.
5077 data->mv_size = NODEDSZ(leaf);
5078 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5079 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5080 DPRINTF(("read overflow page %"Z"u failed", pgno));
5083 data->mv_data = METADATA(omp);
5089 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5090 MDB_val *key, MDB_val *data)
5097 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5099 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5102 if (txn->mt_flags & MDB_TXN_ERROR)
5105 mdb_cursor_init(&mc, txn, dbi, &mx);
5106 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5109 /** Find a sibling for a page.
5110 * Replaces the page at the top of the cursor's stack with the
5111 * specified sibling, if one exists.
5112 * @param[in] mc The cursor for this operation.
5113 * @param[in] move_right Non-zero if the right sibling is requested,
5114 * otherwise the left sibling.
5115 * @return 0 on success, non-zero on failure.
5118 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5124 if (mc->mc_snum < 2) {
5125 return MDB_NOTFOUND; /* root has no siblings */
5129 DPRINTF(("parent page is page %"Z"u, index %u",
5130 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5132 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5133 : (mc->mc_ki[mc->mc_top] == 0)) {
5134 DPRINTF(("no more keys left, moving to %s sibling",
5135 move_right ? "right" : "left"));
5136 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5137 /* undo cursor_pop before returning */
5144 mc->mc_ki[mc->mc_top]++;
5146 mc->mc_ki[mc->mc_top]--;
5147 DPRINTF(("just moving to %s index key %u",
5148 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5150 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5152 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5153 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5154 /* mc will be inconsistent if caller does mc_snum++ as above */
5155 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5159 mdb_cursor_push(mc, mp);
5161 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5166 /** Move the cursor to the next data item. */
5168 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5174 if (mc->mc_flags & C_EOF) {
5175 return MDB_NOTFOUND;
5178 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5180 mp = mc->mc_pg[mc->mc_top];
5182 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5183 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5184 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5185 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5186 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5187 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5188 if (rc == MDB_SUCCESS)
5189 MDB_GET_KEY(leaf, key);
5194 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5195 if (op == MDB_NEXT_DUP)
5196 return MDB_NOTFOUND;
5200 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5201 mdb_dbg_pgno(mp), (void *) mc));
5202 if (mc->mc_flags & C_DEL)
5205 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5206 DPUTS("=====> move to next sibling page");
5207 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5208 mc->mc_flags |= C_EOF;
5211 mp = mc->mc_pg[mc->mc_top];
5212 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5214 mc->mc_ki[mc->mc_top]++;
5217 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5218 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5221 key->mv_size = mc->mc_db->md_pad;
5222 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5226 mdb_cassert(mc, IS_LEAF(mp));
5227 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5229 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5230 mdb_xcursor_init1(mc, leaf);
5233 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5236 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5237 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5238 if (rc != MDB_SUCCESS)
5243 MDB_GET_KEY(leaf, key);
5247 /** Move the cursor to the previous data item. */
5249 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5255 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5257 mp = mc->mc_pg[mc->mc_top];
5259 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5260 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5261 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5262 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5263 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5264 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5265 if (rc == MDB_SUCCESS)
5266 MDB_GET_KEY(leaf, key);
5270 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5271 if (op == MDB_PREV_DUP)
5272 return MDB_NOTFOUND;
5277 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5278 mdb_dbg_pgno(mp), (void *) mc));
5280 if (mc->mc_ki[mc->mc_top] == 0) {
5281 DPUTS("=====> move to prev sibling page");
5282 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5285 mp = mc->mc_pg[mc->mc_top];
5286 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5287 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5289 mc->mc_ki[mc->mc_top]--;
5291 mc->mc_flags &= ~C_EOF;
5293 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5294 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5297 key->mv_size = mc->mc_db->md_pad;
5298 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5302 mdb_cassert(mc, IS_LEAF(mp));
5303 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5305 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5306 mdb_xcursor_init1(mc, leaf);
5309 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5312 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5313 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5314 if (rc != MDB_SUCCESS)
5319 MDB_GET_KEY(leaf, key);
5323 /** Set the cursor on a specific data item. */
5325 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5326 MDB_cursor_op op, int *exactp)
5330 MDB_node *leaf = NULL;
5333 if (key->mv_size == 0)
5334 return MDB_BAD_VALSIZE;
5337 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5339 /* See if we're already on the right page */
5340 if (mc->mc_flags & C_INITIALIZED) {
5343 mp = mc->mc_pg[mc->mc_top];
5345 mc->mc_ki[mc->mc_top] = 0;
5346 return MDB_NOTFOUND;
5348 if (mp->mp_flags & P_LEAF2) {
5349 nodekey.mv_size = mc->mc_db->md_pad;
5350 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5352 leaf = NODEPTR(mp, 0);
5353 MDB_GET_KEY2(leaf, nodekey);
5355 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5357 /* Probably happens rarely, but first node on the page
5358 * was the one we wanted.
5360 mc->mc_ki[mc->mc_top] = 0;
5367 unsigned int nkeys = NUMKEYS(mp);
5369 if (mp->mp_flags & P_LEAF2) {
5370 nodekey.mv_data = LEAF2KEY(mp,
5371 nkeys-1, nodekey.mv_size);
5373 leaf = NODEPTR(mp, nkeys-1);
5374 MDB_GET_KEY2(leaf, nodekey);
5376 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5378 /* last node was the one we wanted */
5379 mc->mc_ki[mc->mc_top] = nkeys-1;
5385 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5386 /* This is definitely the right page, skip search_page */
5387 if (mp->mp_flags & P_LEAF2) {
5388 nodekey.mv_data = LEAF2KEY(mp,
5389 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5391 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5392 MDB_GET_KEY2(leaf, nodekey);
5394 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5396 /* current node was the one we wanted */
5406 /* If any parents have right-sibs, search.
5407 * Otherwise, there's nothing further.
5409 for (i=0; i<mc->mc_top; i++)
5411 NUMKEYS(mc->mc_pg[i])-1)
5413 if (i == mc->mc_top) {
5414 /* There are no other pages */
5415 mc->mc_ki[mc->mc_top] = nkeys;
5416 return MDB_NOTFOUND;
5420 /* There are no other pages */
5421 mc->mc_ki[mc->mc_top] = 0;
5422 if (op == MDB_SET_RANGE && !exactp) {
5426 return MDB_NOTFOUND;
5430 rc = mdb_page_search(mc, key, 0);
5431 if (rc != MDB_SUCCESS)
5434 mp = mc->mc_pg[mc->mc_top];
5435 mdb_cassert(mc, IS_LEAF(mp));
5438 leaf = mdb_node_search(mc, key, exactp);
5439 if (exactp != NULL && !*exactp) {
5440 /* MDB_SET specified and not an exact match. */
5441 return MDB_NOTFOUND;
5445 DPUTS("===> inexact leaf not found, goto sibling");
5446 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5447 return rc; /* no entries matched */
5448 mp = mc->mc_pg[mc->mc_top];
5449 mdb_cassert(mc, IS_LEAF(mp));
5450 leaf = NODEPTR(mp, 0);
5454 mc->mc_flags |= C_INITIALIZED;
5455 mc->mc_flags &= ~C_EOF;
5458 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5459 key->mv_size = mc->mc_db->md_pad;
5460 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5465 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5466 mdb_xcursor_init1(mc, leaf);
5469 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5470 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5471 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5474 if (op == MDB_GET_BOTH) {
5480 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5481 if (rc != MDB_SUCCESS)
5484 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5486 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5488 rc = mc->mc_dbx->md_dcmp(data, &d2);
5490 if (op == MDB_GET_BOTH || rc > 0)
5491 return MDB_NOTFOUND;
5498 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5499 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5504 /* The key already matches in all other cases */
5505 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5506 MDB_GET_KEY(leaf, key);
5507 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5512 /** Move the cursor to the first item in the database. */
5514 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5520 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5522 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5523 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5524 if (rc != MDB_SUCCESS)
5527 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5529 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5530 mc->mc_flags |= C_INITIALIZED;
5531 mc->mc_flags &= ~C_EOF;
5533 mc->mc_ki[mc->mc_top] = 0;
5535 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5536 key->mv_size = mc->mc_db->md_pad;
5537 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5542 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5543 mdb_xcursor_init1(mc, leaf);
5544 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5548 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5552 MDB_GET_KEY(leaf, key);
5556 /** Move the cursor to the last item in the database. */
5558 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5564 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5566 if (!(mc->mc_flags & C_EOF)) {
5568 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5569 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5570 if (rc != MDB_SUCCESS)
5573 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5576 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5577 mc->mc_flags |= C_INITIALIZED|C_EOF;
5578 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5580 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5581 key->mv_size = mc->mc_db->md_pad;
5582 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5587 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5588 mdb_xcursor_init1(mc, leaf);
5589 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5593 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5598 MDB_GET_KEY(leaf, key);
5603 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5608 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5613 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5617 case MDB_GET_CURRENT:
5618 if (!(mc->mc_flags & C_INITIALIZED)) {
5621 MDB_page *mp = mc->mc_pg[mc->mc_top];
5622 int nkeys = NUMKEYS(mp);
5623 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5624 mc->mc_ki[mc->mc_top] = nkeys;
5630 key->mv_size = mc->mc_db->md_pad;
5631 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5633 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5634 MDB_GET_KEY(leaf, key);
5636 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5637 if (mc->mc_flags & C_DEL)
5638 mdb_xcursor_init1(mc, leaf);
5639 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5641 rc = mdb_node_read(mc->mc_txn, leaf, data);
5648 case MDB_GET_BOTH_RANGE:
5653 if (mc->mc_xcursor == NULL) {
5654 rc = MDB_INCOMPATIBLE;
5664 rc = mdb_cursor_set(mc, key, data, op,
5665 op == MDB_SET_RANGE ? NULL : &exact);
5668 case MDB_GET_MULTIPLE:
5669 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5673 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5674 rc = MDB_INCOMPATIBLE;
5678 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5679 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5682 case MDB_NEXT_MULTIPLE:
5687 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5688 rc = MDB_INCOMPATIBLE;
5691 if (!(mc->mc_flags & C_INITIALIZED))
5692 rc = mdb_cursor_first(mc, key, data);
5694 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5695 if (rc == MDB_SUCCESS) {
5696 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5699 mx = &mc->mc_xcursor->mx_cursor;
5700 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5702 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5703 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5711 case MDB_NEXT_NODUP:
5712 if (!(mc->mc_flags & C_INITIALIZED))
5713 rc = mdb_cursor_first(mc, key, data);
5715 rc = mdb_cursor_next(mc, key, data, op);
5719 case MDB_PREV_NODUP:
5720 if (!(mc->mc_flags & C_INITIALIZED)) {
5721 rc = mdb_cursor_last(mc, key, data);
5724 mc->mc_flags |= C_INITIALIZED;
5725 mc->mc_ki[mc->mc_top]++;
5727 rc = mdb_cursor_prev(mc, key, data, op);
5730 rc = mdb_cursor_first(mc, key, data);
5733 mfunc = mdb_cursor_first;
5735 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5739 if (mc->mc_xcursor == NULL) {
5740 rc = MDB_INCOMPATIBLE;
5743 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5747 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5750 rc = mdb_cursor_last(mc, key, data);
5753 mfunc = mdb_cursor_last;
5756 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5761 if (mc->mc_flags & C_DEL)
5762 mc->mc_flags ^= C_DEL;
5767 /** Touch all the pages in the cursor stack. Set mc_top.
5768 * Makes sure all the pages are writable, before attempting a write operation.
5769 * @param[in] mc The cursor to operate on.
5772 mdb_cursor_touch(MDB_cursor *mc)
5774 int rc = MDB_SUCCESS;
5776 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5779 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5780 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5783 *mc->mc_dbflag |= DB_DIRTY;
5788 rc = mdb_page_touch(mc);
5789 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5790 mc->mc_top = mc->mc_snum-1;
5795 /** Do not spill pages to disk if txn is getting full, may fail instead */
5796 #define MDB_NOSPILL 0x8000
5799 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5802 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5804 MDB_node *leaf = NULL;
5807 MDB_val xdata, *rdata, dkey, olddata;
5809 int do_sub = 0, insert_key, insert_data;
5810 unsigned int mcount = 0, dcount = 0, nospill;
5813 unsigned int nflags;
5816 if (mc == NULL || key == NULL)
5819 env = mc->mc_txn->mt_env;
5821 /* Check this first so counter will always be zero on any
5824 if (flags & MDB_MULTIPLE) {
5825 dcount = data[1].mv_size;
5826 data[1].mv_size = 0;
5827 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5828 return MDB_INCOMPATIBLE;
5831 nospill = flags & MDB_NOSPILL;
5832 flags &= ~MDB_NOSPILL;
5834 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5835 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5837 if (key->mv_size-1 >= ENV_MAXKEY(env))
5838 return MDB_BAD_VALSIZE;
5840 #if SIZE_MAX > MAXDATASIZE
5841 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5842 return MDB_BAD_VALSIZE;
5844 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5845 return MDB_BAD_VALSIZE;
5848 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5849 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5853 if (flags == MDB_CURRENT) {
5854 if (!(mc->mc_flags & C_INITIALIZED))
5857 } else if (mc->mc_db->md_root == P_INVALID) {
5858 /* new database, cursor has nothing to point to */
5861 mc->mc_flags &= ~C_INITIALIZED;
5866 if (flags & MDB_APPEND) {
5868 rc = mdb_cursor_last(mc, &k2, &d2);
5870 rc = mc->mc_dbx->md_cmp(key, &k2);
5873 mc->mc_ki[mc->mc_top]++;
5875 /* new key is <= last key */
5880 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5882 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5883 DPRINTF(("duplicate key [%s]", DKEY(key)));
5885 return MDB_KEYEXIST;
5887 if (rc && rc != MDB_NOTFOUND)
5891 if (mc->mc_flags & C_DEL)
5892 mc->mc_flags ^= C_DEL;
5894 /* Cursor is positioned, check for room in the dirty list */
5896 if (flags & MDB_MULTIPLE) {
5898 xdata.mv_size = data->mv_size * dcount;
5902 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5906 if (rc == MDB_NO_ROOT) {
5908 /* new database, write a root leaf page */
5909 DPUTS("allocating new root leaf page");
5910 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5913 mdb_cursor_push(mc, np);
5914 mc->mc_db->md_root = np->mp_pgno;
5915 mc->mc_db->md_depth++;
5916 *mc->mc_dbflag |= DB_DIRTY;
5917 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5919 np->mp_flags |= P_LEAF2;
5920 mc->mc_flags |= C_INITIALIZED;
5922 /* make sure all cursor pages are writable */
5923 rc2 = mdb_cursor_touch(mc);
5928 insert_key = insert_data = rc;
5930 /* The key does not exist */
5931 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5932 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5933 LEAFSIZE(key, data) > env->me_nodemax)
5935 /* Too big for a node, insert in sub-DB */
5936 fp_flags = P_LEAF|P_DIRTY;
5938 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5939 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5943 /* there's only a key anyway, so this is a no-op */
5944 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5946 unsigned int ksize = mc->mc_db->md_pad;
5947 if (key->mv_size != ksize)
5948 return MDB_BAD_VALSIZE;
5949 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5950 memcpy(ptr, key->mv_data, ksize);
5955 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5956 olddata.mv_size = NODEDSZ(leaf);
5957 olddata.mv_data = NODEDATA(leaf);
5960 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5961 /* Prepare (sub-)page/sub-DB to accept the new item,
5962 * if needed. fp: old sub-page or a header faking
5963 * it. mp: new (sub-)page. offset: growth in page
5964 * size. xdata: node data with new page or DB.
5966 unsigned i, offset = 0;
5967 mp = fp = xdata.mv_data = env->me_pbuf;
5968 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5970 /* Was a single item before, must convert now */
5971 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5972 /* Just overwrite the current item */
5973 if (flags == MDB_CURRENT)
5976 #if UINT_MAX < SIZE_MAX
5977 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5978 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
5980 /* does data match? */
5981 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5982 if (flags & MDB_NODUPDATA)
5983 return MDB_KEYEXIST;
5988 /* Back up original data item */
5989 dkey.mv_size = olddata.mv_size;
5990 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5992 /* Make sub-page header for the dup items, with dummy body */
5993 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5994 fp->mp_lower = PAGEHDRSZ;
5995 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5996 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5997 fp->mp_flags |= P_LEAF2;
5998 fp->mp_pad = data->mv_size;
5999 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6001 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6002 (dkey.mv_size & 1) + (data->mv_size & 1);
6004 fp->mp_upper = xdata.mv_size;
6005 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
6006 } else if (leaf->mn_flags & F_SUBDATA) {
6007 /* Data is on sub-DB, just store it */
6008 flags |= F_DUPDATA|F_SUBDATA;
6011 /* Data is on sub-page */
6012 fp = olddata.mv_data;
6015 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6016 offset = EVEN(NODESIZE + sizeof(indx_t) +
6020 offset = fp->mp_pad;
6021 if (SIZELEFT(fp) < offset) {
6022 offset *= 4; /* space for 4 more */
6025 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6027 fp->mp_flags |= P_DIRTY;
6028 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6029 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6033 xdata.mv_size = olddata.mv_size + offset;
6036 fp_flags = fp->mp_flags;
6037 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6038 /* Too big for a sub-page, convert to sub-DB */
6039 fp_flags &= ~P_SUBP;
6041 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6042 fp_flags |= P_LEAF2;
6043 dummy.md_pad = fp->mp_pad;
6044 dummy.md_flags = MDB_DUPFIXED;
6045 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6046 dummy.md_flags |= MDB_INTEGERKEY;
6052 dummy.md_branch_pages = 0;
6053 dummy.md_leaf_pages = 1;
6054 dummy.md_overflow_pages = 0;
6055 dummy.md_entries = NUMKEYS(fp);
6056 xdata.mv_size = sizeof(MDB_db);
6057 xdata.mv_data = &dummy;
6058 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6060 offset = env->me_psize - olddata.mv_size;
6061 flags |= F_DUPDATA|F_SUBDATA;
6062 dummy.md_root = mp->mp_pgno;
6065 mp->mp_flags = fp_flags | P_DIRTY;
6066 mp->mp_pad = fp->mp_pad;
6067 mp->mp_lower = fp->mp_lower;
6068 mp->mp_upper = fp->mp_upper + offset;
6069 if (fp_flags & P_LEAF2) {
6070 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6072 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6073 olddata.mv_size - fp->mp_upper);
6074 for (i=0; i<NUMKEYS(fp); i++)
6075 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6083 mdb_node_del(mc, 0);
6087 /* overflow page overwrites need special handling */
6088 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6091 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6093 memcpy(&pg, olddata.mv_data, sizeof(pg));
6094 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6096 ovpages = omp->mp_pages;
6098 /* Is the ov page large enough? */
6099 if (ovpages >= dpages) {
6100 if (!(omp->mp_flags & P_DIRTY) &&
6101 (level || (env->me_flags & MDB_WRITEMAP)))
6103 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6106 level = 0; /* dirty in this txn or clean */
6109 if (omp->mp_flags & P_DIRTY) {
6110 /* yes, overwrite it. Note in this case we don't
6111 * bother to try shrinking the page if the new data
6112 * is smaller than the overflow threshold.
6115 /* It is writable only in a parent txn */
6116 size_t sz = (size_t) env->me_psize * ovpages, off;
6117 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6123 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6124 mdb_cassert(mc, rc2 == 0);
6125 if (!(flags & MDB_RESERVE)) {
6126 /* Copy end of page, adjusting alignment so
6127 * compiler may copy words instead of bytes.
6129 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6130 memcpy((size_t *)((char *)np + off),
6131 (size_t *)((char *)omp + off), sz - off);
6134 memcpy(np, omp, sz); /* Copy beginning of page */
6137 SETDSZ(leaf, data->mv_size);
6138 if (F_ISSET(flags, MDB_RESERVE))
6139 data->mv_data = METADATA(omp);
6141 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6145 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6147 } else if (data->mv_size == olddata.mv_size) {
6148 /* same size, just replace it. Note that we could
6149 * also reuse this node if the new data is smaller,
6150 * but instead we opt to shrink the node in that case.
6152 if (F_ISSET(flags, MDB_RESERVE))
6153 data->mv_data = olddata.mv_data;
6154 else if (!(mc->mc_flags & C_SUB))
6155 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6157 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6160 mdb_node_del(mc, 0);
6166 nflags = flags & NODE_ADD_FLAGS;
6167 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6168 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6169 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6170 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6172 nflags |= MDB_SPLIT_REPLACE;
6173 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6175 /* There is room already in this leaf page. */
6176 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6177 if (rc == 0 && insert_key) {
6178 /* Adjust other cursors pointing to mp */
6179 MDB_cursor *m2, *m3;
6180 MDB_dbi dbi = mc->mc_dbi;
6181 unsigned i = mc->mc_top;
6182 MDB_page *mp = mc->mc_pg[i];
6184 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6185 if (mc->mc_flags & C_SUB)
6186 m3 = &m2->mc_xcursor->mx_cursor;
6189 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6190 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6197 if (rc == MDB_SUCCESS) {
6198 /* Now store the actual data in the child DB. Note that we're
6199 * storing the user data in the keys field, so there are strict
6200 * size limits on dupdata. The actual data fields of the child
6201 * DB are all zero size.
6209 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6210 if (flags & MDB_CURRENT) {
6211 xflags = MDB_CURRENT|MDB_NOSPILL;
6213 mdb_xcursor_init1(mc, leaf);
6214 xflags = (flags & MDB_NODUPDATA) ?
6215 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6217 /* converted, write the original data first */
6219 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6223 /* Adjust other cursors pointing to mp */
6225 unsigned i = mc->mc_top;
6226 MDB_page *mp = mc->mc_pg[i];
6228 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6229 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6230 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6231 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6232 mdb_xcursor_init1(m2, leaf);
6236 /* we've done our job */
6239 ecount = mc->mc_xcursor->mx_db.md_entries;
6240 if (flags & MDB_APPENDDUP)
6241 xflags |= MDB_APPEND;
6242 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6243 if (flags & F_SUBDATA) {
6244 void *db = NODEDATA(leaf);
6245 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6247 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6249 /* Increment count unless we just replaced an existing item. */
6251 mc->mc_db->md_entries++;
6253 /* Invalidate txn if we created an empty sub-DB */
6256 /* If we succeeded and the key didn't exist before,
6257 * make sure the cursor is marked valid.
6259 mc->mc_flags |= C_INITIALIZED;
6261 if (flags & MDB_MULTIPLE) {
6264 /* let caller know how many succeeded, if any */
6265 data[1].mv_size = mcount;
6266 if (mcount < dcount) {
6267 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6268 insert_key = insert_data = 0;
6275 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6278 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6283 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6289 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6290 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6292 if (!(mc->mc_flags & C_INITIALIZED))
6295 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6296 return MDB_NOTFOUND;
6298 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6301 rc = mdb_cursor_touch(mc);
6305 mp = mc->mc_pg[mc->mc_top];
6308 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6310 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6311 if (flags & MDB_NODUPDATA) {
6312 /* mdb_cursor_del0() will subtract the final entry */
6313 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6315 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6316 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6318 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6321 /* If sub-DB still has entries, we're done */
6322 if (mc->mc_xcursor->mx_db.md_entries) {
6323 if (leaf->mn_flags & F_SUBDATA) {
6324 /* update subDB info */
6325 void *db = NODEDATA(leaf);
6326 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6329 /* shrink fake page */
6330 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6331 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6332 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6333 /* fix other sub-DB cursors pointed at this fake page */
6334 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6335 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6336 if (m2->mc_pg[mc->mc_top] == mp &&
6337 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6338 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6341 mc->mc_db->md_entries--;
6342 mc->mc_flags |= C_DEL;
6345 /* otherwise fall thru and delete the sub-DB */
6346 mc->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
6349 if (leaf->mn_flags & F_SUBDATA) {
6350 /* add all the child DB's pages to the free list */
6351 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6357 /* add overflow pages to free list */
6358 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6362 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6363 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6364 (rc = mdb_ovpage_free(mc, omp)))
6369 return mdb_cursor_del0(mc);
6372 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6376 /** Allocate and initialize new pages for a database.
6377 * @param[in] mc a cursor on the database being added to.
6378 * @param[in] flags flags defining what type of page is being allocated.
6379 * @param[in] num the number of pages to allocate. This is usually 1,
6380 * unless allocating overflow pages for a large record.
6381 * @param[out] mp Address of a page, or NULL on failure.
6382 * @return 0 on success, non-zero on failure.
6385 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6390 if ((rc = mdb_page_alloc(mc, num, &np)))
6392 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6393 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6394 np->mp_flags = flags | P_DIRTY;
6395 np->mp_lower = PAGEHDRSZ;
6396 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6399 mc->mc_db->md_branch_pages++;
6400 else if (IS_LEAF(np))
6401 mc->mc_db->md_leaf_pages++;
6402 else if (IS_OVERFLOW(np)) {
6403 mc->mc_db->md_overflow_pages += num;
6411 /** Calculate the size of a leaf node.
6412 * The size depends on the environment's page size; if a data item
6413 * is too large it will be put onto an overflow page and the node
6414 * size will only include the key and not the data. Sizes are always
6415 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6416 * of the #MDB_node headers.
6417 * @param[in] env The environment handle.
6418 * @param[in] key The key for the node.
6419 * @param[in] data The data for the node.
6420 * @return The number of bytes needed to store the node.
6423 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6427 sz = LEAFSIZE(key, data);
6428 if (sz > env->me_nodemax) {
6429 /* put on overflow page */
6430 sz -= data->mv_size - sizeof(pgno_t);
6433 return EVEN(sz + sizeof(indx_t));
6436 /** Calculate the size of a branch node.
6437 * The size should depend on the environment's page size but since
6438 * we currently don't support spilling large keys onto overflow
6439 * pages, it's simply the size of the #MDB_node header plus the
6440 * size of the key. Sizes are always rounded up to an even number
6441 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6442 * @param[in] env The environment handle.
6443 * @param[in] key The key for the node.
6444 * @return The number of bytes needed to store the node.
6447 mdb_branch_size(MDB_env *env, MDB_val *key)
6452 if (sz > env->me_nodemax) {
6453 /* put on overflow page */
6454 /* not implemented */
6455 /* sz -= key->size - sizeof(pgno_t); */
6458 return sz + sizeof(indx_t);
6461 /** Add a node to the page pointed to by the cursor.
6462 * @param[in] mc The cursor for this operation.
6463 * @param[in] indx The index on the page where the new node should be added.
6464 * @param[in] key The key for the new node.
6465 * @param[in] data The data for the new node, if any.
6466 * @param[in] pgno The page number, if adding a branch node.
6467 * @param[in] flags Flags for the node.
6468 * @return 0 on success, non-zero on failure. Possible errors are:
6470 * <li>ENOMEM - failed to allocate overflow pages for the node.
6471 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6472 * should never happen since all callers already calculate the
6473 * page's free space before calling this function.
6477 mdb_node_add(MDB_cursor *mc, indx_t indx,
6478 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6481 size_t node_size = NODESIZE;
6485 MDB_page *mp = mc->mc_pg[mc->mc_top];
6486 MDB_page *ofp = NULL; /* overflow page */
6489 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6491 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6492 IS_LEAF(mp) ? "leaf" : "branch",
6493 IS_SUBP(mp) ? "sub-" : "",
6494 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6495 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6498 /* Move higher keys up one slot. */
6499 int ksize = mc->mc_db->md_pad, dif;
6500 char *ptr = LEAF2KEY(mp, indx, ksize);
6501 dif = NUMKEYS(mp) - indx;
6503 memmove(ptr+ksize, ptr, dif*ksize);
6504 /* insert new key */
6505 memcpy(ptr, key->mv_data, ksize);
6507 /* Just using these for counting */
6508 mp->mp_lower += sizeof(indx_t);
6509 mp->mp_upper -= ksize - sizeof(indx_t);
6513 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6515 node_size += key->mv_size;
6517 mdb_cassert(mc, data);
6518 if (F_ISSET(flags, F_BIGDATA)) {
6519 /* Data already on overflow page. */
6520 node_size += sizeof(pgno_t);
6521 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6522 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6524 /* Put data on overflow page. */
6525 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6526 data->mv_size, node_size+data->mv_size));
6527 node_size = EVEN(node_size + sizeof(pgno_t));
6528 if ((ssize_t)node_size > room)
6530 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6532 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6536 node_size += data->mv_size;
6539 node_size = EVEN(node_size);
6540 if ((ssize_t)node_size > room)
6544 /* Move higher pointers up one slot. */
6545 for (i = NUMKEYS(mp); i > indx; i--)
6546 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6548 /* Adjust free space offsets. */
6549 ofs = mp->mp_upper - node_size;
6550 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6551 mp->mp_ptrs[indx] = ofs;
6553 mp->mp_lower += sizeof(indx_t);
6555 /* Write the node data. */
6556 node = NODEPTR(mp, indx);
6557 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6558 node->mn_flags = flags;
6560 SETDSZ(node,data->mv_size);
6565 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6568 mdb_cassert(mc, key);
6570 if (F_ISSET(flags, F_BIGDATA))
6571 memcpy(node->mn_data + key->mv_size, data->mv_data,
6573 else if (F_ISSET(flags, MDB_RESERVE))
6574 data->mv_data = node->mn_data + key->mv_size;
6576 memcpy(node->mn_data + key->mv_size, data->mv_data,
6579 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6581 if (F_ISSET(flags, MDB_RESERVE))
6582 data->mv_data = METADATA(ofp);
6584 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6591 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6592 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6593 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6594 DPRINTF(("node size = %"Z"u", node_size));
6595 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6596 return MDB_PAGE_FULL;
6599 /** Delete the specified node from a page.
6600 * @param[in] mc Cursor pointing to the node to delete.
6601 * @param[in] ksize The size of a node. Only used if the page is
6602 * part of a #MDB_DUPFIXED database.
6605 mdb_node_del(MDB_cursor *mc, int ksize)
6607 MDB_page *mp = mc->mc_pg[mc->mc_top];
6608 indx_t indx = mc->mc_ki[mc->mc_top];
6610 indx_t i, j, numkeys, ptr;
6614 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6615 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6616 numkeys = NUMKEYS(mp);
6617 mdb_cassert(mc, indx < numkeys);
6620 int x = numkeys - 1 - indx;
6621 base = LEAF2KEY(mp, indx, ksize);
6623 memmove(base, base + ksize, x * ksize);
6624 mp->mp_lower -= sizeof(indx_t);
6625 mp->mp_upper += ksize - sizeof(indx_t);
6629 node = NODEPTR(mp, indx);
6630 sz = NODESIZE + node->mn_ksize;
6632 if (F_ISSET(node->mn_flags, F_BIGDATA))
6633 sz += sizeof(pgno_t);
6635 sz += NODEDSZ(node);
6639 ptr = mp->mp_ptrs[indx];
6640 for (i = j = 0; i < numkeys; i++) {
6642 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6643 if (mp->mp_ptrs[i] < ptr)
6644 mp->mp_ptrs[j] += sz;
6649 base = (char *)mp + mp->mp_upper;
6650 memmove(base + sz, base, ptr - mp->mp_upper);
6652 mp->mp_lower -= sizeof(indx_t);
6656 /** Compact the main page after deleting a node on a subpage.
6657 * @param[in] mp The main page to operate on.
6658 * @param[in] indx The index of the subpage on the main page.
6661 mdb_node_shrink(MDB_page *mp, indx_t indx)
6667 indx_t i, numkeys, ptr;
6669 node = NODEPTR(mp, indx);
6670 sp = (MDB_page *)NODEDATA(node);
6671 delta = SIZELEFT(sp);
6672 xp = (MDB_page *)((char *)sp + delta);
6674 /* shift subpage upward */
6676 nsize = NUMKEYS(sp) * sp->mp_pad;
6678 return; /* do not make the node uneven-sized */
6679 memmove(METADATA(xp), METADATA(sp), nsize);
6682 numkeys = NUMKEYS(sp);
6683 for (i=numkeys-1; i>=0; i--)
6684 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6686 xp->mp_upper = sp->mp_lower;
6687 xp->mp_lower = sp->mp_lower;
6688 xp->mp_flags = sp->mp_flags;
6689 xp->mp_pad = sp->mp_pad;
6690 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6692 nsize = NODEDSZ(node) - delta;
6693 SETDSZ(node, nsize);
6695 /* shift lower nodes upward */
6696 ptr = mp->mp_ptrs[indx];
6697 numkeys = NUMKEYS(mp);
6698 for (i = 0; i < numkeys; i++) {
6699 if (mp->mp_ptrs[i] <= ptr)
6700 mp->mp_ptrs[i] += delta;
6703 base = (char *)mp + mp->mp_upper;
6704 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6705 mp->mp_upper += delta;
6708 /** Initial setup of a sorted-dups cursor.
6709 * Sorted duplicates are implemented as a sub-database for the given key.
6710 * The duplicate data items are actually keys of the sub-database.
6711 * Operations on the duplicate data items are performed using a sub-cursor
6712 * initialized when the sub-database is first accessed. This function does
6713 * the preliminary setup of the sub-cursor, filling in the fields that
6714 * depend only on the parent DB.
6715 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6718 mdb_xcursor_init0(MDB_cursor *mc)
6720 MDB_xcursor *mx = mc->mc_xcursor;
6722 mx->mx_cursor.mc_xcursor = NULL;
6723 mx->mx_cursor.mc_txn = mc->mc_txn;
6724 mx->mx_cursor.mc_db = &mx->mx_db;
6725 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6726 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6727 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6728 mx->mx_cursor.mc_snum = 0;
6729 mx->mx_cursor.mc_top = 0;
6730 mx->mx_cursor.mc_flags = C_SUB;
6731 mx->mx_dbx.md_name.mv_size = 0;
6732 mx->mx_dbx.md_name.mv_data = NULL;
6733 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6734 mx->mx_dbx.md_dcmp = NULL;
6735 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6738 /** Final setup of a sorted-dups cursor.
6739 * Sets up the fields that depend on the data from the main cursor.
6740 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6741 * @param[in] node The data containing the #MDB_db record for the
6742 * sorted-dup database.
6745 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6747 MDB_xcursor *mx = mc->mc_xcursor;
6749 if (node->mn_flags & F_SUBDATA) {
6750 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6751 mx->mx_cursor.mc_pg[0] = 0;
6752 mx->mx_cursor.mc_snum = 0;
6753 mx->mx_cursor.mc_top = 0;
6754 mx->mx_cursor.mc_flags = C_SUB;
6756 MDB_page *fp = NODEDATA(node);
6757 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6758 mx->mx_db.md_flags = 0;
6759 mx->mx_db.md_depth = 1;
6760 mx->mx_db.md_branch_pages = 0;
6761 mx->mx_db.md_leaf_pages = 1;
6762 mx->mx_db.md_overflow_pages = 0;
6763 mx->mx_db.md_entries = NUMKEYS(fp);
6764 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6765 mx->mx_cursor.mc_snum = 1;
6766 mx->mx_cursor.mc_top = 0;
6767 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6768 mx->mx_cursor.mc_pg[0] = fp;
6769 mx->mx_cursor.mc_ki[0] = 0;
6770 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6771 mx->mx_db.md_flags = MDB_DUPFIXED;
6772 mx->mx_db.md_pad = fp->mp_pad;
6773 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6774 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6777 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6778 mx->mx_db.md_root));
6779 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6780 #if UINT_MAX < SIZE_MAX
6781 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6782 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6786 /** Initialize a cursor for a given transaction and database. */
6788 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6791 mc->mc_backup = NULL;
6794 mc->mc_db = &txn->mt_dbs[dbi];
6795 mc->mc_dbx = &txn->mt_dbxs[dbi];
6796 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6801 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6802 mdb_tassert(txn, mx != NULL);
6803 mc->mc_xcursor = mx;
6804 mdb_xcursor_init0(mc);
6806 mc->mc_xcursor = NULL;
6808 if (*mc->mc_dbflag & DB_STALE) {
6809 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6814 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6817 size_t size = sizeof(MDB_cursor);
6819 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6822 if (txn->mt_flags & MDB_TXN_ERROR)
6825 /* Allow read access to the freelist */
6826 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6829 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6830 size += sizeof(MDB_xcursor);
6832 if ((mc = malloc(size)) != NULL) {
6833 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6834 if (txn->mt_cursors) {
6835 mc->mc_next = txn->mt_cursors[dbi];
6836 txn->mt_cursors[dbi] = mc;
6837 mc->mc_flags |= C_UNTRACK;
6849 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6851 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6854 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6857 if (txn->mt_flags & MDB_TXN_ERROR)
6860 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6864 /* Return the count of duplicate data items for the current key */
6866 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6870 if (mc == NULL || countp == NULL)
6873 if (mc->mc_xcursor == NULL)
6874 return MDB_INCOMPATIBLE;
6876 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6879 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6880 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6883 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6886 *countp = mc->mc_xcursor->mx_db.md_entries;
6892 mdb_cursor_close(MDB_cursor *mc)
6894 if (mc && !mc->mc_backup) {
6895 /* remove from txn, if tracked */
6896 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6897 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6898 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6900 *prev = mc->mc_next;
6907 mdb_cursor_txn(MDB_cursor *mc)
6909 if (!mc) return NULL;
6914 mdb_cursor_dbi(MDB_cursor *mc)
6919 /** Replace the key for a branch node with a new key.
6920 * @param[in] mc Cursor pointing to the node to operate on.
6921 * @param[in] key The new key to use.
6922 * @return 0 on success, non-zero on failure.
6925 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6931 int delta, ksize, oksize;
6932 indx_t ptr, i, numkeys, indx;
6935 indx = mc->mc_ki[mc->mc_top];
6936 mp = mc->mc_pg[mc->mc_top];
6937 node = NODEPTR(mp, indx);
6938 ptr = mp->mp_ptrs[indx];
6942 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6943 k2.mv_data = NODEKEY(node);
6944 k2.mv_size = node->mn_ksize;
6945 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6947 mdb_dkey(&k2, kbuf2),
6953 /* Sizes must be 2-byte aligned. */
6954 ksize = EVEN(key->mv_size);
6955 oksize = EVEN(node->mn_ksize);
6956 delta = ksize - oksize;
6958 /* Shift node contents if EVEN(key length) changed. */
6960 if (delta > 0 && SIZELEFT(mp) < delta) {
6962 /* not enough space left, do a delete and split */
6963 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6964 pgno = NODEPGNO(node);
6965 mdb_node_del(mc, 0);
6966 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6969 numkeys = NUMKEYS(mp);
6970 for (i = 0; i < numkeys; i++) {
6971 if (mp->mp_ptrs[i] <= ptr)
6972 mp->mp_ptrs[i] -= delta;
6975 base = (char *)mp + mp->mp_upper;
6976 len = ptr - mp->mp_upper + NODESIZE;
6977 memmove(base - delta, base, len);
6978 mp->mp_upper -= delta;
6980 node = NODEPTR(mp, indx);
6983 /* But even if no shift was needed, update ksize */
6984 if (node->mn_ksize != key->mv_size)
6985 node->mn_ksize = key->mv_size;
6988 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6994 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6996 /** Move a node from csrc to cdst.
6999 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7006 unsigned short flags;
7010 /* Mark src and dst as dirty. */
7011 if ((rc = mdb_page_touch(csrc)) ||
7012 (rc = mdb_page_touch(cdst)))
7015 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7016 key.mv_size = csrc->mc_db->md_pad;
7017 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7019 data.mv_data = NULL;
7023 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7024 mdb_cassert(csrc, !((size_t)srcnode & 1));
7025 srcpg = NODEPGNO(srcnode);
7026 flags = srcnode->mn_flags;
7027 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7028 unsigned int snum = csrc->mc_snum;
7030 /* must find the lowest key below src */
7031 rc = mdb_page_search_lowest(csrc);
7034 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7035 key.mv_size = csrc->mc_db->md_pad;
7036 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7038 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7039 key.mv_size = NODEKSZ(s2);
7040 key.mv_data = NODEKEY(s2);
7042 csrc->mc_snum = snum--;
7043 csrc->mc_top = snum;
7045 key.mv_size = NODEKSZ(srcnode);
7046 key.mv_data = NODEKEY(srcnode);
7048 data.mv_size = NODEDSZ(srcnode);
7049 data.mv_data = NODEDATA(srcnode);
7051 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7052 unsigned int snum = cdst->mc_snum;
7055 /* must find the lowest key below dst */
7056 mdb_cursor_copy(cdst, &mn);
7057 rc = mdb_page_search_lowest(&mn);
7060 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7061 bkey.mv_size = mn.mc_db->md_pad;
7062 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7064 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7065 bkey.mv_size = NODEKSZ(s2);
7066 bkey.mv_data = NODEKEY(s2);
7068 mn.mc_snum = snum--;
7071 rc = mdb_update_key(&mn, &bkey);
7076 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7077 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7078 csrc->mc_ki[csrc->mc_top],
7080 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7081 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7083 /* Add the node to the destination page.
7085 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7086 if (rc != MDB_SUCCESS)
7089 /* Delete the node from the source page.
7091 mdb_node_del(csrc, key.mv_size);
7094 /* Adjust other cursors pointing to mp */
7095 MDB_cursor *m2, *m3;
7096 MDB_dbi dbi = csrc->mc_dbi;
7097 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7099 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7100 if (csrc->mc_flags & C_SUB)
7101 m3 = &m2->mc_xcursor->mx_cursor;
7104 if (m3 == csrc) continue;
7105 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7106 csrc->mc_ki[csrc->mc_top]) {
7107 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7108 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7113 /* Update the parent separators.
7115 if (csrc->mc_ki[csrc->mc_top] == 0) {
7116 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7117 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7118 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7120 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7121 key.mv_size = NODEKSZ(srcnode);
7122 key.mv_data = NODEKEY(srcnode);
7124 DPRINTF(("update separator for source page %"Z"u to [%s]",
7125 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7126 mdb_cursor_copy(csrc, &mn);
7129 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7132 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7134 indx_t ix = csrc->mc_ki[csrc->mc_top];
7135 nullkey.mv_size = 0;
7136 csrc->mc_ki[csrc->mc_top] = 0;
7137 rc = mdb_update_key(csrc, &nullkey);
7138 csrc->mc_ki[csrc->mc_top] = ix;
7139 mdb_cassert(csrc, rc == MDB_SUCCESS);
7143 if (cdst->mc_ki[cdst->mc_top] == 0) {
7144 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7145 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7146 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7148 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7149 key.mv_size = NODEKSZ(srcnode);
7150 key.mv_data = NODEKEY(srcnode);
7152 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7153 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7154 mdb_cursor_copy(cdst, &mn);
7157 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7160 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7162 indx_t ix = cdst->mc_ki[cdst->mc_top];
7163 nullkey.mv_size = 0;
7164 cdst->mc_ki[cdst->mc_top] = 0;
7165 rc = mdb_update_key(cdst, &nullkey);
7166 cdst->mc_ki[cdst->mc_top] = ix;
7167 mdb_cassert(csrc, rc == MDB_SUCCESS);
7174 /** Merge one page into another.
7175 * The nodes from the page pointed to by \b csrc will
7176 * be copied to the page pointed to by \b cdst and then
7177 * the \b csrc page will be freed.
7178 * @param[in] csrc Cursor pointing to the source page.
7179 * @param[in] cdst Cursor pointing to the destination page.
7180 * @return 0 on success, non-zero on failure.
7183 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7185 MDB_page *psrc, *pdst;
7192 psrc = csrc->mc_pg[csrc->mc_top];
7193 pdst = cdst->mc_pg[cdst->mc_top];
7195 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7197 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7198 mdb_cassert(csrc, cdst->mc_snum > 1);
7200 /* Mark dst as dirty. */
7201 if ((rc = mdb_page_touch(cdst)))
7204 /* Move all nodes from src to dst.
7206 j = nkeys = NUMKEYS(pdst);
7207 if (IS_LEAF2(psrc)) {
7208 key.mv_size = csrc->mc_db->md_pad;
7209 key.mv_data = METADATA(psrc);
7210 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7211 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7212 if (rc != MDB_SUCCESS)
7214 key.mv_data = (char *)key.mv_data + key.mv_size;
7217 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7218 srcnode = NODEPTR(psrc, i);
7219 if (i == 0 && IS_BRANCH(psrc)) {
7222 mdb_cursor_copy(csrc, &mn);
7223 /* must find the lowest key below src */
7224 rc = mdb_page_search_lowest(&mn);
7227 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7228 key.mv_size = mn.mc_db->md_pad;
7229 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7231 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7232 key.mv_size = NODEKSZ(s2);
7233 key.mv_data = NODEKEY(s2);
7236 key.mv_size = srcnode->mn_ksize;
7237 key.mv_data = NODEKEY(srcnode);
7240 data.mv_size = NODEDSZ(srcnode);
7241 data.mv_data = NODEDATA(srcnode);
7242 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7243 if (rc != MDB_SUCCESS)
7248 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7249 pdst->mp_pgno, NUMKEYS(pdst),
7250 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7252 /* Unlink the src page from parent and add to free list.
7255 mdb_node_del(csrc, 0);
7256 if (csrc->mc_ki[csrc->mc_top] == 0) {
7258 rc = mdb_update_key(csrc, &key);
7266 psrc = csrc->mc_pg[csrc->mc_top];
7267 /* If not operating on FreeDB, allow this page to be reused
7268 * in this txn. Otherwise just add to free list.
7270 rc = mdb_page_loose(csrc, psrc);
7274 csrc->mc_db->md_leaf_pages--;
7276 csrc->mc_db->md_branch_pages--;
7278 /* Adjust other cursors pointing to mp */
7279 MDB_cursor *m2, *m3;
7280 MDB_dbi dbi = csrc->mc_dbi;
7282 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7283 if (csrc->mc_flags & C_SUB)
7284 m3 = &m2->mc_xcursor->mx_cursor;
7287 if (m3 == csrc) continue;
7288 if (m3->mc_snum < csrc->mc_snum) continue;
7289 if (m3->mc_pg[csrc->mc_top] == psrc) {
7290 m3->mc_pg[csrc->mc_top] = pdst;
7291 m3->mc_ki[csrc->mc_top] += nkeys;
7296 unsigned int snum = cdst->mc_snum;
7297 uint16_t depth = cdst->mc_db->md_depth;
7298 mdb_cursor_pop(cdst);
7299 rc = mdb_rebalance(cdst);
7300 /* Did the tree shrink? */
7301 if (depth > cdst->mc_db->md_depth)
7303 cdst->mc_snum = snum;
7304 cdst->mc_top = snum-1;
7309 /** Copy the contents of a cursor.
7310 * @param[in] csrc The cursor to copy from.
7311 * @param[out] cdst The cursor to copy to.
7314 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7318 cdst->mc_txn = csrc->mc_txn;
7319 cdst->mc_dbi = csrc->mc_dbi;
7320 cdst->mc_db = csrc->mc_db;
7321 cdst->mc_dbx = csrc->mc_dbx;
7322 cdst->mc_snum = csrc->mc_snum;
7323 cdst->mc_top = csrc->mc_top;
7324 cdst->mc_flags = csrc->mc_flags;
7326 for (i=0; i<csrc->mc_snum; i++) {
7327 cdst->mc_pg[i] = csrc->mc_pg[i];
7328 cdst->mc_ki[i] = csrc->mc_ki[i];
7332 /** Rebalance the tree after a delete operation.
7333 * @param[in] mc Cursor pointing to the page where rebalancing
7335 * @return 0 on success, non-zero on failure.
7338 mdb_rebalance(MDB_cursor *mc)
7342 unsigned int ptop, minkeys;
7346 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7347 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7348 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7349 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7350 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7352 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7353 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7354 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7355 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7359 if (mc->mc_snum < 2) {
7360 MDB_page *mp = mc->mc_pg[0];
7362 DPUTS("Can't rebalance a subpage, ignoring");
7365 if (NUMKEYS(mp) == 0) {
7366 DPUTS("tree is completely empty");
7367 mc->mc_db->md_root = P_INVALID;
7368 mc->mc_db->md_depth = 0;
7369 mc->mc_db->md_leaf_pages = 0;
7370 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7373 /* Adjust cursors pointing to mp */
7376 mc->mc_flags &= ~C_INITIALIZED;
7378 MDB_cursor *m2, *m3;
7379 MDB_dbi dbi = mc->mc_dbi;
7381 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7382 if (mc->mc_flags & C_SUB)
7383 m3 = &m2->mc_xcursor->mx_cursor;
7386 if (m3->mc_snum < mc->mc_snum) continue;
7387 if (m3->mc_pg[0] == mp) {
7390 m3->mc_flags &= ~C_INITIALIZED;
7394 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7396 DPUTS("collapsing root page!");
7397 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7400 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7401 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7404 mc->mc_db->md_depth--;
7405 mc->mc_db->md_branch_pages--;
7406 mc->mc_ki[0] = mc->mc_ki[1];
7407 for (i = 1; i<mc->mc_db->md_depth; i++) {
7408 mc->mc_pg[i] = mc->mc_pg[i+1];
7409 mc->mc_ki[i] = mc->mc_ki[i+1];
7412 /* Adjust other cursors pointing to mp */
7413 MDB_cursor *m2, *m3;
7414 MDB_dbi dbi = mc->mc_dbi;
7416 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7417 if (mc->mc_flags & C_SUB)
7418 m3 = &m2->mc_xcursor->mx_cursor;
7421 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7422 if (m3->mc_pg[0] == mp) {
7425 for (i=0; i<m3->mc_snum; i++) {
7426 m3->mc_pg[i] = m3->mc_pg[i+1];
7427 m3->mc_ki[i] = m3->mc_ki[i+1];
7433 DPUTS("root page doesn't need rebalancing");
7437 /* The parent (branch page) must have at least 2 pointers,
7438 * otherwise the tree is invalid.
7440 ptop = mc->mc_top-1;
7441 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7443 /* Leaf page fill factor is below the threshold.
7444 * Try to move keys from left or right neighbor, or
7445 * merge with a neighbor page.
7450 mdb_cursor_copy(mc, &mn);
7451 mn.mc_xcursor = NULL;
7453 oldki = mc->mc_ki[mc->mc_top];
7454 if (mc->mc_ki[ptop] == 0) {
7455 /* We're the leftmost leaf in our parent.
7457 DPUTS("reading right neighbor");
7459 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7460 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7463 mn.mc_ki[mn.mc_top] = 0;
7464 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7466 /* There is at least one neighbor to the left.
7468 DPUTS("reading left neighbor");
7470 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7471 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7474 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7475 mc->mc_ki[mc->mc_top] = 0;
7478 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7479 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7480 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7482 /* If the neighbor page is above threshold and has enough keys,
7483 * move one key from it. Otherwise we should try to merge them.
7484 * (A branch page must never have less than 2 keys.)
7486 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7487 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7488 rc = mdb_node_move(&mn, mc);
7489 if (mc->mc_ki[ptop]) {
7493 if (mc->mc_ki[ptop] == 0) {
7494 rc = mdb_page_merge(&mn, mc);
7496 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7497 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7498 rc = mdb_page_merge(mc, &mn);
7499 mdb_cursor_copy(&mn, mc);
7501 mc->mc_flags &= ~C_EOF;
7503 mc->mc_ki[mc->mc_top] = oldki;
7507 /** Complete a delete operation started by #mdb_cursor_del(). */
7509 mdb_cursor_del0(MDB_cursor *mc)
7516 ki = mc->mc_ki[mc->mc_top];
7517 mdb_node_del(mc, mc->mc_db->md_pad);
7518 mc->mc_db->md_entries--;
7519 rc = mdb_rebalance(mc);
7521 if (rc == MDB_SUCCESS) {
7522 MDB_cursor *m2, *m3;
7523 MDB_dbi dbi = mc->mc_dbi;
7525 mp = mc->mc_pg[mc->mc_top];
7526 nkeys = NUMKEYS(mp);
7528 /* if mc points past last node in page, find next sibling */
7529 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7530 rc = mdb_cursor_sibling(mc, 1);
7531 if (rc == MDB_NOTFOUND)
7535 /* Adjust other cursors pointing to mp */
7536 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7537 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7538 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7540 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7542 if (m3->mc_pg[mc->mc_top] == mp) {
7543 if (m3->mc_ki[mc->mc_top] >= ki) {
7544 m3->mc_flags |= C_DEL;
7545 if (m3->mc_ki[mc->mc_top] > ki)
7546 m3->mc_ki[mc->mc_top]--;
7548 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7549 rc = mdb_cursor_sibling(m3, 1);
7550 if (rc == MDB_NOTFOUND)
7555 mc->mc_flags |= C_DEL;
7559 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7564 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7565 MDB_val *key, MDB_val *data)
7567 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7570 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7571 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7573 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7574 /* must ignore any data */
7578 return mdb_del0(txn, dbi, key, data, 0);
7582 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7583 MDB_val *key, MDB_val *data, unsigned flags)
7588 MDB_val rdata, *xdata;
7592 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7594 mdb_cursor_init(&mc, txn, dbi, &mx);
7603 flags |= MDB_NODUPDATA;
7605 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7607 /* let mdb_page_split know about this cursor if needed:
7608 * delete will trigger a rebalance; if it needs to move
7609 * a node from one page to another, it will have to
7610 * update the parent's separator key(s). If the new sepkey
7611 * is larger than the current one, the parent page may
7612 * run out of space, triggering a split. We need this
7613 * cursor to be consistent until the end of the rebalance.
7615 mc.mc_flags |= C_UNTRACK;
7616 mc.mc_next = txn->mt_cursors[dbi];
7617 txn->mt_cursors[dbi] = &mc;
7618 rc = mdb_cursor_del(&mc, flags);
7619 txn->mt_cursors[dbi] = mc.mc_next;
7624 /** Split a page and insert a new node.
7625 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7626 * The cursor will be updated to point to the actual page and index where
7627 * the node got inserted after the split.
7628 * @param[in] newkey The key for the newly inserted node.
7629 * @param[in] newdata The data for the newly inserted node.
7630 * @param[in] newpgno The page number, if the new node is a branch node.
7631 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7632 * @return 0 on success, non-zero on failure.
7635 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7636 unsigned int nflags)
7639 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7642 int i, j, split_indx, nkeys, pmax;
7643 MDB_env *env = mc->mc_txn->mt_env;
7645 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7646 MDB_page *copy = NULL;
7647 MDB_page *mp, *rp, *pp;
7652 mp = mc->mc_pg[mc->mc_top];
7653 newindx = mc->mc_ki[mc->mc_top];
7654 nkeys = NUMKEYS(mp);
7656 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7657 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7658 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7660 /* Create a right sibling. */
7661 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7663 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7665 if (mc->mc_snum < 2) {
7666 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7668 /* shift current top to make room for new parent */
7669 mc->mc_pg[1] = mc->mc_pg[0];
7670 mc->mc_ki[1] = mc->mc_ki[0];
7673 mc->mc_db->md_root = pp->mp_pgno;
7674 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7675 mc->mc_db->md_depth++;
7678 /* Add left (implicit) pointer. */
7679 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7680 /* undo the pre-push */
7681 mc->mc_pg[0] = mc->mc_pg[1];
7682 mc->mc_ki[0] = mc->mc_ki[1];
7683 mc->mc_db->md_root = mp->mp_pgno;
7684 mc->mc_db->md_depth--;
7691 ptop = mc->mc_top-1;
7692 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7695 mc->mc_flags |= C_SPLITTING;
7696 mdb_cursor_copy(mc, &mn);
7697 mn.mc_pg[mn.mc_top] = rp;
7698 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7700 if (nflags & MDB_APPEND) {
7701 mn.mc_ki[mn.mc_top] = 0;
7703 split_indx = newindx;
7707 split_indx = (nkeys+1) / 2;
7712 unsigned int lsize, rsize, ksize;
7713 /* Move half of the keys to the right sibling */
7714 x = mc->mc_ki[mc->mc_top] - split_indx;
7715 ksize = mc->mc_db->md_pad;
7716 split = LEAF2KEY(mp, split_indx, ksize);
7717 rsize = (nkeys - split_indx) * ksize;
7718 lsize = (nkeys - split_indx) * sizeof(indx_t);
7719 mp->mp_lower -= lsize;
7720 rp->mp_lower += lsize;
7721 mp->mp_upper += rsize - lsize;
7722 rp->mp_upper -= rsize - lsize;
7723 sepkey.mv_size = ksize;
7724 if (newindx == split_indx) {
7725 sepkey.mv_data = newkey->mv_data;
7727 sepkey.mv_data = split;
7730 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7731 memcpy(rp->mp_ptrs, split, rsize);
7732 sepkey.mv_data = rp->mp_ptrs;
7733 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7734 memcpy(ins, newkey->mv_data, ksize);
7735 mp->mp_lower += sizeof(indx_t);
7736 mp->mp_upper -= ksize - sizeof(indx_t);
7739 memcpy(rp->mp_ptrs, split, x * ksize);
7740 ins = LEAF2KEY(rp, x, ksize);
7741 memcpy(ins, newkey->mv_data, ksize);
7742 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7743 rp->mp_lower += sizeof(indx_t);
7744 rp->mp_upper -= ksize - sizeof(indx_t);
7745 mc->mc_ki[mc->mc_top] = x;
7746 mc->mc_pg[mc->mc_top] = rp;
7749 int psize, nsize, k;
7750 /* Maximum free space in an empty page */
7751 pmax = env->me_psize - PAGEHDRSZ;
7753 nsize = mdb_leaf_size(env, newkey, newdata);
7755 nsize = mdb_branch_size(env, newkey);
7756 nsize = EVEN(nsize);
7758 /* grab a page to hold a temporary copy */
7759 copy = mdb_page_malloc(mc->mc_txn, 1);
7764 copy->mp_pgno = mp->mp_pgno;
7765 copy->mp_flags = mp->mp_flags;
7766 copy->mp_lower = PAGEHDRSZ;
7767 copy->mp_upper = env->me_psize;
7769 /* prepare to insert */
7770 for (i=0, j=0; i<nkeys; i++) {
7772 copy->mp_ptrs[j++] = 0;
7774 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7777 /* When items are relatively large the split point needs
7778 * to be checked, because being off-by-one will make the
7779 * difference between success or failure in mdb_node_add.
7781 * It's also relevant if a page happens to be laid out
7782 * such that one half of its nodes are all "small" and
7783 * the other half of its nodes are "large." If the new
7784 * item is also "large" and falls on the half with
7785 * "large" nodes, it also may not fit.
7787 * As a final tweak, if the new item goes on the last
7788 * spot on the page (and thus, onto the new page), bias
7789 * the split so the new page is emptier than the old page.
7790 * This yields better packing during sequential inserts.
7792 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7793 /* Find split point */
7795 if (newindx <= split_indx || newindx >= nkeys) {
7797 k = newindx >= nkeys ? nkeys : split_indx+2;
7802 for (; i!=k; i+=j) {
7807 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7808 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7810 if (F_ISSET(node->mn_flags, F_BIGDATA))
7811 psize += sizeof(pgno_t);
7813 psize += NODEDSZ(node);
7815 psize = EVEN(psize);
7817 if (psize > pmax || i == k-j) {
7818 split_indx = i + (j<0);
7823 if (split_indx == newindx) {
7824 sepkey.mv_size = newkey->mv_size;
7825 sepkey.mv_data = newkey->mv_data;
7827 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7828 sepkey.mv_size = node->mn_ksize;
7829 sepkey.mv_data = NODEKEY(node);
7834 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7836 /* Copy separator key to the parent.
7838 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7842 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7847 if (mn.mc_snum == mc->mc_snum) {
7848 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7849 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7850 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7851 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7856 /* Right page might now have changed parent.
7857 * Check if left page also changed parent.
7859 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7860 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7861 for (i=0; i<ptop; i++) {
7862 mc->mc_pg[i] = mn.mc_pg[i];
7863 mc->mc_ki[i] = mn.mc_ki[i];
7865 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7866 if (mn.mc_ki[ptop]) {
7867 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7869 /* find right page's left sibling */
7870 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7871 mdb_cursor_sibling(mc, 0);
7876 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7879 mc->mc_flags ^= C_SPLITTING;
7880 if (rc != MDB_SUCCESS) {
7883 if (nflags & MDB_APPEND) {
7884 mc->mc_pg[mc->mc_top] = rp;
7885 mc->mc_ki[mc->mc_top] = 0;
7886 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7889 for (i=0; i<mc->mc_top; i++)
7890 mc->mc_ki[i] = mn.mc_ki[i];
7891 } else if (!IS_LEAF2(mp)) {
7893 mc->mc_pg[mc->mc_top] = rp;
7898 rkey.mv_data = newkey->mv_data;
7899 rkey.mv_size = newkey->mv_size;
7905 /* Update index for the new key. */
7906 mc->mc_ki[mc->mc_top] = j;
7908 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7909 rkey.mv_data = NODEKEY(node);
7910 rkey.mv_size = node->mn_ksize;
7912 xdata.mv_data = NODEDATA(node);
7913 xdata.mv_size = NODEDSZ(node);
7916 pgno = NODEPGNO(node);
7917 flags = node->mn_flags;
7920 if (!IS_LEAF(mp) && j == 0) {
7921 /* First branch index doesn't need key data. */
7925 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7931 mc->mc_pg[mc->mc_top] = copy;
7936 } while (i != split_indx);
7938 nkeys = NUMKEYS(copy);
7939 for (i=0; i<nkeys; i++)
7940 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7941 mp->mp_lower = copy->mp_lower;
7942 mp->mp_upper = copy->mp_upper;
7943 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7944 env->me_psize - copy->mp_upper);
7946 /* reset back to original page */
7947 if (newindx < split_indx) {
7948 mc->mc_pg[mc->mc_top] = mp;
7949 if (nflags & MDB_RESERVE) {
7950 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7951 if (!(node->mn_flags & F_BIGDATA))
7952 newdata->mv_data = NODEDATA(node);
7955 mc->mc_pg[mc->mc_top] = rp;
7957 /* Make sure mc_ki is still valid.
7959 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7960 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7961 for (i=0; i<=ptop; i++) {
7962 mc->mc_pg[i] = mn.mc_pg[i];
7963 mc->mc_ki[i] = mn.mc_ki[i];
7970 /* Adjust other cursors pointing to mp */
7971 MDB_cursor *m2, *m3;
7972 MDB_dbi dbi = mc->mc_dbi;
7973 int fixup = NUMKEYS(mp);
7975 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7976 if (mc->mc_flags & C_SUB)
7977 m3 = &m2->mc_xcursor->mx_cursor;
7982 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7984 if (m3->mc_flags & C_SPLITTING)
7989 for (k=m3->mc_top; k>=0; k--) {
7990 m3->mc_ki[k+1] = m3->mc_ki[k];
7991 m3->mc_pg[k+1] = m3->mc_pg[k];
7993 if (m3->mc_ki[0] >= split_indx) {
7998 m3->mc_pg[0] = mc->mc_pg[0];
8002 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8003 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8004 m3->mc_ki[mc->mc_top]++;
8005 if (m3->mc_ki[mc->mc_top] >= fixup) {
8006 m3->mc_pg[mc->mc_top] = rp;
8007 m3->mc_ki[mc->mc_top] -= fixup;
8008 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8010 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8011 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8016 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8019 if (copy) /* tmp page */
8020 mdb_page_free(env, copy);
8022 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8027 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8028 MDB_val *key, MDB_val *data, unsigned int flags)
8033 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8036 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8039 mdb_cursor_init(&mc, txn, dbi, &mx);
8040 return mdb_cursor_put(&mc, key, data, flags);
8044 #define MDB_WBUF (1024*1024)
8047 /** State needed for a compacting copy. */
8048 typedef struct mdb_copy {
8049 pthread_mutex_t mc_mutex;
8050 pthread_cond_t mc_cond;
8057 pgno_t mc_next_pgno;
8060 volatile int mc_new;
8065 /** Dedicated writer thread for compacting copy. */
8066 static THREAD_RET ESECT
8067 mdb_env_copythr(void *arg)
8071 int toggle = 0, wsize, rc;
8074 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8077 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8080 pthread_mutex_lock(&my->mc_mutex);
8082 pthread_cond_signal(&my->mc_cond);
8085 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8086 if (my->mc_new < 0) {
8091 wsize = my->mc_wlen[toggle];
8092 ptr = my->mc_wbuf[toggle];
8095 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8099 } else if (len > 0) {
8113 /* If there's an overflow page tail, write it too */
8114 if (my->mc_olen[toggle]) {
8115 wsize = my->mc_olen[toggle];
8116 ptr = my->mc_over[toggle];
8117 my->mc_olen[toggle] = 0;
8120 my->mc_wlen[toggle] = 0;
8122 pthread_cond_signal(&my->mc_cond);
8124 pthread_cond_signal(&my->mc_cond);
8125 pthread_mutex_unlock(&my->mc_mutex);
8126 return (THREAD_RET)0;
8130 /** Tell the writer thread there's a buffer ready to write */
8132 mdb_env_cthr_toggle(mdb_copy *my, int st)
8134 int toggle = my->mc_toggle ^ 1;
8135 pthread_mutex_lock(&my->mc_mutex);
8136 if (my->mc_status) {
8137 pthread_mutex_unlock(&my->mc_mutex);
8138 return my->mc_status;
8140 while (my->mc_new == 1)
8141 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8143 my->mc_toggle = toggle;
8144 pthread_cond_signal(&my->mc_cond);
8145 pthread_mutex_unlock(&my->mc_mutex);
8149 /** Depth-first tree traversal for compacting copy. */
8151 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8154 MDB_txn *txn = my->mc_txn;
8156 MDB_page *mo, *mp, *leaf;
8161 /* Empty DB, nothing to do */
8162 if (*pg == P_INVALID)
8169 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8172 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8176 /* Make cursor pages writable */
8177 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8181 for (i=0; i<mc.mc_top; i++) {
8182 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8183 mc.mc_pg[i] = (MDB_page *)ptr;
8184 ptr += my->mc_env->me_psize;
8187 /* This is writable space for a leaf page. Usually not needed. */
8188 leaf = (MDB_page *)ptr;
8190 toggle = my->mc_toggle;
8191 while (mc.mc_snum > 0) {
8193 mp = mc.mc_pg[mc.mc_top];
8197 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8198 for (i=0; i<n; i++) {
8199 ni = NODEPTR(mp, i);
8200 if (ni->mn_flags & F_BIGDATA) {
8204 /* Need writable leaf */
8206 mc.mc_pg[mc.mc_top] = leaf;
8207 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8209 ni = NODEPTR(mp, i);
8212 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8213 rc = mdb_page_get(txn, pg, &omp, NULL);
8216 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8217 rc = mdb_env_cthr_toggle(my, 1);
8220 toggle = my->mc_toggle;
8222 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8223 memcpy(mo, omp, my->mc_env->me_psize);
8224 mo->mp_pgno = my->mc_next_pgno;
8225 my->mc_next_pgno += omp->mp_pages;
8226 my->mc_wlen[toggle] += my->mc_env->me_psize;
8227 if (omp->mp_pages > 1) {
8228 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8229 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8230 rc = mdb_env_cthr_toggle(my, 1);
8233 toggle = my->mc_toggle;
8235 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8236 } else if (ni->mn_flags & F_SUBDATA) {
8239 /* Need writable leaf */
8241 mc.mc_pg[mc.mc_top] = leaf;
8242 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8244 ni = NODEPTR(mp, i);
8247 memcpy(&db, NODEDATA(ni), sizeof(db));
8248 my->mc_toggle = toggle;
8249 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8252 toggle = my->mc_toggle;
8253 memcpy(NODEDATA(ni), &db, sizeof(db));
8258 mc.mc_ki[mc.mc_top]++;
8259 if (mc.mc_ki[mc.mc_top] < n) {
8262 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8264 rc = mdb_page_get(txn, pg, &mp, NULL);
8269 mc.mc_ki[mc.mc_top] = 0;
8270 if (IS_BRANCH(mp)) {
8271 /* Whenever we advance to a sibling branch page,
8272 * we must proceed all the way down to its first leaf.
8274 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8277 mc.mc_pg[mc.mc_top] = mp;
8281 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8282 rc = mdb_env_cthr_toggle(my, 1);
8285 toggle = my->mc_toggle;
8287 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8288 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8289 mo->mp_pgno = my->mc_next_pgno++;
8290 my->mc_wlen[toggle] += my->mc_env->me_psize;
8292 /* Update parent if there is one */
8293 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8294 SETPGNO(ni, mo->mp_pgno);
8295 mdb_cursor_pop(&mc);
8297 /* Otherwise we're done */
8307 /** Copy environment with compaction. */
8309 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8314 MDB_txn *txn = NULL;
8319 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8320 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8321 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize);
8322 if (my.mc_wbuf[0] == NULL)
8325 pthread_mutex_init(&my.mc_mutex, NULL);
8326 pthread_cond_init(&my.mc_cond, NULL);
8327 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2);
8331 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8336 my.mc_next_pgno = 2;
8342 THREAD_CREATE(thr, mdb_env_copythr, &my);
8344 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8348 mp = (MDB_page *)my.mc_wbuf[0];
8349 memset(mp, 0, 2*env->me_psize);
8351 mp->mp_flags = P_META;
8352 mm = (MDB_meta *)METADATA(mp);
8353 mdb_env_init_meta0(env, mm);
8354 mm->mm_address = env->me_metas[0]->mm_address;
8356 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8358 mp->mp_flags = P_META;
8359 *(MDB_meta *)METADATA(mp) = *mm;
8360 mm = (MDB_meta *)METADATA(mp);
8362 /* Count the number of free pages, subtract from lastpg to find
8363 * number of active pages
8366 MDB_ID freecount = 0;
8369 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8370 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8371 freecount += *(MDB_ID *)data.mv_data;
8372 freecount += txn->mt_dbs[0].md_branch_pages +
8373 txn->mt_dbs[0].md_leaf_pages +
8374 txn->mt_dbs[0].md_overflow_pages;
8376 /* Set metapage 1 */
8377 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8378 mm->mm_dbs[1] = txn->mt_dbs[1];
8379 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8382 my.mc_wlen[0] = env->me_psize * 2;
8384 pthread_mutex_lock(&my.mc_mutex);
8386 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8387 pthread_mutex_unlock(&my.mc_mutex);
8388 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8389 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8390 rc = mdb_env_cthr_toggle(&my, 1);
8391 mdb_env_cthr_toggle(&my, -1);
8392 pthread_mutex_lock(&my.mc_mutex);
8394 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8395 pthread_mutex_unlock(&my.mc_mutex);
8400 CloseHandle(my.mc_cond);
8401 CloseHandle(my.mc_mutex);
8402 _aligned_free(my.mc_wbuf[0]);
8404 pthread_cond_destroy(&my.mc_cond);
8405 pthread_mutex_destroy(&my.mc_mutex);
8406 free(my.mc_wbuf[0]);
8411 /** Copy environment as-is. */
8413 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8415 MDB_txn *txn = NULL;
8421 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8425 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8428 /* Do the lock/unlock of the reader mutex before starting the
8429 * write txn. Otherwise other read txns could block writers.
8431 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8436 /* We must start the actual read txn after blocking writers */
8437 mdb_txn_reset0(txn, "reset-stage1");
8439 /* Temporarily block writers until we snapshot the meta pages */
8442 rc = mdb_txn_renew0(txn);
8444 UNLOCK_MUTEX_W(env);
8449 wsize = env->me_psize * 2;
8453 DO_WRITE(rc, fd, ptr, w2, len);
8457 } else if (len > 0) {
8463 /* Non-blocking or async handles are not supported */
8469 UNLOCK_MUTEX_W(env);
8474 w2 = txn->mt_next_pgno * env->me_psize;
8477 LARGE_INTEGER fsize;
8478 GetFileSizeEx(env->me_fd, &fsize);
8479 if (w2 > fsize.QuadPart)
8480 w2 = fsize.QuadPart;
8485 fstat(env->me_fd, &st);
8486 if (w2 > (size_t)st.st_size)
8492 if (wsize > MAX_WRITE)
8496 DO_WRITE(rc, fd, ptr, w2, len);
8500 } else if (len > 0) {
8517 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8519 if (flags & MDB_CP_COMPACT)
8520 return mdb_env_copyfd1(env, fd);
8522 return mdb_env_copyfd0(env, fd);
8526 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8528 return mdb_env_copyfd2(env, fd, 0);
8532 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8536 HANDLE newfd = INVALID_HANDLE_VALUE;
8538 if (env->me_flags & MDB_NOSUBDIR) {
8539 lpath = (char *)path;
8542 len += sizeof(DATANAME);
8543 lpath = malloc(len);
8546 sprintf(lpath, "%s" DATANAME, path);
8549 /* The destination path must exist, but the destination file must not.
8550 * We don't want the OS to cache the writes, since the source data is
8551 * already in the OS cache.
8554 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8555 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8557 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8559 if (newfd == INVALID_HANDLE_VALUE) {
8565 /* Set O_DIRECT if the file system supports it */
8566 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8567 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8569 #ifdef F_NOCACHE /* __APPLE__ */
8570 rc = fcntl(newfd, F_NOCACHE, 1);
8577 rc = mdb_env_copyfd2(env, newfd, flags);
8580 if (!(env->me_flags & MDB_NOSUBDIR))
8582 if (newfd != INVALID_HANDLE_VALUE)
8583 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8590 mdb_env_copy(MDB_env *env, const char *path)
8592 return mdb_env_copy2(env, path, 0);
8596 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8598 if ((flag & CHANGEABLE) != flag)
8601 env->me_flags |= flag;
8603 env->me_flags &= ~flag;
8608 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8613 *arg = env->me_flags;
8618 mdb_env_set_userctx(MDB_env *env, void *ctx)
8622 env->me_userctx = ctx;
8627 mdb_env_get_userctx(MDB_env *env)
8629 return env ? env->me_userctx : NULL;
8633 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8638 env->me_assert_func = func;
8644 mdb_env_get_path(MDB_env *env, const char **arg)
8649 *arg = env->me_path;
8654 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8663 /** Common code for #mdb_stat() and #mdb_env_stat().
8664 * @param[in] env the environment to operate in.
8665 * @param[in] db the #MDB_db record containing the stats to return.
8666 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8667 * @return 0, this function always succeeds.
8670 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8672 arg->ms_psize = env->me_psize;
8673 arg->ms_depth = db->md_depth;
8674 arg->ms_branch_pages = db->md_branch_pages;
8675 arg->ms_leaf_pages = db->md_leaf_pages;
8676 arg->ms_overflow_pages = db->md_overflow_pages;
8677 arg->ms_entries = db->md_entries;
8683 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8687 if (env == NULL || arg == NULL)
8690 toggle = mdb_env_pick_meta(env);
8692 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8696 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8700 if (env == NULL || arg == NULL)
8703 toggle = mdb_env_pick_meta(env);
8704 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8705 arg->me_mapsize = env->me_mapsize;
8706 arg->me_maxreaders = env->me_maxreaders;
8708 /* me_numreaders may be zero if this process never used any readers. Use
8709 * the shared numreader count if it exists.
8711 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8713 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8714 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8718 /** Set the default comparison functions for a database.
8719 * Called immediately after a database is opened to set the defaults.
8720 * The user can then override them with #mdb_set_compare() or
8721 * #mdb_set_dupsort().
8722 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8723 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8726 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8728 uint16_t f = txn->mt_dbs[dbi].md_flags;
8730 txn->mt_dbxs[dbi].md_cmp =
8731 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8732 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8734 txn->mt_dbxs[dbi].md_dcmp =
8735 !(f & MDB_DUPSORT) ? 0 :
8736 ((f & MDB_INTEGERDUP)
8737 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8738 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8741 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8746 int rc, dbflag, exact;
8747 unsigned int unused = 0;
8750 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8751 mdb_default_cmp(txn, FREE_DBI);
8754 if ((flags & VALID_FLAGS) != flags)
8756 if (txn->mt_flags & MDB_TXN_ERROR)
8762 if (flags & PERSISTENT_FLAGS) {
8763 uint16_t f2 = flags & PERSISTENT_FLAGS;
8764 /* make sure flag changes get committed */
8765 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8766 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8767 txn->mt_flags |= MDB_TXN_DIRTY;
8770 mdb_default_cmp(txn, MAIN_DBI);
8774 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8775 mdb_default_cmp(txn, MAIN_DBI);
8778 /* Is the DB already open? */
8780 for (i=2; i<txn->mt_numdbs; i++) {
8781 if (!txn->mt_dbxs[i].md_name.mv_size) {
8782 /* Remember this free slot */
8783 if (!unused) unused = i;
8786 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8787 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8793 /* If no free slot and max hit, fail */
8794 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8795 return MDB_DBS_FULL;
8797 /* Cannot mix named databases with some mainDB flags */
8798 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8799 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8801 /* Find the DB info */
8802 dbflag = DB_NEW|DB_VALID;
8805 key.mv_data = (void *)name;
8806 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8807 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8808 if (rc == MDB_SUCCESS) {
8809 /* make sure this is actually a DB */
8810 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8811 if (!(node->mn_flags & F_SUBDATA))
8812 return MDB_INCOMPATIBLE;
8813 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8814 /* Create if requested */
8816 data.mv_size = sizeof(MDB_db);
8817 data.mv_data = &dummy;
8818 memset(&dummy, 0, sizeof(dummy));
8819 dummy.md_root = P_INVALID;
8820 dummy.md_flags = flags & PERSISTENT_FLAGS;
8821 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8825 /* OK, got info, add to table */
8826 if (rc == MDB_SUCCESS) {
8827 unsigned int slot = unused ? unused : txn->mt_numdbs;
8828 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8829 txn->mt_dbxs[slot].md_name.mv_size = len;
8830 txn->mt_dbxs[slot].md_rel = NULL;
8831 txn->mt_dbflags[slot] = dbflag;
8832 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8834 mdb_default_cmp(txn, slot);
8843 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8845 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8848 if (txn->mt_flags & MDB_TXN_ERROR)
8851 if (txn->mt_dbflags[dbi] & DB_STALE) {
8854 /* Stale, must read the DB's root. cursor_init does it for us. */
8855 mdb_cursor_init(&mc, txn, dbi, &mx);
8857 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8860 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8863 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8865 ptr = env->me_dbxs[dbi].md_name.mv_data;
8866 env->me_dbxs[dbi].md_name.mv_data = NULL;
8867 env->me_dbxs[dbi].md_name.mv_size = 0;
8868 env->me_dbflags[dbi] = 0;
8872 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8874 /* We could return the flags for the FREE_DBI too but what's the point? */
8875 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8877 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8881 /** Add all the DB's pages to the free list.
8882 * @param[in] mc Cursor on the DB to free.
8883 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8884 * @return 0 on success, non-zero on failure.
8887 mdb_drop0(MDB_cursor *mc, int subs)
8891 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8892 if (rc == MDB_SUCCESS) {
8893 MDB_txn *txn = mc->mc_txn;
8898 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8899 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8902 mdb_cursor_copy(mc, &mx);
8903 while (mc->mc_snum > 0) {
8904 MDB_page *mp = mc->mc_pg[mc->mc_top];
8905 unsigned n = NUMKEYS(mp);
8907 for (i=0; i<n; i++) {
8908 ni = NODEPTR(mp, i);
8909 if (ni->mn_flags & F_BIGDATA) {
8912 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8913 rc = mdb_page_get(txn, pg, &omp, NULL);
8916 mdb_cassert(mc, IS_OVERFLOW(omp));
8917 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8921 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8922 mdb_xcursor_init1(mc, ni);
8923 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8929 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8931 for (i=0; i<n; i++) {
8933 ni = NODEPTR(mp, i);
8936 mdb_midl_xappend(txn->mt_free_pgs, pg);
8941 mc->mc_ki[mc->mc_top] = i;
8942 rc = mdb_cursor_sibling(mc, 1);
8944 if (rc != MDB_NOTFOUND)
8946 /* no more siblings, go back to beginning
8947 * of previous level.
8951 for (i=1; i<mc->mc_snum; i++) {
8953 mc->mc_pg[i] = mx.mc_pg[i];
8958 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8961 txn->mt_flags |= MDB_TXN_ERROR;
8962 } else if (rc == MDB_NOTFOUND) {
8968 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8970 MDB_cursor *mc, *m2;
8973 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8976 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8979 rc = mdb_cursor_open(txn, dbi, &mc);
8983 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8984 /* Invalidate the dropped DB's cursors */
8985 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8986 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8990 /* Can't delete the main DB */
8991 if (del && dbi > MAIN_DBI) {
8992 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
8994 txn->mt_dbflags[dbi] = DB_STALE;
8995 mdb_dbi_close(txn->mt_env, dbi);
8997 txn->mt_flags |= MDB_TXN_ERROR;
9000 /* reset the DB record, mark it dirty */
9001 txn->mt_dbflags[dbi] |= DB_DIRTY;
9002 txn->mt_dbs[dbi].md_depth = 0;
9003 txn->mt_dbs[dbi].md_branch_pages = 0;
9004 txn->mt_dbs[dbi].md_leaf_pages = 0;
9005 txn->mt_dbs[dbi].md_overflow_pages = 0;
9006 txn->mt_dbs[dbi].md_entries = 0;
9007 txn->mt_dbs[dbi].md_root = P_INVALID;
9009 txn->mt_flags |= MDB_TXN_DIRTY;
9012 mdb_cursor_close(mc);
9016 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9018 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9021 txn->mt_dbxs[dbi].md_cmp = cmp;
9025 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9027 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9030 txn->mt_dbxs[dbi].md_dcmp = cmp;
9034 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9036 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9039 txn->mt_dbxs[dbi].md_rel = rel;
9043 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9045 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9048 txn->mt_dbxs[dbi].md_relctx = ctx;
9053 mdb_env_get_maxkeysize(MDB_env *env)
9055 return ENV_MAXKEY(env);
9059 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9061 unsigned int i, rdrs;
9064 int rc = 0, first = 1;
9068 if (!env->me_txns) {
9069 return func("(no reader locks)\n", ctx);
9071 rdrs = env->me_txns->mti_numreaders;
9072 mr = env->me_txns->mti_readers;
9073 for (i=0; i<rdrs; i++) {
9075 txnid_t txnid = mr[i].mr_txnid;
9076 sprintf(buf, txnid == (txnid_t)-1 ?
9077 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9078 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9081 rc = func(" pid thread txnid\n", ctx);
9085 rc = func(buf, ctx);
9091 rc = func("(no active readers)\n", ctx);
9096 /** Insert pid into list if not already present.
9097 * return -1 if already present.
9100 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9102 /* binary search of pid in list */
9104 unsigned cursor = 1;
9106 unsigned n = ids[0];
9109 unsigned pivot = n >> 1;
9110 cursor = base + pivot + 1;
9111 val = pid - ids[cursor];
9116 } else if ( val > 0 ) {
9121 /* found, so it's a duplicate */
9130 for (n = ids[0]; n > cursor; n--)
9137 mdb_reader_check(MDB_env *env, int *dead)
9139 unsigned int i, j, rdrs;
9141 MDB_PID_T *pids, pid;
9150 rdrs = env->me_txns->mti_numreaders;
9151 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9155 mr = env->me_txns->mti_readers;
9156 for (i=0; i<rdrs; i++) {
9157 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9159 if (mdb_pid_insert(pids, pid) == 0) {
9160 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9162 /* Recheck, a new process may have reused pid */
9163 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9164 for (j=i; j<rdrs; j++)
9165 if (mr[j].mr_pid == pid) {
9166 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9167 (unsigned) pid, mr[j].mr_txnid));
9172 UNLOCK_MUTEX_R(env);