2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2014 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
83 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
84 #include <netinet/in.h>
85 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
88 #if defined(__APPLE__) || defined (BSD)
89 # define MDB_USE_POSIX_SEM 1
90 # define MDB_FDATASYNC fsync
91 #elif defined(ANDROID)
92 # define MDB_FDATASYNC fsync
97 #ifdef MDB_USE_POSIX_SEM
98 # define MDB_USE_HASH 1
99 #include <semaphore.h>
104 #include <valgrind/memcheck.h>
105 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
106 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
107 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
108 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
109 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
111 #define VGMEMP_CREATE(h,r,z)
112 #define VGMEMP_ALLOC(h,a,s)
113 #define VGMEMP_FREE(h,a)
114 #define VGMEMP_DESTROY(h)
115 #define VGMEMP_DEFINED(a,s)
119 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
120 /* Solaris just defines one or the other */
121 # define LITTLE_ENDIAN 1234
122 # define BIG_ENDIAN 4321
123 # ifdef _LITTLE_ENDIAN
124 # define BYTE_ORDER LITTLE_ENDIAN
126 # define BYTE_ORDER BIG_ENDIAN
129 # define BYTE_ORDER __BYTE_ORDER
133 #ifndef LITTLE_ENDIAN
134 #define LITTLE_ENDIAN __LITTLE_ENDIAN
137 #define BIG_ENDIAN __BIG_ENDIAN
140 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
141 #define MISALIGNED_OK 1
147 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
148 # error "Unknown or unsupported endianness (BYTE_ORDER)"
149 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
150 # error "Two's complement, reasonably sized integer types, please"
154 /** Put infrequently used env functions in separate section */
156 # define ESECT __attribute__ ((section("__TEXT,text_env")))
158 # define ESECT __attribute__ ((section("text_env")))
164 /** @defgroup internal LMDB Internals
167 /** @defgroup compat Compatibility Macros
168 * A bunch of macros to minimize the amount of platform-specific ifdefs
169 * needed throughout the rest of the code. When the features this library
170 * needs are similar enough to POSIX to be hidden in a one-or-two line
171 * replacement, this macro approach is used.
175 /* Features under development */
180 /** Wrapper around __func__, which is a C99 feature */
181 #if __STDC_VERSION__ >= 199901L
182 # define mdb_func_ __func__
183 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
184 # define mdb_func_ __FUNCTION__
186 /* If a debug message says <mdb_unknown>(), update the #if statements above */
187 # define mdb_func_ "<mdb_unknown>"
191 #define MDB_USE_HASH 1
192 #define MDB_PIDLOCK 0
193 #define THREAD_RET DWORD
194 #define pthread_t HANDLE
195 #define pthread_mutex_t HANDLE
196 #define pthread_cond_t HANDLE
197 #define pthread_key_t DWORD
198 #define pthread_self() GetCurrentThreadId()
199 #define pthread_key_create(x,y) \
200 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
201 #define pthread_key_delete(x) TlsFree(x)
202 #define pthread_getspecific(x) TlsGetValue(x)
203 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
204 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
205 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
206 #define pthread_cond_signal(x) SetEvent(*x)
207 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
208 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
209 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
210 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
211 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
212 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
213 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
214 #define getpid() GetCurrentProcessId()
215 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
216 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
217 #define ErrCode() GetLastError()
218 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
219 #define close(fd) (CloseHandle(fd) ? 0 : -1)
220 #define munmap(ptr,len) UnmapViewOfFile(ptr)
221 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
222 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
224 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
228 #define THREAD_RET void *
229 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
230 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
231 #define Z "z" /**< printf format modifier for size_t */
233 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
234 #define MDB_PIDLOCK 1
236 #ifdef MDB_USE_POSIX_SEM
238 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
239 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
240 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
241 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
244 mdb_sem_wait(sem_t *sem)
247 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
252 /** Lock the reader mutex.
254 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
255 /** Unlock the reader mutex.
257 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
259 /** Lock the writer mutex.
260 * Only a single write transaction is allowed at a time. Other writers
261 * will block waiting for this mutex.
263 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
264 /** Unlock the writer mutex.
266 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
267 #endif /* MDB_USE_POSIX_SEM */
269 /** Get the error code for the last failed system function.
271 #define ErrCode() errno
273 /** An abstraction for a file handle.
274 * On POSIX systems file handles are small integers. On Windows
275 * they're opaque pointers.
279 /** A value for an invalid file handle.
280 * Mainly used to initialize file variables and signify that they are
283 #define INVALID_HANDLE_VALUE (-1)
285 /** Get the size of a memory page for the system.
286 * This is the basic size that the platform's memory manager uses, and is
287 * fundamental to the use of memory-mapped files.
289 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
292 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
295 #define MNAME_LEN (sizeof(pthread_mutex_t))
301 /** A flag for opening a file and requesting synchronous data writes.
302 * This is only used when writing a meta page. It's not strictly needed;
303 * we could just do a normal write and then immediately perform a flush.
304 * But if this flag is available it saves us an extra system call.
306 * @note If O_DSYNC is undefined but exists in /usr/include,
307 * preferably set some compiler flag to get the definition.
308 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
311 # define MDB_DSYNC O_DSYNC
315 /** Function for flushing the data of a file. Define this to fsync
316 * if fdatasync() is not supported.
318 #ifndef MDB_FDATASYNC
319 # define MDB_FDATASYNC fdatasync
323 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
334 /** A page number in the database.
335 * Note that 64 bit page numbers are overkill, since pages themselves
336 * already represent 12-13 bits of addressable memory, and the OS will
337 * always limit applications to a maximum of 63 bits of address space.
339 * @note In the #MDB_node structure, we only store 48 bits of this value,
340 * which thus limits us to only 60 bits of addressable data.
342 typedef MDB_ID pgno_t;
344 /** A transaction ID.
345 * See struct MDB_txn.mt_txnid for details.
347 typedef MDB_ID txnid_t;
349 /** @defgroup debug Debug Macros
353 /** Enable debug output. Needs variable argument macros (a C99 feature).
354 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
355 * read from and written to the database (used for free space management).
361 static int mdb_debug;
362 static txnid_t mdb_debug_start;
364 /** Print a debug message with printf formatting.
365 * Requires double parenthesis around 2 or more args.
367 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
368 # define DPRINTF0(fmt, ...) \
369 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
371 # define DPRINTF(args) ((void) 0)
373 /** Print a debug string.
374 * The string is printed literally, with no format processing.
376 #define DPUTS(arg) DPRINTF(("%s", arg))
377 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
379 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
382 /** @brief The maximum size of a database page.
384 * It is 32k or 64k, since value-PAGEBASE must fit in
385 * #MDB_page.%mp_upper.
387 * LMDB will use database pages < OS pages if needed.
388 * That causes more I/O in write transactions: The OS must
389 * know (read) the whole page before writing a partial page.
391 * Note that we don't currently support Huge pages. On Linux,
392 * regular data files cannot use Huge pages, and in general
393 * Huge pages aren't actually pageable. We rely on the OS
394 * demand-pager to read our data and page it out when memory
395 * pressure from other processes is high. So until OSs have
396 * actual paging support for Huge pages, they're not viable.
398 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
400 /** The minimum number of keys required in a database page.
401 * Setting this to a larger value will place a smaller bound on the
402 * maximum size of a data item. Data items larger than this size will
403 * be pushed into overflow pages instead of being stored directly in
404 * the B-tree node. This value used to default to 4. With a page size
405 * of 4096 bytes that meant that any item larger than 1024 bytes would
406 * go into an overflow page. That also meant that on average 2-3KB of
407 * each overflow page was wasted space. The value cannot be lower than
408 * 2 because then there would no longer be a tree structure. With this
409 * value, items larger than 2KB will go into overflow pages, and on
410 * average only 1KB will be wasted.
412 #define MDB_MINKEYS 2
414 /** A stamp that identifies a file as an LMDB file.
415 * There's nothing special about this value other than that it is easily
416 * recognizable, and it will reflect any byte order mismatches.
418 #define MDB_MAGIC 0xBEEFC0DE
420 /** The version number for a database's datafile format. */
421 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
422 /** The version number for a database's lockfile format. */
423 #define MDB_LOCK_VERSION 1
425 /** @brief The max size of a key we can write, or 0 for dynamic max.
427 * Define this as 0 to compute the max from the page size. 511
428 * is default for backwards compat: liblmdb <= 0.9.10 can break
429 * when modifying a DB with keys/dupsort data bigger than its max.
431 * Data items in an #MDB_DUPSORT database are also limited to
432 * this size, since they're actually keys of a sub-DB. Keys and
433 * #MDB_DUPSORT data items must fit on a node in a regular page.
435 #ifndef MDB_MAXKEYSIZE
436 #define MDB_MAXKEYSIZE 511
439 /** The maximum size of a key we can write to the environment. */
441 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
443 #define ENV_MAXKEY(env) ((env)->me_maxkey)
446 /** @brief The maximum size of a data item.
448 * We only store a 32 bit value for node sizes.
450 #define MAXDATASIZE 0xffffffffUL
453 /** Key size which fits in a #DKBUF.
456 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
459 * This is used for printing a hex dump of a key's contents.
461 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
462 /** Display a key in hex.
464 * Invoke a function to display a key in hex.
466 #define DKEY(x) mdb_dkey(x, kbuf)
472 /** An invalid page number.
473 * Mainly used to denote an empty tree.
475 #define P_INVALID (~(pgno_t)0)
477 /** Test if the flags \b f are set in a flag word \b w. */
478 #define F_ISSET(w, f) (((w) & (f)) == (f))
480 /** Round \b n up to an even number. */
481 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
483 /** Used for offsets within a single page.
484 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
487 typedef uint16_t indx_t;
489 /** Default size of memory map.
490 * This is certainly too small for any actual applications. Apps should always set
491 * the size explicitly using #mdb_env_set_mapsize().
493 #define DEFAULT_MAPSIZE 1048576
495 /** @defgroup readers Reader Lock Table
496 * Readers don't acquire any locks for their data access. Instead, they
497 * simply record their transaction ID in the reader table. The reader
498 * mutex is needed just to find an empty slot in the reader table. The
499 * slot's address is saved in thread-specific data so that subsequent read
500 * transactions started by the same thread need no further locking to proceed.
502 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
504 * No reader table is used if the database is on a read-only filesystem, or
505 * if #MDB_NOLOCK is set.
507 * Since the database uses multi-version concurrency control, readers don't
508 * actually need any locking. This table is used to keep track of which
509 * readers are using data from which old transactions, so that we'll know
510 * when a particular old transaction is no longer in use. Old transactions
511 * that have discarded any data pages can then have those pages reclaimed
512 * for use by a later write transaction.
514 * The lock table is constructed such that reader slots are aligned with the
515 * processor's cache line size. Any slot is only ever used by one thread.
516 * This alignment guarantees that there will be no contention or cache
517 * thrashing as threads update their own slot info, and also eliminates
518 * any need for locking when accessing a slot.
520 * A writer thread will scan every slot in the table to determine the oldest
521 * outstanding reader transaction. Any freed pages older than this will be
522 * reclaimed by the writer. The writer doesn't use any locks when scanning
523 * this table. This means that there's no guarantee that the writer will
524 * see the most up-to-date reader info, but that's not required for correct
525 * operation - all we need is to know the upper bound on the oldest reader,
526 * we don't care at all about the newest reader. So the only consequence of
527 * reading stale information here is that old pages might hang around a
528 * while longer before being reclaimed. That's actually good anyway, because
529 * the longer we delay reclaiming old pages, the more likely it is that a
530 * string of contiguous pages can be found after coalescing old pages from
531 * many old transactions together.
534 /** Number of slots in the reader table.
535 * This value was chosen somewhat arbitrarily. 126 readers plus a
536 * couple mutexes fit exactly into 8KB on my development machine.
537 * Applications should set the table size using #mdb_env_set_maxreaders().
539 #define DEFAULT_READERS 126
541 /** The size of a CPU cache line in bytes. We want our lock structures
542 * aligned to this size to avoid false cache line sharing in the
544 * This value works for most CPUs. For Itanium this should be 128.
550 /** The information we store in a single slot of the reader table.
551 * In addition to a transaction ID, we also record the process and
552 * thread ID that owns a slot, so that we can detect stale information,
553 * e.g. threads or processes that went away without cleaning up.
554 * @note We currently don't check for stale records. We simply re-init
555 * the table when we know that we're the only process opening the
558 typedef struct MDB_rxbody {
559 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
560 * Multiple readers that start at the same time will probably have the
561 * same ID here. Again, it's not important to exclude them from
562 * anything; all we need to know is which version of the DB they
563 * started from so we can avoid overwriting any data used in that
564 * particular version.
567 /** The process ID of the process owning this reader txn. */
569 /** The thread ID of the thread owning this txn. */
573 /** The actual reader record, with cacheline padding. */
574 typedef struct MDB_reader {
577 /** shorthand for mrb_txnid */
578 #define mr_txnid mru.mrx.mrb_txnid
579 #define mr_pid mru.mrx.mrb_pid
580 #define mr_tid mru.mrx.mrb_tid
581 /** cache line alignment */
582 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
586 /** The header for the reader table.
587 * The table resides in a memory-mapped file. (This is a different file
588 * than is used for the main database.)
590 * For POSIX the actual mutexes reside in the shared memory of this
591 * mapped file. On Windows, mutexes are named objects allocated by the
592 * kernel; we store the mutex names in this mapped file so that other
593 * processes can grab them. This same approach is also used on
594 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
595 * process-shared POSIX mutexes. For these cases where a named object
596 * is used, the object name is derived from a 64 bit FNV hash of the
597 * environment pathname. As such, naming collisions are extremely
598 * unlikely. If a collision occurs, the results are unpredictable.
600 typedef struct MDB_txbody {
601 /** Stamp identifying this as an LMDB file. It must be set
604 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
606 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
607 char mtb_rmname[MNAME_LEN];
609 /** Mutex protecting access to this table.
610 * This is the reader lock that #LOCK_MUTEX_R acquires.
612 pthread_mutex_t mtb_mutex;
614 /** The ID of the last transaction committed to the database.
615 * This is recorded here only for convenience; the value can always
616 * be determined by reading the main database meta pages.
619 /** The number of slots that have been used in the reader table.
620 * This always records the maximum count, it is not decremented
621 * when readers release their slots.
623 unsigned mtb_numreaders;
626 /** The actual reader table definition. */
627 typedef struct MDB_txninfo {
630 #define mti_magic mt1.mtb.mtb_magic
631 #define mti_format mt1.mtb.mtb_format
632 #define mti_mutex mt1.mtb.mtb_mutex
633 #define mti_rmname mt1.mtb.mtb_rmname
634 #define mti_txnid mt1.mtb.mtb_txnid
635 #define mti_numreaders mt1.mtb.mtb_numreaders
636 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
639 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
640 char mt2_wmname[MNAME_LEN];
641 #define mti_wmname mt2.mt2_wmname
643 pthread_mutex_t mt2_wmutex;
644 #define mti_wmutex mt2.mt2_wmutex
646 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
648 MDB_reader mti_readers[1];
651 /** Lockfile format signature: version, features and field layout */
652 #define MDB_LOCK_FORMAT \
654 ((MDB_LOCK_VERSION) \
655 /* Flags which describe functionality */ \
656 + (((MDB_PIDLOCK) != 0) << 16)))
659 /** Common header for all page types.
660 * Overflow records occupy a number of contiguous pages with no
661 * headers on any page after the first.
663 typedef struct MDB_page {
664 #define mp_pgno mp_p.p_pgno
665 #define mp_next mp_p.p_next
667 pgno_t p_pgno; /**< page number */
668 struct MDB_page *p_next; /**< for in-memory list of freed pages */
671 /** @defgroup mdb_page Page Flags
673 * Flags for the page headers.
676 #define P_BRANCH 0x01 /**< branch page */
677 #define P_LEAF 0x02 /**< leaf page */
678 #define P_OVERFLOW 0x04 /**< overflow page */
679 #define P_META 0x08 /**< meta page */
680 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
681 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
682 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
683 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
684 #define P_KEEP 0x8000 /**< leave this page alone during spill */
686 uint16_t mp_flags; /**< @ref mdb_page */
687 #define mp_lower mp_pb.pb.pb_lower
688 #define mp_upper mp_pb.pb.pb_upper
689 #define mp_pages mp_pb.pb_pages
692 indx_t pb_lower; /**< lower bound of free space */
693 indx_t pb_upper; /**< upper bound of free space */
695 uint32_t pb_pages; /**< number of overflow pages */
697 indx_t mp_ptrs[1]; /**< dynamic size */
700 /** Size of the page header, excluding dynamic data at the end */
701 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
703 /** Address of first usable data byte in a page, after the header */
704 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
706 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
707 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
709 /** Number of nodes on a page */
710 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
712 /** The amount of space remaining in the page */
713 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
715 /** The percentage of space used in the page, in tenths of a percent. */
716 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
717 ((env)->me_psize - PAGEHDRSZ))
718 /** The minimum page fill factor, in tenths of a percent.
719 * Pages emptier than this are candidates for merging.
721 #define FILL_THRESHOLD 250
723 /** Test if a page is a leaf page */
724 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
725 /** Test if a page is a LEAF2 page */
726 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
727 /** Test if a page is a branch page */
728 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
729 /** Test if a page is an overflow page */
730 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
731 /** Test if a page is a sub page */
732 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
734 /** The number of overflow pages needed to store the given size. */
735 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
737 /** Link in #MDB_txn.%mt_loose_pages list */
738 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
740 /** Header for a single key/data pair within a page.
741 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
742 * We guarantee 2-byte alignment for 'MDB_node's.
744 typedef struct MDB_node {
745 /** lo and hi are used for data size on leaf nodes and for
746 * child pgno on branch nodes. On 64 bit platforms, flags
747 * is also used for pgno. (Branch nodes have no flags).
748 * They are in host byte order in case that lets some
749 * accesses be optimized into a 32-bit word access.
751 #if BYTE_ORDER == LITTLE_ENDIAN
752 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
754 unsigned short mn_hi, mn_lo;
756 /** @defgroup mdb_node Node Flags
758 * Flags for node headers.
761 #define F_BIGDATA 0x01 /**< data put on overflow page */
762 #define F_SUBDATA 0x02 /**< data is a sub-database */
763 #define F_DUPDATA 0x04 /**< data has duplicates */
765 /** valid flags for #mdb_node_add() */
766 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
769 unsigned short mn_flags; /**< @ref mdb_node */
770 unsigned short mn_ksize; /**< key size */
771 char mn_data[1]; /**< key and data are appended here */
774 /** Size of the node header, excluding dynamic data at the end */
775 #define NODESIZE offsetof(MDB_node, mn_data)
777 /** Bit position of top word in page number, for shifting mn_flags */
778 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
780 /** Size of a node in a branch page with a given key.
781 * This is just the node header plus the key, there is no data.
783 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
785 /** Size of a node in a leaf page with a given key and data.
786 * This is node header plus key plus data size.
788 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
790 /** Address of node \b i in page \b p */
791 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
793 /** Address of the key for the node */
794 #define NODEKEY(node) (void *)((node)->mn_data)
796 /** Address of the data for a node */
797 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
799 /** Get the page number pointed to by a branch node */
800 #define NODEPGNO(node) \
801 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
802 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
803 /** Set the page number in a branch node */
804 #define SETPGNO(node,pgno) do { \
805 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
806 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
808 /** Get the size of the data in a leaf node */
809 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
810 /** Set the size of the data for a leaf node */
811 #define SETDSZ(node,size) do { \
812 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
813 /** The size of a key in a node */
814 #define NODEKSZ(node) ((node)->mn_ksize)
816 /** Copy a page number from src to dst */
818 #define COPY_PGNO(dst,src) dst = src
820 #if SIZE_MAX > 4294967295UL
821 #define COPY_PGNO(dst,src) do { \
822 unsigned short *s, *d; \
823 s = (unsigned short *)&(src); \
824 d = (unsigned short *)&(dst); \
831 #define COPY_PGNO(dst,src) do { \
832 unsigned short *s, *d; \
833 s = (unsigned short *)&(src); \
834 d = (unsigned short *)&(dst); \
840 /** The address of a key in a LEAF2 page.
841 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
842 * There are no node headers, keys are stored contiguously.
844 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
846 /** Set the \b node's key into \b keyptr, if requested. */
847 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
848 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
850 /** Set the \b node's key into \b key. */
851 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
853 /** Information about a single database in the environment. */
854 typedef struct MDB_db {
855 uint32_t md_pad; /**< also ksize for LEAF2 pages */
856 uint16_t md_flags; /**< @ref mdb_dbi_open */
857 uint16_t md_depth; /**< depth of this tree */
858 pgno_t md_branch_pages; /**< number of internal pages */
859 pgno_t md_leaf_pages; /**< number of leaf pages */
860 pgno_t md_overflow_pages; /**< number of overflow pages */
861 size_t md_entries; /**< number of data items */
862 pgno_t md_root; /**< the root page of this tree */
865 /** mdb_dbi_open flags */
866 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
867 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
868 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
869 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
871 /** Handle for the DB used to track free pages. */
873 /** Handle for the default DB. */
876 /** Meta page content.
877 * A meta page is the start point for accessing a database snapshot.
878 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
880 typedef struct MDB_meta {
881 /** Stamp identifying this as an LMDB file. It must be set
884 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
886 void *mm_address; /**< address for fixed mapping */
887 size_t mm_mapsize; /**< size of mmap region */
888 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
889 /** The size of pages used in this DB */
890 #define mm_psize mm_dbs[0].md_pad
891 /** Any persistent environment flags. @ref mdb_env */
892 #define mm_flags mm_dbs[0].md_flags
893 pgno_t mm_last_pg; /**< last used page in file */
894 txnid_t mm_txnid; /**< txnid that committed this page */
897 /** Buffer for a stack-allocated meta page.
898 * The members define size and alignment, and silence type
899 * aliasing warnings. They are not used directly; that could
900 * mean incorrectly using several union members in parallel.
902 typedef union MDB_metabuf {
905 char mm_pad[PAGEHDRSZ];
910 /** Auxiliary DB info.
911 * The information here is mostly static/read-only. There is
912 * only a single copy of this record in the environment.
914 typedef struct MDB_dbx {
915 MDB_val md_name; /**< name of the database */
916 MDB_cmp_func *md_cmp; /**< function for comparing keys */
917 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
918 MDB_rel_func *md_rel; /**< user relocate function */
919 void *md_relctx; /**< user-provided context for md_rel */
922 /** A database transaction.
923 * Every operation requires a transaction handle.
926 MDB_txn *mt_parent; /**< parent of a nested txn */
927 MDB_txn *mt_child; /**< nested txn under this txn */
928 pgno_t mt_next_pgno; /**< next unallocated page */
929 /** The ID of this transaction. IDs are integers incrementing from 1.
930 * Only committed write transactions increment the ID. If a transaction
931 * aborts, the ID may be re-used by the next writer.
934 MDB_env *mt_env; /**< the DB environment */
935 /** The list of pages that became unused during this transaction.
938 /** The list of loose pages that became unused and may be reused
939 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
941 MDB_page *mt_loose_pgs;
942 /** The sorted list of dirty pages we temporarily wrote to disk
943 * because the dirty list was full. page numbers in here are
944 * shifted left by 1, deleted slots have the LSB set.
946 MDB_IDL mt_spill_pgs;
948 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
950 /** For read txns: This thread/txn's reader table slot, or NULL. */
953 /** Array of records for each DB known in the environment. */
955 /** Array of MDB_db records for each known DB */
957 /** @defgroup mt_dbflag Transaction DB Flags
961 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
962 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
963 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
964 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
966 /** In write txns, array of cursors for each DB */
967 MDB_cursor **mt_cursors;
968 /** Array of flags for each DB */
969 unsigned char *mt_dbflags;
970 /** Number of DB records in use. This number only ever increments;
971 * we don't decrement it when individual DB handles are closed.
975 /** @defgroup mdb_txn Transaction Flags
979 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
980 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
981 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
982 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
984 unsigned int mt_flags; /**< @ref mdb_txn */
985 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
986 * Includes ancestor txns' dirty pages not hidden by other txns'
987 * dirty/spilled pages. Thus commit(nested txn) has room to merge
988 * dirty_list into mt_parent after freeing hidden mt_parent pages.
990 unsigned int mt_dirty_room;
993 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
994 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
995 * raise this on a 64 bit machine.
997 #define CURSOR_STACK 32
1001 /** Cursors are used for all DB operations.
1002 * A cursor holds a path of (page pointer, key index) from the DB
1003 * root to a position in the DB, plus other state. #MDB_DUPSORT
1004 * cursors include an xcursor to the current data item. Write txns
1005 * track their cursors and keep them up to date when data moves.
1006 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1007 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1010 /** Next cursor on this DB in this txn */
1011 MDB_cursor *mc_next;
1012 /** Backup of the original cursor if this cursor is a shadow */
1013 MDB_cursor *mc_backup;
1014 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1015 struct MDB_xcursor *mc_xcursor;
1016 /** The transaction that owns this cursor */
1018 /** The database handle this cursor operates on */
1020 /** The database record for this cursor */
1022 /** The database auxiliary record for this cursor */
1024 /** The @ref mt_dbflag for this database */
1025 unsigned char *mc_dbflag;
1026 unsigned short mc_snum; /**< number of pushed pages */
1027 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1028 /** @defgroup mdb_cursor Cursor Flags
1030 * Cursor state flags.
1033 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1034 #define C_EOF 0x02 /**< No more data */
1035 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1036 #define C_DEL 0x08 /**< last op was a cursor_del */
1037 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1038 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1040 unsigned int mc_flags; /**< @ref mdb_cursor */
1041 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1042 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1045 /** Context for sorted-dup records.
1046 * We could have gone to a fully recursive design, with arbitrarily
1047 * deep nesting of sub-databases. But for now we only handle these
1048 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1050 typedef struct MDB_xcursor {
1051 /** A sub-cursor for traversing the Dup DB */
1052 MDB_cursor mx_cursor;
1053 /** The database record for this Dup DB */
1055 /** The auxiliary DB record for this Dup DB */
1057 /** The @ref mt_dbflag for this Dup DB */
1058 unsigned char mx_dbflag;
1061 /** State of FreeDB old pages, stored in the MDB_env */
1062 typedef struct MDB_pgstate {
1063 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1064 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1067 /** The database environment. */
1069 HANDLE me_fd; /**< The main data file */
1070 HANDLE me_lfd; /**< The lock file */
1071 HANDLE me_mfd; /**< just for writing the meta pages */
1072 /** Failed to update the meta page. Probably an I/O error. */
1073 #define MDB_FATAL_ERROR 0x80000000U
1074 /** Some fields are initialized. */
1075 #define MDB_ENV_ACTIVE 0x20000000U
1076 /** me_txkey is set */
1077 #define MDB_ENV_TXKEY 0x10000000U
1078 uint32_t me_flags; /**< @ref mdb_env */
1079 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1080 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1081 unsigned int me_maxreaders; /**< size of the reader table */
1082 unsigned int me_numreaders; /**< max numreaders set by this env */
1083 MDB_dbi me_numdbs; /**< number of DBs opened */
1084 MDB_dbi me_maxdbs; /**< size of the DB table */
1085 MDB_PID_T me_pid; /**< process ID of this env */
1086 char *me_path; /**< path to the DB files */
1087 char *me_map; /**< the memory map of the data file */
1088 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1089 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1090 void *me_pbuf; /**< scratch area for DUPSORT put() */
1091 MDB_txn *me_txn; /**< current write transaction */
1092 size_t me_mapsize; /**< size of the data memory map */
1093 off_t me_size; /**< current file size */
1094 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1095 MDB_dbx *me_dbxs; /**< array of static DB info */
1096 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1097 pthread_key_t me_txkey; /**< thread-key for readers */
1098 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1099 # define me_pglast me_pgstate.mf_pglast
1100 # define me_pghead me_pgstate.mf_pghead
1101 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1102 /** IDL of pages that became unused in a write txn */
1103 MDB_IDL me_free_pgs;
1104 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1105 MDB_ID2L me_dirty_list;
1106 /** Max number of freelist items that can fit in a single overflow page */
1108 /** Max size of a node on a page */
1109 unsigned int me_nodemax;
1110 #if !(MDB_MAXKEYSIZE)
1111 unsigned int me_maxkey; /**< max size of a key */
1113 int me_live_reader; /**< have liveness lock in reader table */
1115 int me_pidquery; /**< Used in OpenProcess */
1116 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1118 #elif defined(MDB_USE_POSIX_SEM)
1119 sem_t *me_rmutex; /* Shared mutexes are not supported */
1122 void *me_userctx; /**< User-settable context */
1123 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1126 /** Nested transaction */
1127 typedef struct MDB_ntxn {
1128 MDB_txn mnt_txn; /**< the transaction */
1129 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1132 /** max number of pages to commit in one writev() call */
1133 #define MDB_COMMIT_PAGES 64
1134 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1135 #undef MDB_COMMIT_PAGES
1136 #define MDB_COMMIT_PAGES IOV_MAX
1139 /** max bytes to write in one call */
1140 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1142 /** Check \b txn and \b dbi arguments to a function */
1143 #define TXN_DBI_EXIST(txn, dbi) \
1144 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1146 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1147 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1148 static int mdb_page_touch(MDB_cursor *mc);
1150 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1151 static int mdb_page_search_root(MDB_cursor *mc,
1152 MDB_val *key, int modify);
1153 #define MDB_PS_MODIFY 1
1154 #define MDB_PS_ROOTONLY 2
1155 #define MDB_PS_FIRST 4
1156 #define MDB_PS_LAST 8
1157 static int mdb_page_search(MDB_cursor *mc,
1158 MDB_val *key, int flags);
1159 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1161 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1162 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1163 pgno_t newpgno, unsigned int nflags);
1165 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1166 static int mdb_env_pick_meta(const MDB_env *env);
1167 static int mdb_env_write_meta(MDB_txn *txn);
1168 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1169 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1171 static void mdb_env_close0(MDB_env *env, int excl);
1173 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1174 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1175 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1176 static void mdb_node_del(MDB_cursor *mc, int ksize);
1177 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1178 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1179 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1180 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1181 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1183 static int mdb_rebalance(MDB_cursor *mc);
1184 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1186 static void mdb_cursor_pop(MDB_cursor *mc);
1187 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1189 static int mdb_cursor_del0(MDB_cursor *mc);
1190 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1191 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1192 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1193 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1194 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1196 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1197 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1199 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1200 static void mdb_xcursor_init0(MDB_cursor *mc);
1201 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1203 static int mdb_drop0(MDB_cursor *mc, int subs);
1204 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1207 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1211 static SECURITY_DESCRIPTOR mdb_null_sd;
1212 static SECURITY_ATTRIBUTES mdb_all_sa;
1213 static int mdb_sec_inited;
1216 /** Return the library version info. */
1218 mdb_version(int *major, int *minor, int *patch)
1220 if (major) *major = MDB_VERSION_MAJOR;
1221 if (minor) *minor = MDB_VERSION_MINOR;
1222 if (patch) *patch = MDB_VERSION_PATCH;
1223 return MDB_VERSION_STRING;
1226 /** Table of descriptions for LMDB @ref errors */
1227 static char *const mdb_errstr[] = {
1228 "MDB_KEYEXIST: Key/data pair already exists",
1229 "MDB_NOTFOUND: No matching key/data pair found",
1230 "MDB_PAGE_NOTFOUND: Requested page not found",
1231 "MDB_CORRUPTED: Located page was wrong type",
1232 "MDB_PANIC: Update of meta page failed",
1233 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1234 "MDB_INVALID: File is not an LMDB file",
1235 "MDB_MAP_FULL: Environment mapsize limit reached",
1236 "MDB_DBS_FULL: Environment maxdbs limit reached",
1237 "MDB_READERS_FULL: Environment maxreaders limit reached",
1238 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1239 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1240 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1241 "MDB_PAGE_FULL: Internal error - page has no more space",
1242 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1243 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1244 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1245 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1246 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1250 mdb_strerror(int err)
1254 return ("Successful return: 0");
1256 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1257 i = err - MDB_KEYEXIST;
1258 return mdb_errstr[i];
1261 return strerror(err);
1264 /** assert(3) variant in cursor context */
1265 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1266 /** assert(3) variant in transaction context */
1267 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1268 /** assert(3) variant in environment context */
1269 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1272 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1273 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1276 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1277 const char *func, const char *file, int line)
1280 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1281 file, line, expr_txt, func);
1282 if (env->me_assert_func)
1283 env->me_assert_func(env, buf);
1284 fprintf(stderr, "%s\n", buf);
1288 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1292 /** Return the page number of \b mp which may be sub-page, for debug output */
1294 mdb_dbg_pgno(MDB_page *mp)
1297 COPY_PGNO(ret, mp->mp_pgno);
1301 /** Display a key in hexadecimal and return the address of the result.
1302 * @param[in] key the key to display
1303 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1304 * @return The key in hexadecimal form.
1307 mdb_dkey(MDB_val *key, char *buf)
1310 unsigned char *c = key->mv_data;
1316 if (key->mv_size > DKBUF_MAXKEYSIZE)
1317 return "MDB_MAXKEYSIZE";
1318 /* may want to make this a dynamic check: if the key is mostly
1319 * printable characters, print it as-is instead of converting to hex.
1323 for (i=0; i<key->mv_size; i++)
1324 ptr += sprintf(ptr, "%02x", *c++);
1326 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1332 mdb_leafnode_type(MDB_node *n)
1334 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1335 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1336 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1339 /** Display all the keys in the page. */
1341 mdb_page_list(MDB_page *mp)
1343 pgno_t pgno = mdb_dbg_pgno(mp);
1344 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1346 unsigned int i, nkeys, nsize, total = 0;
1350 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1351 case P_BRANCH: type = "Branch page"; break;
1352 case P_LEAF: type = "Leaf page"; break;
1353 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1354 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1355 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1357 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1358 pgno, mp->mp_pages, state);
1361 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1362 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1365 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1369 nkeys = NUMKEYS(mp);
1370 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1372 for (i=0; i<nkeys; i++) {
1373 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1374 key.mv_size = nsize = mp->mp_pad;
1375 key.mv_data = LEAF2KEY(mp, i, nsize);
1377 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1380 node = NODEPTR(mp, i);
1381 key.mv_size = node->mn_ksize;
1382 key.mv_data = node->mn_data;
1383 nsize = NODESIZE + key.mv_size;
1384 if (IS_BRANCH(mp)) {
1385 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1389 if (F_ISSET(node->mn_flags, F_BIGDATA))
1390 nsize += sizeof(pgno_t);
1392 nsize += NODEDSZ(node);
1394 nsize += sizeof(indx_t);
1395 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1396 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1398 total = EVEN(total);
1400 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1401 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1405 mdb_cursor_chk(MDB_cursor *mc)
1411 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1412 for (i=0; i<mc->mc_top; i++) {
1414 node = NODEPTR(mp, mc->mc_ki[i]);
1415 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1418 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1424 /** Count all the pages in each DB and in the freelist
1425 * and make sure it matches the actual number of pages
1427 * All named DBs must be open for a correct count.
1429 static void mdb_audit(MDB_txn *txn)
1433 MDB_ID freecount, count;
1438 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1439 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1440 freecount += *(MDB_ID *)data.mv_data;
1441 mdb_tassert(txn, rc == MDB_NOTFOUND);
1444 for (i = 0; i<txn->mt_numdbs; i++) {
1446 if (!(txn->mt_dbflags[i] & DB_VALID))
1448 mdb_cursor_init(&mc, txn, i, &mx);
1449 if (txn->mt_dbs[i].md_root == P_INVALID)
1451 count += txn->mt_dbs[i].md_branch_pages +
1452 txn->mt_dbs[i].md_leaf_pages +
1453 txn->mt_dbs[i].md_overflow_pages;
1454 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1455 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1456 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1459 mp = mc.mc_pg[mc.mc_top];
1460 for (j=0; j<NUMKEYS(mp); j++) {
1461 MDB_node *leaf = NODEPTR(mp, j);
1462 if (leaf->mn_flags & F_SUBDATA) {
1464 memcpy(&db, NODEDATA(leaf), sizeof(db));
1465 count += db.md_branch_pages + db.md_leaf_pages +
1466 db.md_overflow_pages;
1470 mdb_tassert(txn, rc == MDB_NOTFOUND);
1473 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1474 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1475 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1481 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1483 return txn->mt_dbxs[dbi].md_cmp(a, b);
1487 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1489 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1492 /** Allocate memory for a page.
1493 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1496 mdb_page_malloc(MDB_txn *txn, unsigned num)
1498 MDB_env *env = txn->mt_env;
1499 MDB_page *ret = env->me_dpages;
1500 size_t psize = env->me_psize, sz = psize, off;
1501 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1502 * For a single page alloc, we init everything after the page header.
1503 * For multi-page, we init the final page; if the caller needed that
1504 * many pages they will be filling in at least up to the last page.
1508 VGMEMP_ALLOC(env, ret, sz);
1509 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1510 env->me_dpages = ret->mp_next;
1513 psize -= off = PAGEHDRSZ;
1518 if ((ret = malloc(sz)) != NULL) {
1519 VGMEMP_ALLOC(env, ret, sz);
1520 if (!(env->me_flags & MDB_NOMEMINIT)) {
1521 memset((char *)ret + off, 0, psize);
1525 txn->mt_flags |= MDB_TXN_ERROR;
1529 /** Free a single page.
1530 * Saves single pages to a list, for future reuse.
1531 * (This is not used for multi-page overflow pages.)
1534 mdb_page_free(MDB_env *env, MDB_page *mp)
1536 mp->mp_next = env->me_dpages;
1537 VGMEMP_FREE(env, mp);
1538 env->me_dpages = mp;
1541 /** Free a dirty page */
1543 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1545 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1546 mdb_page_free(env, dp);
1548 /* large pages just get freed directly */
1549 VGMEMP_FREE(env, dp);
1554 /** Return all dirty pages to dpage list */
1556 mdb_dlist_free(MDB_txn *txn)
1558 MDB_env *env = txn->mt_env;
1559 MDB_ID2L dl = txn->mt_u.dirty_list;
1560 unsigned i, n = dl[0].mid;
1562 for (i = 1; i <= n; i++) {
1563 mdb_dpage_free(env, dl[i].mptr);
1568 /** Loosen or free a single page.
1569 * Saves single pages to a list for future reuse
1570 * in this same txn. It has been pulled from the freeDB
1571 * and already resides on the dirty list, but has been
1572 * deleted. Use these pages first before pulling again
1575 * If the page wasn't dirtied in this txn, just add it
1576 * to this txn's free list.
1579 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1582 pgno_t pgno = mp->mp_pgno;
1584 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1585 if (mc->mc_txn->mt_parent) {
1586 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1587 /* If txn has a parent, make sure the page is in our
1591 unsigned x = mdb_mid2l_search(dl, pgno);
1592 if (x <= dl[0].mid && dl[x].mid == pgno) {
1593 if (mp != dl[x].mptr) { /* bad cursor? */
1594 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1595 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1596 return MDB_CORRUPTED;
1603 /* no parent txn, so it's just ours */
1608 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1610 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1611 mc->mc_txn->mt_loose_pgs = mp;
1612 mp->mp_flags |= P_LOOSE;
1614 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1622 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1623 * @param[in] mc A cursor handle for the current operation.
1624 * @param[in] pflags Flags of the pages to update:
1625 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1626 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1627 * @return 0 on success, non-zero on failure.
1630 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1632 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1633 MDB_txn *txn = mc->mc_txn;
1639 int rc = MDB_SUCCESS, level;
1641 /* Mark pages seen by cursors */
1642 if (mc->mc_flags & C_UNTRACK)
1643 mc = NULL; /* will find mc in mt_cursors */
1644 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1645 for (; mc; mc=mc->mc_next) {
1646 if (!(mc->mc_flags & C_INITIALIZED))
1648 for (m3 = mc;; m3 = &mx->mx_cursor) {
1650 for (j=0; j<m3->mc_snum; j++) {
1652 if ((mp->mp_flags & Mask) == pflags)
1653 mp->mp_flags ^= P_KEEP;
1655 mx = m3->mc_xcursor;
1656 /* Proceed to mx if it is at a sub-database */
1657 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1659 if (! (mp && (mp->mp_flags & P_LEAF)))
1661 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1662 if (!(leaf->mn_flags & F_SUBDATA))
1671 /* Mark dirty root pages */
1672 for (i=0; i<txn->mt_numdbs; i++) {
1673 if (txn->mt_dbflags[i] & DB_DIRTY) {
1674 pgno_t pgno = txn->mt_dbs[i].md_root;
1675 if (pgno == P_INVALID)
1677 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1679 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1680 dp->mp_flags ^= P_KEEP;
1688 static int mdb_page_flush(MDB_txn *txn, int keep);
1690 /** Spill pages from the dirty list back to disk.
1691 * This is intended to prevent running into #MDB_TXN_FULL situations,
1692 * but note that they may still occur in a few cases:
1693 * 1) our estimate of the txn size could be too small. Currently this
1694 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1695 * 2) child txns may run out of space if their parents dirtied a
1696 * lot of pages and never spilled them. TODO: we probably should do
1697 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1698 * the parent's dirty_room is below a given threshold.
1700 * Otherwise, if not using nested txns, it is expected that apps will
1701 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1702 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1703 * If the txn never references them again, they can be left alone.
1704 * If the txn only reads them, they can be used without any fuss.
1705 * If the txn writes them again, they can be dirtied immediately without
1706 * going thru all of the work of #mdb_page_touch(). Such references are
1707 * handled by #mdb_page_unspill().
1709 * Also note, we never spill DB root pages, nor pages of active cursors,
1710 * because we'll need these back again soon anyway. And in nested txns,
1711 * we can't spill a page in a child txn if it was already spilled in a
1712 * parent txn. That would alter the parent txns' data even though
1713 * the child hasn't committed yet, and we'd have no way to undo it if
1714 * the child aborted.
1716 * @param[in] m0 cursor A cursor handle identifying the transaction and
1717 * database for which we are checking space.
1718 * @param[in] key For a put operation, the key being stored.
1719 * @param[in] data For a put operation, the data being stored.
1720 * @return 0 on success, non-zero on failure.
1723 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1725 MDB_txn *txn = m0->mc_txn;
1727 MDB_ID2L dl = txn->mt_u.dirty_list;
1728 unsigned int i, j, need;
1731 if (m0->mc_flags & C_SUB)
1734 /* Estimate how much space this op will take */
1735 i = m0->mc_db->md_depth;
1736 /* Named DBs also dirty the main DB */
1737 if (m0->mc_dbi > MAIN_DBI)
1738 i += txn->mt_dbs[MAIN_DBI].md_depth;
1739 /* For puts, roughly factor in the key+data size */
1741 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1742 i += i; /* double it for good measure */
1745 if (txn->mt_dirty_room > i)
1748 if (!txn->mt_spill_pgs) {
1749 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1750 if (!txn->mt_spill_pgs)
1753 /* purge deleted slots */
1754 MDB_IDL sl = txn->mt_spill_pgs;
1755 unsigned int num = sl[0];
1757 for (i=1; i<=num; i++) {
1764 /* Preserve pages which may soon be dirtied again */
1765 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1768 /* Less aggressive spill - we originally spilled the entire dirty list,
1769 * with a few exceptions for cursor pages and DB root pages. But this
1770 * turns out to be a lot of wasted effort because in a large txn many
1771 * of those pages will need to be used again. So now we spill only 1/8th
1772 * of the dirty pages. Testing revealed this to be a good tradeoff,
1773 * better than 1/2, 1/4, or 1/10.
1775 if (need < MDB_IDL_UM_MAX / 8)
1776 need = MDB_IDL_UM_MAX / 8;
1778 /* Save the page IDs of all the pages we're flushing */
1779 /* flush from the tail forward, this saves a lot of shifting later on. */
1780 for (i=dl[0].mid; i && need; i--) {
1781 MDB_ID pn = dl[i].mid << 1;
1783 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1785 /* Can't spill twice, make sure it's not already in a parent's
1788 if (txn->mt_parent) {
1790 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1791 if (tx2->mt_spill_pgs) {
1792 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1793 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1794 dp->mp_flags |= P_KEEP;
1802 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1806 mdb_midl_sort(txn->mt_spill_pgs);
1808 /* Flush the spilled part of dirty list */
1809 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1812 /* Reset any dirty pages we kept that page_flush didn't see */
1813 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1816 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1820 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1822 mdb_find_oldest(MDB_txn *txn)
1825 txnid_t mr, oldest = txn->mt_txnid - 1;
1826 if (txn->mt_env->me_txns) {
1827 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1828 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1839 /** Add a page to the txn's dirty list */
1841 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1844 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1846 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1847 insert = mdb_mid2l_append;
1849 insert = mdb_mid2l_insert;
1851 mid.mid = mp->mp_pgno;
1853 rc = insert(txn->mt_u.dirty_list, &mid);
1854 mdb_tassert(txn, rc == 0);
1855 txn->mt_dirty_room--;
1858 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1859 * me_pghead and mt_next_pgno.
1861 * If there are free pages available from older transactions, they
1862 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1863 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1864 * and move me_pglast to say which records were consumed. Only this
1865 * function can create me_pghead and move me_pglast/mt_next_pgno.
1866 * @param[in] mc cursor A cursor handle identifying the transaction and
1867 * database for which we are allocating.
1868 * @param[in] num the number of pages to allocate.
1869 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1870 * will always be satisfied by a single contiguous chunk of memory.
1871 * @return 0 on success, non-zero on failure.
1874 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1876 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1877 /* Get at most <Max_retries> more freeDB records once me_pghead
1878 * has enough pages. If not enough, use new pages from the map.
1879 * If <Paranoid> and mc is updating the freeDB, only get new
1880 * records if me_pghead is empty. Then the freelist cannot play
1881 * catch-up with itself by growing while trying to save it.
1883 enum { Paranoid = 1, Max_retries = 500 };
1885 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1887 int rc, retry = num * 20;
1888 MDB_txn *txn = mc->mc_txn;
1889 MDB_env *env = txn->mt_env;
1890 pgno_t pgno, *mop = env->me_pghead;
1891 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1893 txnid_t oldest = 0, last;
1897 /* If there are any loose pages, just use them */
1898 if (num == 1 && txn->mt_loose_pgs) {
1899 np = txn->mt_loose_pgs;
1900 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1901 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1909 /* If our dirty list is already full, we can't do anything */
1910 if (txn->mt_dirty_room == 0) {
1915 for (op = MDB_FIRST;; op = MDB_NEXT) {
1918 pgno_t *idl, old_id, new_id;
1920 /* Seek a big enough contiguous page range. Prefer
1921 * pages at the tail, just truncating the list.
1927 if (mop[i-n2] == pgno+n2)
1934 if (op == MDB_FIRST) { /* 1st iteration */
1935 /* Prepare to fetch more and coalesce */
1936 oldest = mdb_find_oldest(txn);
1937 last = env->me_pglast;
1938 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1941 key.mv_data = &last; /* will look up last+1 */
1942 key.mv_size = sizeof(last);
1944 if (Paranoid && mc->mc_dbi == FREE_DBI)
1947 if (Paranoid && retry < 0 && mop_len)
1951 /* Do not fetch more if the record will be too recent */
1954 rc = mdb_cursor_get(&m2, &key, NULL, op);
1956 if (rc == MDB_NOTFOUND)
1960 last = *(txnid_t*)key.mv_data;
1963 np = m2.mc_pg[m2.mc_top];
1964 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1965 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1968 idl = (MDB_ID *) data.mv_data;
1971 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1976 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1978 mop = env->me_pghead;
1980 env->me_pglast = last;
1982 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1983 last, txn->mt_dbs[FREE_DBI].md_root, i));
1985 DPRINTF(("IDL %"Z"u", idl[k]));
1987 /* Merge in descending sorted order */
1990 mop[0] = (pgno_t)-1;
1994 for (; old_id < new_id; old_id = mop[--j])
2001 /* Use new pages from the map when nothing suitable in the freeDB */
2003 pgno = txn->mt_next_pgno;
2004 if (pgno + num >= env->me_maxpg) {
2005 DPUTS("DB size maxed out");
2011 if (env->me_flags & MDB_WRITEMAP) {
2012 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2014 if (!(np = mdb_page_malloc(txn, num))) {
2020 mop[0] = mop_len -= num;
2021 /* Move any stragglers down */
2022 for (j = i-num; j < mop_len; )
2023 mop[++j] = mop[++i];
2025 txn->mt_next_pgno = pgno + num;
2028 mdb_page_dirty(txn, np);
2034 txn->mt_flags |= MDB_TXN_ERROR;
2038 /** Copy the used portions of a non-overflow page.
2039 * @param[in] dst page to copy into
2040 * @param[in] src page to copy from
2041 * @param[in] psize size of a page
2044 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2046 enum { Align = sizeof(pgno_t) };
2047 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2049 /* If page isn't full, just copy the used portion. Adjust
2050 * alignment so memcpy may copy words instead of bytes.
2052 if ((unused &= -Align) && !IS_LEAF2(src)) {
2053 upper = (upper + PAGEBASE) & -Align;
2054 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2055 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2058 memcpy(dst, src, psize - unused);
2062 /** Pull a page off the txn's spill list, if present.
2063 * If a page being referenced was spilled to disk in this txn, bring
2064 * it back and make it dirty/writable again.
2065 * @param[in] txn the transaction handle.
2066 * @param[in] mp the page being referenced. It must not be dirty.
2067 * @param[out] ret the writable page, if any. ret is unchanged if
2068 * mp wasn't spilled.
2071 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2073 MDB_env *env = txn->mt_env;
2076 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2078 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2079 if (!tx2->mt_spill_pgs)
2081 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2082 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2085 if (txn->mt_dirty_room == 0)
2086 return MDB_TXN_FULL;
2087 if (IS_OVERFLOW(mp))
2091 if (env->me_flags & MDB_WRITEMAP) {
2094 np = mdb_page_malloc(txn, num);
2098 memcpy(np, mp, num * env->me_psize);
2100 mdb_page_copy(np, mp, env->me_psize);
2103 /* If in current txn, this page is no longer spilled.
2104 * If it happens to be the last page, truncate the spill list.
2105 * Otherwise mark it as deleted by setting the LSB.
2107 if (x == txn->mt_spill_pgs[0])
2108 txn->mt_spill_pgs[0]--;
2110 txn->mt_spill_pgs[x] |= 1;
2111 } /* otherwise, if belonging to a parent txn, the
2112 * page remains spilled until child commits
2115 mdb_page_dirty(txn, np);
2116 np->mp_flags |= P_DIRTY;
2124 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2125 * @param[in] mc cursor pointing to the page to be touched
2126 * @return 0 on success, non-zero on failure.
2129 mdb_page_touch(MDB_cursor *mc)
2131 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2132 MDB_txn *txn = mc->mc_txn;
2133 MDB_cursor *m2, *m3;
2137 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2138 if (txn->mt_flags & MDB_TXN_SPILLS) {
2140 rc = mdb_page_unspill(txn, mp, &np);
2146 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2147 (rc = mdb_page_alloc(mc, 1, &np)))
2150 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2151 mp->mp_pgno, pgno));
2152 mdb_cassert(mc, mp->mp_pgno != pgno);
2153 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2154 /* Update the parent page, if any, to point to the new page */
2156 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2157 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2158 SETPGNO(node, pgno);
2160 mc->mc_db->md_root = pgno;
2162 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2163 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2165 /* If txn has a parent, make sure the page is in our
2169 unsigned x = mdb_mid2l_search(dl, pgno);
2170 if (x <= dl[0].mid && dl[x].mid == pgno) {
2171 if (mp != dl[x].mptr) { /* bad cursor? */
2172 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2173 txn->mt_flags |= MDB_TXN_ERROR;
2174 return MDB_CORRUPTED;
2179 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2181 np = mdb_page_malloc(txn, 1);
2186 rc = mdb_mid2l_insert(dl, &mid);
2187 mdb_cassert(mc, rc == 0);
2192 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2194 np->mp_flags |= P_DIRTY;
2197 /* Adjust cursors pointing to mp */
2198 mc->mc_pg[mc->mc_top] = np;
2199 m2 = txn->mt_cursors[mc->mc_dbi];
2200 if (mc->mc_flags & C_SUB) {
2201 for (; m2; m2=m2->mc_next) {
2202 m3 = &m2->mc_xcursor->mx_cursor;
2203 if (m3->mc_snum < mc->mc_snum) continue;
2204 if (m3->mc_pg[mc->mc_top] == mp)
2205 m3->mc_pg[mc->mc_top] = np;
2208 for (; m2; m2=m2->mc_next) {
2209 if (m2->mc_snum < mc->mc_snum) continue;
2210 if (m2->mc_pg[mc->mc_top] == mp) {
2211 m2->mc_pg[mc->mc_top] = np;
2212 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2214 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2216 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2217 if (!(leaf->mn_flags & F_SUBDATA))
2218 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2226 txn->mt_flags |= MDB_TXN_ERROR;
2231 mdb_env_sync(MDB_env *env, int force)
2234 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2235 if (env->me_flags & MDB_WRITEMAP) {
2236 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2237 ? MS_ASYNC : MS_SYNC;
2238 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2241 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2245 if (MDB_FDATASYNC(env->me_fd))
2252 /** Back up parent txn's cursors, then grab the originals for tracking */
2254 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2256 MDB_cursor *mc, *bk;
2261 for (i = src->mt_numdbs; --i >= 0; ) {
2262 if ((mc = src->mt_cursors[i]) != NULL) {
2263 size = sizeof(MDB_cursor);
2265 size += sizeof(MDB_xcursor);
2266 for (; mc; mc = bk->mc_next) {
2272 mc->mc_db = &dst->mt_dbs[i];
2273 /* Kill pointers into src - and dst to reduce abuse: The
2274 * user may not use mc until dst ends. Otherwise we'd...
2276 mc->mc_txn = NULL; /* ...set this to dst */
2277 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2278 if ((mx = mc->mc_xcursor) != NULL) {
2279 *(MDB_xcursor *)(bk+1) = *mx;
2280 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2282 mc->mc_next = dst->mt_cursors[i];
2283 dst->mt_cursors[i] = mc;
2290 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2291 * @param[in] txn the transaction handle.
2292 * @param[in] merge true to keep changes to parent cursors, false to revert.
2293 * @return 0 on success, non-zero on failure.
2296 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2298 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2302 for (i = txn->mt_numdbs; --i >= 0; ) {
2303 for (mc = cursors[i]; mc; mc = next) {
2305 if ((bk = mc->mc_backup) != NULL) {
2307 /* Commit changes to parent txn */
2308 mc->mc_next = bk->mc_next;
2309 mc->mc_backup = bk->mc_backup;
2310 mc->mc_txn = bk->mc_txn;
2311 mc->mc_db = bk->mc_db;
2312 mc->mc_dbflag = bk->mc_dbflag;
2313 if ((mx = mc->mc_xcursor) != NULL)
2314 mx->mx_cursor.mc_txn = bk->mc_txn;
2316 /* Abort nested txn */
2318 if ((mx = mc->mc_xcursor) != NULL)
2319 *mx = *(MDB_xcursor *)(bk+1);
2323 /* Only malloced cursors are permanently tracked. */
2331 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2334 mdb_txn_reset0(MDB_txn *txn, const char *act);
2336 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2342 Pidset = F_SETLK, Pidcheck = F_GETLK
2346 /** Set or check a pid lock. Set returns 0 on success.
2347 * Check returns 0 if the process is certainly dead, nonzero if it may
2348 * be alive (the lock exists or an error happened so we do not know).
2350 * On Windows Pidset is a no-op, we merely check for the existence
2351 * of the process with the given pid. On POSIX we use a single byte
2352 * lock on the lockfile, set at an offset equal to the pid.
2355 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2357 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2360 if (op == Pidcheck) {
2361 h = OpenProcess(env->me_pidquery, FALSE, pid);
2362 /* No documented "no such process" code, but other program use this: */
2364 return ErrCode() != ERROR_INVALID_PARAMETER;
2365 /* A process exists until all handles to it close. Has it exited? */
2366 ret = WaitForSingleObject(h, 0) != 0;
2373 struct flock lock_info;
2374 memset(&lock_info, 0, sizeof(lock_info));
2375 lock_info.l_type = F_WRLCK;
2376 lock_info.l_whence = SEEK_SET;
2377 lock_info.l_start = pid;
2378 lock_info.l_len = 1;
2379 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2380 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2382 } else if ((rc = ErrCode()) == EINTR) {
2390 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2391 * @param[in] txn the transaction handle to initialize
2392 * @return 0 on success, non-zero on failure.
2395 mdb_txn_renew0(MDB_txn *txn)
2397 MDB_env *env = txn->mt_env;
2398 MDB_txninfo *ti = env->me_txns;
2402 int rc, new_notls = 0;
2405 txn->mt_numdbs = env->me_numdbs;
2406 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2408 if (txn->mt_flags & MDB_TXN_RDONLY) {
2410 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2411 txn->mt_txnid = meta->mm_txnid;
2412 txn->mt_u.reader = NULL;
2414 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2415 pthread_getspecific(env->me_txkey);
2417 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2418 return MDB_BAD_RSLOT;
2420 MDB_PID_T pid = env->me_pid;
2421 MDB_THR_T tid = pthread_self();
2423 if (!env->me_live_reader) {
2424 rc = mdb_reader_pid(env, Pidset, pid);
2427 env->me_live_reader = 1;
2431 nr = ti->mti_numreaders;
2432 for (i=0; i<nr; i++)
2433 if (ti->mti_readers[i].mr_pid == 0)
2435 if (i == env->me_maxreaders) {
2436 UNLOCK_MUTEX_R(env);
2437 return MDB_READERS_FULL;
2439 ti->mti_readers[i].mr_pid = pid;
2440 ti->mti_readers[i].mr_tid = tid;
2442 ti->mti_numreaders = ++nr;
2443 /* Save numreaders for un-mutexed mdb_env_close() */
2444 env->me_numreaders = nr;
2445 UNLOCK_MUTEX_R(env);
2447 r = &ti->mti_readers[i];
2448 new_notls = (env->me_flags & MDB_NOTLS);
2449 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2454 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2455 txn->mt_u.reader = r;
2456 meta = env->me_metas[txn->mt_txnid & 1];
2462 txn->mt_txnid = ti->mti_txnid;
2463 meta = env->me_metas[txn->mt_txnid & 1];
2465 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2466 txn->mt_txnid = meta->mm_txnid;
2470 if (txn->mt_txnid == mdb_debug_start)
2473 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2474 txn->mt_u.dirty_list = env->me_dirty_list;
2475 txn->mt_u.dirty_list[0].mid = 0;
2476 txn->mt_free_pgs = env->me_free_pgs;
2477 txn->mt_free_pgs[0] = 0;
2478 txn->mt_spill_pgs = NULL;
2482 /* Copy the DB info and flags */
2483 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2485 /* Moved to here to avoid a data race in read TXNs */
2486 txn->mt_next_pgno = meta->mm_last_pg+1;
2488 for (i=2; i<txn->mt_numdbs; i++) {
2489 x = env->me_dbflags[i];
2490 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2491 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2493 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2495 if (env->me_maxpg < txn->mt_next_pgno) {
2496 mdb_txn_reset0(txn, "renew0-mapfail");
2498 txn->mt_u.reader->mr_pid = 0;
2499 txn->mt_u.reader = NULL;
2501 return MDB_MAP_RESIZED;
2508 mdb_txn_renew(MDB_txn *txn)
2512 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2515 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2516 DPUTS("environment had fatal error, must shutdown!");
2520 rc = mdb_txn_renew0(txn);
2521 if (rc == MDB_SUCCESS) {
2522 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2523 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2524 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2530 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2534 int rc, size, tsize = sizeof(MDB_txn);
2536 if (env->me_flags & MDB_FATAL_ERROR) {
2537 DPUTS("environment had fatal error, must shutdown!");
2540 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2543 /* Nested transactions: Max 1 child, write txns only, no writemap */
2544 if (parent->mt_child ||
2545 (flags & MDB_RDONLY) ||
2546 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2547 (env->me_flags & MDB_WRITEMAP))
2549 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2551 tsize = sizeof(MDB_ntxn);
2553 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2554 if (!(flags & MDB_RDONLY))
2555 size += env->me_maxdbs * sizeof(MDB_cursor *);
2557 if ((txn = calloc(1, size)) == NULL) {
2558 DPRINTF(("calloc: %s", strerror(ErrCode())));
2561 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2562 if (flags & MDB_RDONLY) {
2563 txn->mt_flags |= MDB_TXN_RDONLY;
2564 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2566 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2567 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2573 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2574 if (!txn->mt_u.dirty_list ||
2575 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2577 free(txn->mt_u.dirty_list);
2581 txn->mt_txnid = parent->mt_txnid;
2582 txn->mt_dirty_room = parent->mt_dirty_room;
2583 txn->mt_u.dirty_list[0].mid = 0;
2584 txn->mt_spill_pgs = NULL;
2585 txn->mt_next_pgno = parent->mt_next_pgno;
2586 parent->mt_child = txn;
2587 txn->mt_parent = parent;
2588 txn->mt_numdbs = parent->mt_numdbs;
2589 txn->mt_flags = parent->mt_flags;
2590 txn->mt_dbxs = parent->mt_dbxs;
2591 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2592 /* Copy parent's mt_dbflags, but clear DB_NEW */
2593 for (i=0; i<txn->mt_numdbs; i++)
2594 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2596 ntxn = (MDB_ntxn *)txn;
2597 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2598 if (env->me_pghead) {
2599 size = MDB_IDL_SIZEOF(env->me_pghead);
2600 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2602 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2607 rc = mdb_cursor_shadow(parent, txn);
2609 mdb_txn_reset0(txn, "beginchild-fail");
2611 rc = mdb_txn_renew0(txn);
2617 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2618 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2619 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2626 mdb_txn_env(MDB_txn *txn)
2628 if(!txn) return NULL;
2632 /** Export or close DBI handles opened in this txn. */
2634 mdb_dbis_update(MDB_txn *txn, int keep)
2637 MDB_dbi n = txn->mt_numdbs;
2638 MDB_env *env = txn->mt_env;
2639 unsigned char *tdbflags = txn->mt_dbflags;
2641 for (i = n; --i >= 2;) {
2642 if (tdbflags[i] & DB_NEW) {
2644 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2646 char *ptr = env->me_dbxs[i].md_name.mv_data;
2647 env->me_dbxs[i].md_name.mv_data = NULL;
2648 env->me_dbxs[i].md_name.mv_size = 0;
2649 env->me_dbflags[i] = 0;
2654 if (keep && env->me_numdbs < n)
2658 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2659 * May be called twice for readonly txns: First reset it, then abort.
2660 * @param[in] txn the transaction handle to reset
2661 * @param[in] act why the transaction is being reset
2664 mdb_txn_reset0(MDB_txn *txn, const char *act)
2666 MDB_env *env = txn->mt_env;
2668 /* Close any DBI handles opened in this txn */
2669 mdb_dbis_update(txn, 0);
2671 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2672 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2673 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2675 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2676 if (txn->mt_u.reader) {
2677 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2678 if (!(env->me_flags & MDB_NOTLS))
2679 txn->mt_u.reader = NULL; /* txn does not own reader */
2681 txn->mt_numdbs = 0; /* close nothing if called again */
2682 txn->mt_dbxs = NULL; /* mark txn as reset */
2684 mdb_cursors_close(txn, 0);
2686 if (!(env->me_flags & MDB_WRITEMAP)) {
2687 mdb_dlist_free(txn);
2689 mdb_midl_free(env->me_pghead);
2691 if (txn->mt_parent) {
2692 txn->mt_parent->mt_child = NULL;
2693 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2694 mdb_midl_free(txn->mt_free_pgs);
2695 mdb_midl_free(txn->mt_spill_pgs);
2696 free(txn->mt_u.dirty_list);
2700 if (mdb_midl_shrink(&txn->mt_free_pgs))
2701 env->me_free_pgs = txn->mt_free_pgs;
2702 env->me_pghead = NULL;
2706 /* The writer mutex was locked in mdb_txn_begin. */
2708 UNLOCK_MUTEX_W(env);
2713 mdb_txn_reset(MDB_txn *txn)
2718 /* This call is only valid for read-only txns */
2719 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2722 mdb_txn_reset0(txn, "reset");
2726 mdb_txn_abort(MDB_txn *txn)
2732 mdb_txn_abort(txn->mt_child);
2734 mdb_txn_reset0(txn, "abort");
2735 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2736 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2737 txn->mt_u.reader->mr_pid = 0;
2742 /** Save the freelist as of this transaction to the freeDB.
2743 * This changes the freelist. Keep trying until it stabilizes.
2746 mdb_freelist_save(MDB_txn *txn)
2748 /* env->me_pghead[] can grow and shrink during this call.
2749 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2750 * Page numbers cannot disappear from txn->mt_free_pgs[].
2753 MDB_env *env = txn->mt_env;
2754 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2755 txnid_t pglast = 0, head_id = 0;
2756 pgno_t freecnt = 0, *free_pgs, *mop;
2757 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2759 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2761 if (env->me_pghead) {
2762 /* Make sure first page of freeDB is touched and on freelist */
2763 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2764 if (rc && rc != MDB_NOTFOUND)
2768 /* Dispose of loose pages. Usually they will have all
2769 * been used up by the time we get here.
2771 if (txn->mt_loose_pgs) {
2772 MDB_page *mp = txn->mt_loose_pgs;
2773 /* Just return them to freeDB */
2774 if (env->me_pghead) {
2776 mop = env->me_pghead;
2777 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2778 pgno_t pg = mp->mp_pgno;
2780 for (i = mop[0]; i && mop[i] < pg; i--)
2786 /* Oh well, they were wasted. Put on freelist */
2787 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2788 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
2791 txn->mt_loose_pgs = NULL;
2794 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2795 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2796 ? SSIZE_MAX : maxfree_1pg;
2799 /* Come back here after each Put() in case freelist changed */
2804 /* If using records from freeDB which we have not yet
2805 * deleted, delete them and any we reserved for me_pghead.
2807 while (pglast < env->me_pglast) {
2808 rc = mdb_cursor_first(&mc, &key, NULL);
2811 pglast = head_id = *(txnid_t *)key.mv_data;
2812 total_room = head_room = 0;
2813 mdb_tassert(txn, pglast <= env->me_pglast);
2814 rc = mdb_cursor_del(&mc, 0);
2819 /* Save the IDL of pages freed by this txn, to a single record */
2820 if (freecnt < txn->mt_free_pgs[0]) {
2822 /* Make sure last page of freeDB is touched and on freelist */
2823 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2824 if (rc && rc != MDB_NOTFOUND)
2827 free_pgs = txn->mt_free_pgs;
2828 /* Write to last page of freeDB */
2829 key.mv_size = sizeof(txn->mt_txnid);
2830 key.mv_data = &txn->mt_txnid;
2832 freecnt = free_pgs[0];
2833 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2834 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2837 /* Retry if mt_free_pgs[] grew during the Put() */
2838 free_pgs = txn->mt_free_pgs;
2839 } while (freecnt < free_pgs[0]);
2840 mdb_midl_sort(free_pgs);
2841 memcpy(data.mv_data, free_pgs, data.mv_size);
2844 unsigned int i = free_pgs[0];
2845 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2846 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2848 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2854 mop = env->me_pghead;
2855 mop_len = mop ? mop[0] : 0;
2857 /* Reserve records for me_pghead[]. Split it if multi-page,
2858 * to avoid searching freeDB for a page range. Use keys in
2859 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2861 if (total_room >= mop_len) {
2862 if (total_room == mop_len || --more < 0)
2864 } else if (head_room >= maxfree_1pg && head_id > 1) {
2865 /* Keep current record (overflow page), add a new one */
2869 /* (Re)write {key = head_id, IDL length = head_room} */
2870 total_room -= head_room;
2871 head_room = mop_len - total_room;
2872 if (head_room > maxfree_1pg && head_id > 1) {
2873 /* Overflow multi-page for part of me_pghead */
2874 head_room /= head_id; /* amortize page sizes */
2875 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2876 } else if (head_room < 0) {
2877 /* Rare case, not bothering to delete this record */
2880 key.mv_size = sizeof(head_id);
2881 key.mv_data = &head_id;
2882 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2883 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2886 /* IDL is initially empty, zero out at least the length */
2887 pgs = (pgno_t *)data.mv_data;
2888 j = head_room > clean_limit ? head_room : 0;
2892 total_room += head_room;
2895 /* Fill in the reserved me_pghead records */
2901 rc = mdb_cursor_first(&mc, &key, &data);
2902 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2903 txnid_t id = *(txnid_t *)key.mv_data;
2904 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2907 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2909 if (len > mop_len) {
2911 data.mv_size = (len + 1) * sizeof(MDB_ID);
2913 data.mv_data = mop -= len;
2916 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2918 if (rc || !(mop_len -= len))
2925 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2926 * @param[in] txn the transaction that's being committed
2927 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2928 * @return 0 on success, non-zero on failure.
2931 mdb_page_flush(MDB_txn *txn, int keep)
2933 MDB_env *env = txn->mt_env;
2934 MDB_ID2L dl = txn->mt_u.dirty_list;
2935 unsigned psize = env->me_psize, j;
2936 int i, pagecount = dl[0].mid, rc;
2937 size_t size = 0, pos = 0;
2939 MDB_page *dp = NULL;
2943 struct iovec iov[MDB_COMMIT_PAGES];
2944 ssize_t wpos = 0, wsize = 0, wres;
2945 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2951 if (env->me_flags & MDB_WRITEMAP) {
2952 /* Clear dirty flags */
2953 while (++i <= pagecount) {
2955 /* Don't flush this page yet */
2956 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
2957 dp->mp_flags &= ~P_KEEP;
2961 dp->mp_flags &= ~P_DIRTY;
2966 /* Write the pages */
2968 if (++i <= pagecount) {
2970 /* Don't flush this page yet */
2971 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
2972 dp->mp_flags &= ~P_KEEP;
2977 /* clear dirty flag */
2978 dp->mp_flags &= ~P_DIRTY;
2981 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2986 /* Windows actually supports scatter/gather I/O, but only on
2987 * unbuffered file handles. Since we're relying on the OS page
2988 * cache for all our data, that's self-defeating. So we just
2989 * write pages one at a time. We use the ov structure to set
2990 * the write offset, to at least save the overhead of a Seek
2993 DPRINTF(("committing page %"Z"u", pgno));
2994 memset(&ov, 0, sizeof(ov));
2995 ov.Offset = pos & 0xffffffff;
2996 ov.OffsetHigh = pos >> 16 >> 16;
2997 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2999 DPRINTF(("WriteFile: %d", rc));
3003 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3004 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3006 /* Write previous page(s) */
3007 #ifdef MDB_USE_PWRITEV
3008 wres = pwritev(env->me_fd, iov, n, wpos);
3011 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3013 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3015 DPRINTF(("lseek: %s", strerror(rc)));
3018 wres = writev(env->me_fd, iov, n);
3021 if (wres != wsize) {
3024 DPRINTF(("Write error: %s", strerror(rc)));
3026 rc = EIO; /* TODO: Use which error code? */
3027 DPUTS("short write, filesystem full?");
3038 DPRINTF(("committing page %"Z"u", pgno));
3039 next_pos = pos + size;
3040 iov[n].iov_len = size;
3041 iov[n].iov_base = (char *)dp;
3047 for (i = keep; ++i <= pagecount; ) {
3049 /* This is a page we skipped above */
3052 dl[j].mid = dp->mp_pgno;
3055 mdb_dpage_free(env, dp);
3060 txn->mt_dirty_room += i - j;
3066 mdb_txn_commit(MDB_txn *txn)
3072 if (txn == NULL || txn->mt_env == NULL)
3075 if (txn->mt_child) {
3076 rc = mdb_txn_commit(txn->mt_child);
3077 txn->mt_child = NULL;
3084 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3085 mdb_dbis_update(txn, 1);
3086 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3091 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3092 DPUTS("error flag is set, can't commit");
3094 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3099 if (txn->mt_parent) {
3100 MDB_txn *parent = txn->mt_parent;
3104 unsigned x, y, len, ps_len;
3106 /* Append our free list to parent's */
3107 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3110 mdb_midl_free(txn->mt_free_pgs);
3111 /* Failures after this must either undo the changes
3112 * to the parent or set MDB_TXN_ERROR in the parent.
3115 parent->mt_next_pgno = txn->mt_next_pgno;
3116 parent->mt_flags = txn->mt_flags;
3118 /* Merge our cursors into parent's and close them */
3119 mdb_cursors_close(txn, 1);
3121 /* Update parent's DB table. */
3122 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3123 parent->mt_numdbs = txn->mt_numdbs;
3124 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3125 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3126 for (i=2; i<txn->mt_numdbs; i++) {
3127 /* preserve parent's DB_NEW status */
3128 x = parent->mt_dbflags[i] & DB_NEW;
3129 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3132 dst = parent->mt_u.dirty_list;
3133 src = txn->mt_u.dirty_list;
3134 /* Remove anything in our dirty list from parent's spill list */
3135 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3137 pspill[0] = (pgno_t)-1;
3138 /* Mark our dirty pages as deleted in parent spill list */
3139 for (i=0, len=src[0].mid; ++i <= len; ) {
3140 MDB_ID pn = src[i].mid << 1;
3141 while (pn > pspill[x])
3143 if (pn == pspill[x]) {
3148 /* Squash deleted pagenums if we deleted any */
3149 for (x=y; ++x <= ps_len; )
3150 if (!(pspill[x] & 1))
3151 pspill[++y] = pspill[x];
3155 /* Find len = length of merging our dirty list with parent's */
3157 dst[0].mid = 0; /* simplify loops */
3158 if (parent->mt_parent) {
3159 len = x + src[0].mid;
3160 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3161 for (i = x; y && i; y--) {
3162 pgno_t yp = src[y].mid;
3163 while (yp < dst[i].mid)
3165 if (yp == dst[i].mid) {
3170 } else { /* Simplify the above for single-ancestor case */
3171 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3173 /* Merge our dirty list with parent's */
3175 for (i = len; y; dst[i--] = src[y--]) {
3176 pgno_t yp = src[y].mid;
3177 while (yp < dst[x].mid)
3178 dst[i--] = dst[x--];
3179 if (yp == dst[x].mid)
3180 free(dst[x--].mptr);
3182 mdb_tassert(txn, i == x);
3184 free(txn->mt_u.dirty_list);
3185 parent->mt_dirty_room = txn->mt_dirty_room;
3186 if (txn->mt_spill_pgs) {
3187 if (parent->mt_spill_pgs) {
3188 /* TODO: Prevent failure here, so parent does not fail */
3189 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3191 parent->mt_flags |= MDB_TXN_ERROR;
3192 mdb_midl_free(txn->mt_spill_pgs);
3193 mdb_midl_sort(parent->mt_spill_pgs);
3195 parent->mt_spill_pgs = txn->mt_spill_pgs;
3199 /* Append our loose page list to parent's */
3200 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3202 *lp = txn->mt_loose_pgs;
3204 parent->mt_child = NULL;
3205 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3210 if (txn != env->me_txn) {
3211 DPUTS("attempt to commit unknown transaction");
3216 mdb_cursors_close(txn, 0);
3218 if (!txn->mt_u.dirty_list[0].mid &&
3219 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3222 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3223 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3225 /* Update DB root pointers */
3226 if (txn->mt_numdbs > 2) {
3230 data.mv_size = sizeof(MDB_db);
3232 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3233 for (i = 2; i < txn->mt_numdbs; i++) {
3234 if (txn->mt_dbflags[i] & DB_DIRTY) {
3235 data.mv_data = &txn->mt_dbs[i];
3236 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3243 rc = mdb_freelist_save(txn);
3247 mdb_midl_free(env->me_pghead);
3248 env->me_pghead = NULL;
3249 if (mdb_midl_shrink(&txn->mt_free_pgs))
3250 env->me_free_pgs = txn->mt_free_pgs;
3256 if ((rc = mdb_page_flush(txn, 0)) ||
3257 (rc = mdb_env_sync(env, 0)) ||
3258 (rc = mdb_env_write_meta(txn)))
3264 mdb_dbis_update(txn, 1);
3267 UNLOCK_MUTEX_W(env);
3277 /** Read the environment parameters of a DB environment before
3278 * mapping it into memory.
3279 * @param[in] env the environment handle
3280 * @param[out] meta address of where to store the meta information
3281 * @return 0 on success, non-zero on failure.
3284 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3290 enum { Size = sizeof(pbuf) };
3292 /* We don't know the page size yet, so use a minimum value.
3293 * Read both meta pages so we can use the latest one.
3296 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3300 memset(&ov, 0, sizeof(ov));
3302 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3303 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3306 rc = pread(env->me_fd, &pbuf, Size, off);
3309 if (rc == 0 && off == 0)
3311 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3312 DPRINTF(("read: %s", mdb_strerror(rc)));
3316 p = (MDB_page *)&pbuf;
3318 if (!F_ISSET(p->mp_flags, P_META)) {
3319 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3324 if (m->mm_magic != MDB_MAGIC) {
3325 DPUTS("meta has invalid magic");
3329 if (m->mm_version != MDB_DATA_VERSION) {
3330 DPRINTF(("database is version %u, expected version %u",
3331 m->mm_version, MDB_DATA_VERSION));
3332 return MDB_VERSION_MISMATCH;
3335 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3342 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3344 meta->mm_magic = MDB_MAGIC;
3345 meta->mm_version = MDB_DATA_VERSION;
3346 meta->mm_mapsize = env->me_mapsize;
3347 meta->mm_psize = env->me_psize;
3348 meta->mm_last_pg = 1;
3349 meta->mm_flags = env->me_flags & 0xffff;
3350 meta->mm_flags |= MDB_INTEGERKEY;
3351 meta->mm_dbs[0].md_root = P_INVALID;
3352 meta->mm_dbs[1].md_root = P_INVALID;
3355 /** Write the environment parameters of a freshly created DB environment.
3356 * @param[in] env the environment handle
3357 * @param[out] meta address of where to store the meta information
3358 * @return 0 on success, non-zero on failure.
3361 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3369 memset(&ov, 0, sizeof(ov));
3370 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3372 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3375 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3376 len = pwrite(fd, ptr, size, pos); \
3377 rc = (len >= 0); } while(0)
3380 DPUTS("writing new meta page");
3382 psize = env->me_psize;
3384 mdb_env_init_meta0(env, meta);
3386 p = calloc(2, psize);
3388 p->mp_flags = P_META;
3389 *(MDB_meta *)METADATA(p) = *meta;
3391 q = (MDB_page *)((char *)p + psize);
3393 q->mp_flags = P_META;
3394 *(MDB_meta *)METADATA(q) = *meta;
3396 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3399 else if ((unsigned) len == psize * 2)
3407 /** Update the environment info to commit a transaction.
3408 * @param[in] txn the transaction that's being committed
3409 * @return 0 on success, non-zero on failure.
3412 mdb_env_write_meta(MDB_txn *txn)
3415 MDB_meta meta, metab, *mp;
3417 int rc, len, toggle;
3426 toggle = txn->mt_txnid & 1;
3427 DPRINTF(("writing meta page %d for root page %"Z"u",
3428 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3431 mp = env->me_metas[toggle];
3433 if (env->me_flags & MDB_WRITEMAP) {
3434 /* Persist any increases of mapsize config */
3435 if (env->me_mapsize > mp->mm_mapsize)
3436 mp->mm_mapsize = env->me_mapsize;
3437 mp->mm_dbs[0] = txn->mt_dbs[0];
3438 mp->mm_dbs[1] = txn->mt_dbs[1];
3439 mp->mm_last_pg = txn->mt_next_pgno - 1;
3440 mp->mm_txnid = txn->mt_txnid;
3441 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3442 unsigned meta_size = env->me_psize;
3443 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3446 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3447 if (meta_size < env->me_os_psize)
3448 meta_size += meta_size;
3453 if (MDB_MSYNC(ptr, meta_size, rc)) {
3460 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3461 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3463 ptr = (char *)&meta;
3464 if (env->me_mapsize > mp->mm_mapsize) {
3465 /* Persist any increases of mapsize config */
3466 meta.mm_mapsize = env->me_mapsize;
3467 off = offsetof(MDB_meta, mm_mapsize);
3469 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3471 len = sizeof(MDB_meta) - off;
3474 meta.mm_dbs[0] = txn->mt_dbs[0];
3475 meta.mm_dbs[1] = txn->mt_dbs[1];
3476 meta.mm_last_pg = txn->mt_next_pgno - 1;
3477 meta.mm_txnid = txn->mt_txnid;
3480 off += env->me_psize;
3483 /* Write to the SYNC fd */
3484 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3485 env->me_fd : env->me_mfd;
3488 memset(&ov, 0, sizeof(ov));
3490 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3494 rc = pwrite(mfd, ptr, len, off);
3497 rc = rc < 0 ? ErrCode() : EIO;
3498 DPUTS("write failed, disk error?");
3499 /* On a failure, the pagecache still contains the new data.
3500 * Write some old data back, to prevent it from being used.
3501 * Use the non-SYNC fd; we know it will fail anyway.
3503 meta.mm_last_pg = metab.mm_last_pg;
3504 meta.mm_txnid = metab.mm_txnid;
3506 memset(&ov, 0, sizeof(ov));
3508 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3510 r2 = pwrite(env->me_fd, ptr, len, off);
3511 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3514 env->me_flags |= MDB_FATAL_ERROR;
3518 /* Memory ordering issues are irrelevant; since the entire writer
3519 * is wrapped by wmutex, all of these changes will become visible
3520 * after the wmutex is unlocked. Since the DB is multi-version,
3521 * readers will get consistent data regardless of how fresh or
3522 * how stale their view of these values is.
3525 env->me_txns->mti_txnid = txn->mt_txnid;
3530 /** Check both meta pages to see which one is newer.
3531 * @param[in] env the environment handle
3532 * @return meta toggle (0 or 1).
3535 mdb_env_pick_meta(const MDB_env *env)
3537 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3541 mdb_env_create(MDB_env **env)
3545 e = calloc(1, sizeof(MDB_env));
3549 e->me_maxreaders = DEFAULT_READERS;
3550 e->me_maxdbs = e->me_numdbs = 2;
3551 e->me_fd = INVALID_HANDLE_VALUE;
3552 e->me_lfd = INVALID_HANDLE_VALUE;
3553 e->me_mfd = INVALID_HANDLE_VALUE;
3554 #ifdef MDB_USE_POSIX_SEM
3555 e->me_rmutex = SEM_FAILED;
3556 e->me_wmutex = SEM_FAILED;
3558 e->me_pid = getpid();
3559 GET_PAGESIZE(e->me_os_psize);
3560 VGMEMP_CREATE(e,0,0);
3566 mdb_env_map(MDB_env *env, void *addr, int newsize)
3569 unsigned int flags = env->me_flags;
3573 LONG sizelo, sizehi;
3576 if (flags & MDB_RDONLY) {
3581 msize = env->me_mapsize;
3582 sizelo = msize & 0xffffffff;
3583 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3586 /* Windows won't create mappings for zero length files.
3587 * Just allocate the maxsize right now.
3590 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3591 || !SetEndOfFile(env->me_fd)
3592 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3595 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3596 PAGE_READWRITE : PAGE_READONLY,
3597 sizehi, sizelo, NULL);
3600 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3601 FILE_MAP_WRITE : FILE_MAP_READ,
3603 rc = env->me_map ? 0 : ErrCode();
3608 int prot = PROT_READ;
3609 if (flags & MDB_WRITEMAP) {
3611 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3614 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3616 if (env->me_map == MAP_FAILED) {
3621 if (flags & MDB_NORDAHEAD) {
3622 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3624 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3626 #ifdef POSIX_MADV_RANDOM
3627 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3628 #endif /* POSIX_MADV_RANDOM */
3629 #endif /* MADV_RANDOM */
3633 /* Can happen because the address argument to mmap() is just a
3634 * hint. mmap() can pick another, e.g. if the range is in use.
3635 * The MAP_FIXED flag would prevent that, but then mmap could
3636 * instead unmap existing pages to make room for the new map.
3638 if (addr && env->me_map != addr)
3639 return EBUSY; /* TODO: Make a new MDB_* error code? */
3641 p = (MDB_page *)env->me_map;
3642 env->me_metas[0] = METADATA(p);
3643 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3649 mdb_env_set_mapsize(MDB_env *env, size_t size)
3651 /* If env is already open, caller is responsible for making
3652 * sure there are no active txns.
3660 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3661 else if (size < env->me_mapsize) {
3662 /* If the configured size is smaller, make sure it's
3663 * still big enough. Silently round up to minimum if not.
3665 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3669 munmap(env->me_map, env->me_mapsize);
3670 env->me_mapsize = size;
3671 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3672 rc = mdb_env_map(env, old, 1);
3676 env->me_mapsize = size;
3678 env->me_maxpg = env->me_mapsize / env->me_psize;
3683 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3687 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3692 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3694 if (env->me_map || readers < 1)
3696 env->me_maxreaders = readers;
3701 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3703 if (!env || !readers)
3705 *readers = env->me_maxreaders;
3709 /** Further setup required for opening an LMDB environment
3712 mdb_env_open2(MDB_env *env)
3714 unsigned int flags = env->me_flags;
3715 int i, newenv = 0, rc;
3719 /* See if we should use QueryLimited */
3721 if ((rc & 0xff) > 5)
3722 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3724 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3727 memset(&meta, 0, sizeof(meta));
3729 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3732 DPUTS("new mdbenv");
3734 env->me_psize = env->me_os_psize;
3735 if (env->me_psize > MAX_PAGESIZE)
3736 env->me_psize = MAX_PAGESIZE;
3738 env->me_psize = meta.mm_psize;
3741 /* Was a mapsize configured? */
3742 if (!env->me_mapsize) {
3743 /* If this is a new environment, take the default,
3744 * else use the size recorded in the existing env.
3746 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3747 } else if (env->me_mapsize < meta.mm_mapsize) {
3748 /* If the configured size is smaller, make sure it's
3749 * still big enough. Silently round up to minimum if not.
3751 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3752 if (env->me_mapsize < minsize)
3753 env->me_mapsize = minsize;
3756 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3761 if (flags & MDB_FIXEDMAP)
3762 meta.mm_address = env->me_map;
3763 i = mdb_env_init_meta(env, &meta);
3764 if (i != MDB_SUCCESS) {
3769 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3770 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3772 #if !(MDB_MAXKEYSIZE)
3773 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3775 env->me_maxpg = env->me_mapsize / env->me_psize;
3779 int toggle = mdb_env_pick_meta(env);
3780 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3782 DPRINTF(("opened database version %u, pagesize %u",
3783 env->me_metas[0]->mm_version, env->me_psize));
3784 DPRINTF(("using meta page %d", toggle));
3785 DPRINTF(("depth: %u", db->md_depth));
3786 DPRINTF(("entries: %"Z"u", db->md_entries));
3787 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3788 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3789 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3790 DPRINTF(("root: %"Z"u", db->md_root));
3798 /** Release a reader thread's slot in the reader lock table.
3799 * This function is called automatically when a thread exits.
3800 * @param[in] ptr This points to the slot in the reader lock table.
3803 mdb_env_reader_dest(void *ptr)
3805 MDB_reader *reader = ptr;
3811 /** Junk for arranging thread-specific callbacks on Windows. This is
3812 * necessarily platform and compiler-specific. Windows supports up
3813 * to 1088 keys. Let's assume nobody opens more than 64 environments
3814 * in a single process, for now. They can override this if needed.
3816 #ifndef MAX_TLS_KEYS
3817 #define MAX_TLS_KEYS 64
3819 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3820 static int mdb_tls_nkeys;
3822 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3826 case DLL_PROCESS_ATTACH: break;
3827 case DLL_THREAD_ATTACH: break;
3828 case DLL_THREAD_DETACH:
3829 for (i=0; i<mdb_tls_nkeys; i++) {
3830 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3832 mdb_env_reader_dest(r);
3836 case DLL_PROCESS_DETACH: break;
3841 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3843 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3847 /* Force some symbol references.
3848 * _tls_used forces the linker to create the TLS directory if not already done
3849 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3851 #pragma comment(linker, "/INCLUDE:_tls_used")
3852 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3853 #pragma const_seg(".CRT$XLB")
3854 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3855 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3858 #pragma comment(linker, "/INCLUDE:__tls_used")
3859 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3860 #pragma data_seg(".CRT$XLB")
3861 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3863 #endif /* WIN 32/64 */
3864 #endif /* !__GNUC__ */
3867 /** Downgrade the exclusive lock on the region back to shared */
3869 mdb_env_share_locks(MDB_env *env, int *excl)
3871 int rc = 0, toggle = mdb_env_pick_meta(env);
3873 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3878 /* First acquire a shared lock. The Unlock will
3879 * then release the existing exclusive lock.
3881 memset(&ov, 0, sizeof(ov));
3882 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3885 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3891 struct flock lock_info;
3892 /* The shared lock replaces the existing lock */
3893 memset((void *)&lock_info, 0, sizeof(lock_info));
3894 lock_info.l_type = F_RDLCK;
3895 lock_info.l_whence = SEEK_SET;
3896 lock_info.l_start = 0;
3897 lock_info.l_len = 1;
3898 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3899 (rc = ErrCode()) == EINTR) ;
3900 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3907 /** Try to get exlusive lock, otherwise shared.
3908 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3911 mdb_env_excl_lock(MDB_env *env, int *excl)
3915 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3919 memset(&ov, 0, sizeof(ov));
3920 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3927 struct flock lock_info;
3928 memset((void *)&lock_info, 0, sizeof(lock_info));
3929 lock_info.l_type = F_WRLCK;
3930 lock_info.l_whence = SEEK_SET;
3931 lock_info.l_start = 0;
3932 lock_info.l_len = 1;
3933 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3934 (rc = ErrCode()) == EINTR) ;
3938 # ifdef MDB_USE_POSIX_SEM
3939 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3942 lock_info.l_type = F_RDLCK;
3943 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3944 (rc = ErrCode()) == EINTR) ;
3954 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3956 * @(#) $Revision: 5.1 $
3957 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3958 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3960 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3964 * Please do not copyright this code. This code is in the public domain.
3966 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3967 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3968 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3969 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3970 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3971 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3972 * PERFORMANCE OF THIS SOFTWARE.
3975 * chongo <Landon Curt Noll> /\oo/\
3976 * http://www.isthe.com/chongo/
3978 * Share and Enjoy! :-)
3981 typedef unsigned long long mdb_hash_t;
3982 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3984 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3985 * @param[in] val value to hash
3986 * @param[in] hval initial value for hash
3987 * @return 64 bit hash
3989 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3990 * hval arg on the first call.
3993 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3995 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3996 unsigned char *end = s + val->mv_size;
3998 * FNV-1a hash each octet of the string
4001 /* xor the bottom with the current octet */
4002 hval ^= (mdb_hash_t)*s++;
4004 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4005 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4006 (hval << 7) + (hval << 8) + (hval << 40);
4008 /* return our new hash value */
4012 /** Hash the string and output the encoded hash.
4013 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4014 * very short name limits. We don't care about the encoding being reversible,
4015 * we just want to preserve as many bits of the input as possible in a
4016 * small printable string.
4017 * @param[in] str string to hash
4018 * @param[out] encbuf an array of 11 chars to hold the hash
4020 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4023 mdb_pack85(unsigned long l, char *out)
4027 for (i=0; i<5; i++) {
4028 *out++ = mdb_a85[l % 85];
4034 mdb_hash_enc(MDB_val *val, char *encbuf)
4036 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4038 mdb_pack85(h, encbuf);
4039 mdb_pack85(h>>32, encbuf+5);
4044 /** Open and/or initialize the lock region for the environment.
4045 * @param[in] env The LMDB environment.
4046 * @param[in] lpath The pathname of the file used for the lock region.
4047 * @param[in] mode The Unix permissions for the file, if we create it.
4048 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4049 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4050 * @return 0 on success, non-zero on failure.
4053 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4056 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4058 # define MDB_ERRCODE_ROFS EROFS
4059 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4060 # define MDB_CLOEXEC O_CLOEXEC
4063 # define MDB_CLOEXEC 0
4070 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4071 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4072 FILE_ATTRIBUTE_NORMAL, NULL);
4074 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4076 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4078 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4083 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4084 /* Lose record locks when exec*() */
4085 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4086 fcntl(env->me_lfd, F_SETFD, fdflags);
4089 if (!(env->me_flags & MDB_NOTLS)) {
4090 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4093 env->me_flags |= MDB_ENV_TXKEY;
4095 /* Windows TLS callbacks need help finding their TLS info. */
4096 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4100 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4104 /* Try to get exclusive lock. If we succeed, then
4105 * nobody is using the lock region and we should initialize it.
4107 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4110 size = GetFileSize(env->me_lfd, NULL);
4112 size = lseek(env->me_lfd, 0, SEEK_END);
4113 if (size == -1) goto fail_errno;
4115 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4116 if (size < rsize && *excl > 0) {
4118 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4119 || !SetEndOfFile(env->me_lfd))
4122 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4126 size = rsize - sizeof(MDB_txninfo);
4127 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4132 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4134 if (!mh) goto fail_errno;
4135 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4137 if (!env->me_txns) goto fail_errno;
4139 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4141 if (m == MAP_FAILED) goto fail_errno;
4147 BY_HANDLE_FILE_INFORMATION stbuf;
4156 if (!mdb_sec_inited) {
4157 InitializeSecurityDescriptor(&mdb_null_sd,
4158 SECURITY_DESCRIPTOR_REVISION);
4159 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4160 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4161 mdb_all_sa.bInheritHandle = FALSE;
4162 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4165 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4166 idbuf.volume = stbuf.dwVolumeSerialNumber;
4167 idbuf.nhigh = stbuf.nFileIndexHigh;
4168 idbuf.nlow = stbuf.nFileIndexLow;
4169 val.mv_data = &idbuf;
4170 val.mv_size = sizeof(idbuf);
4171 mdb_hash_enc(&val, encbuf);
4172 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4173 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4174 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4175 if (!env->me_rmutex) goto fail_errno;
4176 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4177 if (!env->me_wmutex) goto fail_errno;
4178 #elif defined(MDB_USE_POSIX_SEM)
4187 #if defined(__NetBSD__)
4188 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4190 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4191 idbuf.dev = stbuf.st_dev;
4192 idbuf.ino = stbuf.st_ino;
4193 val.mv_data = &idbuf;
4194 val.mv_size = sizeof(idbuf);
4195 mdb_hash_enc(&val, encbuf);
4196 #ifdef MDB_SHORT_SEMNAMES
4197 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4199 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4200 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4201 /* Clean up after a previous run, if needed: Try to
4202 * remove both semaphores before doing anything else.
4204 sem_unlink(env->me_txns->mti_rmname);
4205 sem_unlink(env->me_txns->mti_wmname);
4206 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4207 O_CREAT|O_EXCL, mode, 1);
4208 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4209 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4210 O_CREAT|O_EXCL, mode, 1);
4211 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4212 #else /* MDB_USE_POSIX_SEM */
4213 pthread_mutexattr_t mattr;
4215 if ((rc = pthread_mutexattr_init(&mattr))
4216 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4217 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4218 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4220 pthread_mutexattr_destroy(&mattr);
4221 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4223 env->me_txns->mti_magic = MDB_MAGIC;
4224 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4225 env->me_txns->mti_txnid = 0;
4226 env->me_txns->mti_numreaders = 0;
4229 if (env->me_txns->mti_magic != MDB_MAGIC) {
4230 DPUTS("lock region has invalid magic");
4234 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4235 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4236 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4237 rc = MDB_VERSION_MISMATCH;
4241 if (rc && rc != EACCES && rc != EAGAIN) {
4245 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4246 if (!env->me_rmutex) goto fail_errno;
4247 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4248 if (!env->me_wmutex) goto fail_errno;
4249 #elif defined(MDB_USE_POSIX_SEM)
4250 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4251 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4252 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4253 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4264 /** The name of the lock file in the DB environment */
4265 #define LOCKNAME "/lock.mdb"
4266 /** The name of the data file in the DB environment */
4267 #define DATANAME "/data.mdb"
4268 /** The suffix of the lock file when no subdir is used */
4269 #define LOCKSUFF "-lock"
4270 /** Only a subset of the @ref mdb_env flags can be changed
4271 * at runtime. Changing other flags requires closing the
4272 * environment and re-opening it with the new flags.
4274 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4275 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4276 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4278 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4279 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4283 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4285 int oflags, rc, len, excl = -1;
4286 char *lpath, *dpath;
4288 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4292 if (flags & MDB_NOSUBDIR) {
4293 rc = len + sizeof(LOCKSUFF) + len + 1;
4295 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4300 if (flags & MDB_NOSUBDIR) {
4301 dpath = lpath + len + sizeof(LOCKSUFF);
4302 sprintf(lpath, "%s" LOCKSUFF, path);
4303 strcpy(dpath, path);
4305 dpath = lpath + len + sizeof(LOCKNAME);
4306 sprintf(lpath, "%s" LOCKNAME, path);
4307 sprintf(dpath, "%s" DATANAME, path);
4311 flags |= env->me_flags;
4312 if (flags & MDB_RDONLY) {
4313 /* silently ignore WRITEMAP when we're only getting read access */
4314 flags &= ~MDB_WRITEMAP;
4316 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4317 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4320 env->me_flags = flags |= MDB_ENV_ACTIVE;
4324 env->me_path = strdup(path);
4325 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4326 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4327 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4332 /* For RDONLY, get lockfile after we know datafile exists */
4333 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4334 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4340 if (F_ISSET(flags, MDB_RDONLY)) {
4341 oflags = GENERIC_READ;
4342 len = OPEN_EXISTING;
4344 oflags = GENERIC_READ|GENERIC_WRITE;
4347 mode = FILE_ATTRIBUTE_NORMAL;
4348 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4349 NULL, len, mode, NULL);
4351 if (F_ISSET(flags, MDB_RDONLY))
4354 oflags = O_RDWR | O_CREAT;
4356 env->me_fd = open(dpath, oflags, mode);
4358 if (env->me_fd == INVALID_HANDLE_VALUE) {
4363 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4364 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4369 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4370 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4371 env->me_mfd = env->me_fd;
4373 /* Synchronous fd for meta writes. Needed even with
4374 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4377 len = OPEN_EXISTING;
4378 env->me_mfd = CreateFile(dpath, oflags,
4379 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4380 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4383 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4385 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4390 DPRINTF(("opened dbenv %p", (void *) env));
4392 rc = mdb_env_share_locks(env, &excl);
4396 if (!((flags & MDB_RDONLY) ||
4397 (env->me_pbuf = calloc(1, env->me_psize))))
4403 mdb_env_close0(env, excl);
4409 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4411 mdb_env_close0(MDB_env *env, int excl)
4415 if (!(env->me_flags & MDB_ENV_ACTIVE))
4418 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4419 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4420 free(env->me_dbxs[i].md_name.mv_data);
4423 free(env->me_dbflags);
4426 free(env->me_dirty_list);
4427 mdb_midl_free(env->me_free_pgs);
4429 if (env->me_flags & MDB_ENV_TXKEY) {
4430 pthread_key_delete(env->me_txkey);
4432 /* Delete our key from the global list */
4433 for (i=0; i<mdb_tls_nkeys; i++)
4434 if (mdb_tls_keys[i] == env->me_txkey) {
4435 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4443 munmap(env->me_map, env->me_mapsize);
4445 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4446 (void) close(env->me_mfd);
4447 if (env->me_fd != INVALID_HANDLE_VALUE)
4448 (void) close(env->me_fd);
4450 MDB_PID_T pid = env->me_pid;
4451 /* Clearing readers is done in this function because
4452 * me_txkey with its destructor must be disabled first.
4454 for (i = env->me_numreaders; --i >= 0; )
4455 if (env->me_txns->mti_readers[i].mr_pid == pid)
4456 env->me_txns->mti_readers[i].mr_pid = 0;
4458 if (env->me_rmutex) {
4459 CloseHandle(env->me_rmutex);
4460 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4462 /* Windows automatically destroys the mutexes when
4463 * the last handle closes.
4465 #elif defined(MDB_USE_POSIX_SEM)
4466 if (env->me_rmutex != SEM_FAILED) {
4467 sem_close(env->me_rmutex);
4468 if (env->me_wmutex != SEM_FAILED)
4469 sem_close(env->me_wmutex);
4470 /* If we have the filelock: If we are the
4471 * only remaining user, clean up semaphores.
4474 mdb_env_excl_lock(env, &excl);
4476 sem_unlink(env->me_txns->mti_rmname);
4477 sem_unlink(env->me_txns->mti_wmname);
4481 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4483 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4486 /* Unlock the lockfile. Windows would have unlocked it
4487 * after closing anyway, but not necessarily at once.
4489 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4492 (void) close(env->me_lfd);
4495 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4500 mdb_env_close(MDB_env *env)
4507 VGMEMP_DESTROY(env);
4508 while ((dp = env->me_dpages) != NULL) {
4509 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4510 env->me_dpages = dp->mp_next;
4514 mdb_env_close0(env, 0);
4518 /** Compare two items pointing at aligned size_t's */
4520 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4522 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4523 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4526 /** Compare two items pointing at aligned unsigned int's */
4528 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4530 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4531 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4534 /** Compare two items pointing at unsigned ints of unknown alignment.
4535 * Nodes and keys are guaranteed to be 2-byte aligned.
4538 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4540 #if BYTE_ORDER == LITTLE_ENDIAN
4541 unsigned short *u, *c;
4544 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4545 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4548 } while(!x && u > (unsigned short *)a->mv_data);
4551 unsigned short *u, *c, *end;
4554 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4555 u = (unsigned short *)a->mv_data;
4556 c = (unsigned short *)b->mv_data;
4559 } while(!x && u < end);
4564 /** Compare two items pointing at size_t's of unknown alignment. */
4565 #ifdef MISALIGNED_OK
4566 # define mdb_cmp_clong mdb_cmp_long
4568 # define mdb_cmp_clong mdb_cmp_cint
4571 /** Compare two items lexically */
4573 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4580 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4586 diff = memcmp(a->mv_data, b->mv_data, len);
4587 return diff ? diff : len_diff<0 ? -1 : len_diff;
4590 /** Compare two items in reverse byte order */
4592 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4594 const unsigned char *p1, *p2, *p1_lim;
4598 p1_lim = (const unsigned char *)a->mv_data;
4599 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4600 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4602 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4608 while (p1 > p1_lim) {
4609 diff = *--p1 - *--p2;
4613 return len_diff<0 ? -1 : len_diff;
4616 /** Search for key within a page, using binary search.
4617 * Returns the smallest entry larger or equal to the key.
4618 * If exactp is non-null, stores whether the found entry was an exact match
4619 * in *exactp (1 or 0).
4620 * Updates the cursor index with the index of the found entry.
4621 * If no entry larger or equal to the key is found, returns NULL.
4624 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4626 unsigned int i = 0, nkeys;
4629 MDB_page *mp = mc->mc_pg[mc->mc_top];
4630 MDB_node *node = NULL;
4635 nkeys = NUMKEYS(mp);
4637 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4638 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4641 low = IS_LEAF(mp) ? 0 : 1;
4643 cmp = mc->mc_dbx->md_cmp;
4645 /* Branch pages have no data, so if using integer keys,
4646 * alignment is guaranteed. Use faster mdb_cmp_int.
4648 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4649 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4656 nodekey.mv_size = mc->mc_db->md_pad;
4657 node = NODEPTR(mp, 0); /* fake */
4658 while (low <= high) {
4659 i = (low + high) >> 1;
4660 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4661 rc = cmp(key, &nodekey);
4662 DPRINTF(("found leaf index %u [%s], rc = %i",
4663 i, DKEY(&nodekey), rc));
4672 while (low <= high) {
4673 i = (low + high) >> 1;
4675 node = NODEPTR(mp, i);
4676 nodekey.mv_size = NODEKSZ(node);
4677 nodekey.mv_data = NODEKEY(node);
4679 rc = cmp(key, &nodekey);
4682 DPRINTF(("found leaf index %u [%s], rc = %i",
4683 i, DKEY(&nodekey), rc));
4685 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4686 i, DKEY(&nodekey), NODEPGNO(node), rc));
4697 if (rc > 0) { /* Found entry is less than the key. */
4698 i++; /* Skip to get the smallest entry larger than key. */
4700 node = NODEPTR(mp, i);
4703 *exactp = (rc == 0 && nkeys > 0);
4704 /* store the key index */
4705 mc->mc_ki[mc->mc_top] = i;
4707 /* There is no entry larger or equal to the key. */
4710 /* nodeptr is fake for LEAF2 */
4716 mdb_cursor_adjust(MDB_cursor *mc, func)
4720 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4721 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4728 /** Pop a page off the top of the cursor's stack. */
4730 mdb_cursor_pop(MDB_cursor *mc)
4734 MDB_page *top = mc->mc_pg[mc->mc_top];
4740 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4741 DDBI(mc), (void *) mc));
4745 /** Push a page onto the top of the cursor's stack. */
4747 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4749 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4750 DDBI(mc), (void *) mc));
4752 if (mc->mc_snum >= CURSOR_STACK) {
4753 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4754 return MDB_CURSOR_FULL;
4757 mc->mc_top = mc->mc_snum++;
4758 mc->mc_pg[mc->mc_top] = mp;
4759 mc->mc_ki[mc->mc_top] = 0;
4764 /** Find the address of the page corresponding to a given page number.
4765 * @param[in] txn the transaction for this access.
4766 * @param[in] pgno the page number for the page to retrieve.
4767 * @param[out] ret address of a pointer where the page's address will be stored.
4768 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4769 * @return 0 on success, non-zero on failure.
4772 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4774 MDB_env *env = txn->mt_env;
4778 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4782 MDB_ID2L dl = tx2->mt_u.dirty_list;
4784 /* Spilled pages were dirtied in this txn and flushed
4785 * because the dirty list got full. Bring this page
4786 * back in from the map (but don't unspill it here,
4787 * leave that unless page_touch happens again).
4789 if (tx2->mt_spill_pgs) {
4790 MDB_ID pn = pgno << 1;
4791 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4792 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4793 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4798 unsigned x = mdb_mid2l_search(dl, pgno);
4799 if (x <= dl[0].mid && dl[x].mid == pgno) {
4805 } while ((tx2 = tx2->mt_parent) != NULL);
4808 if (pgno < txn->mt_next_pgno) {
4810 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4812 DPRINTF(("page %"Z"u not found", pgno));
4813 txn->mt_flags |= MDB_TXN_ERROR;
4814 return MDB_PAGE_NOTFOUND;
4824 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4825 * The cursor is at the root page, set up the rest of it.
4828 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4830 MDB_page *mp = mc->mc_pg[mc->mc_top];
4834 while (IS_BRANCH(mp)) {
4838 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4839 mdb_cassert(mc, NUMKEYS(mp) > 1);
4840 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4842 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4844 if (flags & MDB_PS_LAST)
4845 i = NUMKEYS(mp) - 1;
4848 node = mdb_node_search(mc, key, &exact);
4850 i = NUMKEYS(mp) - 1;
4852 i = mc->mc_ki[mc->mc_top];
4854 mdb_cassert(mc, i > 0);
4858 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4861 mdb_cassert(mc, i < NUMKEYS(mp));
4862 node = NODEPTR(mp, i);
4864 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4867 mc->mc_ki[mc->mc_top] = i;
4868 if ((rc = mdb_cursor_push(mc, mp)))
4871 if (flags & MDB_PS_MODIFY) {
4872 if ((rc = mdb_page_touch(mc)) != 0)
4874 mp = mc->mc_pg[mc->mc_top];
4879 DPRINTF(("internal error, index points to a %02X page!?",
4881 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4882 return MDB_CORRUPTED;
4885 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4886 key ? DKEY(key) : "null"));
4887 mc->mc_flags |= C_INITIALIZED;
4888 mc->mc_flags &= ~C_EOF;
4893 /** Search for the lowest key under the current branch page.
4894 * This just bypasses a NUMKEYS check in the current page
4895 * before calling mdb_page_search_root(), because the callers
4896 * are all in situations where the current page is known to
4900 mdb_page_search_lowest(MDB_cursor *mc)
4902 MDB_page *mp = mc->mc_pg[mc->mc_top];
4903 MDB_node *node = NODEPTR(mp, 0);
4906 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4909 mc->mc_ki[mc->mc_top] = 0;
4910 if ((rc = mdb_cursor_push(mc, mp)))
4912 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4915 /** Search for the page a given key should be in.
4916 * Push it and its parent pages on the cursor stack.
4917 * @param[in,out] mc the cursor for this operation.
4918 * @param[in] key the key to search for, or NULL for first/last page.
4919 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4920 * are touched (updated with new page numbers).
4921 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4922 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4923 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4924 * @return 0 on success, non-zero on failure.
4927 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4932 /* Make sure the txn is still viable, then find the root from
4933 * the txn's db table and set it as the root of the cursor's stack.
4935 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4936 DPUTS("transaction has failed, must abort");
4939 /* Make sure we're using an up-to-date root */
4940 if (*mc->mc_dbflag & DB_STALE) {
4942 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4943 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4950 MDB_node *leaf = mdb_node_search(&mc2,
4951 &mc->mc_dbx->md_name, &exact);
4953 return MDB_NOTFOUND;
4954 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4957 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4959 /* The txn may not know this DBI, or another process may
4960 * have dropped and recreated the DB with other flags.
4962 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4963 return MDB_INCOMPATIBLE;
4964 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4966 *mc->mc_dbflag &= ~DB_STALE;
4968 root = mc->mc_db->md_root;
4970 if (root == P_INVALID) { /* Tree is empty. */
4971 DPUTS("tree is empty");
4972 return MDB_NOTFOUND;
4976 mdb_cassert(mc, root > 1);
4977 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4978 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4984 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4985 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4987 if (flags & MDB_PS_MODIFY) {
4988 if ((rc = mdb_page_touch(mc)))
4992 if (flags & MDB_PS_ROOTONLY)
4995 return mdb_page_search_root(mc, key, flags);
4999 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5001 MDB_txn *txn = mc->mc_txn;
5002 pgno_t pg = mp->mp_pgno;
5003 unsigned x = 0, ovpages = mp->mp_pages;
5004 MDB_env *env = txn->mt_env;
5005 MDB_IDL sl = txn->mt_spill_pgs;
5006 MDB_ID pn = pg << 1;
5009 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5010 /* If the page is dirty or on the spill list we just acquired it,
5011 * so we should give it back to our current free list, if any.
5012 * Otherwise put it onto the list of pages we freed in this txn.
5014 * Won't create me_pghead: me_pglast must be inited along with it.
5015 * Unsupported in nested txns: They would need to hide the page
5016 * range in ancestor txns' dirty and spilled lists.
5018 if (env->me_pghead &&
5020 ((mp->mp_flags & P_DIRTY) ||
5021 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5025 MDB_ID2 *dl, ix, iy;
5026 rc = mdb_midl_need(&env->me_pghead, ovpages);
5029 if (!(mp->mp_flags & P_DIRTY)) {
5030 /* This page is no longer spilled */
5037 /* Remove from dirty list */
5038 dl = txn->mt_u.dirty_list;
5040 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5046 mdb_cassert(mc, x > 1);
5048 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5049 txn->mt_flags |= MDB_TXN_ERROR;
5050 return MDB_CORRUPTED;
5053 if (!(env->me_flags & MDB_WRITEMAP))
5054 mdb_dpage_free(env, mp);
5056 /* Insert in me_pghead */
5057 mop = env->me_pghead;
5058 j = mop[0] + ovpages;
5059 for (i = mop[0]; i && mop[i] < pg; i--)
5065 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5069 mc->mc_db->md_overflow_pages -= ovpages;
5073 /** Return the data associated with a given node.
5074 * @param[in] txn The transaction for this operation.
5075 * @param[in] leaf The node being read.
5076 * @param[out] data Updated to point to the node's data.
5077 * @return 0 on success, non-zero on failure.
5080 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5082 MDB_page *omp; /* overflow page */
5086 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5087 data->mv_size = NODEDSZ(leaf);
5088 data->mv_data = NODEDATA(leaf);
5092 /* Read overflow data.
5094 data->mv_size = NODEDSZ(leaf);
5095 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5096 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5097 DPRINTF(("read overflow page %"Z"u failed", pgno));
5100 data->mv_data = METADATA(omp);
5106 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5107 MDB_val *key, MDB_val *data)
5114 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5116 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5119 if (txn->mt_flags & MDB_TXN_ERROR)
5122 mdb_cursor_init(&mc, txn, dbi, &mx);
5123 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5126 /** Find a sibling for a page.
5127 * Replaces the page at the top of the cursor's stack with the
5128 * specified sibling, if one exists.
5129 * @param[in] mc The cursor for this operation.
5130 * @param[in] move_right Non-zero if the right sibling is requested,
5131 * otherwise the left sibling.
5132 * @return 0 on success, non-zero on failure.
5135 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5141 if (mc->mc_snum < 2) {
5142 return MDB_NOTFOUND; /* root has no siblings */
5146 DPRINTF(("parent page is page %"Z"u, index %u",
5147 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5149 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5150 : (mc->mc_ki[mc->mc_top] == 0)) {
5151 DPRINTF(("no more keys left, moving to %s sibling",
5152 move_right ? "right" : "left"));
5153 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5154 /* undo cursor_pop before returning */
5161 mc->mc_ki[mc->mc_top]++;
5163 mc->mc_ki[mc->mc_top]--;
5164 DPRINTF(("just moving to %s index key %u",
5165 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5167 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5169 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5170 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5171 /* mc will be inconsistent if caller does mc_snum++ as above */
5172 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5176 mdb_cursor_push(mc, mp);
5178 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5183 /** Move the cursor to the next data item. */
5185 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5191 if (mc->mc_flags & C_EOF) {
5192 return MDB_NOTFOUND;
5195 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5197 mp = mc->mc_pg[mc->mc_top];
5199 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5200 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5201 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5202 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5203 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5204 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5205 if (rc == MDB_SUCCESS)
5206 MDB_GET_KEY(leaf, key);
5211 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5212 if (op == MDB_NEXT_DUP)
5213 return MDB_NOTFOUND;
5217 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5218 mdb_dbg_pgno(mp), (void *) mc));
5219 if (mc->mc_flags & C_DEL)
5222 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5223 DPUTS("=====> move to next sibling page");
5224 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5225 mc->mc_flags |= C_EOF;
5228 mp = mc->mc_pg[mc->mc_top];
5229 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5231 mc->mc_ki[mc->mc_top]++;
5234 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5235 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5238 key->mv_size = mc->mc_db->md_pad;
5239 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5243 mdb_cassert(mc, IS_LEAF(mp));
5244 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5246 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5247 mdb_xcursor_init1(mc, leaf);
5250 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5253 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5254 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5255 if (rc != MDB_SUCCESS)
5260 MDB_GET_KEY(leaf, key);
5264 /** Move the cursor to the previous data item. */
5266 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5272 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5274 mp = mc->mc_pg[mc->mc_top];
5276 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5277 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5278 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5279 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5280 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5281 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5282 if (rc == MDB_SUCCESS)
5283 MDB_GET_KEY(leaf, key);
5287 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5288 if (op == MDB_PREV_DUP)
5289 return MDB_NOTFOUND;
5294 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5295 mdb_dbg_pgno(mp), (void *) mc));
5297 if (mc->mc_ki[mc->mc_top] == 0) {
5298 DPUTS("=====> move to prev sibling page");
5299 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5302 mp = mc->mc_pg[mc->mc_top];
5303 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5304 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5306 mc->mc_ki[mc->mc_top]--;
5308 mc->mc_flags &= ~C_EOF;
5310 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5311 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5314 key->mv_size = mc->mc_db->md_pad;
5315 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5319 mdb_cassert(mc, IS_LEAF(mp));
5320 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5322 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5323 mdb_xcursor_init1(mc, leaf);
5326 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5329 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5330 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5331 if (rc != MDB_SUCCESS)
5336 MDB_GET_KEY(leaf, key);
5340 /** Set the cursor on a specific data item. */
5342 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5343 MDB_cursor_op op, int *exactp)
5347 MDB_node *leaf = NULL;
5350 if (key->mv_size == 0)
5351 return MDB_BAD_VALSIZE;
5354 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5356 /* See if we're already on the right page */
5357 if (mc->mc_flags & C_INITIALIZED) {
5360 mp = mc->mc_pg[mc->mc_top];
5362 mc->mc_ki[mc->mc_top] = 0;
5363 return MDB_NOTFOUND;
5365 if (mp->mp_flags & P_LEAF2) {
5366 nodekey.mv_size = mc->mc_db->md_pad;
5367 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5369 leaf = NODEPTR(mp, 0);
5370 MDB_GET_KEY2(leaf, nodekey);
5372 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5374 /* Probably happens rarely, but first node on the page
5375 * was the one we wanted.
5377 mc->mc_ki[mc->mc_top] = 0;
5384 unsigned int nkeys = NUMKEYS(mp);
5386 if (mp->mp_flags & P_LEAF2) {
5387 nodekey.mv_data = LEAF2KEY(mp,
5388 nkeys-1, nodekey.mv_size);
5390 leaf = NODEPTR(mp, nkeys-1);
5391 MDB_GET_KEY2(leaf, nodekey);
5393 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5395 /* last node was the one we wanted */
5396 mc->mc_ki[mc->mc_top] = nkeys-1;
5402 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5403 /* This is definitely the right page, skip search_page */
5404 if (mp->mp_flags & P_LEAF2) {
5405 nodekey.mv_data = LEAF2KEY(mp,
5406 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5408 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5409 MDB_GET_KEY2(leaf, nodekey);
5411 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5413 /* current node was the one we wanted */
5423 /* If any parents have right-sibs, search.
5424 * Otherwise, there's nothing further.
5426 for (i=0; i<mc->mc_top; i++)
5428 NUMKEYS(mc->mc_pg[i])-1)
5430 if (i == mc->mc_top) {
5431 /* There are no other pages */
5432 mc->mc_ki[mc->mc_top] = nkeys;
5433 return MDB_NOTFOUND;
5437 /* There are no other pages */
5438 mc->mc_ki[mc->mc_top] = 0;
5439 if (op == MDB_SET_RANGE && !exactp) {
5443 return MDB_NOTFOUND;
5447 rc = mdb_page_search(mc, key, 0);
5448 if (rc != MDB_SUCCESS)
5451 mp = mc->mc_pg[mc->mc_top];
5452 mdb_cassert(mc, IS_LEAF(mp));
5455 leaf = mdb_node_search(mc, key, exactp);
5456 if (exactp != NULL && !*exactp) {
5457 /* MDB_SET specified and not an exact match. */
5458 return MDB_NOTFOUND;
5462 DPUTS("===> inexact leaf not found, goto sibling");
5463 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5464 return rc; /* no entries matched */
5465 mp = mc->mc_pg[mc->mc_top];
5466 mdb_cassert(mc, IS_LEAF(mp));
5467 leaf = NODEPTR(mp, 0);
5471 mc->mc_flags |= C_INITIALIZED;
5472 mc->mc_flags &= ~C_EOF;
5475 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5476 key->mv_size = mc->mc_db->md_pad;
5477 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5482 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5483 mdb_xcursor_init1(mc, leaf);
5486 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5487 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5488 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5491 if (op == MDB_GET_BOTH) {
5497 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5498 if (rc != MDB_SUCCESS)
5501 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5503 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5505 rc = mc->mc_dbx->md_dcmp(data, &d2);
5507 if (op == MDB_GET_BOTH || rc > 0)
5508 return MDB_NOTFOUND;
5515 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5516 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5521 /* The key already matches in all other cases */
5522 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5523 MDB_GET_KEY(leaf, key);
5524 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5529 /** Move the cursor to the first item in the database. */
5531 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5537 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5539 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5540 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5541 if (rc != MDB_SUCCESS)
5544 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5546 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5547 mc->mc_flags |= C_INITIALIZED;
5548 mc->mc_flags &= ~C_EOF;
5550 mc->mc_ki[mc->mc_top] = 0;
5552 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5553 key->mv_size = mc->mc_db->md_pad;
5554 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5559 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5560 mdb_xcursor_init1(mc, leaf);
5561 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5565 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5569 MDB_GET_KEY(leaf, key);
5573 /** Move the cursor to the last item in the database. */
5575 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5581 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5583 if (!(mc->mc_flags & C_EOF)) {
5585 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5586 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5587 if (rc != MDB_SUCCESS)
5590 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5593 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5594 mc->mc_flags |= C_INITIALIZED|C_EOF;
5595 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5597 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5598 key->mv_size = mc->mc_db->md_pad;
5599 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5604 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5605 mdb_xcursor_init1(mc, leaf);
5606 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5610 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5615 MDB_GET_KEY(leaf, key);
5620 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5625 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5630 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5634 case MDB_GET_CURRENT:
5635 if (!(mc->mc_flags & C_INITIALIZED)) {
5638 MDB_page *mp = mc->mc_pg[mc->mc_top];
5639 int nkeys = NUMKEYS(mp);
5640 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5641 mc->mc_ki[mc->mc_top] = nkeys;
5647 key->mv_size = mc->mc_db->md_pad;
5648 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5650 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5651 MDB_GET_KEY(leaf, key);
5653 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5654 if (mc->mc_flags & C_DEL)
5655 mdb_xcursor_init1(mc, leaf);
5656 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5658 rc = mdb_node_read(mc->mc_txn, leaf, data);
5665 case MDB_GET_BOTH_RANGE:
5670 if (mc->mc_xcursor == NULL) {
5671 rc = MDB_INCOMPATIBLE;
5681 rc = mdb_cursor_set(mc, key, data, op,
5682 op == MDB_SET_RANGE ? NULL : &exact);
5685 case MDB_GET_MULTIPLE:
5686 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5690 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5691 rc = MDB_INCOMPATIBLE;
5695 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5696 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5699 case MDB_NEXT_MULTIPLE:
5704 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5705 rc = MDB_INCOMPATIBLE;
5708 if (!(mc->mc_flags & C_INITIALIZED))
5709 rc = mdb_cursor_first(mc, key, data);
5711 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5712 if (rc == MDB_SUCCESS) {
5713 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5716 mx = &mc->mc_xcursor->mx_cursor;
5717 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5719 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5720 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5728 case MDB_NEXT_NODUP:
5729 if (!(mc->mc_flags & C_INITIALIZED))
5730 rc = mdb_cursor_first(mc, key, data);
5732 rc = mdb_cursor_next(mc, key, data, op);
5736 case MDB_PREV_NODUP:
5737 if (!(mc->mc_flags & C_INITIALIZED)) {
5738 rc = mdb_cursor_last(mc, key, data);
5741 mc->mc_flags |= C_INITIALIZED;
5742 mc->mc_ki[mc->mc_top]++;
5744 rc = mdb_cursor_prev(mc, key, data, op);
5747 rc = mdb_cursor_first(mc, key, data);
5750 mfunc = mdb_cursor_first;
5752 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5756 if (mc->mc_xcursor == NULL) {
5757 rc = MDB_INCOMPATIBLE;
5760 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5764 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5767 rc = mdb_cursor_last(mc, key, data);
5770 mfunc = mdb_cursor_last;
5773 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5778 if (mc->mc_flags & C_DEL)
5779 mc->mc_flags ^= C_DEL;
5784 /** Touch all the pages in the cursor stack. Set mc_top.
5785 * Makes sure all the pages are writable, before attempting a write operation.
5786 * @param[in] mc The cursor to operate on.
5789 mdb_cursor_touch(MDB_cursor *mc)
5791 int rc = MDB_SUCCESS;
5793 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5796 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5797 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5800 *mc->mc_dbflag |= DB_DIRTY;
5805 rc = mdb_page_touch(mc);
5806 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5807 mc->mc_top = mc->mc_snum-1;
5812 /** Do not spill pages to disk if txn is getting full, may fail instead */
5813 #define MDB_NOSPILL 0x8000
5816 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5819 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5821 MDB_node *leaf = NULL;
5824 MDB_val xdata, *rdata, dkey, olddata;
5826 int do_sub = 0, insert_key, insert_data;
5827 unsigned int mcount = 0, dcount = 0, nospill;
5830 unsigned int nflags;
5833 if (mc == NULL || key == NULL)
5836 env = mc->mc_txn->mt_env;
5838 /* Check this first so counter will always be zero on any
5841 if (flags & MDB_MULTIPLE) {
5842 dcount = data[1].mv_size;
5843 data[1].mv_size = 0;
5844 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5845 return MDB_INCOMPATIBLE;
5848 nospill = flags & MDB_NOSPILL;
5849 flags &= ~MDB_NOSPILL;
5851 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5852 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5854 if (key->mv_size-1 >= ENV_MAXKEY(env))
5855 return MDB_BAD_VALSIZE;
5857 #if SIZE_MAX > MAXDATASIZE
5858 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5859 return MDB_BAD_VALSIZE;
5861 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5862 return MDB_BAD_VALSIZE;
5865 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5866 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5870 if (flags == MDB_CURRENT) {
5871 if (!(mc->mc_flags & C_INITIALIZED))
5874 } else if (mc->mc_db->md_root == P_INVALID) {
5875 /* new database, cursor has nothing to point to */
5878 mc->mc_flags &= ~C_INITIALIZED;
5883 if (flags & MDB_APPEND) {
5885 rc = mdb_cursor_last(mc, &k2, &d2);
5887 rc = mc->mc_dbx->md_cmp(key, &k2);
5890 mc->mc_ki[mc->mc_top]++;
5892 /* new key is <= last key */
5897 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5899 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5900 DPRINTF(("duplicate key [%s]", DKEY(key)));
5902 return MDB_KEYEXIST;
5904 if (rc && rc != MDB_NOTFOUND)
5908 if (mc->mc_flags & C_DEL)
5909 mc->mc_flags ^= C_DEL;
5911 /* Cursor is positioned, check for room in the dirty list */
5913 if (flags & MDB_MULTIPLE) {
5915 xdata.mv_size = data->mv_size * dcount;
5919 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5923 if (rc == MDB_NO_ROOT) {
5925 /* new database, write a root leaf page */
5926 DPUTS("allocating new root leaf page");
5927 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5930 mdb_cursor_push(mc, np);
5931 mc->mc_db->md_root = np->mp_pgno;
5932 mc->mc_db->md_depth++;
5933 *mc->mc_dbflag |= DB_DIRTY;
5934 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5936 np->mp_flags |= P_LEAF2;
5937 mc->mc_flags |= C_INITIALIZED;
5939 /* make sure all cursor pages are writable */
5940 rc2 = mdb_cursor_touch(mc);
5945 insert_key = insert_data = rc;
5947 /* The key does not exist */
5948 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5949 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5950 LEAFSIZE(key, data) > env->me_nodemax)
5952 /* Too big for a node, insert in sub-DB. Set up an empty
5953 * "old sub-page" for prep_subDB to expand to a full page.
5955 fp_flags = P_LEAF|P_DIRTY;
5957 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5958 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
5959 olddata.mv_size = PAGEHDRSZ;
5963 /* there's only a key anyway, so this is a no-op */
5964 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5966 unsigned int ksize = mc->mc_db->md_pad;
5967 if (key->mv_size != ksize)
5968 return MDB_BAD_VALSIZE;
5969 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5970 memcpy(ptr, key->mv_data, ksize);
5975 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5976 olddata.mv_size = NODEDSZ(leaf);
5977 olddata.mv_data = NODEDATA(leaf);
5980 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5981 /* Prepare (sub-)page/sub-DB to accept the new item,
5982 * if needed. fp: old sub-page or a header faking
5983 * it. mp: new (sub-)page. offset: growth in page
5984 * size. xdata: node data with new page or DB.
5986 unsigned i, offset = 0;
5987 mp = fp = xdata.mv_data = env->me_pbuf;
5988 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5990 /* Was a single item before, must convert now */
5991 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5992 /* Just overwrite the current item */
5993 if (flags == MDB_CURRENT)
5996 #if UINT_MAX < SIZE_MAX
5997 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5998 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6000 /* does data match? */
6001 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6002 if (flags & MDB_NODUPDATA)
6003 return MDB_KEYEXIST;
6008 /* Back up original data item */
6009 dkey.mv_size = olddata.mv_size;
6010 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6012 /* Make sub-page header for the dup items, with dummy body */
6013 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6014 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6015 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6016 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6017 fp->mp_flags |= P_LEAF2;
6018 fp->mp_pad = data->mv_size;
6019 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6021 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6022 (dkey.mv_size & 1) + (data->mv_size & 1);
6024 fp->mp_upper = xdata.mv_size - PAGEBASE;
6025 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6026 } else if (leaf->mn_flags & F_SUBDATA) {
6027 /* Data is on sub-DB, just store it */
6028 flags |= F_DUPDATA|F_SUBDATA;
6031 /* Data is on sub-page */
6032 fp = olddata.mv_data;
6035 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6036 offset = EVEN(NODESIZE + sizeof(indx_t) +
6040 offset = fp->mp_pad;
6041 if (SIZELEFT(fp) < offset) {
6042 offset *= 4; /* space for 4 more */
6045 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6047 fp->mp_flags |= P_DIRTY;
6048 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6049 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6053 xdata.mv_size = olddata.mv_size + offset;
6056 fp_flags = fp->mp_flags;
6057 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6058 /* Too big for a sub-page, convert to sub-DB */
6059 fp_flags &= ~P_SUBP;
6061 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6062 fp_flags |= P_LEAF2;
6063 dummy.md_pad = fp->mp_pad;
6064 dummy.md_flags = MDB_DUPFIXED;
6065 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6066 dummy.md_flags |= MDB_INTEGERKEY;
6072 dummy.md_branch_pages = 0;
6073 dummy.md_leaf_pages = 1;
6074 dummy.md_overflow_pages = 0;
6075 dummy.md_entries = NUMKEYS(fp);
6076 xdata.mv_size = sizeof(MDB_db);
6077 xdata.mv_data = &dummy;
6078 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6080 offset = env->me_psize - olddata.mv_size;
6081 flags |= F_DUPDATA|F_SUBDATA;
6082 dummy.md_root = mp->mp_pgno;
6085 mp->mp_flags = fp_flags | P_DIRTY;
6086 mp->mp_pad = fp->mp_pad;
6087 mp->mp_lower = fp->mp_lower;
6088 mp->mp_upper = fp->mp_upper + offset;
6089 if (fp_flags & P_LEAF2) {
6090 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6092 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6093 olddata.mv_size - fp->mp_upper - PAGEBASE);
6094 for (i=0; i<NUMKEYS(fp); i++)
6095 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6103 mdb_node_del(mc, 0);
6107 /* overflow page overwrites need special handling */
6108 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6111 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6113 memcpy(&pg, olddata.mv_data, sizeof(pg));
6114 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6116 ovpages = omp->mp_pages;
6118 /* Is the ov page large enough? */
6119 if (ovpages >= dpages) {
6120 if (!(omp->mp_flags & P_DIRTY) &&
6121 (level || (env->me_flags & MDB_WRITEMAP)))
6123 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6126 level = 0; /* dirty in this txn or clean */
6129 if (omp->mp_flags & P_DIRTY) {
6130 /* yes, overwrite it. Note in this case we don't
6131 * bother to try shrinking the page if the new data
6132 * is smaller than the overflow threshold.
6135 /* It is writable only in a parent txn */
6136 size_t sz = (size_t) env->me_psize * ovpages, off;
6137 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6143 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6144 mdb_cassert(mc, rc2 == 0);
6145 if (!(flags & MDB_RESERVE)) {
6146 /* Copy end of page, adjusting alignment so
6147 * compiler may copy words instead of bytes.
6149 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6150 memcpy((size_t *)((char *)np + off),
6151 (size_t *)((char *)omp + off), sz - off);
6154 memcpy(np, omp, sz); /* Copy beginning of page */
6157 SETDSZ(leaf, data->mv_size);
6158 if (F_ISSET(flags, MDB_RESERVE))
6159 data->mv_data = METADATA(omp);
6161 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6165 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6167 } else if (data->mv_size == olddata.mv_size) {
6168 /* same size, just replace it. Note that we could
6169 * also reuse this node if the new data is smaller,
6170 * but instead we opt to shrink the node in that case.
6172 if (F_ISSET(flags, MDB_RESERVE))
6173 data->mv_data = olddata.mv_data;
6174 else if (!(mc->mc_flags & C_SUB))
6175 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6177 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6180 mdb_node_del(mc, 0);
6186 nflags = flags & NODE_ADD_FLAGS;
6187 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6188 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6189 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6190 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6192 nflags |= MDB_SPLIT_REPLACE;
6193 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6195 /* There is room already in this leaf page. */
6196 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6197 if (rc == 0 && insert_key) {
6198 /* Adjust other cursors pointing to mp */
6199 MDB_cursor *m2, *m3;
6200 MDB_dbi dbi = mc->mc_dbi;
6201 unsigned i = mc->mc_top;
6202 MDB_page *mp = mc->mc_pg[i];
6204 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6205 if (mc->mc_flags & C_SUB)
6206 m3 = &m2->mc_xcursor->mx_cursor;
6209 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6210 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6217 if (rc == MDB_SUCCESS) {
6218 /* Now store the actual data in the child DB. Note that we're
6219 * storing the user data in the keys field, so there are strict
6220 * size limits on dupdata. The actual data fields of the child
6221 * DB are all zero size.
6229 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6230 if (flags & MDB_CURRENT) {
6231 xflags = MDB_CURRENT|MDB_NOSPILL;
6233 mdb_xcursor_init1(mc, leaf);
6234 xflags = (flags & MDB_NODUPDATA) ?
6235 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6237 /* converted, write the original data first */
6239 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6243 /* Adjust other cursors pointing to mp */
6245 unsigned i = mc->mc_top;
6246 MDB_page *mp = mc->mc_pg[i];
6248 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6249 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6250 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6251 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6252 mdb_xcursor_init1(m2, leaf);
6256 /* we've done our job */
6259 ecount = mc->mc_xcursor->mx_db.md_entries;
6260 if (flags & MDB_APPENDDUP)
6261 xflags |= MDB_APPEND;
6262 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6263 if (flags & F_SUBDATA) {
6264 void *db = NODEDATA(leaf);
6265 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6267 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6269 /* Increment count unless we just replaced an existing item. */
6271 mc->mc_db->md_entries++;
6273 /* Invalidate txn if we created an empty sub-DB */
6276 /* If we succeeded and the key didn't exist before,
6277 * make sure the cursor is marked valid.
6279 mc->mc_flags |= C_INITIALIZED;
6281 if (flags & MDB_MULTIPLE) {
6284 /* let caller know how many succeeded, if any */
6285 data[1].mv_size = mcount;
6286 if (mcount < dcount) {
6287 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6288 insert_key = insert_data = 0;
6295 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6298 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6303 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6309 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6310 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6312 if (!(mc->mc_flags & C_INITIALIZED))
6315 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6316 return MDB_NOTFOUND;
6318 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6321 rc = mdb_cursor_touch(mc);
6325 mp = mc->mc_pg[mc->mc_top];
6328 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6330 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6331 if (flags & MDB_NODUPDATA) {
6332 /* mdb_cursor_del0() will subtract the final entry */
6333 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6335 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6336 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6338 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6341 /* If sub-DB still has entries, we're done */
6342 if (mc->mc_xcursor->mx_db.md_entries) {
6343 if (leaf->mn_flags & F_SUBDATA) {
6344 /* update subDB info */
6345 void *db = NODEDATA(leaf);
6346 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6349 /* shrink fake page */
6350 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6351 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6352 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6353 /* fix other sub-DB cursors pointed at this fake page */
6354 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6355 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6356 if (m2->mc_pg[mc->mc_top] == mp &&
6357 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6358 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6361 mc->mc_db->md_entries--;
6362 mc->mc_flags |= C_DEL;
6365 /* otherwise fall thru and delete the sub-DB */
6368 if (leaf->mn_flags & F_SUBDATA) {
6369 /* add all the child DB's pages to the free list */
6370 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6376 /* add overflow pages to free list */
6377 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6381 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6382 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6383 (rc = mdb_ovpage_free(mc, omp)))
6388 return mdb_cursor_del0(mc);
6391 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6395 /** Allocate and initialize new pages for a database.
6396 * @param[in] mc a cursor on the database being added to.
6397 * @param[in] flags flags defining what type of page is being allocated.
6398 * @param[in] num the number of pages to allocate. This is usually 1,
6399 * unless allocating overflow pages for a large record.
6400 * @param[out] mp Address of a page, or NULL on failure.
6401 * @return 0 on success, non-zero on failure.
6404 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6409 if ((rc = mdb_page_alloc(mc, num, &np)))
6411 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6412 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6413 np->mp_flags = flags | P_DIRTY;
6414 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6415 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6418 mc->mc_db->md_branch_pages++;
6419 else if (IS_LEAF(np))
6420 mc->mc_db->md_leaf_pages++;
6421 else if (IS_OVERFLOW(np)) {
6422 mc->mc_db->md_overflow_pages += num;
6430 /** Calculate the size of a leaf node.
6431 * The size depends on the environment's page size; if a data item
6432 * is too large it will be put onto an overflow page and the node
6433 * size will only include the key and not the data. Sizes are always
6434 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6435 * of the #MDB_node headers.
6436 * @param[in] env The environment handle.
6437 * @param[in] key The key for the node.
6438 * @param[in] data The data for the node.
6439 * @return The number of bytes needed to store the node.
6442 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6446 sz = LEAFSIZE(key, data);
6447 if (sz > env->me_nodemax) {
6448 /* put on overflow page */
6449 sz -= data->mv_size - sizeof(pgno_t);
6452 return EVEN(sz + sizeof(indx_t));
6455 /** Calculate the size of a branch node.
6456 * The size should depend on the environment's page size but since
6457 * we currently don't support spilling large keys onto overflow
6458 * pages, it's simply the size of the #MDB_node header plus the
6459 * size of the key. Sizes are always rounded up to an even number
6460 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6461 * @param[in] env The environment handle.
6462 * @param[in] key The key for the node.
6463 * @return The number of bytes needed to store the node.
6466 mdb_branch_size(MDB_env *env, MDB_val *key)
6471 if (sz > env->me_nodemax) {
6472 /* put on overflow page */
6473 /* not implemented */
6474 /* sz -= key->size - sizeof(pgno_t); */
6477 return sz + sizeof(indx_t);
6480 /** Add a node to the page pointed to by the cursor.
6481 * @param[in] mc The cursor for this operation.
6482 * @param[in] indx The index on the page where the new node should be added.
6483 * @param[in] key The key for the new node.
6484 * @param[in] data The data for the new node, if any.
6485 * @param[in] pgno The page number, if adding a branch node.
6486 * @param[in] flags Flags for the node.
6487 * @return 0 on success, non-zero on failure. Possible errors are:
6489 * <li>ENOMEM - failed to allocate overflow pages for the node.
6490 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6491 * should never happen since all callers already calculate the
6492 * page's free space before calling this function.
6496 mdb_node_add(MDB_cursor *mc, indx_t indx,
6497 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6500 size_t node_size = NODESIZE;
6504 MDB_page *mp = mc->mc_pg[mc->mc_top];
6505 MDB_page *ofp = NULL; /* overflow page */
6508 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6510 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6511 IS_LEAF(mp) ? "leaf" : "branch",
6512 IS_SUBP(mp) ? "sub-" : "",
6513 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6514 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6517 /* Move higher keys up one slot. */
6518 int ksize = mc->mc_db->md_pad, dif;
6519 char *ptr = LEAF2KEY(mp, indx, ksize);
6520 dif = NUMKEYS(mp) - indx;
6522 memmove(ptr+ksize, ptr, dif*ksize);
6523 /* insert new key */
6524 memcpy(ptr, key->mv_data, ksize);
6526 /* Just using these for counting */
6527 mp->mp_lower += sizeof(indx_t);
6528 mp->mp_upper -= ksize - sizeof(indx_t);
6532 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6534 node_size += key->mv_size;
6536 mdb_cassert(mc, data);
6537 if (F_ISSET(flags, F_BIGDATA)) {
6538 /* Data already on overflow page. */
6539 node_size += sizeof(pgno_t);
6540 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6541 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6543 /* Put data on overflow page. */
6544 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6545 data->mv_size, node_size+data->mv_size));
6546 node_size = EVEN(node_size + sizeof(pgno_t));
6547 if ((ssize_t)node_size > room)
6549 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6551 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6555 node_size += data->mv_size;
6558 node_size = EVEN(node_size);
6559 if ((ssize_t)node_size > room)
6563 /* Move higher pointers up one slot. */
6564 for (i = NUMKEYS(mp); i > indx; i--)
6565 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6567 /* Adjust free space offsets. */
6568 ofs = mp->mp_upper - node_size;
6569 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6570 mp->mp_ptrs[indx] = ofs;
6572 mp->mp_lower += sizeof(indx_t);
6574 /* Write the node data. */
6575 node = NODEPTR(mp, indx);
6576 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6577 node->mn_flags = flags;
6579 SETDSZ(node,data->mv_size);
6584 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6587 mdb_cassert(mc, key);
6589 if (F_ISSET(flags, F_BIGDATA))
6590 memcpy(node->mn_data + key->mv_size, data->mv_data,
6592 else if (F_ISSET(flags, MDB_RESERVE))
6593 data->mv_data = node->mn_data + key->mv_size;
6595 memcpy(node->mn_data + key->mv_size, data->mv_data,
6598 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6600 if (F_ISSET(flags, MDB_RESERVE))
6601 data->mv_data = METADATA(ofp);
6603 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6610 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6611 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6612 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6613 DPRINTF(("node size = %"Z"u", node_size));
6614 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6615 return MDB_PAGE_FULL;
6618 /** Delete the specified node from a page.
6619 * @param[in] mc Cursor pointing to the node to delete.
6620 * @param[in] ksize The size of a node. Only used if the page is
6621 * part of a #MDB_DUPFIXED database.
6624 mdb_node_del(MDB_cursor *mc, int ksize)
6626 MDB_page *mp = mc->mc_pg[mc->mc_top];
6627 indx_t indx = mc->mc_ki[mc->mc_top];
6629 indx_t i, j, numkeys, ptr;
6633 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6634 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6635 numkeys = NUMKEYS(mp);
6636 mdb_cassert(mc, indx < numkeys);
6639 int x = numkeys - 1 - indx;
6640 base = LEAF2KEY(mp, indx, ksize);
6642 memmove(base, base + ksize, x * ksize);
6643 mp->mp_lower -= sizeof(indx_t);
6644 mp->mp_upper += ksize - sizeof(indx_t);
6648 node = NODEPTR(mp, indx);
6649 sz = NODESIZE + node->mn_ksize;
6651 if (F_ISSET(node->mn_flags, F_BIGDATA))
6652 sz += sizeof(pgno_t);
6654 sz += NODEDSZ(node);
6658 ptr = mp->mp_ptrs[indx];
6659 for (i = j = 0; i < numkeys; i++) {
6661 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6662 if (mp->mp_ptrs[i] < ptr)
6663 mp->mp_ptrs[j] += sz;
6668 base = (char *)mp + mp->mp_upper + PAGEBASE;
6669 memmove(base + sz, base, ptr - mp->mp_upper);
6671 mp->mp_lower -= sizeof(indx_t);
6675 /** Compact the main page after deleting a node on a subpage.
6676 * @param[in] mp The main page to operate on.
6677 * @param[in] indx The index of the subpage on the main page.
6680 mdb_node_shrink(MDB_page *mp, indx_t indx)
6686 indx_t i, numkeys, ptr;
6688 node = NODEPTR(mp, indx);
6689 sp = (MDB_page *)NODEDATA(node);
6690 delta = SIZELEFT(sp);
6691 xp = (MDB_page *)((char *)sp + delta);
6693 /* shift subpage upward */
6695 nsize = NUMKEYS(sp) * sp->mp_pad;
6697 return; /* do not make the node uneven-sized */
6698 memmove(METADATA(xp), METADATA(sp), nsize);
6701 numkeys = NUMKEYS(sp);
6702 for (i=numkeys-1; i>=0; i--)
6703 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6705 xp->mp_upper = sp->mp_lower;
6706 xp->mp_lower = sp->mp_lower;
6707 xp->mp_flags = sp->mp_flags;
6708 xp->mp_pad = sp->mp_pad;
6709 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6711 nsize = NODEDSZ(node) - delta;
6712 SETDSZ(node, nsize);
6714 /* shift lower nodes upward */
6715 ptr = mp->mp_ptrs[indx];
6716 numkeys = NUMKEYS(mp);
6717 for (i = 0; i < numkeys; i++) {
6718 if (mp->mp_ptrs[i] <= ptr)
6719 mp->mp_ptrs[i] += delta;
6722 base = (char *)mp + mp->mp_upper + PAGEBASE;
6723 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6724 mp->mp_upper += delta;
6727 /** Initial setup of a sorted-dups cursor.
6728 * Sorted duplicates are implemented as a sub-database for the given key.
6729 * The duplicate data items are actually keys of the sub-database.
6730 * Operations on the duplicate data items are performed using a sub-cursor
6731 * initialized when the sub-database is first accessed. This function does
6732 * the preliminary setup of the sub-cursor, filling in the fields that
6733 * depend only on the parent DB.
6734 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6737 mdb_xcursor_init0(MDB_cursor *mc)
6739 MDB_xcursor *mx = mc->mc_xcursor;
6741 mx->mx_cursor.mc_xcursor = NULL;
6742 mx->mx_cursor.mc_txn = mc->mc_txn;
6743 mx->mx_cursor.mc_db = &mx->mx_db;
6744 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6745 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6746 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6747 mx->mx_cursor.mc_snum = 0;
6748 mx->mx_cursor.mc_top = 0;
6749 mx->mx_cursor.mc_flags = C_SUB;
6750 mx->mx_dbx.md_name.mv_size = 0;
6751 mx->mx_dbx.md_name.mv_data = NULL;
6752 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6753 mx->mx_dbx.md_dcmp = NULL;
6754 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6757 /** Final setup of a sorted-dups cursor.
6758 * Sets up the fields that depend on the data from the main cursor.
6759 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6760 * @param[in] node The data containing the #MDB_db record for the
6761 * sorted-dup database.
6764 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6766 MDB_xcursor *mx = mc->mc_xcursor;
6768 if (node->mn_flags & F_SUBDATA) {
6769 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6770 mx->mx_cursor.mc_pg[0] = 0;
6771 mx->mx_cursor.mc_snum = 0;
6772 mx->mx_cursor.mc_top = 0;
6773 mx->mx_cursor.mc_flags = C_SUB;
6775 MDB_page *fp = NODEDATA(node);
6776 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6777 mx->mx_db.md_flags = 0;
6778 mx->mx_db.md_depth = 1;
6779 mx->mx_db.md_branch_pages = 0;
6780 mx->mx_db.md_leaf_pages = 1;
6781 mx->mx_db.md_overflow_pages = 0;
6782 mx->mx_db.md_entries = NUMKEYS(fp);
6783 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6784 mx->mx_cursor.mc_snum = 1;
6785 mx->mx_cursor.mc_top = 0;
6786 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6787 mx->mx_cursor.mc_pg[0] = fp;
6788 mx->mx_cursor.mc_ki[0] = 0;
6789 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6790 mx->mx_db.md_flags = MDB_DUPFIXED;
6791 mx->mx_db.md_pad = fp->mp_pad;
6792 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6793 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6796 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6797 mx->mx_db.md_root));
6798 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6799 #if UINT_MAX < SIZE_MAX
6800 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6801 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6805 /** Initialize a cursor for a given transaction and database. */
6807 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6810 mc->mc_backup = NULL;
6813 mc->mc_db = &txn->mt_dbs[dbi];
6814 mc->mc_dbx = &txn->mt_dbxs[dbi];
6815 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6820 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6821 mdb_tassert(txn, mx != NULL);
6822 mc->mc_xcursor = mx;
6823 mdb_xcursor_init0(mc);
6825 mc->mc_xcursor = NULL;
6827 if (*mc->mc_dbflag & DB_STALE) {
6828 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6833 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6836 size_t size = sizeof(MDB_cursor);
6838 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6841 if (txn->mt_flags & MDB_TXN_ERROR)
6844 /* Allow read access to the freelist */
6845 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6848 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6849 size += sizeof(MDB_xcursor);
6851 if ((mc = malloc(size)) != NULL) {
6852 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6853 if (txn->mt_cursors) {
6854 mc->mc_next = txn->mt_cursors[dbi];
6855 txn->mt_cursors[dbi] = mc;
6856 mc->mc_flags |= C_UNTRACK;
6868 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6870 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6873 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6876 if (txn->mt_flags & MDB_TXN_ERROR)
6879 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6883 /* Return the count of duplicate data items for the current key */
6885 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6889 if (mc == NULL || countp == NULL)
6892 if (mc->mc_xcursor == NULL)
6893 return MDB_INCOMPATIBLE;
6895 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6898 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6899 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6902 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6905 *countp = mc->mc_xcursor->mx_db.md_entries;
6911 mdb_cursor_close(MDB_cursor *mc)
6913 if (mc && !mc->mc_backup) {
6914 /* remove from txn, if tracked */
6915 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6916 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6917 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6919 *prev = mc->mc_next;
6926 mdb_cursor_txn(MDB_cursor *mc)
6928 if (!mc) return NULL;
6933 mdb_cursor_dbi(MDB_cursor *mc)
6938 /** Replace the key for a branch node with a new key.
6939 * @param[in] mc Cursor pointing to the node to operate on.
6940 * @param[in] key The new key to use.
6941 * @return 0 on success, non-zero on failure.
6944 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6950 int delta, ksize, oksize;
6951 indx_t ptr, i, numkeys, indx;
6954 indx = mc->mc_ki[mc->mc_top];
6955 mp = mc->mc_pg[mc->mc_top];
6956 node = NODEPTR(mp, indx);
6957 ptr = mp->mp_ptrs[indx];
6961 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6962 k2.mv_data = NODEKEY(node);
6963 k2.mv_size = node->mn_ksize;
6964 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6966 mdb_dkey(&k2, kbuf2),
6972 /* Sizes must be 2-byte aligned. */
6973 ksize = EVEN(key->mv_size);
6974 oksize = EVEN(node->mn_ksize);
6975 delta = ksize - oksize;
6977 /* Shift node contents if EVEN(key length) changed. */
6979 if (delta > 0 && SIZELEFT(mp) < delta) {
6981 /* not enough space left, do a delete and split */
6982 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6983 pgno = NODEPGNO(node);
6984 mdb_node_del(mc, 0);
6985 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6988 numkeys = NUMKEYS(mp);
6989 for (i = 0; i < numkeys; i++) {
6990 if (mp->mp_ptrs[i] <= ptr)
6991 mp->mp_ptrs[i] -= delta;
6994 base = (char *)mp + mp->mp_upper + PAGEBASE;
6995 len = ptr - mp->mp_upper + NODESIZE;
6996 memmove(base - delta, base, len);
6997 mp->mp_upper -= delta;
6999 node = NODEPTR(mp, indx);
7002 /* But even if no shift was needed, update ksize */
7003 if (node->mn_ksize != key->mv_size)
7004 node->mn_ksize = key->mv_size;
7007 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7013 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7015 /** Move a node from csrc to cdst.
7018 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7025 unsigned short flags;
7029 /* Mark src and dst as dirty. */
7030 if ((rc = mdb_page_touch(csrc)) ||
7031 (rc = mdb_page_touch(cdst)))
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], csrc->mc_ki[csrc->mc_top], key.mv_size);
7038 data.mv_data = NULL;
7042 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7043 mdb_cassert(csrc, !((size_t)srcnode & 1));
7044 srcpg = NODEPGNO(srcnode);
7045 flags = srcnode->mn_flags;
7046 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7047 unsigned int snum = csrc->mc_snum;
7049 /* must find the lowest key below src */
7050 rc = mdb_page_search_lowest(csrc);
7053 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7054 key.mv_size = csrc->mc_db->md_pad;
7055 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7057 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7058 key.mv_size = NODEKSZ(s2);
7059 key.mv_data = NODEKEY(s2);
7061 csrc->mc_snum = snum--;
7062 csrc->mc_top = snum;
7064 key.mv_size = NODEKSZ(srcnode);
7065 key.mv_data = NODEKEY(srcnode);
7067 data.mv_size = NODEDSZ(srcnode);
7068 data.mv_data = NODEDATA(srcnode);
7070 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7071 unsigned int snum = cdst->mc_snum;
7074 /* must find the lowest key below dst */
7075 mdb_cursor_copy(cdst, &mn);
7076 rc = mdb_page_search_lowest(&mn);
7079 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7080 bkey.mv_size = mn.mc_db->md_pad;
7081 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7083 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7084 bkey.mv_size = NODEKSZ(s2);
7085 bkey.mv_data = NODEKEY(s2);
7087 mn.mc_snum = snum--;
7090 rc = mdb_update_key(&mn, &bkey);
7095 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7096 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7097 csrc->mc_ki[csrc->mc_top],
7099 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7100 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7102 /* Add the node to the destination page.
7104 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7105 if (rc != MDB_SUCCESS)
7108 /* Delete the node from the source page.
7110 mdb_node_del(csrc, key.mv_size);
7113 /* Adjust other cursors pointing to mp */
7114 MDB_cursor *m2, *m3;
7115 MDB_dbi dbi = csrc->mc_dbi;
7116 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7118 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7119 if (csrc->mc_flags & C_SUB)
7120 m3 = &m2->mc_xcursor->mx_cursor;
7123 if (m3 == csrc) continue;
7124 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7125 csrc->mc_ki[csrc->mc_top]) {
7126 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7127 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7132 /* Update the parent separators.
7134 if (csrc->mc_ki[csrc->mc_top] == 0) {
7135 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7136 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7137 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7139 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7140 key.mv_size = NODEKSZ(srcnode);
7141 key.mv_data = NODEKEY(srcnode);
7143 DPRINTF(("update separator for source page %"Z"u to [%s]",
7144 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7145 mdb_cursor_copy(csrc, &mn);
7148 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7151 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7153 indx_t ix = csrc->mc_ki[csrc->mc_top];
7154 nullkey.mv_size = 0;
7155 csrc->mc_ki[csrc->mc_top] = 0;
7156 rc = mdb_update_key(csrc, &nullkey);
7157 csrc->mc_ki[csrc->mc_top] = ix;
7158 mdb_cassert(csrc, rc == MDB_SUCCESS);
7162 if (cdst->mc_ki[cdst->mc_top] == 0) {
7163 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7164 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7165 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7167 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7168 key.mv_size = NODEKSZ(srcnode);
7169 key.mv_data = NODEKEY(srcnode);
7171 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7172 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7173 mdb_cursor_copy(cdst, &mn);
7176 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7179 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7181 indx_t ix = cdst->mc_ki[cdst->mc_top];
7182 nullkey.mv_size = 0;
7183 cdst->mc_ki[cdst->mc_top] = 0;
7184 rc = mdb_update_key(cdst, &nullkey);
7185 cdst->mc_ki[cdst->mc_top] = ix;
7186 mdb_cassert(csrc, rc == MDB_SUCCESS);
7193 /** Merge one page into another.
7194 * The nodes from the page pointed to by \b csrc will
7195 * be copied to the page pointed to by \b cdst and then
7196 * the \b csrc page will be freed.
7197 * @param[in] csrc Cursor pointing to the source page.
7198 * @param[in] cdst Cursor pointing to the destination page.
7199 * @return 0 on success, non-zero on failure.
7202 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7204 MDB_page *psrc, *pdst;
7211 psrc = csrc->mc_pg[csrc->mc_top];
7212 pdst = cdst->mc_pg[cdst->mc_top];
7214 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7216 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7217 mdb_cassert(csrc, cdst->mc_snum > 1);
7219 /* Mark dst as dirty. */
7220 if ((rc = mdb_page_touch(cdst)))
7223 /* Move all nodes from src to dst.
7225 j = nkeys = NUMKEYS(pdst);
7226 if (IS_LEAF2(psrc)) {
7227 key.mv_size = csrc->mc_db->md_pad;
7228 key.mv_data = METADATA(psrc);
7229 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7230 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7231 if (rc != MDB_SUCCESS)
7233 key.mv_data = (char *)key.mv_data + key.mv_size;
7236 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7237 srcnode = NODEPTR(psrc, i);
7238 if (i == 0 && IS_BRANCH(psrc)) {
7241 mdb_cursor_copy(csrc, &mn);
7242 /* must find the lowest key below src */
7243 rc = mdb_page_search_lowest(&mn);
7246 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7247 key.mv_size = mn.mc_db->md_pad;
7248 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7250 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7251 key.mv_size = NODEKSZ(s2);
7252 key.mv_data = NODEKEY(s2);
7255 key.mv_size = srcnode->mn_ksize;
7256 key.mv_data = NODEKEY(srcnode);
7259 data.mv_size = NODEDSZ(srcnode);
7260 data.mv_data = NODEDATA(srcnode);
7261 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7262 if (rc != MDB_SUCCESS)
7267 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7268 pdst->mp_pgno, NUMKEYS(pdst),
7269 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7271 /* Unlink the src page from parent and add to free list.
7274 mdb_node_del(csrc, 0);
7275 if (csrc->mc_ki[csrc->mc_top] == 0) {
7277 rc = mdb_update_key(csrc, &key);
7285 psrc = csrc->mc_pg[csrc->mc_top];
7286 /* If not operating on FreeDB, allow this page to be reused
7287 * in this txn. Otherwise just add to free list.
7289 rc = mdb_page_loose(csrc, psrc);
7293 csrc->mc_db->md_leaf_pages--;
7295 csrc->mc_db->md_branch_pages--;
7297 /* Adjust other cursors pointing to mp */
7298 MDB_cursor *m2, *m3;
7299 MDB_dbi dbi = csrc->mc_dbi;
7301 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7302 if (csrc->mc_flags & C_SUB)
7303 m3 = &m2->mc_xcursor->mx_cursor;
7306 if (m3 == csrc) continue;
7307 if (m3->mc_snum < csrc->mc_snum) continue;
7308 if (m3->mc_pg[csrc->mc_top] == psrc) {
7309 m3->mc_pg[csrc->mc_top] = pdst;
7310 m3->mc_ki[csrc->mc_top] += nkeys;
7315 unsigned int snum = cdst->mc_snum;
7316 uint16_t depth = cdst->mc_db->md_depth;
7317 mdb_cursor_pop(cdst);
7318 rc = mdb_rebalance(cdst);
7319 /* Did the tree shrink? */
7320 if (depth > cdst->mc_db->md_depth)
7322 cdst->mc_snum = snum;
7323 cdst->mc_top = snum-1;
7328 /** Copy the contents of a cursor.
7329 * @param[in] csrc The cursor to copy from.
7330 * @param[out] cdst The cursor to copy to.
7333 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7337 cdst->mc_txn = csrc->mc_txn;
7338 cdst->mc_dbi = csrc->mc_dbi;
7339 cdst->mc_db = csrc->mc_db;
7340 cdst->mc_dbx = csrc->mc_dbx;
7341 cdst->mc_snum = csrc->mc_snum;
7342 cdst->mc_top = csrc->mc_top;
7343 cdst->mc_flags = csrc->mc_flags;
7345 for (i=0; i<csrc->mc_snum; i++) {
7346 cdst->mc_pg[i] = csrc->mc_pg[i];
7347 cdst->mc_ki[i] = csrc->mc_ki[i];
7351 /** Rebalance the tree after a delete operation.
7352 * @param[in] mc Cursor pointing to the page where rebalancing
7354 * @return 0 on success, non-zero on failure.
7357 mdb_rebalance(MDB_cursor *mc)
7361 unsigned int ptop, minkeys;
7365 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7366 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7367 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7368 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7369 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7371 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7372 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7373 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7374 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7378 if (mc->mc_snum < 2) {
7379 MDB_page *mp = mc->mc_pg[0];
7381 DPUTS("Can't rebalance a subpage, ignoring");
7384 if (NUMKEYS(mp) == 0) {
7385 DPUTS("tree is completely empty");
7386 mc->mc_db->md_root = P_INVALID;
7387 mc->mc_db->md_depth = 0;
7388 mc->mc_db->md_leaf_pages = 0;
7389 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7392 /* Adjust cursors pointing to mp */
7395 mc->mc_flags &= ~C_INITIALIZED;
7397 MDB_cursor *m2, *m3;
7398 MDB_dbi dbi = mc->mc_dbi;
7400 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7401 if (mc->mc_flags & C_SUB)
7402 m3 = &m2->mc_xcursor->mx_cursor;
7405 if (m3->mc_snum < mc->mc_snum) continue;
7406 if (m3->mc_pg[0] == mp) {
7409 m3->mc_flags &= ~C_INITIALIZED;
7413 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7415 DPUTS("collapsing root page!");
7416 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7419 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7420 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7423 mc->mc_db->md_depth--;
7424 mc->mc_db->md_branch_pages--;
7425 mc->mc_ki[0] = mc->mc_ki[1];
7426 for (i = 1; i<mc->mc_db->md_depth; i++) {
7427 mc->mc_pg[i] = mc->mc_pg[i+1];
7428 mc->mc_ki[i] = mc->mc_ki[i+1];
7431 /* Adjust other cursors pointing to mp */
7432 MDB_cursor *m2, *m3;
7433 MDB_dbi dbi = mc->mc_dbi;
7435 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7436 if (mc->mc_flags & C_SUB)
7437 m3 = &m2->mc_xcursor->mx_cursor;
7440 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7441 if (m3->mc_pg[0] == mp) {
7444 for (i=0; i<m3->mc_snum; i++) {
7445 m3->mc_pg[i] = m3->mc_pg[i+1];
7446 m3->mc_ki[i] = m3->mc_ki[i+1];
7452 DPUTS("root page doesn't need rebalancing");
7456 /* The parent (branch page) must have at least 2 pointers,
7457 * otherwise the tree is invalid.
7459 ptop = mc->mc_top-1;
7460 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7462 /* Leaf page fill factor is below the threshold.
7463 * Try to move keys from left or right neighbor, or
7464 * merge with a neighbor page.
7469 mdb_cursor_copy(mc, &mn);
7470 mn.mc_xcursor = NULL;
7472 oldki = mc->mc_ki[mc->mc_top];
7473 if (mc->mc_ki[ptop] == 0) {
7474 /* We're the leftmost leaf in our parent.
7476 DPUTS("reading right neighbor");
7478 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7479 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7482 mn.mc_ki[mn.mc_top] = 0;
7483 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7485 /* There is at least one neighbor to the left.
7487 DPUTS("reading left neighbor");
7489 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7490 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7493 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7494 mc->mc_ki[mc->mc_top] = 0;
7497 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7498 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7499 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7501 /* If the neighbor page is above threshold and has enough keys,
7502 * move one key from it. Otherwise we should try to merge them.
7503 * (A branch page must never have less than 2 keys.)
7505 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7506 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7507 rc = mdb_node_move(&mn, mc);
7508 if (mc->mc_ki[ptop]) {
7512 if (mc->mc_ki[ptop] == 0) {
7513 rc = mdb_page_merge(&mn, mc);
7515 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7516 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7517 rc = mdb_page_merge(mc, &mn);
7518 mdb_cursor_copy(&mn, mc);
7520 mc->mc_flags &= ~C_EOF;
7522 mc->mc_ki[mc->mc_top] = oldki;
7526 /** Complete a delete operation started by #mdb_cursor_del(). */
7528 mdb_cursor_del0(MDB_cursor *mc)
7535 ki = mc->mc_ki[mc->mc_top];
7536 mdb_node_del(mc, mc->mc_db->md_pad);
7537 mc->mc_db->md_entries--;
7538 rc = mdb_rebalance(mc);
7540 if (rc == MDB_SUCCESS) {
7541 MDB_cursor *m2, *m3;
7542 MDB_dbi dbi = mc->mc_dbi;
7544 mp = mc->mc_pg[mc->mc_top];
7545 nkeys = NUMKEYS(mp);
7547 /* if mc points past last node in page, find next sibling */
7548 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7549 rc = mdb_cursor_sibling(mc, 1);
7550 if (rc == MDB_NOTFOUND) {
7551 mc->mc_flags |= C_EOF;
7556 /* Adjust other cursors pointing to mp */
7557 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7558 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7559 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7561 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7563 if (m3->mc_pg[mc->mc_top] == mp) {
7564 if (m3->mc_ki[mc->mc_top] >= ki) {
7565 m3->mc_flags |= C_DEL;
7566 if (m3->mc_ki[mc->mc_top] > ki)
7567 m3->mc_ki[mc->mc_top]--;
7569 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7570 rc = mdb_cursor_sibling(m3, 1);
7571 if (rc == MDB_NOTFOUND) {
7572 m3->mc_flags |= C_EOF;
7578 mc->mc_flags |= C_DEL;
7582 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7587 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7588 MDB_val *key, MDB_val *data)
7590 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7593 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7594 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7596 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7597 /* must ignore any data */
7601 return mdb_del0(txn, dbi, key, data, 0);
7605 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7606 MDB_val *key, MDB_val *data, unsigned flags)
7611 MDB_val rdata, *xdata;
7615 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7617 mdb_cursor_init(&mc, txn, dbi, &mx);
7626 flags |= MDB_NODUPDATA;
7628 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7630 /* let mdb_page_split know about this cursor if needed:
7631 * delete will trigger a rebalance; if it needs to move
7632 * a node from one page to another, it will have to
7633 * update the parent's separator key(s). If the new sepkey
7634 * is larger than the current one, the parent page may
7635 * run out of space, triggering a split. We need this
7636 * cursor to be consistent until the end of the rebalance.
7638 mc.mc_flags |= C_UNTRACK;
7639 mc.mc_next = txn->mt_cursors[dbi];
7640 txn->mt_cursors[dbi] = &mc;
7641 rc = mdb_cursor_del(&mc, flags);
7642 txn->mt_cursors[dbi] = mc.mc_next;
7647 /** Split a page and insert a new node.
7648 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7649 * The cursor will be updated to point to the actual page and index where
7650 * the node got inserted after the split.
7651 * @param[in] newkey The key for the newly inserted node.
7652 * @param[in] newdata The data for the newly inserted node.
7653 * @param[in] newpgno The page number, if the new node is a branch node.
7654 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7655 * @return 0 on success, non-zero on failure.
7658 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7659 unsigned int nflags)
7662 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7665 int i, j, split_indx, nkeys, pmax;
7666 MDB_env *env = mc->mc_txn->mt_env;
7668 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7669 MDB_page *copy = NULL;
7670 MDB_page *mp, *rp, *pp;
7675 mp = mc->mc_pg[mc->mc_top];
7676 newindx = mc->mc_ki[mc->mc_top];
7677 nkeys = NUMKEYS(mp);
7679 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7680 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7681 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7683 /* Create a right sibling. */
7684 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7686 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7688 if (mc->mc_snum < 2) {
7689 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7691 /* shift current top to make room for new parent */
7692 mc->mc_pg[1] = mc->mc_pg[0];
7693 mc->mc_ki[1] = mc->mc_ki[0];
7696 mc->mc_db->md_root = pp->mp_pgno;
7697 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7698 mc->mc_db->md_depth++;
7701 /* Add left (implicit) pointer. */
7702 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7703 /* undo the pre-push */
7704 mc->mc_pg[0] = mc->mc_pg[1];
7705 mc->mc_ki[0] = mc->mc_ki[1];
7706 mc->mc_db->md_root = mp->mp_pgno;
7707 mc->mc_db->md_depth--;
7714 ptop = mc->mc_top-1;
7715 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7718 mc->mc_flags |= C_SPLITTING;
7719 mdb_cursor_copy(mc, &mn);
7720 mn.mc_pg[mn.mc_top] = rp;
7721 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7723 if (nflags & MDB_APPEND) {
7724 mn.mc_ki[mn.mc_top] = 0;
7726 split_indx = newindx;
7730 split_indx = (nkeys+1) / 2;
7735 unsigned int lsize, rsize, ksize;
7736 /* Move half of the keys to the right sibling */
7737 x = mc->mc_ki[mc->mc_top] - split_indx;
7738 ksize = mc->mc_db->md_pad;
7739 split = LEAF2KEY(mp, split_indx, ksize);
7740 rsize = (nkeys - split_indx) * ksize;
7741 lsize = (nkeys - split_indx) * sizeof(indx_t);
7742 mp->mp_lower -= lsize;
7743 rp->mp_lower += lsize;
7744 mp->mp_upper += rsize - lsize;
7745 rp->mp_upper -= rsize - lsize;
7746 sepkey.mv_size = ksize;
7747 if (newindx == split_indx) {
7748 sepkey.mv_data = newkey->mv_data;
7750 sepkey.mv_data = split;
7753 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7754 memcpy(rp->mp_ptrs, split, rsize);
7755 sepkey.mv_data = rp->mp_ptrs;
7756 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7757 memcpy(ins, newkey->mv_data, ksize);
7758 mp->mp_lower += sizeof(indx_t);
7759 mp->mp_upper -= ksize - sizeof(indx_t);
7762 memcpy(rp->mp_ptrs, split, x * ksize);
7763 ins = LEAF2KEY(rp, x, ksize);
7764 memcpy(ins, newkey->mv_data, ksize);
7765 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7766 rp->mp_lower += sizeof(indx_t);
7767 rp->mp_upper -= ksize - sizeof(indx_t);
7768 mc->mc_ki[mc->mc_top] = x;
7769 mc->mc_pg[mc->mc_top] = rp;
7772 int psize, nsize, k;
7773 /* Maximum free space in an empty page */
7774 pmax = env->me_psize - PAGEHDRSZ;
7776 nsize = mdb_leaf_size(env, newkey, newdata);
7778 nsize = mdb_branch_size(env, newkey);
7779 nsize = EVEN(nsize);
7781 /* grab a page to hold a temporary copy */
7782 copy = mdb_page_malloc(mc->mc_txn, 1);
7787 copy->mp_pgno = mp->mp_pgno;
7788 copy->mp_flags = mp->mp_flags;
7789 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7790 copy->mp_upper = env->me_psize - PAGEBASE;
7792 /* prepare to insert */
7793 for (i=0, j=0; i<nkeys; i++) {
7795 copy->mp_ptrs[j++] = 0;
7797 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7800 /* When items are relatively large the split point needs
7801 * to be checked, because being off-by-one will make the
7802 * difference between success or failure in mdb_node_add.
7804 * It's also relevant if a page happens to be laid out
7805 * such that one half of its nodes are all "small" and
7806 * the other half of its nodes are "large." If the new
7807 * item is also "large" and falls on the half with
7808 * "large" nodes, it also may not fit.
7810 * As a final tweak, if the new item goes on the last
7811 * spot on the page (and thus, onto the new page), bias
7812 * the split so the new page is emptier than the old page.
7813 * This yields better packing during sequential inserts.
7815 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7816 /* Find split point */
7818 if (newindx <= split_indx || newindx >= nkeys) {
7820 k = newindx >= nkeys ? nkeys : split_indx+2;
7825 for (; i!=k; i+=j) {
7830 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7831 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7833 if (F_ISSET(node->mn_flags, F_BIGDATA))
7834 psize += sizeof(pgno_t);
7836 psize += NODEDSZ(node);
7838 psize = EVEN(psize);
7840 if (psize > pmax || i == k-j) {
7841 split_indx = i + (j<0);
7846 if (split_indx == newindx) {
7847 sepkey.mv_size = newkey->mv_size;
7848 sepkey.mv_data = newkey->mv_data;
7850 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
7851 sepkey.mv_size = node->mn_ksize;
7852 sepkey.mv_data = NODEKEY(node);
7857 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7859 /* Copy separator key to the parent.
7861 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7865 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7870 if (mn.mc_snum == mc->mc_snum) {
7871 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7872 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7873 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7874 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7879 /* Right page might now have changed parent.
7880 * Check if left page also changed parent.
7882 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7883 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7884 for (i=0; i<ptop; i++) {
7885 mc->mc_pg[i] = mn.mc_pg[i];
7886 mc->mc_ki[i] = mn.mc_ki[i];
7888 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7889 if (mn.mc_ki[ptop]) {
7890 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7892 /* find right page's left sibling */
7893 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7894 mdb_cursor_sibling(mc, 0);
7899 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7902 mc->mc_flags ^= C_SPLITTING;
7903 if (rc != MDB_SUCCESS) {
7906 if (nflags & MDB_APPEND) {
7907 mc->mc_pg[mc->mc_top] = rp;
7908 mc->mc_ki[mc->mc_top] = 0;
7909 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7912 for (i=0; i<mc->mc_top; i++)
7913 mc->mc_ki[i] = mn.mc_ki[i];
7914 } else if (!IS_LEAF2(mp)) {
7916 mc->mc_pg[mc->mc_top] = rp;
7921 rkey.mv_data = newkey->mv_data;
7922 rkey.mv_size = newkey->mv_size;
7928 /* Update index for the new key. */
7929 mc->mc_ki[mc->mc_top] = j;
7931 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7932 rkey.mv_data = NODEKEY(node);
7933 rkey.mv_size = node->mn_ksize;
7935 xdata.mv_data = NODEDATA(node);
7936 xdata.mv_size = NODEDSZ(node);
7939 pgno = NODEPGNO(node);
7940 flags = node->mn_flags;
7943 if (!IS_LEAF(mp) && j == 0) {
7944 /* First branch index doesn't need key data. */
7948 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7954 mc->mc_pg[mc->mc_top] = copy;
7959 } while (i != split_indx);
7961 nkeys = NUMKEYS(copy);
7962 for (i=0; i<nkeys; i++)
7963 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7964 mp->mp_lower = copy->mp_lower;
7965 mp->mp_upper = copy->mp_upper;
7966 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7967 env->me_psize - copy->mp_upper - PAGEBASE);
7969 /* reset back to original page */
7970 if (newindx < split_indx) {
7971 mc->mc_pg[mc->mc_top] = mp;
7972 if (nflags & MDB_RESERVE) {
7973 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7974 if (!(node->mn_flags & F_BIGDATA))
7975 newdata->mv_data = NODEDATA(node);
7978 mc->mc_pg[mc->mc_top] = rp;
7980 /* Make sure mc_ki is still valid.
7982 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7983 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7984 for (i=0; i<=ptop; i++) {
7985 mc->mc_pg[i] = mn.mc_pg[i];
7986 mc->mc_ki[i] = mn.mc_ki[i];
7993 /* Adjust other cursors pointing to mp */
7994 MDB_cursor *m2, *m3;
7995 MDB_dbi dbi = mc->mc_dbi;
7996 int fixup = NUMKEYS(mp);
7998 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7999 if (mc->mc_flags & C_SUB)
8000 m3 = &m2->mc_xcursor->mx_cursor;
8005 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8007 if (m3->mc_flags & C_SPLITTING)
8012 for (k=m3->mc_top; k>=0; k--) {
8013 m3->mc_ki[k+1] = m3->mc_ki[k];
8014 m3->mc_pg[k+1] = m3->mc_pg[k];
8016 if (m3->mc_ki[0] >= split_indx) {
8021 m3->mc_pg[0] = mc->mc_pg[0];
8025 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8026 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8027 m3->mc_ki[mc->mc_top]++;
8028 if (m3->mc_ki[mc->mc_top] >= fixup) {
8029 m3->mc_pg[mc->mc_top] = rp;
8030 m3->mc_ki[mc->mc_top] -= fixup;
8031 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8033 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8034 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8039 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8042 if (copy) /* tmp page */
8043 mdb_page_free(env, copy);
8045 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8050 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8051 MDB_val *key, MDB_val *data, unsigned int flags)
8056 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8059 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8062 mdb_cursor_init(&mc, txn, dbi, &mx);
8063 return mdb_cursor_put(&mc, key, data, flags);
8067 #define MDB_WBUF (1024*1024)
8070 /** State needed for a compacting copy. */
8071 typedef struct mdb_copy {
8072 pthread_mutex_t mc_mutex;
8073 pthread_cond_t mc_cond;
8080 pgno_t mc_next_pgno;
8083 volatile int mc_new;
8088 /** Dedicated writer thread for compacting copy. */
8089 static THREAD_RET ESECT
8090 mdb_env_copythr(void *arg)
8094 int toggle = 0, wsize, rc;
8097 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8100 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8103 pthread_mutex_lock(&my->mc_mutex);
8105 pthread_cond_signal(&my->mc_cond);
8108 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8109 if (my->mc_new < 0) {
8114 wsize = my->mc_wlen[toggle];
8115 ptr = my->mc_wbuf[toggle];
8118 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8122 } else if (len > 0) {
8136 /* If there's an overflow page tail, write it too */
8137 if (my->mc_olen[toggle]) {
8138 wsize = my->mc_olen[toggle];
8139 ptr = my->mc_over[toggle];
8140 my->mc_olen[toggle] = 0;
8143 my->mc_wlen[toggle] = 0;
8145 pthread_cond_signal(&my->mc_cond);
8147 pthread_cond_signal(&my->mc_cond);
8148 pthread_mutex_unlock(&my->mc_mutex);
8149 return (THREAD_RET)0;
8153 /** Tell the writer thread there's a buffer ready to write */
8155 mdb_env_cthr_toggle(mdb_copy *my, int st)
8157 int toggle = my->mc_toggle ^ 1;
8158 pthread_mutex_lock(&my->mc_mutex);
8159 if (my->mc_status) {
8160 pthread_mutex_unlock(&my->mc_mutex);
8161 return my->mc_status;
8163 while (my->mc_new == 1)
8164 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8166 my->mc_toggle = toggle;
8167 pthread_cond_signal(&my->mc_cond);
8168 pthread_mutex_unlock(&my->mc_mutex);
8172 /** Depth-first tree traversal for compacting copy. */
8174 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8177 MDB_txn *txn = my->mc_txn;
8179 MDB_page *mo, *mp, *leaf;
8184 /* Empty DB, nothing to do */
8185 if (*pg == P_INVALID)
8192 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8195 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8199 /* Make cursor pages writable */
8200 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8204 for (i=0; i<mc.mc_top; i++) {
8205 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8206 mc.mc_pg[i] = (MDB_page *)ptr;
8207 ptr += my->mc_env->me_psize;
8210 /* This is writable space for a leaf page. Usually not needed. */
8211 leaf = (MDB_page *)ptr;
8213 toggle = my->mc_toggle;
8214 while (mc.mc_snum > 0) {
8216 mp = mc.mc_pg[mc.mc_top];
8220 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8221 for (i=0; i<n; i++) {
8222 ni = NODEPTR(mp, i);
8223 if (ni->mn_flags & F_BIGDATA) {
8227 /* Need writable leaf */
8229 mc.mc_pg[mc.mc_top] = leaf;
8230 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8232 ni = NODEPTR(mp, i);
8235 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8236 rc = mdb_page_get(txn, pg, &omp, NULL);
8239 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8240 rc = mdb_env_cthr_toggle(my, 1);
8243 toggle = my->mc_toggle;
8245 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8246 memcpy(mo, omp, my->mc_env->me_psize);
8247 mo->mp_pgno = my->mc_next_pgno;
8248 my->mc_next_pgno += omp->mp_pages;
8249 my->mc_wlen[toggle] += my->mc_env->me_psize;
8250 if (omp->mp_pages > 1) {
8251 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8252 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8253 rc = mdb_env_cthr_toggle(my, 1);
8256 toggle = my->mc_toggle;
8258 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8259 } else if (ni->mn_flags & F_SUBDATA) {
8262 /* Need writable leaf */
8264 mc.mc_pg[mc.mc_top] = leaf;
8265 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8267 ni = NODEPTR(mp, i);
8270 memcpy(&db, NODEDATA(ni), sizeof(db));
8271 my->mc_toggle = toggle;
8272 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8275 toggle = my->mc_toggle;
8276 memcpy(NODEDATA(ni), &db, sizeof(db));
8281 mc.mc_ki[mc.mc_top]++;
8282 if (mc.mc_ki[mc.mc_top] < n) {
8285 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8287 rc = mdb_page_get(txn, pg, &mp, NULL);
8292 mc.mc_ki[mc.mc_top] = 0;
8293 if (IS_BRANCH(mp)) {
8294 /* Whenever we advance to a sibling branch page,
8295 * we must proceed all the way down to its first leaf.
8297 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8300 mc.mc_pg[mc.mc_top] = mp;
8304 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8305 rc = mdb_env_cthr_toggle(my, 1);
8308 toggle = my->mc_toggle;
8310 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8311 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8312 mo->mp_pgno = my->mc_next_pgno++;
8313 my->mc_wlen[toggle] += my->mc_env->me_psize;
8315 /* Update parent if there is one */
8316 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8317 SETPGNO(ni, mo->mp_pgno);
8318 mdb_cursor_pop(&mc);
8320 /* Otherwise we're done */
8330 /** Copy environment with compaction. */
8332 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8337 MDB_txn *txn = NULL;
8342 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8343 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8344 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize);
8345 if (my.mc_wbuf[0] == NULL)
8348 pthread_mutex_init(&my.mc_mutex, NULL);
8349 pthread_cond_init(&my.mc_cond, NULL);
8350 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2);
8354 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8359 my.mc_next_pgno = 2;
8365 THREAD_CREATE(thr, mdb_env_copythr, &my);
8367 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8371 mp = (MDB_page *)my.mc_wbuf[0];
8372 memset(mp, 0, 2*env->me_psize);
8374 mp->mp_flags = P_META;
8375 mm = (MDB_meta *)METADATA(mp);
8376 mdb_env_init_meta0(env, mm);
8377 mm->mm_address = env->me_metas[0]->mm_address;
8379 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8381 mp->mp_flags = P_META;
8382 *(MDB_meta *)METADATA(mp) = *mm;
8383 mm = (MDB_meta *)METADATA(mp);
8385 /* Count the number of free pages, subtract from lastpg to find
8386 * number of active pages
8389 MDB_ID freecount = 0;
8392 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8393 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8394 freecount += *(MDB_ID *)data.mv_data;
8395 freecount += txn->mt_dbs[0].md_branch_pages +
8396 txn->mt_dbs[0].md_leaf_pages +
8397 txn->mt_dbs[0].md_overflow_pages;
8399 /* Set metapage 1 */
8400 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8401 mm->mm_dbs[1] = txn->mt_dbs[1];
8402 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8405 my.mc_wlen[0] = env->me_psize * 2;
8407 pthread_mutex_lock(&my.mc_mutex);
8409 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8410 pthread_mutex_unlock(&my.mc_mutex);
8411 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8412 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8413 rc = mdb_env_cthr_toggle(&my, 1);
8414 mdb_env_cthr_toggle(&my, -1);
8415 pthread_mutex_lock(&my.mc_mutex);
8417 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8418 pthread_mutex_unlock(&my.mc_mutex);
8423 CloseHandle(my.mc_cond);
8424 CloseHandle(my.mc_mutex);
8425 _aligned_free(my.mc_wbuf[0]);
8427 pthread_cond_destroy(&my.mc_cond);
8428 pthread_mutex_destroy(&my.mc_mutex);
8429 free(my.mc_wbuf[0]);
8434 /** Copy environment as-is. */
8436 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8438 MDB_txn *txn = NULL;
8444 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8448 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8451 /* Do the lock/unlock of the reader mutex before starting the
8452 * write txn. Otherwise other read txns could block writers.
8454 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8459 /* We must start the actual read txn after blocking writers */
8460 mdb_txn_reset0(txn, "reset-stage1");
8462 /* Temporarily block writers until we snapshot the meta pages */
8465 rc = mdb_txn_renew0(txn);
8467 UNLOCK_MUTEX_W(env);
8472 wsize = env->me_psize * 2;
8476 DO_WRITE(rc, fd, ptr, w2, len);
8480 } else if (len > 0) {
8486 /* Non-blocking or async handles are not supported */
8492 UNLOCK_MUTEX_W(env);
8497 w2 = txn->mt_next_pgno * env->me_psize;
8500 LARGE_INTEGER fsize;
8501 GetFileSizeEx(env->me_fd, &fsize);
8502 if (w2 > fsize.QuadPart)
8503 w2 = fsize.QuadPart;
8508 fstat(env->me_fd, &st);
8509 if (w2 > (size_t)st.st_size)
8515 if (wsize > MAX_WRITE)
8519 DO_WRITE(rc, fd, ptr, w2, len);
8523 } else if (len > 0) {
8540 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8542 if (flags & MDB_CP_COMPACT)
8543 return mdb_env_copyfd1(env, fd);
8545 return mdb_env_copyfd0(env, fd);
8549 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8551 return mdb_env_copyfd2(env, fd, 0);
8555 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8559 HANDLE newfd = INVALID_HANDLE_VALUE;
8561 if (env->me_flags & MDB_NOSUBDIR) {
8562 lpath = (char *)path;
8565 len += sizeof(DATANAME);
8566 lpath = malloc(len);
8569 sprintf(lpath, "%s" DATANAME, path);
8572 /* The destination path must exist, but the destination file must not.
8573 * We don't want the OS to cache the writes, since the source data is
8574 * already in the OS cache.
8577 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8578 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8580 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8582 if (newfd == INVALID_HANDLE_VALUE) {
8588 /* Set O_DIRECT if the file system supports it */
8589 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8590 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8592 #ifdef F_NOCACHE /* __APPLE__ */
8593 rc = fcntl(newfd, F_NOCACHE, 1);
8600 rc = mdb_env_copyfd2(env, newfd, flags);
8603 if (!(env->me_flags & MDB_NOSUBDIR))
8605 if (newfd != INVALID_HANDLE_VALUE)
8606 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8613 mdb_env_copy(MDB_env *env, const char *path)
8615 return mdb_env_copy2(env, path, 0);
8619 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8621 if ((flag & CHANGEABLE) != flag)
8624 env->me_flags |= flag;
8626 env->me_flags &= ~flag;
8631 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8636 *arg = env->me_flags;
8641 mdb_env_set_userctx(MDB_env *env, void *ctx)
8645 env->me_userctx = ctx;
8650 mdb_env_get_userctx(MDB_env *env)
8652 return env ? env->me_userctx : NULL;
8656 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8661 env->me_assert_func = func;
8667 mdb_env_get_path(MDB_env *env, const char **arg)
8672 *arg = env->me_path;
8677 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8686 /** Common code for #mdb_stat() and #mdb_env_stat().
8687 * @param[in] env the environment to operate in.
8688 * @param[in] db the #MDB_db record containing the stats to return.
8689 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8690 * @return 0, this function always succeeds.
8693 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8695 arg->ms_psize = env->me_psize;
8696 arg->ms_depth = db->md_depth;
8697 arg->ms_branch_pages = db->md_branch_pages;
8698 arg->ms_leaf_pages = db->md_leaf_pages;
8699 arg->ms_overflow_pages = db->md_overflow_pages;
8700 arg->ms_entries = db->md_entries;
8706 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8710 if (env == NULL || arg == NULL)
8713 toggle = mdb_env_pick_meta(env);
8715 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8719 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8723 if (env == NULL || arg == NULL)
8726 toggle = mdb_env_pick_meta(env);
8727 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8728 arg->me_mapsize = env->me_mapsize;
8729 arg->me_maxreaders = env->me_maxreaders;
8731 /* me_numreaders may be zero if this process never used any readers. Use
8732 * the shared numreader count if it exists.
8734 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8736 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8737 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8741 /** Set the default comparison functions for a database.
8742 * Called immediately after a database is opened to set the defaults.
8743 * The user can then override them with #mdb_set_compare() or
8744 * #mdb_set_dupsort().
8745 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8746 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8749 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8751 uint16_t f = txn->mt_dbs[dbi].md_flags;
8753 txn->mt_dbxs[dbi].md_cmp =
8754 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8755 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8757 txn->mt_dbxs[dbi].md_dcmp =
8758 !(f & MDB_DUPSORT) ? 0 :
8759 ((f & MDB_INTEGERDUP)
8760 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8761 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8764 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8769 int rc, dbflag, exact;
8770 unsigned int unused = 0;
8773 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8774 mdb_default_cmp(txn, FREE_DBI);
8777 if ((flags & VALID_FLAGS) != flags)
8779 if (txn->mt_flags & MDB_TXN_ERROR)
8785 if (flags & PERSISTENT_FLAGS) {
8786 uint16_t f2 = flags & PERSISTENT_FLAGS;
8787 /* make sure flag changes get committed */
8788 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8789 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8790 txn->mt_flags |= MDB_TXN_DIRTY;
8793 mdb_default_cmp(txn, MAIN_DBI);
8797 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8798 mdb_default_cmp(txn, MAIN_DBI);
8801 /* Is the DB already open? */
8803 for (i=2; i<txn->mt_numdbs; i++) {
8804 if (!txn->mt_dbxs[i].md_name.mv_size) {
8805 /* Remember this free slot */
8806 if (!unused) unused = i;
8809 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8810 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8816 /* If no free slot and max hit, fail */
8817 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8818 return MDB_DBS_FULL;
8820 /* Cannot mix named databases with some mainDB flags */
8821 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8822 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8824 /* Find the DB info */
8825 dbflag = DB_NEW|DB_VALID;
8828 key.mv_data = (void *)name;
8829 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8830 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8831 if (rc == MDB_SUCCESS) {
8832 /* make sure this is actually a DB */
8833 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8834 if (!(node->mn_flags & F_SUBDATA))
8835 return MDB_INCOMPATIBLE;
8836 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8837 /* Create if requested */
8839 data.mv_size = sizeof(MDB_db);
8840 data.mv_data = &dummy;
8841 memset(&dummy, 0, sizeof(dummy));
8842 dummy.md_root = P_INVALID;
8843 dummy.md_flags = flags & PERSISTENT_FLAGS;
8844 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8848 /* OK, got info, add to table */
8849 if (rc == MDB_SUCCESS) {
8850 unsigned int slot = unused ? unused : txn->mt_numdbs;
8851 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8852 txn->mt_dbxs[slot].md_name.mv_size = len;
8853 txn->mt_dbxs[slot].md_rel = NULL;
8854 txn->mt_dbflags[slot] = dbflag;
8855 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8857 mdb_default_cmp(txn, slot);
8866 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8868 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8871 if (txn->mt_flags & MDB_TXN_ERROR)
8874 if (txn->mt_dbflags[dbi] & DB_STALE) {
8877 /* Stale, must read the DB's root. cursor_init does it for us. */
8878 mdb_cursor_init(&mc, txn, dbi, &mx);
8880 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8883 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8886 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8888 ptr = env->me_dbxs[dbi].md_name.mv_data;
8889 env->me_dbxs[dbi].md_name.mv_data = NULL;
8890 env->me_dbxs[dbi].md_name.mv_size = 0;
8891 env->me_dbflags[dbi] = 0;
8895 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8897 /* We could return the flags for the FREE_DBI too but what's the point? */
8898 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8900 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8904 /** Add all the DB's pages to the free list.
8905 * @param[in] mc Cursor on the DB to free.
8906 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8907 * @return 0 on success, non-zero on failure.
8910 mdb_drop0(MDB_cursor *mc, int subs)
8914 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8915 if (rc == MDB_SUCCESS) {
8916 MDB_txn *txn = mc->mc_txn;
8921 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8922 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8925 mdb_cursor_copy(mc, &mx);
8926 while (mc->mc_snum > 0) {
8927 MDB_page *mp = mc->mc_pg[mc->mc_top];
8928 unsigned n = NUMKEYS(mp);
8930 for (i=0; i<n; i++) {
8931 ni = NODEPTR(mp, i);
8932 if (ni->mn_flags & F_BIGDATA) {
8935 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8936 rc = mdb_page_get(txn, pg, &omp, NULL);
8939 mdb_cassert(mc, IS_OVERFLOW(omp));
8940 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8944 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8945 mdb_xcursor_init1(mc, ni);
8946 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8952 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8954 for (i=0; i<n; i++) {
8956 ni = NODEPTR(mp, i);
8959 mdb_midl_xappend(txn->mt_free_pgs, pg);
8964 mc->mc_ki[mc->mc_top] = i;
8965 rc = mdb_cursor_sibling(mc, 1);
8967 if (rc != MDB_NOTFOUND)
8969 /* no more siblings, go back to beginning
8970 * of previous level.
8974 for (i=1; i<mc->mc_snum; i++) {
8976 mc->mc_pg[i] = mx.mc_pg[i];
8981 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8984 txn->mt_flags |= MDB_TXN_ERROR;
8985 } else if (rc == MDB_NOTFOUND) {
8991 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8993 MDB_cursor *mc, *m2;
8996 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8999 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9002 rc = mdb_cursor_open(txn, dbi, &mc);
9006 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9007 /* Invalidate the dropped DB's cursors */
9008 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9009 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9013 /* Can't delete the main DB */
9014 if (del && dbi > MAIN_DBI) {
9015 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9017 txn->mt_dbflags[dbi] = DB_STALE;
9018 mdb_dbi_close(txn->mt_env, dbi);
9020 txn->mt_flags |= MDB_TXN_ERROR;
9023 /* reset the DB record, mark it dirty */
9024 txn->mt_dbflags[dbi] |= DB_DIRTY;
9025 txn->mt_dbs[dbi].md_depth = 0;
9026 txn->mt_dbs[dbi].md_branch_pages = 0;
9027 txn->mt_dbs[dbi].md_leaf_pages = 0;
9028 txn->mt_dbs[dbi].md_overflow_pages = 0;
9029 txn->mt_dbs[dbi].md_entries = 0;
9030 txn->mt_dbs[dbi].md_root = P_INVALID;
9032 txn->mt_flags |= MDB_TXN_DIRTY;
9035 mdb_cursor_close(mc);
9039 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9041 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9044 txn->mt_dbxs[dbi].md_cmp = cmp;
9048 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9050 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9053 txn->mt_dbxs[dbi].md_dcmp = cmp;
9057 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9059 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9062 txn->mt_dbxs[dbi].md_rel = rel;
9066 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9068 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9071 txn->mt_dbxs[dbi].md_relctx = ctx;
9076 mdb_env_get_maxkeysize(MDB_env *env)
9078 return ENV_MAXKEY(env);
9082 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9084 unsigned int i, rdrs;
9087 int rc = 0, first = 1;
9091 if (!env->me_txns) {
9092 return func("(no reader locks)\n", ctx);
9094 rdrs = env->me_txns->mti_numreaders;
9095 mr = env->me_txns->mti_readers;
9096 for (i=0; i<rdrs; i++) {
9098 txnid_t txnid = mr[i].mr_txnid;
9099 sprintf(buf, txnid == (txnid_t)-1 ?
9100 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9101 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9104 rc = func(" pid thread txnid\n", ctx);
9108 rc = func(buf, ctx);
9114 rc = func("(no active readers)\n", ctx);
9119 /** Insert pid into list if not already present.
9120 * return -1 if already present.
9123 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9125 /* binary search of pid in list */
9127 unsigned cursor = 1;
9129 unsigned n = ids[0];
9132 unsigned pivot = n >> 1;
9133 cursor = base + pivot + 1;
9134 val = pid - ids[cursor];
9139 } else if ( val > 0 ) {
9144 /* found, so it's a duplicate */
9153 for (n = ids[0]; n > cursor; n--)
9160 mdb_reader_check(MDB_env *env, int *dead)
9162 unsigned int i, j, rdrs;
9164 MDB_PID_T *pids, pid;
9173 rdrs = env->me_txns->mti_numreaders;
9174 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9178 mr = env->me_txns->mti_readers;
9179 for (i=0; i<rdrs; i++) {
9180 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9182 if (mdb_pid_insert(pids, pid) == 0) {
9183 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9185 /* Recheck, a new process may have reused pid */
9186 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9187 for (j=i; j<rdrs; j++)
9188 if (mr[j].mr_pid == pid) {
9189 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9190 (unsigned) pid, mr[j].mr_txnid));
9195 UNLOCK_MUTEX_R(env);