2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 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.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define MDB_PIDLOCK 0
144 #define pthread_t DWORD
145 #define pthread_mutex_t HANDLE
146 #define pthread_key_t DWORD
147 #define pthread_self() GetCurrentThreadId()
148 #define pthread_key_create(x,y) \
149 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
150 #define pthread_key_delete(x) TlsFree(x)
151 #define pthread_getspecific(x) TlsGetValue(x)
152 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
153 #define pthread_mutex_unlock(x) ReleaseMutex(x)
154 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
155 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
156 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
157 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
158 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
159 #define getpid() GetCurrentProcessId()
160 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
161 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
162 #define ErrCode() GetLastError()
163 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
164 #define close(fd) (CloseHandle(fd) ? 0 : -1)
165 #define munmap(ptr,len) UnmapViewOfFile(ptr)
166 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
167 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
169 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
174 #define Z "z" /**< printf format modifier for size_t */
176 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
177 #define MDB_PIDLOCK 1
179 #ifdef MDB_USE_POSIX_SEM
181 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
182 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
183 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
184 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
187 mdb_sem_wait(sem_t *sem)
190 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
195 /** Lock the reader mutex.
197 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
198 /** Unlock the reader mutex.
200 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
202 /** Lock the writer mutex.
203 * Only a single write transaction is allowed at a time. Other writers
204 * will block waiting for this mutex.
206 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
207 /** Unlock the writer mutex.
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
210 #endif /* MDB_USE_POSIX_SEM */
212 /** Get the error code for the last failed system function.
214 #define ErrCode() errno
216 /** An abstraction for a file handle.
217 * On POSIX systems file handles are small integers. On Windows
218 * they're opaque pointers.
222 /** A value for an invalid file handle.
223 * Mainly used to initialize file variables and signify that they are
226 #define INVALID_HANDLE_VALUE (-1)
228 /** Get the size of a memory page for the system.
229 * This is the basic size that the platform's memory manager uses, and is
230 * fundamental to the use of memory-mapped files.
232 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
235 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
238 #define MNAME_LEN (sizeof(pthread_mutex_t))
244 /** A flag for opening a file and requesting synchronous data writes.
245 * This is only used when writing a meta page. It's not strictly needed;
246 * we could just do a normal write and then immediately perform a flush.
247 * But if this flag is available it saves us an extra system call.
249 * @note If O_DSYNC is undefined but exists in /usr/include,
250 * preferably set some compiler flag to get the definition.
251 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
254 # define MDB_DSYNC O_DSYNC
258 /** Function for flushing the data of a file. Define this to fsync
259 * if fdatasync() is not supported.
261 #ifndef MDB_FDATASYNC
262 # define MDB_FDATASYNC fdatasync
266 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
277 /** A page number in the database.
278 * Note that 64 bit page numbers are overkill, since pages themselves
279 * already represent 12-13 bits of addressable memory, and the OS will
280 * always limit applications to a maximum of 63 bits of address space.
282 * @note In the #MDB_node structure, we only store 48 bits of this value,
283 * which thus limits us to only 60 bits of addressable data.
285 typedef MDB_ID pgno_t;
287 /** A transaction ID.
288 * See struct MDB_txn.mt_txnid for details.
290 typedef MDB_ID txnid_t;
292 /** @defgroup debug Debug Macros
296 /** Enable debug output. Needs variable argument macros (a C99 feature).
297 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
298 * read from and written to the database (used for free space management).
304 static int mdb_debug;
305 static txnid_t mdb_debug_start;
307 /** Print a debug message with printf formatting.
308 * Requires double parenthesis around 2 or more args.
310 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
311 # define DPRINTF0(fmt, ...) \
312 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
314 # define DPRINTF(args) ((void) 0)
316 /** Print a debug string.
317 * The string is printed literally, with no format processing.
319 #define DPUTS(arg) DPRINTF(("%s", arg))
322 /** A default memory page size.
323 * The actual size is platform-dependent, but we use this for
324 * boot-strapping. We probably should not be using this any more.
325 * The #GET_PAGESIZE() macro is used to get the actual size.
327 * Note that we don't currently support Huge pages. On Linux,
328 * regular data files cannot use Huge pages, and in general
329 * Huge pages aren't actually pageable. We rely on the OS
330 * demand-pager to read our data and page it out when memory
331 * pressure from other processes is high. So until OSs have
332 * actual paging support for Huge pages, they're not viable.
334 #define MDB_PAGESIZE 4096
336 /** The minimum number of keys required in a database page.
337 * Setting this to a larger value will place a smaller bound on the
338 * maximum size of a data item. Data items larger than this size will
339 * be pushed into overflow pages instead of being stored directly in
340 * the B-tree node. This value used to default to 4. With a page size
341 * of 4096 bytes that meant that any item larger than 1024 bytes would
342 * go into an overflow page. That also meant that on average 2-3KB of
343 * each overflow page was wasted space. The value cannot be lower than
344 * 2 because then there would no longer be a tree structure. With this
345 * value, items larger than 2KB will go into overflow pages, and on
346 * average only 1KB will be wasted.
348 #define MDB_MINKEYS 2
350 /** A stamp that identifies a file as an MDB file.
351 * There's nothing special about this value other than that it is easily
352 * recognizable, and it will reflect any byte order mismatches.
354 #define MDB_MAGIC 0xBEEFC0DE
356 /** The version number for a database's datafile format. */
357 #define MDB_DATA_VERSION 1
358 /** The version number for a database's lockfile format. */
359 #define MDB_LOCK_VERSION 1
361 /** @brief The maximum size of a key in the database.
363 * The library rejects bigger keys, and cannot deal with records
364 * with bigger keys stored by a library with bigger max keysize.
366 * We require that keys all fit onto a regular page. This limit
367 * could be raised a bit further if needed; to something just
368 * under #MDB_PAGESIZE / #MDB_MINKEYS.
370 * Note that data items in an #MDB_DUPSORT database are actually keys
371 * of a subDB, so they're also limited to this size.
373 #ifndef MDB_MAXKEYSIZE
374 #define MDB_MAXKEYSIZE 511
377 /** @brief The maximum size of a data item.
379 * We only store a 32 bit value for node sizes.
381 #define MAXDATASIZE 0xffffffffUL
386 * This is used for printing a hex dump of a key's contents.
388 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
389 /** Display a key in hex.
391 * Invoke a function to display a key in hex.
393 #define DKEY(x) mdb_dkey(x, kbuf)
399 /** An invalid page number.
400 * Mainly used to denote an empty tree.
402 #define P_INVALID (~(pgno_t)0)
404 /** Test if the flags \b f are set in a flag word \b w. */
405 #define F_ISSET(w, f) (((w) & (f)) == (f))
407 /** Used for offsets within a single page.
408 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
411 typedef uint16_t indx_t;
413 /** Default size of memory map.
414 * This is certainly too small for any actual applications. Apps should always set
415 * the size explicitly using #mdb_env_set_mapsize().
417 #define DEFAULT_MAPSIZE 1048576
419 /** @defgroup readers Reader Lock Table
420 * Readers don't acquire any locks for their data access. Instead, they
421 * simply record their transaction ID in the reader table. The reader
422 * mutex is needed just to find an empty slot in the reader table. The
423 * slot's address is saved in thread-specific data so that subsequent read
424 * transactions started by the same thread need no further locking to proceed.
426 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
428 * No reader table is used if the database is on a read-only filesystem.
430 * Since the database uses multi-version concurrency control, readers don't
431 * actually need any locking. This table is used to keep track of which
432 * readers are using data from which old transactions, so that we'll know
433 * when a particular old transaction is no longer in use. Old transactions
434 * that have discarded any data pages can then have those pages reclaimed
435 * for use by a later write transaction.
437 * The lock table is constructed such that reader slots are aligned with the
438 * processor's cache line size. Any slot is only ever used by one thread.
439 * This alignment guarantees that there will be no contention or cache
440 * thrashing as threads update their own slot info, and also eliminates
441 * any need for locking when accessing a slot.
443 * A writer thread will scan every slot in the table to determine the oldest
444 * outstanding reader transaction. Any freed pages older than this will be
445 * reclaimed by the writer. The writer doesn't use any locks when scanning
446 * this table. This means that there's no guarantee that the writer will
447 * see the most up-to-date reader info, but that's not required for correct
448 * operation - all we need is to know the upper bound on the oldest reader,
449 * we don't care at all about the newest reader. So the only consequence of
450 * reading stale information here is that old pages might hang around a
451 * while longer before being reclaimed. That's actually good anyway, because
452 * the longer we delay reclaiming old pages, the more likely it is that a
453 * string of contiguous pages can be found after coalescing old pages from
454 * many old transactions together.
457 /** Number of slots in the reader table.
458 * This value was chosen somewhat arbitrarily. 126 readers plus a
459 * couple mutexes fit exactly into 8KB on my development machine.
460 * Applications should set the table size using #mdb_env_set_maxreaders().
462 #define DEFAULT_READERS 126
464 /** The size of a CPU cache line in bytes. We want our lock structures
465 * aligned to this size to avoid false cache line sharing in the
467 * This value works for most CPUs. For Itanium this should be 128.
473 /** The information we store in a single slot of the reader table.
474 * In addition to a transaction ID, we also record the process and
475 * thread ID that owns a slot, so that we can detect stale information,
476 * e.g. threads or processes that went away without cleaning up.
477 * @note We currently don't check for stale records. We simply re-init
478 * the table when we know that we're the only process opening the
481 typedef struct MDB_rxbody {
482 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
483 * Multiple readers that start at the same time will probably have the
484 * same ID here. Again, it's not important to exclude them from
485 * anything; all we need to know is which version of the DB they
486 * started from so we can avoid overwriting any data used in that
487 * particular version.
490 /** The process ID of the process owning this reader txn. */
492 /** The thread ID of the thread owning this txn. */
496 /** The actual reader record, with cacheline padding. */
497 typedef struct MDB_reader {
500 /** shorthand for mrb_txnid */
501 #define mr_txnid mru.mrx.mrb_txnid
502 #define mr_pid mru.mrx.mrb_pid
503 #define mr_tid mru.mrx.mrb_tid
504 /** cache line alignment */
505 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
509 /** The header for the reader table.
510 * The table resides in a memory-mapped file. (This is a different file
511 * than is used for the main database.)
513 * For POSIX the actual mutexes reside in the shared memory of this
514 * mapped file. On Windows, mutexes are named objects allocated by the
515 * kernel; we store the mutex names in this mapped file so that other
516 * processes can grab them. This same approach is also used on
517 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
518 * process-shared POSIX mutexes. For these cases where a named object
519 * is used, the object name is derived from a 64 bit FNV hash of the
520 * environment pathname. As such, naming collisions are extremely
521 * unlikely. If a collision occurs, the results are unpredictable.
523 typedef struct MDB_txbody {
524 /** Stamp identifying this as an MDB file. It must be set
527 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
529 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
530 char mtb_rmname[MNAME_LEN];
532 /** Mutex protecting access to this table.
533 * This is the reader lock that #LOCK_MUTEX_R acquires.
535 pthread_mutex_t mtb_mutex;
537 /** The ID of the last transaction committed to the database.
538 * This is recorded here only for convenience; the value can always
539 * be determined by reading the main database meta pages.
542 /** The number of slots that have been used in the reader table.
543 * This always records the maximum count, it is not decremented
544 * when readers release their slots.
546 unsigned mtb_numreaders;
549 /** The actual reader table definition. */
550 typedef struct MDB_txninfo {
553 #define mti_magic mt1.mtb.mtb_magic
554 #define mti_format mt1.mtb.mtb_format
555 #define mti_mutex mt1.mtb.mtb_mutex
556 #define mti_rmname mt1.mtb.mtb_rmname
557 #define mti_txnid mt1.mtb.mtb_txnid
558 #define mti_numreaders mt1.mtb.mtb_numreaders
559 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
562 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
563 char mt2_wmname[MNAME_LEN];
564 #define mti_wmname mt2.mt2_wmname
566 pthread_mutex_t mt2_wmutex;
567 #define mti_wmutex mt2.mt2_wmutex
569 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
571 MDB_reader mti_readers[1];
574 /** Lockfile format signature: version, features and field layout */
575 #define MDB_LOCK_FORMAT \
577 ((MDB_LOCK_VERSION) \
578 /* Flags which describe functionality */ \
579 + (((MDB_PIDLOCK) != 0) << 16)))
582 /** Common header for all page types.
583 * Overflow records occupy a number of contiguous pages with no
584 * headers on any page after the first.
586 typedef struct MDB_page {
587 #define mp_pgno mp_p.p_pgno
588 #define mp_next mp_p.p_next
590 pgno_t p_pgno; /**< page number */
591 void * p_next; /**< for in-memory list of freed structs */
594 /** @defgroup mdb_page Page Flags
596 * Flags for the page headers.
599 #define P_BRANCH 0x01 /**< branch page */
600 #define P_LEAF 0x02 /**< leaf page */
601 #define P_OVERFLOW 0x04 /**< overflow page */
602 #define P_META 0x08 /**< meta page */
603 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
604 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
605 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
606 #define P_KEEP 0x8000 /**< leave this page alone during spill */
608 uint16_t mp_flags; /**< @ref mdb_page */
609 #define mp_lower mp_pb.pb.pb_lower
610 #define mp_upper mp_pb.pb.pb_upper
611 #define mp_pages mp_pb.pb_pages
614 indx_t pb_lower; /**< lower bound of free space */
615 indx_t pb_upper; /**< upper bound of free space */
617 uint32_t pb_pages; /**< number of overflow pages */
619 indx_t mp_ptrs[1]; /**< dynamic size */
622 /** Size of the page header, excluding dynamic data at the end */
623 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
625 /** Address of first usable data byte in a page, after the header */
626 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
628 /** Number of nodes on a page */
629 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
631 /** The amount of space remaining in the page */
632 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
634 /** The percentage of space used in the page, in tenths of a percent. */
635 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
636 ((env)->me_psize - PAGEHDRSZ))
637 /** The minimum page fill factor, in tenths of a percent.
638 * Pages emptier than this are candidates for merging.
640 #define FILL_THRESHOLD 250
642 /** Test if a page is a leaf page */
643 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
644 /** Test if a page is a LEAF2 page */
645 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
646 /** Test if a page is a branch page */
647 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
648 /** Test if a page is an overflow page */
649 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
650 /** Test if a page is a sub page */
651 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
653 /** The number of overflow pages needed to store the given size. */
654 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
656 /** Header for a single key/data pair within a page.
657 * We guarantee 2-byte alignment for nodes.
659 typedef struct MDB_node {
660 /** lo and hi are used for data size on leaf nodes and for
661 * child pgno on branch nodes. On 64 bit platforms, flags
662 * is also used for pgno. (Branch nodes have no flags).
663 * They are in host byte order in case that lets some
664 * accesses be optimized into a 32-bit word access.
666 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
667 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
668 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
669 /** @defgroup mdb_node Node Flags
671 * Flags for node headers.
674 #define F_BIGDATA 0x01 /**< data put on overflow page */
675 #define F_SUBDATA 0x02 /**< data is a sub-database */
676 #define F_DUPDATA 0x04 /**< data has duplicates */
678 /** valid flags for #mdb_node_add() */
679 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
682 unsigned short mn_flags; /**< @ref mdb_node */
683 unsigned short mn_ksize; /**< key size */
684 char mn_data[1]; /**< key and data are appended here */
687 /** Size of the node header, excluding dynamic data at the end */
688 #define NODESIZE offsetof(MDB_node, mn_data)
690 /** Bit position of top word in page number, for shifting mn_flags */
691 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
693 /** Size of a node in a branch page with a given key.
694 * This is just the node header plus the key, there is no data.
696 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
698 /** Size of a node in a leaf page with a given key and data.
699 * This is node header plus key plus data size.
701 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
703 /** Address of node \b i in page \b p */
704 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
706 /** Address of the key for the node */
707 #define NODEKEY(node) (void *)((node)->mn_data)
709 /** Address of the data for a node */
710 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
712 /** Get the page number pointed to by a branch node */
713 #define NODEPGNO(node) \
714 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
715 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
716 /** Set the page number in a branch node */
717 #define SETPGNO(node,pgno) do { \
718 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
719 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
721 /** Get the size of the data in a leaf node */
722 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
723 /** Set the size of the data for a leaf node */
724 #define SETDSZ(node,size) do { \
725 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
726 /** The size of a key in a node */
727 #define NODEKSZ(node) ((node)->mn_ksize)
729 /** Copy a page number from src to dst */
731 #define COPY_PGNO(dst,src) dst = src
733 #if SIZE_MAX > 4294967295UL
734 #define COPY_PGNO(dst,src) do { \
735 unsigned short *s, *d; \
736 s = (unsigned short *)&(src); \
737 d = (unsigned short *)&(dst); \
744 #define COPY_PGNO(dst,src) do { \
745 unsigned short *s, *d; \
746 s = (unsigned short *)&(src); \
747 d = (unsigned short *)&(dst); \
753 /** The address of a key in a LEAF2 page.
754 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
755 * There are no node headers, keys are stored contiguously.
757 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
759 /** Set the \b node's key into \b keyptr, if requested. */
760 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
761 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
763 /** Set the \b node's key into \b key. */
764 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
766 /** Information about a single database in the environment. */
767 typedef struct MDB_db {
768 uint32_t md_pad; /**< also ksize for LEAF2 pages */
769 uint16_t md_flags; /**< @ref mdb_dbi_open */
770 uint16_t md_depth; /**< depth of this tree */
771 pgno_t md_branch_pages; /**< number of internal pages */
772 pgno_t md_leaf_pages; /**< number of leaf pages */
773 pgno_t md_overflow_pages; /**< number of overflow pages */
774 size_t md_entries; /**< number of data items */
775 pgno_t md_root; /**< the root page of this tree */
778 /** mdb_dbi_open flags */
779 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
780 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
781 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
782 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
784 /** Handle for the DB used to track free pages. */
786 /** Handle for the default DB. */
789 /** Meta page content.
790 * A meta page is the start point for accessing a database snapshot.
791 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
793 typedef struct MDB_meta {
794 /** Stamp identifying this as an MDB file. It must be set
797 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
799 void *mm_address; /**< address for fixed mapping */
800 size_t mm_mapsize; /**< size of mmap region */
801 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
802 /** The size of pages used in this DB */
803 #define mm_psize mm_dbs[0].md_pad
804 /** Any persistent environment flags. @ref mdb_env */
805 #define mm_flags mm_dbs[0].md_flags
806 pgno_t mm_last_pg; /**< last used page in file */
807 txnid_t mm_txnid; /**< txnid that committed this page */
810 /** Buffer for a stack-allocated dirty page.
811 * The members define size and alignment, and silence type
812 * aliasing warnings. They are not used directly; that could
813 * mean incorrectly using several union members in parallel.
815 typedef union MDB_pagebuf {
816 char mb_raw[MDB_PAGESIZE];
819 char mm_pad[PAGEHDRSZ];
824 /** Auxiliary DB info.
825 * The information here is mostly static/read-only. There is
826 * only a single copy of this record in the environment.
828 typedef struct MDB_dbx {
829 MDB_val md_name; /**< name of the database */
830 MDB_cmp_func *md_cmp; /**< function for comparing keys */
831 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
832 MDB_rel_func *md_rel; /**< user relocate function */
833 void *md_relctx; /**< user-provided context for md_rel */
836 /** A database transaction.
837 * Every operation requires a transaction handle.
840 MDB_txn *mt_parent; /**< parent of a nested txn */
841 MDB_txn *mt_child; /**< nested txn under this txn */
842 pgno_t mt_next_pgno; /**< next unallocated page */
843 /** The ID of this transaction. IDs are integers incrementing from 1.
844 * Only committed write transactions increment the ID. If a transaction
845 * aborts, the ID may be re-used by the next writer.
848 MDB_env *mt_env; /**< the DB environment */
849 /** The list of pages that became unused during this transaction.
852 /** The sorted list of dirty pages we temporarily wrote to disk
853 * because the dirty list was full. page numbers in here are
854 * shifted left by 1, deleted slots have the LSB set.
856 MDB_IDL mt_spill_pgs;
858 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
860 /** For read txns: This thread/txn's reader table slot, or NULL. */
863 /** Array of records for each DB known in the environment. */
865 /** Array of MDB_db records for each known DB */
867 /** @defgroup mt_dbflag Transaction DB Flags
871 #define DB_DIRTY 0x01 /**< DB was written in this txn */
872 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
873 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
874 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
876 /** In write txns, array of cursors for each DB */
877 MDB_cursor **mt_cursors;
878 /** Array of flags for each DB */
879 unsigned char *mt_dbflags;
880 /** Number of DB records in use. This number only ever increments;
881 * we don't decrement it when individual DB handles are closed.
885 /** @defgroup mdb_txn Transaction Flags
889 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
890 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
891 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
892 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
894 unsigned int mt_flags; /**< @ref mdb_txn */
895 /** dirty_list room: Array size - #dirty pages visible to this txn.
896 * Includes ancestor txns' dirty pages not hidden by other txns'
897 * dirty/spilled pages. Thus commit(nested txn) has room to merge
898 * dirty_list into mt_parent after freeing hidden mt_parent pages.
900 unsigned int mt_dirty_room;
903 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
904 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
905 * raise this on a 64 bit machine.
907 #define CURSOR_STACK 32
911 /** Cursors are used for all DB operations.
912 * A cursor holds a path of (page pointer, key index) from the DB
913 * root to a position in the DB, plus other state. #MDB_DUPSORT
914 * cursors include an xcursor to the current data item. Write txns
915 * track their cursors and keep them up to date when data moves.
916 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
917 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
920 /** Next cursor on this DB in this txn */
922 /** Backup of the original cursor if this cursor is a shadow */
923 MDB_cursor *mc_backup;
924 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
925 struct MDB_xcursor *mc_xcursor;
926 /** The transaction that owns this cursor */
928 /** The database handle this cursor operates on */
930 /** The database record for this cursor */
932 /** The database auxiliary record for this cursor */
934 /** The @ref mt_dbflag for this database */
935 unsigned char *mc_dbflag;
936 unsigned short mc_snum; /**< number of pushed pages */
937 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
938 /** @defgroup mdb_cursor Cursor Flags
940 * Cursor state flags.
943 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
944 #define C_EOF 0x02 /**< No more data */
945 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
946 #define C_DEL 0x08 /**< last op was a cursor_del */
947 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
948 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
950 unsigned int mc_flags; /**< @ref mdb_cursor */
951 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
952 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
955 /** Context for sorted-dup records.
956 * We could have gone to a fully recursive design, with arbitrarily
957 * deep nesting of sub-databases. But for now we only handle these
958 * levels - main DB, optional sub-DB, sorted-duplicate DB.
960 typedef struct MDB_xcursor {
961 /** A sub-cursor for traversing the Dup DB */
962 MDB_cursor mx_cursor;
963 /** The database record for this Dup DB */
965 /** The auxiliary DB record for this Dup DB */
967 /** The @ref mt_dbflag for this Dup DB */
968 unsigned char mx_dbflag;
971 /** State of FreeDB old pages, stored in the MDB_env */
972 typedef struct MDB_pgstate {
973 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
974 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
977 /** The database environment. */
979 HANDLE me_fd; /**< The main data file */
980 HANDLE me_lfd; /**< The lock file */
981 HANDLE me_mfd; /**< just for writing the meta pages */
982 /** Failed to update the meta page. Probably an I/O error. */
983 #define MDB_FATAL_ERROR 0x80000000U
984 /** Some fields are initialized. */
985 #define MDB_ENV_ACTIVE 0x20000000U
986 /** me_txkey is set */
987 #define MDB_ENV_TXKEY 0x10000000U
988 /** Have liveness lock in reader table */
989 #define MDB_LIVE_READER 0x08000000U
990 uint32_t me_flags; /**< @ref mdb_env */
991 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
992 unsigned int me_maxreaders; /**< size of the reader table */
993 unsigned int me_numreaders; /**< max numreaders set by this env */
994 MDB_dbi me_numdbs; /**< number of DBs opened */
995 MDB_dbi me_maxdbs; /**< size of the DB table */
996 pid_t me_pid; /**< process ID of this env */
997 char *me_path; /**< path to the DB files */
998 char *me_map; /**< the memory map of the data file */
999 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1000 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1001 MDB_txn *me_txn; /**< current write transaction */
1002 size_t me_mapsize; /**< size of the data memory map */
1003 off_t me_size; /**< current file size */
1004 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1005 MDB_dbx *me_dbxs; /**< array of static DB info */
1006 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1007 pthread_key_t me_txkey; /**< thread-key for readers */
1008 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1009 # define me_pglast me_pgstate.mf_pglast
1010 # define me_pghead me_pgstate.mf_pghead
1011 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1012 /** IDL of pages that became unused in a write txn */
1013 MDB_IDL me_free_pgs;
1014 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1015 MDB_ID2L me_dirty_list;
1016 /** Max number of freelist items that can fit in a single overflow page */
1018 /** Max size of a node on a page */
1019 unsigned int me_nodemax;
1021 int me_pidquery; /**< Used in OpenProcess */
1022 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1024 #elif defined(MDB_USE_POSIX_SEM)
1025 sem_t *me_rmutex; /* Shared mutexes are not supported */
1030 /** Nested transaction */
1031 typedef struct MDB_ntxn {
1032 MDB_txn mnt_txn; /**< the transaction */
1033 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1036 /** max number of pages to commit in one writev() call */
1037 #define MDB_COMMIT_PAGES 64
1038 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1039 #undef MDB_COMMIT_PAGES
1040 #define MDB_COMMIT_PAGES IOV_MAX
1043 /* max bytes to write in one call */
1044 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1046 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1047 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1048 static int mdb_page_touch(MDB_cursor *mc);
1050 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1051 static int mdb_page_search_root(MDB_cursor *mc,
1052 MDB_val *key, int modify);
1053 #define MDB_PS_MODIFY 1
1054 #define MDB_PS_ROOTONLY 2
1055 #define MDB_PS_FIRST 4
1056 #define MDB_PS_LAST 8
1057 static int mdb_page_search(MDB_cursor *mc,
1058 MDB_val *key, int flags);
1059 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1061 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1062 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1063 pgno_t newpgno, unsigned int nflags);
1065 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1066 static int mdb_env_pick_meta(const MDB_env *env);
1067 static int mdb_env_write_meta(MDB_txn *txn);
1068 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1069 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1071 static void mdb_env_close0(MDB_env *env, int excl);
1073 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1074 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1075 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1076 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1077 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1078 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1079 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1080 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1081 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1083 static int mdb_rebalance(MDB_cursor *mc);
1084 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1086 static void mdb_cursor_pop(MDB_cursor *mc);
1087 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1089 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1090 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1091 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1092 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1093 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1095 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1096 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1098 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1099 static void mdb_xcursor_init0(MDB_cursor *mc);
1100 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1102 static int mdb_drop0(MDB_cursor *mc, int subs);
1103 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1106 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1110 static SECURITY_DESCRIPTOR mdb_null_sd;
1111 static SECURITY_ATTRIBUTES mdb_all_sa;
1112 static int mdb_sec_inited;
1115 /** Return the library version info. */
1117 mdb_version(int *major, int *minor, int *patch)
1119 if (major) *major = MDB_VERSION_MAJOR;
1120 if (minor) *minor = MDB_VERSION_MINOR;
1121 if (patch) *patch = MDB_VERSION_PATCH;
1122 return MDB_VERSION_STRING;
1125 /** Table of descriptions for MDB @ref errors */
1126 static char *const mdb_errstr[] = {
1127 "MDB_KEYEXIST: Key/data pair already exists",
1128 "MDB_NOTFOUND: No matching key/data pair found",
1129 "MDB_PAGE_NOTFOUND: Requested page not found",
1130 "MDB_CORRUPTED: Located page was wrong type",
1131 "MDB_PANIC: Update of meta page failed",
1132 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1133 "MDB_INVALID: File is not an MDB file",
1134 "MDB_MAP_FULL: Environment mapsize limit reached",
1135 "MDB_DBS_FULL: Environment maxdbs limit reached",
1136 "MDB_READERS_FULL: Environment maxreaders limit reached",
1137 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1138 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1139 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1140 "MDB_PAGE_FULL: Internal error - page has no more space",
1141 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1142 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1143 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1144 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1145 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1149 mdb_strerror(int err)
1153 return ("Successful return: 0");
1155 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1156 i = err - MDB_KEYEXIST;
1157 return mdb_errstr[i];
1160 return strerror(err);
1164 /** Display a key in hexadecimal and return the address of the result.
1165 * @param[in] key the key to display
1166 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1167 * @return The key in hexadecimal form.
1170 mdb_dkey(MDB_val *key, char *buf)
1173 unsigned char *c = key->mv_data;
1179 if (key->mv_size > MDB_MAXKEYSIZE)
1180 return "MDB_MAXKEYSIZE";
1181 /* may want to make this a dynamic check: if the key is mostly
1182 * printable characters, print it as-is instead of converting to hex.
1186 for (i=0; i<key->mv_size; i++)
1187 ptr += sprintf(ptr, "%02x", *c++);
1189 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1194 /** Display all the keys in the page. */
1196 mdb_page_list(MDB_page *mp)
1199 unsigned int i, nkeys, nsize;
1203 nkeys = NUMKEYS(mp);
1204 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1205 for (i=0; i<nkeys; i++) {
1206 node = NODEPTR(mp, i);
1207 key.mv_size = node->mn_ksize;
1208 key.mv_data = node->mn_data;
1209 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1210 if (IS_BRANCH(mp)) {
1211 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1214 if (F_ISSET(node->mn_flags, F_BIGDATA))
1215 nsize += sizeof(pgno_t);
1217 nsize += NODEDSZ(node);
1218 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1224 mdb_cursor_chk(MDB_cursor *mc)
1230 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1231 for (i=0; i<mc->mc_top; i++) {
1233 node = NODEPTR(mp, mc->mc_ki[i]);
1234 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1237 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1243 /** Count all the pages in each DB and in the freelist
1244 * and make sure it matches the actual number of pages
1247 static void mdb_audit(MDB_txn *txn)
1251 MDB_ID freecount, count;
1256 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1257 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1258 freecount += *(MDB_ID *)data.mv_data;
1261 for (i = 0; i<txn->mt_numdbs; i++) {
1263 mdb_cursor_init(&mc, txn, i, &mx);
1264 if (txn->mt_dbs[i].md_root == P_INVALID)
1266 count += txn->mt_dbs[i].md_branch_pages +
1267 txn->mt_dbs[i].md_leaf_pages +
1268 txn->mt_dbs[i].md_overflow_pages;
1269 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1270 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1274 mp = mc.mc_pg[mc.mc_top];
1275 for (j=0; j<NUMKEYS(mp); j++) {
1276 MDB_node *leaf = NODEPTR(mp, j);
1277 if (leaf->mn_flags & F_SUBDATA) {
1279 memcpy(&db, NODEDATA(leaf), sizeof(db));
1280 count += db.md_branch_pages + db.md_leaf_pages +
1281 db.md_overflow_pages;
1285 while (mdb_cursor_sibling(&mc, 1) == 0);
1288 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1289 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1290 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1296 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1298 return txn->mt_dbxs[dbi].md_cmp(a, b);
1302 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1304 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1307 /** Allocate memory for a page.
1308 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1311 mdb_page_malloc(MDB_txn *txn, unsigned num)
1313 MDB_env *env = txn->mt_env;
1314 MDB_page *ret = env->me_dpages;
1315 size_t sz = env->me_psize;
1318 VGMEMP_ALLOC(env, ret, sz);
1319 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1320 env->me_dpages = ret->mp_next;
1326 if ((ret = malloc(sz)) != NULL) {
1327 VGMEMP_ALLOC(env, ret, sz);
1332 /** Free a single page.
1333 * Saves single pages to a list, for future reuse.
1334 * (This is not used for multi-page overflow pages.)
1337 mdb_page_free(MDB_env *env, MDB_page *mp)
1339 mp->mp_next = env->me_dpages;
1340 VGMEMP_FREE(env, mp);
1341 env->me_dpages = mp;
1344 /** Free a dirty page */
1346 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1348 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1349 mdb_page_free(env, dp);
1351 /* large pages just get freed directly */
1352 VGMEMP_FREE(env, dp);
1357 /** Return all dirty pages to dpage list */
1359 mdb_dlist_free(MDB_txn *txn)
1361 MDB_env *env = txn->mt_env;
1362 MDB_ID2L dl = txn->mt_u.dirty_list;
1363 unsigned i, n = dl[0].mid;
1365 for (i = 1; i <= n; i++) {
1366 mdb_dpage_free(env, dl[i].mptr);
1371 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1372 * @param[in] mc A cursor handle for the current operation.
1373 * @param[in] pflags Flags of the pages to update:
1374 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1375 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1376 * @return 0 on success, non-zero on failure.
1379 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1381 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1382 MDB_txn *txn = mc->mc_txn;
1388 int rc = MDB_SUCCESS, level;
1390 /* Mark pages seen by cursors */
1391 if (mc->mc_flags & C_UNTRACK)
1392 mc = NULL; /* will find mc in mt_cursors */
1393 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1394 for (; mc; mc=mc->mc_next) {
1395 if (!(mc->mc_flags & C_INITIALIZED))
1397 for (m3 = mc;; m3 = &mx->mx_cursor) {
1399 for (j=0; j<m3->mc_snum; j++) {
1401 if ((mp->mp_flags & Mask) == pflags)
1402 mp->mp_flags ^= P_KEEP;
1404 mx = m3->mc_xcursor;
1405 /* Proceed to mx if it is at a sub-database */
1406 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1408 if (! (mp && (mp->mp_flags & P_LEAF)))
1410 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1411 if (!(leaf->mn_flags & F_SUBDATA))
1420 /* Mark dirty root pages */
1421 for (i=0; i<txn->mt_numdbs; i++) {
1422 if (txn->mt_dbflags[i] & DB_DIRTY) {
1423 pgno_t pgno = txn->mt_dbs[i].md_root;
1424 if (pgno == P_INVALID)
1426 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1428 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1429 dp->mp_flags ^= P_KEEP;
1437 static int mdb_page_flush(MDB_txn *txn, int keep);
1439 /** Spill pages from the dirty list back to disk.
1440 * This is intended to prevent running into #MDB_TXN_FULL situations,
1441 * but note that they may still occur in a few cases:
1442 * 1) our estimate of the txn size could be too small. Currently this
1443 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1444 * 2) child txns may run out of space if their parents dirtied a
1445 * lot of pages and never spilled them. TODO: we probably should do
1446 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1447 * the parent's dirty_room is below a given threshold.
1449 * Otherwise, if not using nested txns, it is expected that apps will
1450 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1451 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1452 * If the txn never references them again, they can be left alone.
1453 * If the txn only reads them, they can be used without any fuss.
1454 * If the txn writes them again, they can be dirtied immediately without
1455 * going thru all of the work of #mdb_page_touch(). Such references are
1456 * handled by #mdb_page_unspill().
1458 * Also note, we never spill DB root pages, nor pages of active cursors,
1459 * because we'll need these back again soon anyway. And in nested txns,
1460 * we can't spill a page in a child txn if it was already spilled in a
1461 * parent txn. That would alter the parent txns' data even though
1462 * the child hasn't committed yet, and we'd have no way to undo it if
1463 * the child aborted.
1465 * @param[in] m0 cursor A cursor handle identifying the transaction and
1466 * database for which we are checking space.
1467 * @param[in] key For a put operation, the key being stored.
1468 * @param[in] data For a put operation, the data being stored.
1469 * @return 0 on success, non-zero on failure.
1472 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1474 MDB_txn *txn = m0->mc_txn;
1476 MDB_ID2L dl = txn->mt_u.dirty_list;
1477 unsigned int i, j, need;
1480 if (m0->mc_flags & C_SUB)
1483 /* Estimate how much space this op will take */
1484 i = m0->mc_db->md_depth;
1485 /* Named DBs also dirty the main DB */
1486 if (m0->mc_dbi > MAIN_DBI)
1487 i += txn->mt_dbs[MAIN_DBI].md_depth;
1488 /* For puts, roughly factor in the key+data size */
1490 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1491 i += i; /* double it for good measure */
1494 if (txn->mt_dirty_room > i)
1497 if (!txn->mt_spill_pgs) {
1498 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1499 if (!txn->mt_spill_pgs)
1502 /* purge deleted slots */
1503 MDB_IDL sl = txn->mt_spill_pgs;
1504 unsigned int num = sl[0];
1506 for (i=1; i<=num; i++) {
1513 /* Preserve pages which may soon be dirtied again */
1514 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1517 /* Less aggressive spill - we originally spilled the entire dirty list,
1518 * with a few exceptions for cursor pages and DB root pages. But this
1519 * turns out to be a lot of wasted effort because in a large txn many
1520 * of those pages will need to be used again. So now we spill only 1/8th
1521 * of the dirty pages. Testing revealed this to be a good tradeoff,
1522 * better than 1/2, 1/4, or 1/10.
1524 if (need < MDB_IDL_UM_MAX / 8)
1525 need = MDB_IDL_UM_MAX / 8;
1527 /* Save the page IDs of all the pages we're flushing */
1528 /* flush from the tail forward, this saves a lot of shifting later on. */
1529 for (i=dl[0].mid; i && need; i--) {
1530 MDB_ID pn = dl[i].mid << 1;
1532 if (dp->mp_flags & P_KEEP)
1534 /* Can't spill twice, make sure it's not already in a parent's
1537 if (txn->mt_parent) {
1539 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1540 if (tx2->mt_spill_pgs) {
1541 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1542 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1543 dp->mp_flags |= P_KEEP;
1551 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1555 mdb_midl_sort(txn->mt_spill_pgs);
1557 /* Flush the spilled part of dirty list */
1558 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1561 /* Reset any dirty pages we kept that page_flush didn't see */
1562 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1565 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1569 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1571 mdb_find_oldest(MDB_txn *txn)
1574 txnid_t mr, oldest = txn->mt_txnid - 1;
1575 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1576 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1586 /** Add a page to the txn's dirty list */
1588 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1591 int (*insert)(MDB_ID2L, MDB_ID2 *);
1593 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1594 insert = mdb_mid2l_append;
1596 insert = mdb_mid2l_insert;
1598 mid.mid = mp->mp_pgno;
1600 insert(txn->mt_u.dirty_list, &mid);
1601 txn->mt_dirty_room--;
1604 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1605 * me_pghead and mt_next_pgno.
1607 * If there are free pages available from older transactions, they
1608 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1609 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1610 * and move me_pglast to say which records were consumed. Only this
1611 * function can create me_pghead and move me_pglast/mt_next_pgno.
1612 * @param[in] mc cursor A cursor handle identifying the transaction and
1613 * database for which we are allocating.
1614 * @param[in] num the number of pages to allocate.
1615 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1616 * will always be satisfied by a single contiguous chunk of memory.
1617 * @return 0 on success, non-zero on failure.
1620 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1622 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1623 /* Get at most <Max_retries> more freeDB records once me_pghead
1624 * has enough pages. If not enough, use new pages from the map.
1625 * If <Paranoid> and mc is updating the freeDB, only get new
1626 * records if me_pghead is empty. Then the freelist cannot play
1627 * catch-up with itself by growing while trying to save it.
1629 enum { Paranoid = 1, Max_retries = 500 };
1631 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1633 int rc, n2 = num-1, retry = Max_retries;
1634 MDB_txn *txn = mc->mc_txn;
1635 MDB_env *env = txn->mt_env;
1636 pgno_t pgno, *mop = env->me_pghead;
1637 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1639 txnid_t oldest = 0, last;
1645 /* If our dirty list is already full, we can't do anything */
1646 if (txn->mt_dirty_room == 0)
1647 return MDB_TXN_FULL;
1649 for (op = MDB_FIRST;; op = MDB_NEXT) {
1652 pgno_t *idl, old_id, new_id;
1654 /* Seek a big enough contiguous page range. Prefer
1655 * pages at the tail, just truncating the list.
1657 if (mop_len >= (unsigned)num) {
1661 if (mop[i-n2] == pgno+n2)
1663 } while (--i >= (unsigned)num);
1664 if (Max_retries < INT_MAX && --retry < 0)
1668 if (op == MDB_FIRST) { /* 1st iteration */
1669 /* Prepare to fetch more and coalesce */
1670 oldest = mdb_find_oldest(txn);
1671 last = env->me_pglast;
1672 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1675 key.mv_data = &last; /* will look up last+1 */
1676 key.mv_size = sizeof(last);
1678 if (Paranoid && mc->mc_dbi == FREE_DBI)
1681 if (Paranoid && retry < 0 && mop_len)
1685 /* Do not fetch more if the record will be too recent */
1688 rc = mdb_cursor_get(&m2, &key, NULL, op);
1690 if (rc == MDB_NOTFOUND)
1694 last = *(txnid_t*)key.mv_data;
1697 np = m2.mc_pg[m2.mc_top];
1698 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1699 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1702 idl = (MDB_ID *) data.mv_data;
1705 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1708 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1710 mop = env->me_pghead;
1712 env->me_pglast = last;
1714 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1715 last, txn->mt_dbs[FREE_DBI].md_root, i));
1717 DPRINTF(("IDL %"Z"u", idl[k]));
1719 /* Merge in descending sorted order */
1722 mop[0] = (pgno_t)-1;
1726 for (; old_id < new_id; old_id = mop[--j])
1733 /* Use new pages from the map when nothing suitable in the freeDB */
1735 pgno = txn->mt_next_pgno;
1736 if (pgno + num >= env->me_maxpg) {
1737 DPUTS("DB size maxed out");
1738 return MDB_MAP_FULL;
1742 if (env->me_flags & MDB_WRITEMAP) {
1743 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1745 if (!(np = mdb_page_malloc(txn, num)))
1749 mop[0] = mop_len -= num;
1750 /* Move any stragglers down */
1751 for (j = i-num; j < mop_len; )
1752 mop[++j] = mop[++i];
1754 txn->mt_next_pgno = pgno + num;
1757 mdb_page_dirty(txn, np);
1763 /** Copy the used portions of a non-overflow page.
1764 * @param[in] dst page to copy into
1765 * @param[in] src page to copy from
1766 * @param[in] psize size of a page
1769 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1771 enum { Align = sizeof(pgno_t) };
1772 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1774 /* If page isn't full, just copy the used portion. Adjust
1775 * alignment so memcpy may copy words instead of bytes.
1777 if ((unused &= -Align) && !IS_LEAF2(src)) {
1779 memcpy(dst, src, (lower + (Align-1)) & -Align);
1780 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1783 memcpy(dst, src, psize - unused);
1787 /** Pull a page off the txn's spill list, if present.
1788 * If a page being referenced was spilled to disk in this txn, bring
1789 * it back and make it dirty/writable again.
1790 * @param[in] txn the transaction handle.
1791 * @param[in] mp the page being referenced.
1792 * @param[out] ret the writable page, if any. ret is unchanged if
1793 * mp wasn't spilled.
1796 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1798 MDB_env *env = txn->mt_env;
1801 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1803 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1804 if (!tx2->mt_spill_pgs)
1806 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1807 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1810 if (txn->mt_dirty_room == 0)
1811 return MDB_TXN_FULL;
1812 if (IS_OVERFLOW(mp))
1816 if (env->me_flags & MDB_WRITEMAP) {
1819 np = mdb_page_malloc(txn, num);
1823 memcpy(np, mp, num * env->me_psize);
1825 mdb_page_copy(np, mp, env->me_psize);
1828 /* If in current txn, this page is no longer spilled.
1829 * If it happens to be the last page, truncate the spill list.
1830 * Otherwise mark it as deleted by setting the LSB.
1832 if (x == txn->mt_spill_pgs[0])
1833 txn->mt_spill_pgs[0]--;
1835 txn->mt_spill_pgs[x] |= 1;
1836 } /* otherwise, if belonging to a parent txn, the
1837 * page remains spilled until child commits
1840 mdb_page_dirty(txn, np);
1841 np->mp_flags |= P_DIRTY;
1849 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1850 * @param[in] mc cursor pointing to the page to be touched
1851 * @return 0 on success, non-zero on failure.
1854 mdb_page_touch(MDB_cursor *mc)
1856 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1857 MDB_txn *txn = mc->mc_txn;
1858 MDB_cursor *m2, *m3;
1863 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1864 if (txn->mt_flags & MDB_TXN_SPILLS) {
1866 rc = mdb_page_unspill(txn, mp, &np);
1872 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1873 (rc = mdb_page_alloc(mc, 1, &np)))
1876 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1877 assert(mp->mp_pgno != pgno);
1878 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1879 /* Update the parent page, if any, to point to the new page */
1881 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1882 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1883 SETPGNO(node, pgno);
1885 mc->mc_db->md_root = pgno;
1887 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1888 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1890 /* If txn has a parent, make sure the page is in our
1894 unsigned x = mdb_mid2l_search(dl, pgno);
1895 if (x <= dl[0].mid && dl[x].mid == pgno) {
1896 if (mp != dl[x].mptr) { /* bad cursor? */
1897 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1898 return MDB_CORRUPTED;
1903 assert(dl[0].mid < MDB_IDL_UM_MAX);
1905 np = mdb_page_malloc(txn, 1);
1910 mdb_mid2l_insert(dl, &mid);
1915 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1917 np->mp_flags |= P_DIRTY;
1920 /* Adjust cursors pointing to mp */
1921 mc->mc_pg[mc->mc_top] = np;
1923 if (mc->mc_flags & C_SUB) {
1925 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1926 m3 = &m2->mc_xcursor->mx_cursor;
1927 if (m3->mc_snum < mc->mc_snum) continue;
1928 if (m3->mc_pg[mc->mc_top] == mp)
1929 m3->mc_pg[mc->mc_top] = np;
1932 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1933 if (m2->mc_snum < mc->mc_snum) continue;
1934 if (m2->mc_pg[mc->mc_top] == mp) {
1935 m2->mc_pg[mc->mc_top] = np;
1936 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1937 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1939 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1940 if (!(leaf->mn_flags & F_SUBDATA))
1941 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1950 mdb_env_sync(MDB_env *env, int force)
1953 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1954 if (env->me_flags & MDB_WRITEMAP) {
1955 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1956 ? MS_ASYNC : MS_SYNC;
1957 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1960 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1964 if (MDB_FDATASYNC(env->me_fd))
1971 /** Back up parent txn's cursors, then grab the originals for tracking */
1973 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1975 MDB_cursor *mc, *bk;
1980 for (i = src->mt_numdbs; --i >= 0; ) {
1981 if ((mc = src->mt_cursors[i]) != NULL) {
1982 size = sizeof(MDB_cursor);
1984 size += sizeof(MDB_xcursor);
1985 for (; mc; mc = bk->mc_next) {
1991 mc->mc_db = &dst->mt_dbs[i];
1992 /* Kill pointers into src - and dst to reduce abuse: The
1993 * user may not use mc until dst ends. Otherwise we'd...
1995 mc->mc_txn = NULL; /* ...set this to dst */
1996 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1997 if ((mx = mc->mc_xcursor) != NULL) {
1998 *(MDB_xcursor *)(bk+1) = *mx;
1999 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2001 mc->mc_next = dst->mt_cursors[i];
2002 dst->mt_cursors[i] = mc;
2009 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2010 * @param[in] txn the transaction handle.
2011 * @param[in] merge true to keep changes to parent cursors, false to revert.
2012 * @return 0 on success, non-zero on failure.
2015 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2017 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2021 for (i = txn->mt_numdbs; --i >= 0; ) {
2022 for (mc = cursors[i]; mc; mc = next) {
2024 if ((bk = mc->mc_backup) != NULL) {
2026 /* Commit changes to parent txn */
2027 mc->mc_next = bk->mc_next;
2028 mc->mc_backup = bk->mc_backup;
2029 mc->mc_txn = bk->mc_txn;
2030 mc->mc_db = bk->mc_db;
2031 mc->mc_dbflag = bk->mc_dbflag;
2032 if ((mx = mc->mc_xcursor) != NULL)
2033 mx->mx_cursor.mc_txn = bk->mc_txn;
2035 /* Abort nested txn */
2037 if ((mx = mc->mc_xcursor) != NULL)
2038 *mx = *(MDB_xcursor *)(bk+1);
2042 /* Only malloced cursors are permanently tracked. */
2050 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2053 mdb_txn_reset0(MDB_txn *txn, const char *act);
2055 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2061 Pidset = F_SETLK, Pidcheck = F_GETLK
2065 /** Set or check a pid lock. Set returns 0 on success.
2066 * Check returns 0 if the process is certainly dead, nonzero if it may
2067 * be alive (the lock exists or an error happened so we do not know).
2069 * On Windows Pidset is a no-op, we merely check for the existence
2070 * of the process with the given pid. On POSIX we use a single byte
2071 * lock on the lockfile, set at an offset equal to the pid.
2074 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2076 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2079 if (op == Pidcheck) {
2080 h = OpenProcess(env->me_pidquery, FALSE, pid);
2081 /* No documented "no such process" code, but other program use this: */
2083 return ErrCode() != ERROR_INVALID_PARAMETER;
2084 /* A process exists until all handles to it close. Has it exited? */
2085 ret = WaitForSingleObject(h, 0) != 0;
2092 struct flock lock_info;
2093 memset(&lock_info, 0, sizeof(lock_info));
2094 lock_info.l_type = F_WRLCK;
2095 lock_info.l_whence = SEEK_SET;
2096 lock_info.l_start = pid;
2097 lock_info.l_len = 1;
2098 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2099 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2101 } else if ((rc = ErrCode()) == EINTR) {
2109 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2110 * @param[in] txn the transaction handle to initialize
2111 * @return 0 on success, non-zero on failure.
2114 mdb_txn_renew0(MDB_txn *txn)
2116 MDB_env *env = txn->mt_env;
2120 int rc, new_notls = 0;
2123 txn->mt_numdbs = env->me_numdbs;
2124 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2126 if (txn->mt_flags & MDB_TXN_RDONLY) {
2127 if (!env->me_txns) {
2128 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2129 txn->mt_txnid = meta->mm_txnid;
2130 txn->mt_u.reader = NULL;
2132 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2133 pthread_getspecific(env->me_txkey);
2135 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2136 return MDB_BAD_RSLOT;
2138 pid_t pid = env->me_pid;
2139 pthread_t tid = pthread_self();
2141 if (!(env->me_flags & MDB_LIVE_READER)) {
2142 rc = mdb_reader_pid(env, Pidset, pid);
2144 UNLOCK_MUTEX_R(env);
2147 env->me_flags |= MDB_LIVE_READER;
2151 for (i=0; i<env->me_txns->mti_numreaders; i++)
2152 if (env->me_txns->mti_readers[i].mr_pid == 0)
2154 if (i == env->me_maxreaders) {
2155 UNLOCK_MUTEX_R(env);
2156 return MDB_READERS_FULL;
2158 env->me_txns->mti_readers[i].mr_pid = pid;
2159 env->me_txns->mti_readers[i].mr_tid = tid;
2160 if (i >= env->me_txns->mti_numreaders)
2161 env->me_txns->mti_numreaders = i+1;
2162 /* Save numreaders for un-mutexed mdb_env_close() */
2163 env->me_numreaders = env->me_txns->mti_numreaders;
2164 UNLOCK_MUTEX_R(env);
2165 r = &env->me_txns->mti_readers[i];
2166 new_notls = (env->me_flags & MDB_NOTLS);
2167 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2172 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2173 txn->mt_u.reader = r;
2174 meta = env->me_metas[txn->mt_txnid & 1];
2179 txn->mt_txnid = env->me_txns->mti_txnid;
2180 meta = env->me_metas[txn->mt_txnid & 1];
2183 if (txn->mt_txnid == mdb_debug_start)
2186 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2187 txn->mt_u.dirty_list = env->me_dirty_list;
2188 txn->mt_u.dirty_list[0].mid = 0;
2189 txn->mt_free_pgs = env->me_free_pgs;
2190 txn->mt_free_pgs[0] = 0;
2191 txn->mt_spill_pgs = NULL;
2195 /* Copy the DB info and flags */
2196 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2198 /* Moved to here to avoid a data race in read TXNs */
2199 txn->mt_next_pgno = meta->mm_last_pg+1;
2201 for (i=2; i<txn->mt_numdbs; i++) {
2202 x = env->me_dbflags[i];
2203 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2204 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2206 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2208 if (env->me_maxpg < txn->mt_next_pgno) {
2209 mdb_txn_reset0(txn, "renew0-mapfail");
2211 txn->mt_u.reader->mr_pid = 0;
2212 txn->mt_u.reader = NULL;
2214 return MDB_MAP_RESIZED;
2221 mdb_txn_renew(MDB_txn *txn)
2225 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2228 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2229 DPUTS("environment had fatal error, must shutdown!");
2233 rc = mdb_txn_renew0(txn);
2234 if (rc == MDB_SUCCESS) {
2235 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2236 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2237 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2243 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2247 int rc, size, tsize = sizeof(MDB_txn);
2249 if (env->me_flags & MDB_FATAL_ERROR) {
2250 DPUTS("environment had fatal error, must shutdown!");
2253 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2256 /* Nested transactions: Max 1 child, write txns only, no writemap */
2257 if (parent->mt_child ||
2258 (flags & MDB_RDONLY) ||
2259 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2260 (env->me_flags & MDB_WRITEMAP))
2262 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2264 tsize = sizeof(MDB_ntxn);
2266 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2267 if (!(flags & MDB_RDONLY))
2268 size += env->me_maxdbs * sizeof(MDB_cursor *);
2270 if ((txn = calloc(1, size)) == NULL) {
2271 DPRINTF(("calloc: %s", strerror(ErrCode())));
2274 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2275 if (flags & MDB_RDONLY) {
2276 txn->mt_flags |= MDB_TXN_RDONLY;
2277 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2279 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2280 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2286 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2287 if (!txn->mt_u.dirty_list ||
2288 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2290 free(txn->mt_u.dirty_list);
2294 txn->mt_txnid = parent->mt_txnid;
2295 txn->mt_dirty_room = parent->mt_dirty_room;
2296 txn->mt_u.dirty_list[0].mid = 0;
2297 txn->mt_spill_pgs = NULL;
2298 txn->mt_next_pgno = parent->mt_next_pgno;
2299 parent->mt_child = txn;
2300 txn->mt_parent = parent;
2301 txn->mt_numdbs = parent->mt_numdbs;
2302 txn->mt_flags = parent->mt_flags;
2303 txn->mt_dbxs = parent->mt_dbxs;
2304 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2305 /* Copy parent's mt_dbflags, but clear DB_NEW */
2306 for (i=0; i<txn->mt_numdbs; i++)
2307 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2309 ntxn = (MDB_ntxn *)txn;
2310 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2311 if (env->me_pghead) {
2312 size = MDB_IDL_SIZEOF(env->me_pghead);
2313 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2315 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2320 rc = mdb_cursor_shadow(parent, txn);
2322 mdb_txn_reset0(txn, "beginchild-fail");
2324 rc = mdb_txn_renew0(txn);
2330 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2331 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2332 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2339 mdb_txn_env(MDB_txn *txn)
2341 if(!txn) return NULL;
2345 /** Export or close DBI handles opened in this txn. */
2347 mdb_dbis_update(MDB_txn *txn, int keep)
2350 MDB_dbi n = txn->mt_numdbs;
2351 MDB_env *env = txn->mt_env;
2352 unsigned char *tdbflags = txn->mt_dbflags;
2354 for (i = n; --i >= 2;) {
2355 if (tdbflags[i] & DB_NEW) {
2357 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2359 char *ptr = env->me_dbxs[i].md_name.mv_data;
2360 env->me_dbxs[i].md_name.mv_data = NULL;
2361 env->me_dbxs[i].md_name.mv_size = 0;
2362 env->me_dbflags[i] = 0;
2367 if (keep && env->me_numdbs < n)
2371 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2372 * May be called twice for readonly txns: First reset it, then abort.
2373 * @param[in] txn the transaction handle to reset
2374 * @param[in] act why the transaction is being reset
2377 mdb_txn_reset0(MDB_txn *txn, const char *act)
2379 MDB_env *env = txn->mt_env;
2381 /* Close any DBI handles opened in this txn */
2382 mdb_dbis_update(txn, 0);
2384 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2385 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2386 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2388 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2389 if (txn->mt_u.reader) {
2390 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2391 if (!(env->me_flags & MDB_NOTLS))
2392 txn->mt_u.reader = NULL; /* txn does not own reader */
2394 txn->mt_numdbs = 0; /* close nothing if called again */
2395 txn->mt_dbxs = NULL; /* mark txn as reset */
2397 mdb_cursors_close(txn, 0);
2399 if (!(env->me_flags & MDB_WRITEMAP)) {
2400 mdb_dlist_free(txn);
2402 mdb_midl_free(env->me_pghead);
2404 if (txn->mt_parent) {
2405 txn->mt_parent->mt_child = NULL;
2406 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2407 mdb_midl_free(txn->mt_free_pgs);
2408 mdb_midl_free(txn->mt_spill_pgs);
2409 free(txn->mt_u.dirty_list);
2413 if (mdb_midl_shrink(&txn->mt_free_pgs))
2414 env->me_free_pgs = txn->mt_free_pgs;
2415 env->me_pghead = NULL;
2419 /* The writer mutex was locked in mdb_txn_begin. */
2420 UNLOCK_MUTEX_W(env);
2425 mdb_txn_reset(MDB_txn *txn)
2430 /* This call is only valid for read-only txns */
2431 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2434 mdb_txn_reset0(txn, "reset");
2438 mdb_txn_abort(MDB_txn *txn)
2444 mdb_txn_abort(txn->mt_child);
2446 mdb_txn_reset0(txn, "abort");
2447 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2448 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2449 txn->mt_u.reader->mr_pid = 0;
2454 /** Save the freelist as of this transaction to the freeDB.
2455 * This changes the freelist. Keep trying until it stabilizes.
2458 mdb_freelist_save(MDB_txn *txn)
2460 /* env->me_pghead[] can grow and shrink during this call.
2461 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2462 * Page numbers cannot disappear from txn->mt_free_pgs[].
2465 MDB_env *env = txn->mt_env;
2466 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2467 txnid_t pglast = 0, head_id = 0;
2468 pgno_t freecnt = 0, *free_pgs, *mop;
2469 ssize_t head_room = 0, total_room = 0, mop_len;
2471 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2473 if (env->me_pghead) {
2474 /* Make sure first page of freeDB is touched and on freelist */
2475 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2476 if (rc && rc != MDB_NOTFOUND)
2481 /* Come back here after each Put() in case freelist changed */
2484 /* If using records from freeDB which we have not yet
2485 * deleted, delete them and any we reserved for me_pghead.
2487 while (pglast < env->me_pglast) {
2488 rc = mdb_cursor_first(&mc, &key, NULL);
2491 pglast = head_id = *(txnid_t *)key.mv_data;
2492 total_room = head_room = 0;
2493 assert(pglast <= env->me_pglast);
2494 rc = mdb_cursor_del(&mc, 0);
2499 /* Save the IDL of pages freed by this txn, to a single record */
2500 if (freecnt < txn->mt_free_pgs[0]) {
2502 /* Make sure last page of freeDB is touched and on freelist */
2503 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2504 if (rc && rc != MDB_NOTFOUND)
2507 free_pgs = txn->mt_free_pgs;
2508 /* Write to last page of freeDB */
2509 key.mv_size = sizeof(txn->mt_txnid);
2510 key.mv_data = &txn->mt_txnid;
2512 freecnt = free_pgs[0];
2513 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2514 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2517 /* Retry if mt_free_pgs[] grew during the Put() */
2518 free_pgs = txn->mt_free_pgs;
2519 } while (freecnt < free_pgs[0]);
2520 mdb_midl_sort(free_pgs);
2521 memcpy(data.mv_data, free_pgs, data.mv_size);
2524 unsigned int i = free_pgs[0];
2525 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2526 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2528 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2534 mop = env->me_pghead;
2535 mop_len = mop ? mop[0] : 0;
2537 /* Reserve records for me_pghead[]. Split it if multi-page,
2538 * to avoid searching freeDB for a page range. Use keys in
2539 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2541 if (total_room >= mop_len) {
2542 if (total_room == mop_len || --more < 0)
2544 } else if (head_room >= maxfree_1pg && head_id > 1) {
2545 /* Keep current record (overflow page), add a new one */
2549 /* (Re)write {key = head_id, IDL length = head_room} */
2550 total_room -= head_room;
2551 head_room = mop_len - total_room;
2552 if (head_room > maxfree_1pg && head_id > 1) {
2553 /* Overflow multi-page for part of me_pghead */
2554 head_room /= head_id; /* amortize page sizes */
2555 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2556 } else if (head_room < 0) {
2557 /* Rare case, not bothering to delete this record */
2560 key.mv_size = sizeof(head_id);
2561 key.mv_data = &head_id;
2562 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2563 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2566 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2567 total_room += head_room;
2570 /* Fill in the reserved me_pghead records */
2576 rc = mdb_cursor_first(&mc, &key, &data);
2577 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2578 unsigned flags = MDB_CURRENT;
2579 txnid_t id = *(txnid_t *)key.mv_data;
2580 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2583 assert(len >= 0 && id <= env->me_pglast);
2585 if (len > mop_len) {
2587 data.mv_size = (len + 1) * sizeof(MDB_ID);
2590 data.mv_data = mop -= len;
2593 rc = mdb_cursor_put(&mc, &key, &data, flags);
2595 if (rc || !(mop_len -= len))
2602 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2603 * @param[in] txn the transaction that's being committed
2604 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2605 * @return 0 on success, non-zero on failure.
2608 mdb_page_flush(MDB_txn *txn, int keep)
2610 MDB_env *env = txn->mt_env;
2611 MDB_ID2L dl = txn->mt_u.dirty_list;
2612 unsigned psize = env->me_psize, j;
2613 int i, pagecount = dl[0].mid, rc;
2614 size_t size = 0, pos = 0;
2616 MDB_page *dp = NULL;
2620 struct iovec iov[MDB_COMMIT_PAGES];
2621 ssize_t wpos = 0, wsize = 0, wres;
2622 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2628 if (env->me_flags & MDB_WRITEMAP) {
2629 /* Clear dirty flags */
2630 while (++i <= pagecount) {
2632 /* Don't flush this page yet */
2633 if (dp->mp_flags & P_KEEP) {
2634 dp->mp_flags ^= P_KEEP;
2638 dp->mp_flags &= ~P_DIRTY;
2643 /* Write the pages */
2645 if (++i <= pagecount) {
2647 /* Don't flush this page yet */
2648 if (dp->mp_flags & P_KEEP) {
2649 dp->mp_flags ^= P_KEEP;
2654 /* clear dirty flag */
2655 dp->mp_flags &= ~P_DIRTY;
2658 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2663 /* Windows actually supports scatter/gather I/O, but only on
2664 * unbuffered file handles. Since we're relying on the OS page
2665 * cache for all our data, that's self-defeating. So we just
2666 * write pages one at a time. We use the ov structure to set
2667 * the write offset, to at least save the overhead of a Seek
2670 DPRINTF(("committing page %"Z"u", pgno));
2671 memset(&ov, 0, sizeof(ov));
2672 ov.Offset = pos & 0xffffffff;
2673 ov.OffsetHigh = pos >> 16 >> 16;
2674 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2676 DPRINTF(("WriteFile: %d", rc));
2680 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2681 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2683 /* Write previous page(s) */
2684 #ifdef MDB_USE_PWRITEV
2685 wres = pwritev(env->me_fd, iov, n, wpos);
2688 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2690 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2692 DPRINTF(("lseek: %s", strerror(rc)));
2695 wres = writev(env->me_fd, iov, n);
2698 if (wres != wsize) {
2701 DPRINTF(("Write error: %s", strerror(rc)));
2703 rc = EIO; /* TODO: Use which error code? */
2704 DPUTS("short write, filesystem full?");
2715 DPRINTF(("committing page %"Z"u", pgno));
2716 next_pos = pos + size;
2717 iov[n].iov_len = size;
2718 iov[n].iov_base = (char *)dp;
2724 for (i = keep; ++i <= pagecount; ) {
2726 /* This is a page we skipped above */
2729 dl[j].mid = dp->mp_pgno;
2732 mdb_dpage_free(env, dp);
2737 txn->mt_dirty_room += i - j;
2743 mdb_txn_commit(MDB_txn *txn)
2749 assert(txn != NULL);
2750 assert(txn->mt_env != NULL);
2752 if (txn->mt_child) {
2753 rc = mdb_txn_commit(txn->mt_child);
2754 txn->mt_child = NULL;
2761 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2762 mdb_dbis_update(txn, 1);
2763 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2768 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2769 DPUTS("error flag is set, can't commit");
2771 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2776 if (txn->mt_parent) {
2777 MDB_txn *parent = txn->mt_parent;
2780 unsigned x, y, len, ps_len;
2782 /* Append our free list to parent's */
2783 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2786 mdb_midl_free(txn->mt_free_pgs);
2787 /* Failures after this must either undo the changes
2788 * to the parent or set MDB_TXN_ERROR in the parent.
2791 parent->mt_next_pgno = txn->mt_next_pgno;
2792 parent->mt_flags = txn->mt_flags;
2794 /* Merge our cursors into parent's and close them */
2795 mdb_cursors_close(txn, 1);
2797 /* Update parent's DB table. */
2798 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2799 parent->mt_numdbs = txn->mt_numdbs;
2800 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2801 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2802 for (i=2; i<txn->mt_numdbs; i++) {
2803 /* preserve parent's DB_NEW status */
2804 x = parent->mt_dbflags[i] & DB_NEW;
2805 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2808 dst = parent->mt_u.dirty_list;
2809 src = txn->mt_u.dirty_list;
2810 /* Remove anything in our dirty list from parent's spill list */
2811 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2813 pspill[0] = (pgno_t)-1;
2814 /* Mark our dirty pages as deleted in parent spill list */
2815 for (i=0, len=src[0].mid; ++i <= len; ) {
2816 MDB_ID pn = src[i].mid << 1;
2817 while (pn > pspill[x])
2819 if (pn == pspill[x]) {
2824 /* Squash deleted pagenums if we deleted any */
2825 for (x=y; ++x <= ps_len; )
2826 if (!(pspill[x] & 1))
2827 pspill[++y] = pspill[x];
2831 /* Find len = length of merging our dirty list with parent's */
2833 dst[0].mid = 0; /* simplify loops */
2834 if (parent->mt_parent) {
2835 len = x + src[0].mid;
2836 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2837 for (i = x; y && i; y--) {
2838 pgno_t yp = src[y].mid;
2839 while (yp < dst[i].mid)
2841 if (yp == dst[i].mid) {
2846 } else { /* Simplify the above for single-ancestor case */
2847 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2849 /* Merge our dirty list with parent's */
2851 for (i = len; y; dst[i--] = src[y--]) {
2852 pgno_t yp = src[y].mid;
2853 while (yp < dst[x].mid)
2854 dst[i--] = dst[x--];
2855 if (yp == dst[x].mid)
2856 free(dst[x--].mptr);
2860 free(txn->mt_u.dirty_list);
2861 parent->mt_dirty_room = txn->mt_dirty_room;
2862 if (txn->mt_spill_pgs) {
2863 if (parent->mt_spill_pgs) {
2864 /* TODO: Prevent failure here, so parent does not fail */
2865 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2867 parent->mt_flags |= MDB_TXN_ERROR;
2868 mdb_midl_free(txn->mt_spill_pgs);
2869 mdb_midl_sort(parent->mt_spill_pgs);
2871 parent->mt_spill_pgs = txn->mt_spill_pgs;
2875 parent->mt_child = NULL;
2876 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2881 if (txn != env->me_txn) {
2882 DPUTS("attempt to commit unknown transaction");
2887 mdb_cursors_close(txn, 0);
2889 if (!txn->mt_u.dirty_list[0].mid &&
2890 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2893 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2894 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2896 /* Update DB root pointers */
2897 if (txn->mt_numdbs > 2) {
2901 data.mv_size = sizeof(MDB_db);
2903 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2904 for (i = 2; i < txn->mt_numdbs; i++) {
2905 if (txn->mt_dbflags[i] & DB_DIRTY) {
2906 data.mv_data = &txn->mt_dbs[i];
2907 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2914 rc = mdb_freelist_save(txn);
2918 mdb_midl_free(env->me_pghead);
2919 env->me_pghead = NULL;
2920 if (mdb_midl_shrink(&txn->mt_free_pgs))
2921 env->me_free_pgs = txn->mt_free_pgs;
2927 if ((rc = mdb_page_flush(txn, 0)) ||
2928 (rc = mdb_env_sync(env, 0)) ||
2929 (rc = mdb_env_write_meta(txn)))
2935 mdb_dbis_update(txn, 1);
2937 UNLOCK_MUTEX_W(env);
2947 /** Read the environment parameters of a DB environment before
2948 * mapping it into memory.
2949 * @param[in] env the environment handle
2950 * @param[out] meta address of where to store the meta information
2951 * @return 0 on success, non-zero on failure.
2954 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2961 /* We don't know the page size yet, so use a minimum value.
2962 * Read both meta pages so we can use the latest one.
2965 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2969 memset(&ov, 0, sizeof(ov));
2971 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2972 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2975 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2977 if (rc != MDB_PAGESIZE) {
2978 if (rc == 0 && off == 0)
2980 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2981 DPRINTF(("read: %s", mdb_strerror(rc)));
2985 p = (MDB_page *)&pbuf;
2987 if (!F_ISSET(p->mp_flags, P_META)) {
2988 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
2993 if (m->mm_magic != MDB_MAGIC) {
2994 DPUTS("meta has invalid magic");
2998 if (m->mm_version != MDB_DATA_VERSION) {
2999 DPRINTF(("database is version %u, expected version %u",
3000 m->mm_version, MDB_DATA_VERSION));
3001 return MDB_VERSION_MISMATCH;
3004 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3010 /** Write the environment parameters of a freshly created DB environment.
3011 * @param[in] env the environment handle
3012 * @param[out] meta address of where to store the meta information
3013 * @return 0 on success, non-zero on failure.
3016 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3024 memset(&ov, 0, sizeof(ov));
3025 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3027 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3030 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3031 len = pwrite(fd, ptr, size, pos); \
3032 rc = (len >= 0); } while(0)
3035 DPUTS("writing new meta page");
3037 psize = env->me_psize;
3039 meta->mm_magic = MDB_MAGIC;
3040 meta->mm_version = MDB_DATA_VERSION;
3041 meta->mm_mapsize = env->me_mapsize;
3042 meta->mm_psize = psize;
3043 meta->mm_last_pg = 1;
3044 meta->mm_flags = env->me_flags & 0xffff;
3045 meta->mm_flags |= MDB_INTEGERKEY;
3046 meta->mm_dbs[0].md_root = P_INVALID;
3047 meta->mm_dbs[1].md_root = P_INVALID;
3049 p = calloc(2, psize);
3051 p->mp_flags = P_META;
3052 *(MDB_meta *)METADATA(p) = *meta;
3054 q = (MDB_page *)((char *)p + psize);
3056 q->mp_flags = P_META;
3057 *(MDB_meta *)METADATA(q) = *meta;
3059 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3062 else if ((unsigned) len == psize * 2)
3070 /** Update the environment info to commit a transaction.
3071 * @param[in] txn the transaction that's being committed
3072 * @return 0 on success, non-zero on failure.
3075 mdb_env_write_meta(MDB_txn *txn)
3078 MDB_meta meta, metab, *mp;
3080 int rc, len, toggle;
3089 assert(txn != NULL);
3090 assert(txn->mt_env != NULL);
3092 toggle = txn->mt_txnid & 1;
3093 DPRINTF(("writing meta page %d for root page %"Z"u",
3094 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3097 mp = env->me_metas[toggle];
3099 if (env->me_flags & MDB_WRITEMAP) {
3100 /* Persist any increases of mapsize config */
3101 if (env->me_mapsize > mp->mm_mapsize)
3102 mp->mm_mapsize = env->me_mapsize;
3103 mp->mm_dbs[0] = txn->mt_dbs[0];
3104 mp->mm_dbs[1] = txn->mt_dbs[1];
3105 mp->mm_last_pg = txn->mt_next_pgno - 1;
3106 mp->mm_txnid = txn->mt_txnid;
3107 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3108 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3111 ptr += env->me_psize;
3112 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3119 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3120 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3122 ptr = (char *)&meta;
3123 if (env->me_mapsize > mp->mm_mapsize) {
3124 /* Persist any increases of mapsize config */
3125 meta.mm_mapsize = env->me_mapsize;
3126 off = offsetof(MDB_meta, mm_mapsize);
3128 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3130 len = sizeof(MDB_meta) - off;
3133 meta.mm_dbs[0] = txn->mt_dbs[0];
3134 meta.mm_dbs[1] = txn->mt_dbs[1];
3135 meta.mm_last_pg = txn->mt_next_pgno - 1;
3136 meta.mm_txnid = txn->mt_txnid;
3139 off += env->me_psize;
3142 /* Write to the SYNC fd */
3143 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3144 env->me_fd : env->me_mfd;
3147 memset(&ov, 0, sizeof(ov));
3149 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3153 rc = pwrite(mfd, ptr, len, off);
3156 rc = rc < 0 ? ErrCode() : EIO;
3157 DPUTS("write failed, disk error?");
3158 /* On a failure, the pagecache still contains the new data.
3159 * Write some old data back, to prevent it from being used.
3160 * Use the non-SYNC fd; we know it will fail anyway.
3162 meta.mm_last_pg = metab.mm_last_pg;
3163 meta.mm_txnid = metab.mm_txnid;
3165 memset(&ov, 0, sizeof(ov));
3167 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3169 r2 = pwrite(env->me_fd, ptr, len, off);
3170 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3173 env->me_flags |= MDB_FATAL_ERROR;
3177 /* Memory ordering issues are irrelevant; since the entire writer
3178 * is wrapped by wmutex, all of these changes will become visible
3179 * after the wmutex is unlocked. Since the DB is multi-version,
3180 * readers will get consistent data regardless of how fresh or
3181 * how stale their view of these values is.
3183 env->me_txns->mti_txnid = txn->mt_txnid;
3188 /** Check both meta pages to see which one is newer.
3189 * @param[in] env the environment handle
3190 * @return meta toggle (0 or 1).
3193 mdb_env_pick_meta(const MDB_env *env)
3195 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3199 mdb_env_create(MDB_env **env)
3203 e = calloc(1, sizeof(MDB_env));
3207 e->me_maxreaders = DEFAULT_READERS;
3208 e->me_maxdbs = e->me_numdbs = 2;
3209 e->me_fd = INVALID_HANDLE_VALUE;
3210 e->me_lfd = INVALID_HANDLE_VALUE;
3211 e->me_mfd = INVALID_HANDLE_VALUE;
3212 #ifdef MDB_USE_POSIX_SEM
3213 e->me_rmutex = SEM_FAILED;
3214 e->me_wmutex = SEM_FAILED;
3216 e->me_pid = getpid();
3217 VGMEMP_CREATE(e,0,0);
3223 mdb_env_map(MDB_env *env, void *addr, int newsize)
3226 unsigned int flags = env->me_flags;
3230 LONG sizelo, sizehi;
3231 sizelo = env->me_mapsize & 0xffffffff;
3232 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3234 /* Windows won't create mappings for zero length files.
3235 * Just allocate the maxsize right now.
3238 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3239 || !SetEndOfFile(env->me_fd)
3240 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3243 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3244 PAGE_READWRITE : PAGE_READONLY,
3245 sizehi, sizelo, NULL);
3248 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3249 FILE_MAP_WRITE : FILE_MAP_READ,
3250 0, 0, env->me_mapsize, addr);
3251 rc = env->me_map ? 0 : ErrCode();
3256 int prot = PROT_READ;
3257 if (flags & MDB_WRITEMAP) {
3259 if (newsize && ftruncate(env->me_fd, env->me_mapsize) < 0)
3262 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3264 if (env->me_map == MAP_FAILED) {
3268 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3270 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3272 #ifdef POSIX_MADV_RANDOM
3273 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3274 #endif /* POSIX_MADV_RANDOM */
3275 #endif /* MADV_RANDOM */
3278 /* Can happen because the address argument to mmap() is just a
3279 * hint. mmap() can pick another, e.g. if the range is in use.
3280 * The MAP_FIXED flag would prevent that, but then mmap could
3281 * instead unmap existing pages to make room for the new map.
3283 if (addr && env->me_map != addr)
3284 return EBUSY; /* TODO: Make a new MDB_* error code? */
3286 p = (MDB_page *)env->me_map;
3287 env->me_metas[0] = METADATA(p);
3288 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3294 mdb_env_set_mapsize(MDB_env *env, size_t size)
3296 /* If env is already open, caller is responsible for making
3297 * sure there are no active txns.
3305 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3306 munmap(env->me_map, env->me_mapsize);
3307 env->me_mapsize = size;
3308 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3309 rc = mdb_env_map(env, old, 1);
3313 env->me_mapsize = size;
3315 env->me_maxpg = env->me_mapsize / env->me_psize;
3320 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3324 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3329 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3331 if (env->me_map || readers < 1)
3333 env->me_maxreaders = readers;
3338 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3340 if (!env || !readers)
3342 *readers = env->me_maxreaders;
3346 /** Further setup required for opening an MDB environment
3349 mdb_env_open2(MDB_env *env)
3351 unsigned int flags = env->me_flags;
3352 int i, newenv = 0, rc;
3356 /* See if we should use QueryLimited */
3358 if ((rc & 0xff) > 5)
3359 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3361 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3364 memset(&meta, 0, sizeof(meta));
3366 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3369 DPUTS("new mdbenv");
3371 GET_PAGESIZE(env->me_psize);
3373 env->me_psize = meta.mm_psize;
3376 /* Was a mapsize configured? */
3377 if (!env->me_mapsize) {
3378 /* If this is a new environment, take the default,
3379 * else use the size recorded in the existing env.
3381 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3382 } else if (env->me_mapsize < meta.mm_mapsize) {
3383 /* If the configured size is smaller, make sure it's
3384 * still big enough. Silently round up to minimum if not.
3386 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3387 if (env->me_mapsize < minsize)
3388 env->me_mapsize = minsize;
3391 rc = mdb_env_map(env, meta.mm_address, newenv);
3396 if (flags & MDB_FIXEDMAP)
3397 meta.mm_address = env->me_map;
3398 i = mdb_env_init_meta(env, &meta);
3399 if (i != MDB_SUCCESS) {
3403 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3404 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3406 env->me_maxpg = env->me_mapsize / env->me_psize;
3409 int toggle = mdb_env_pick_meta(env);
3410 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3412 DPRINTF(("opened database version %u, pagesize %u",
3413 env->me_metas[0]->mm_version, env->me_psize));
3414 DPRINTF(("using meta page %d", toggle));
3415 DPRINTF(("depth: %u", db->md_depth));
3416 DPRINTF(("entries: %"Z"u", db->md_entries));
3417 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3418 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3419 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3420 DPRINTF(("root: %"Z"u", db->md_root));
3428 /** Release a reader thread's slot in the reader lock table.
3429 * This function is called automatically when a thread exits.
3430 * @param[in] ptr This points to the slot in the reader lock table.
3433 mdb_env_reader_dest(void *ptr)
3435 MDB_reader *reader = ptr;
3441 /** Junk for arranging thread-specific callbacks on Windows. This is
3442 * necessarily platform and compiler-specific. Windows supports up
3443 * to 1088 keys. Let's assume nobody opens more than 64 environments
3444 * in a single process, for now. They can override this if needed.
3446 #ifndef MAX_TLS_KEYS
3447 #define MAX_TLS_KEYS 64
3449 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3450 static int mdb_tls_nkeys;
3452 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3456 case DLL_PROCESS_ATTACH: break;
3457 case DLL_THREAD_ATTACH: break;
3458 case DLL_THREAD_DETACH:
3459 for (i=0; i<mdb_tls_nkeys; i++) {
3460 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3461 mdb_env_reader_dest(r);
3464 case DLL_PROCESS_DETACH: break;
3469 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3471 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3475 /* Force some symbol references.
3476 * _tls_used forces the linker to create the TLS directory if not already done
3477 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3479 #pragma comment(linker, "/INCLUDE:_tls_used")
3480 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3481 #pragma const_seg(".CRT$XLB")
3482 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3483 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3486 #pragma comment(linker, "/INCLUDE:__tls_used")
3487 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3488 #pragma data_seg(".CRT$XLB")
3489 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3491 #endif /* WIN 32/64 */
3492 #endif /* !__GNUC__ */
3495 /** Downgrade the exclusive lock on the region back to shared */
3497 mdb_env_share_locks(MDB_env *env, int *excl)
3499 int rc = 0, toggle = mdb_env_pick_meta(env);
3501 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3506 /* First acquire a shared lock. The Unlock will
3507 * then release the existing exclusive lock.
3509 memset(&ov, 0, sizeof(ov));
3510 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3513 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3519 struct flock lock_info;
3520 /* The shared lock replaces the existing lock */
3521 memset((void *)&lock_info, 0, sizeof(lock_info));
3522 lock_info.l_type = F_RDLCK;
3523 lock_info.l_whence = SEEK_SET;
3524 lock_info.l_start = 0;
3525 lock_info.l_len = 1;
3526 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3527 (rc = ErrCode()) == EINTR) ;
3528 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3535 /** Try to get exlusive lock, otherwise shared.
3536 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3539 mdb_env_excl_lock(MDB_env *env, int *excl)
3543 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3547 memset(&ov, 0, sizeof(ov));
3548 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3555 struct flock lock_info;
3556 memset((void *)&lock_info, 0, sizeof(lock_info));
3557 lock_info.l_type = F_WRLCK;
3558 lock_info.l_whence = SEEK_SET;
3559 lock_info.l_start = 0;
3560 lock_info.l_len = 1;
3561 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3562 (rc = ErrCode()) == EINTR) ;
3566 # ifdef MDB_USE_POSIX_SEM
3567 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3570 lock_info.l_type = F_RDLCK;
3571 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3572 (rc = ErrCode()) == EINTR) ;
3580 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3582 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3584 * @(#) $Revision: 5.1 $
3585 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3586 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3588 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3592 * Please do not copyright this code. This code is in the public domain.
3594 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3595 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3596 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3597 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3598 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3599 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3600 * PERFORMANCE OF THIS SOFTWARE.
3603 * chongo <Landon Curt Noll> /\oo/\
3604 * http://www.isthe.com/chongo/
3606 * Share and Enjoy! :-)
3609 typedef unsigned long long mdb_hash_t;
3610 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3612 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3613 * @param[in] val value to hash
3614 * @param[in] hval initial value for hash
3615 * @return 64 bit hash
3617 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3618 * hval arg on the first call.
3621 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3623 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3624 unsigned char *end = s + val->mv_size;
3626 * FNV-1a hash each octet of the string
3629 /* xor the bottom with the current octet */
3630 hval ^= (mdb_hash_t)*s++;
3632 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3633 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3634 (hval << 7) + (hval << 8) + (hval << 40);
3636 /* return our new hash value */
3640 /** Hash the string and output the encoded hash.
3641 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3642 * very short name limits. We don't care about the encoding being reversible,
3643 * we just want to preserve as many bits of the input as possible in a
3644 * small printable string.
3645 * @param[in] str string to hash
3646 * @param[out] encbuf an array of 11 chars to hold the hash
3648 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3651 mdb_pack85(unsigned long l, char *out)
3655 for (i=0; i<5; i++) {
3656 *out++ = mdb_a85[l % 85];
3662 mdb_hash_enc(MDB_val *val, char *encbuf)
3664 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3666 mdb_pack85(h, encbuf);
3667 mdb_pack85(h>>32, encbuf+5);
3672 /** Open and/or initialize the lock region for the environment.
3673 * @param[in] env The MDB environment.
3674 * @param[in] lpath The pathname of the file used for the lock region.
3675 * @param[in] mode The Unix permissions for the file, if we create it.
3676 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3677 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3678 * @return 0 on success, non-zero on failure.
3681 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3684 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3686 # define MDB_ERRCODE_ROFS EROFS
3687 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3688 # define MDB_CLOEXEC O_CLOEXEC
3691 # define MDB_CLOEXEC 0
3698 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3699 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3700 FILE_ATTRIBUTE_NORMAL, NULL);
3702 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3704 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3706 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3711 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3712 /* Lose record locks when exec*() */
3713 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3714 fcntl(env->me_lfd, F_SETFD, fdflags);
3717 if (!(env->me_flags & MDB_NOTLS)) {
3718 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3721 env->me_flags |= MDB_ENV_TXKEY;
3723 /* Windows TLS callbacks need help finding their TLS info. */
3724 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3728 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3732 /* Try to get exclusive lock. If we succeed, then
3733 * nobody is using the lock region and we should initialize it.
3735 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3738 size = GetFileSize(env->me_lfd, NULL);
3740 size = lseek(env->me_lfd, 0, SEEK_END);
3741 if (size == -1) goto fail_errno;
3743 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3744 if (size < rsize && *excl > 0) {
3746 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3747 || !SetEndOfFile(env->me_lfd))
3750 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3754 size = rsize - sizeof(MDB_txninfo);
3755 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3760 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3762 if (!mh) goto fail_errno;
3763 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3765 if (!env->me_txns) goto fail_errno;
3767 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3769 if (m == MAP_FAILED) goto fail_errno;
3775 BY_HANDLE_FILE_INFORMATION stbuf;
3784 if (!mdb_sec_inited) {
3785 InitializeSecurityDescriptor(&mdb_null_sd,
3786 SECURITY_DESCRIPTOR_REVISION);
3787 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3788 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3789 mdb_all_sa.bInheritHandle = FALSE;
3790 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3793 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3794 idbuf.volume = stbuf.dwVolumeSerialNumber;
3795 idbuf.nhigh = stbuf.nFileIndexHigh;
3796 idbuf.nlow = stbuf.nFileIndexLow;
3797 val.mv_data = &idbuf;
3798 val.mv_size = sizeof(idbuf);
3799 mdb_hash_enc(&val, encbuf);
3800 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3801 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3802 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3803 if (!env->me_rmutex) goto fail_errno;
3804 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3805 if (!env->me_wmutex) goto fail_errno;
3806 #elif defined(MDB_USE_POSIX_SEM)
3815 #if defined(__NetBSD__)
3816 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3818 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3819 idbuf.dev = stbuf.st_dev;
3820 idbuf.ino = stbuf.st_ino;
3821 val.mv_data = &idbuf;
3822 val.mv_size = sizeof(idbuf);
3823 mdb_hash_enc(&val, encbuf);
3824 #ifdef MDB_SHORT_SEMNAMES
3825 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3827 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3828 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3829 /* Clean up after a previous run, if needed: Try to
3830 * remove both semaphores before doing anything else.
3832 sem_unlink(env->me_txns->mti_rmname);
3833 sem_unlink(env->me_txns->mti_wmname);
3834 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3835 O_CREAT|O_EXCL, mode, 1);
3836 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3837 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3838 O_CREAT|O_EXCL, mode, 1);
3839 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3840 #else /* MDB_USE_POSIX_SEM */
3841 pthread_mutexattr_t mattr;
3843 if ((rc = pthread_mutexattr_init(&mattr))
3844 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3845 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3846 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3848 pthread_mutexattr_destroy(&mattr);
3849 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3851 env->me_txns->mti_magic = MDB_MAGIC;
3852 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3853 env->me_txns->mti_txnid = 0;
3854 env->me_txns->mti_numreaders = 0;
3857 if (env->me_txns->mti_magic != MDB_MAGIC) {
3858 DPUTS("lock region has invalid magic");
3862 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3863 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3864 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3865 rc = MDB_VERSION_MISMATCH;
3869 if (rc && rc != EACCES && rc != EAGAIN) {
3873 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3874 if (!env->me_rmutex) goto fail_errno;
3875 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3876 if (!env->me_wmutex) goto fail_errno;
3877 #elif defined(MDB_USE_POSIX_SEM)
3878 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3879 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3880 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3881 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3892 /** The name of the lock file in the DB environment */
3893 #define LOCKNAME "/lock.mdb"
3894 /** The name of the data file in the DB environment */
3895 #define DATANAME "/data.mdb"
3896 /** The suffix of the lock file when no subdir is used */
3897 #define LOCKSUFF "-lock"
3898 /** Only a subset of the @ref mdb_env flags can be changed
3899 * at runtime. Changing other flags requires closing the
3900 * environment and re-opening it with the new flags.
3902 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3903 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3906 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3908 int oflags, rc, len, excl = -1;
3909 char *lpath, *dpath;
3911 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3915 if (flags & MDB_NOSUBDIR) {
3916 rc = len + sizeof(LOCKSUFF) + len + 1;
3918 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3923 if (flags & MDB_NOSUBDIR) {
3924 dpath = lpath + len + sizeof(LOCKSUFF);
3925 sprintf(lpath, "%s" LOCKSUFF, path);
3926 strcpy(dpath, path);
3928 dpath = lpath + len + sizeof(LOCKNAME);
3929 sprintf(lpath, "%s" LOCKNAME, path);
3930 sprintf(dpath, "%s" DATANAME, path);
3934 flags |= env->me_flags;
3935 if (flags & MDB_RDONLY) {
3936 /* silently ignore WRITEMAP when we're only getting read access */
3937 flags &= ~MDB_WRITEMAP;
3939 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3940 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3943 env->me_flags = flags |= MDB_ENV_ACTIVE;
3947 env->me_path = strdup(path);
3948 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3949 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3950 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3955 /* For RDONLY, get lockfile after we know datafile exists */
3956 if (!F_ISSET(flags, MDB_RDONLY)) {
3957 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3963 if (F_ISSET(flags, MDB_RDONLY)) {
3964 oflags = GENERIC_READ;
3965 len = OPEN_EXISTING;
3967 oflags = GENERIC_READ|GENERIC_WRITE;
3970 mode = FILE_ATTRIBUTE_NORMAL;
3971 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3972 NULL, len, mode, NULL);
3974 if (F_ISSET(flags, MDB_RDONLY))
3977 oflags = O_RDWR | O_CREAT;
3979 env->me_fd = open(dpath, oflags, mode);
3981 if (env->me_fd == INVALID_HANDLE_VALUE) {
3986 if (F_ISSET(flags, MDB_RDONLY)) {
3987 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3992 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3993 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3994 env->me_mfd = env->me_fd;
3996 /* Synchronous fd for meta writes. Needed even with
3997 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4000 len = OPEN_EXISTING;
4001 env->me_mfd = CreateFile(dpath, oflags,
4002 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4003 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4006 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4008 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4013 DPRINTF(("opened dbenv %p", (void *) env));
4015 rc = mdb_env_share_locks(env, &excl);
4021 mdb_env_close0(env, excl);
4027 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4029 mdb_env_close0(MDB_env *env, int excl)
4033 if (!(env->me_flags & MDB_ENV_ACTIVE))
4036 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4037 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4038 free(env->me_dbxs[i].md_name.mv_data);
4040 free(env->me_dbflags);
4043 free(env->me_dirty_list);
4044 mdb_midl_free(env->me_free_pgs);
4046 if (env->me_flags & MDB_ENV_TXKEY) {
4047 pthread_key_delete(env->me_txkey);
4049 /* Delete our key from the global list */
4050 for (i=0; i<mdb_tls_nkeys; i++)
4051 if (mdb_tls_keys[i] == env->me_txkey) {
4052 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4060 munmap(env->me_map, env->me_mapsize);
4062 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4063 (void) close(env->me_mfd);
4064 if (env->me_fd != INVALID_HANDLE_VALUE)
4065 (void) close(env->me_fd);
4067 pid_t pid = env->me_pid;
4068 /* Clearing readers is done in this function because
4069 * me_txkey with its destructor must be disabled first.
4071 for (i = env->me_numreaders; --i >= 0; )
4072 if (env->me_txns->mti_readers[i].mr_pid == pid)
4073 env->me_txns->mti_readers[i].mr_pid = 0;
4075 if (env->me_rmutex) {
4076 CloseHandle(env->me_rmutex);
4077 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4079 /* Windows automatically destroys the mutexes when
4080 * the last handle closes.
4082 #elif defined(MDB_USE_POSIX_SEM)
4083 if (env->me_rmutex != SEM_FAILED) {
4084 sem_close(env->me_rmutex);
4085 if (env->me_wmutex != SEM_FAILED)
4086 sem_close(env->me_wmutex);
4087 /* If we have the filelock: If we are the
4088 * only remaining user, clean up semaphores.
4091 mdb_env_excl_lock(env, &excl);
4093 sem_unlink(env->me_txns->mti_rmname);
4094 sem_unlink(env->me_txns->mti_wmname);
4098 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4100 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4103 /* Unlock the lockfile. Windows would have unlocked it
4104 * after closing anyway, but not necessarily at once.
4106 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4109 (void) close(env->me_lfd);
4112 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4116 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4118 MDB_txn *txn = NULL;
4124 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4128 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4131 /* Do the lock/unlock of the reader mutex before starting the
4132 * write txn. Otherwise other read txns could block writers.
4134 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4139 /* We must start the actual read txn after blocking writers */
4140 mdb_txn_reset0(txn, "reset-stage1");
4142 /* Temporarily block writers until we snapshot the meta pages */
4145 rc = mdb_txn_renew0(txn);
4147 UNLOCK_MUTEX_W(env);
4152 wsize = env->me_psize * 2;
4156 DO_WRITE(rc, fd, ptr, w2, len);
4160 } else if (len > 0) {
4166 /* Non-blocking or async handles are not supported */
4172 UNLOCK_MUTEX_W(env);
4177 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4179 if (wsize > MAX_WRITE)
4183 DO_WRITE(rc, fd, ptr, w2, len);
4187 } else if (len > 0) {
4204 mdb_env_copy(MDB_env *env, const char *path)
4208 HANDLE newfd = INVALID_HANDLE_VALUE;
4210 if (env->me_flags & MDB_NOSUBDIR) {
4211 lpath = (char *)path;
4214 len += sizeof(DATANAME);
4215 lpath = malloc(len);
4218 sprintf(lpath, "%s" DATANAME, path);
4221 /* The destination path must exist, but the destination file must not.
4222 * We don't want the OS to cache the writes, since the source data is
4223 * already in the OS cache.
4226 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4227 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4229 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4231 if (newfd == INVALID_HANDLE_VALUE) {
4237 /* Set O_DIRECT if the file system supports it */
4238 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4239 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4241 #ifdef F_NOCACHE /* __APPLE__ */
4242 rc = fcntl(newfd, F_NOCACHE, 1);
4249 rc = mdb_env_copyfd(env, newfd);
4252 if (!(env->me_flags & MDB_NOSUBDIR))
4254 if (newfd != INVALID_HANDLE_VALUE)
4255 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4262 mdb_env_close(MDB_env *env)
4269 VGMEMP_DESTROY(env);
4270 while ((dp = env->me_dpages) != NULL) {
4271 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4272 env->me_dpages = dp->mp_next;
4276 mdb_env_close0(env, 0);
4280 /** Compare two items pointing at aligned size_t's */
4282 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4284 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4285 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4288 /** Compare two items pointing at aligned unsigned int's */
4290 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4292 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4293 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4296 /** Compare two items pointing at unsigned ints of unknown alignment.
4297 * Nodes and keys are guaranteed to be 2-byte aligned.
4300 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4302 #if BYTE_ORDER == LITTLE_ENDIAN
4303 unsigned short *u, *c;
4306 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4307 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4310 } while(!x && u > (unsigned short *)a->mv_data);
4313 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4317 /** Compare two items lexically */
4319 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4326 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4332 diff = memcmp(a->mv_data, b->mv_data, len);
4333 return diff ? diff : len_diff<0 ? -1 : len_diff;
4336 /** Compare two items in reverse byte order */
4338 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4340 const unsigned char *p1, *p2, *p1_lim;
4344 p1_lim = (const unsigned char *)a->mv_data;
4345 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4346 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4348 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4354 while (p1 > p1_lim) {
4355 diff = *--p1 - *--p2;
4359 return len_diff<0 ? -1 : len_diff;
4362 /** Search for key within a page, using binary search.
4363 * Returns the smallest entry larger or equal to the key.
4364 * If exactp is non-null, stores whether the found entry was an exact match
4365 * in *exactp (1 or 0).
4366 * Updates the cursor index with the index of the found entry.
4367 * If no entry larger or equal to the key is found, returns NULL.
4370 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4372 unsigned int i = 0, nkeys;
4375 MDB_page *mp = mc->mc_pg[mc->mc_top];
4376 MDB_node *node = NULL;
4381 nkeys = NUMKEYS(mp);
4386 COPY_PGNO(pgno, mp->mp_pgno);
4387 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4388 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4395 low = IS_LEAF(mp) ? 0 : 1;
4397 cmp = mc->mc_dbx->md_cmp;
4399 /* Branch pages have no data, so if using integer keys,
4400 * alignment is guaranteed. Use faster mdb_cmp_int.
4402 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4403 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4410 nodekey.mv_size = mc->mc_db->md_pad;
4411 node = NODEPTR(mp, 0); /* fake */
4412 while (low <= high) {
4413 i = (low + high) >> 1;
4414 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4415 rc = cmp(key, &nodekey);
4416 DPRINTF(("found leaf index %u [%s], rc = %i",
4417 i, DKEY(&nodekey), rc));
4426 while (low <= high) {
4427 i = (low + high) >> 1;
4429 node = NODEPTR(mp, i);
4430 nodekey.mv_size = NODEKSZ(node);
4431 nodekey.mv_data = NODEKEY(node);
4433 rc = cmp(key, &nodekey);
4436 DPRINTF(("found leaf index %u [%s], rc = %i",
4437 i, DKEY(&nodekey), rc));
4439 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4440 i, DKEY(&nodekey), NODEPGNO(node), rc));
4451 if (rc > 0) { /* Found entry is less than the key. */
4452 i++; /* Skip to get the smallest entry larger than key. */
4454 node = NODEPTR(mp, i);
4457 *exactp = (rc == 0);
4458 /* store the key index */
4459 mc->mc_ki[mc->mc_top] = i;
4461 /* There is no entry larger or equal to the key. */
4464 /* nodeptr is fake for LEAF2 */
4470 mdb_cursor_adjust(MDB_cursor *mc, func)
4474 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4475 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4482 /** Pop a page off the top of the cursor's stack. */
4484 mdb_cursor_pop(MDB_cursor *mc)
4488 MDB_page *top = mc->mc_pg[mc->mc_top];
4494 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4495 mc->mc_dbi, (void *) mc));
4499 /** Push a page onto the top of the cursor's stack. */
4501 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4503 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4504 mc->mc_dbi, (void *) mc));
4506 if (mc->mc_snum >= CURSOR_STACK) {
4507 assert(mc->mc_snum < CURSOR_STACK);
4508 return MDB_CURSOR_FULL;
4511 mc->mc_top = mc->mc_snum++;
4512 mc->mc_pg[mc->mc_top] = mp;
4513 mc->mc_ki[mc->mc_top] = 0;
4518 /** Find the address of the page corresponding to a given page number.
4519 * @param[in] txn the transaction for this access.
4520 * @param[in] pgno the page number for the page to retrieve.
4521 * @param[out] ret address of a pointer where the page's address will be stored.
4522 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4523 * @return 0 on success, non-zero on failure.
4526 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4528 MDB_env *env = txn->mt_env;
4532 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4536 MDB_ID2L dl = tx2->mt_u.dirty_list;
4538 /* Spilled pages were dirtied in this txn and flushed
4539 * because the dirty list got full. Bring this page
4540 * back in from the map (but don't unspill it here,
4541 * leave that unless page_touch happens again).
4543 if (tx2->mt_spill_pgs) {
4544 MDB_ID pn = pgno << 1;
4545 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4546 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4547 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4552 unsigned x = mdb_mid2l_search(dl, pgno);
4553 if (x <= dl[0].mid && dl[x].mid == pgno) {
4559 } while ((tx2 = tx2->mt_parent) != NULL);
4562 if (pgno < txn->mt_next_pgno) {
4564 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4566 DPRINTF(("page %"Z"u not found", pgno));
4568 return MDB_PAGE_NOTFOUND;
4578 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4579 * The cursor is at the root page, set up the rest of it.
4582 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4584 MDB_page *mp = mc->mc_pg[mc->mc_top];
4588 while (IS_BRANCH(mp)) {
4592 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4593 assert(NUMKEYS(mp) > 1);
4594 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4596 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4598 if (flags & MDB_PS_LAST)
4599 i = NUMKEYS(mp) - 1;
4602 node = mdb_node_search(mc, key, &exact);
4604 i = NUMKEYS(mp) - 1;
4606 i = mc->mc_ki[mc->mc_top];
4612 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4615 assert(i < NUMKEYS(mp));
4616 node = NODEPTR(mp, i);
4618 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4621 mc->mc_ki[mc->mc_top] = i;
4622 if ((rc = mdb_cursor_push(mc, mp)))
4625 if (flags & MDB_PS_MODIFY) {
4626 if ((rc = mdb_page_touch(mc)) != 0)
4628 mp = mc->mc_pg[mc->mc_top];
4633 DPRINTF(("internal error, index points to a %02X page!?",
4635 return MDB_CORRUPTED;
4638 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4639 key ? DKEY(key) : "null"));
4640 mc->mc_flags |= C_INITIALIZED;
4641 mc->mc_flags &= ~C_EOF;
4646 /** Search for the lowest key under the current branch page.
4647 * This just bypasses a NUMKEYS check in the current page
4648 * before calling mdb_page_search_root(), because the callers
4649 * are all in situations where the current page is known to
4653 mdb_page_search_lowest(MDB_cursor *mc)
4655 MDB_page *mp = mc->mc_pg[mc->mc_top];
4656 MDB_node *node = NODEPTR(mp, 0);
4659 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4662 mc->mc_ki[mc->mc_top] = 0;
4663 if ((rc = mdb_cursor_push(mc, mp)))
4665 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4668 /** Search for the page a given key should be in.
4669 * Push it and its parent pages on the cursor stack.
4670 * @param[in,out] mc the cursor for this operation.
4671 * @param[in] key the key to search for, or NULL for first/last page.
4672 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4673 * are touched (updated with new page numbers).
4674 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4675 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4676 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4677 * @return 0 on success, non-zero on failure.
4680 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4685 /* Make sure the txn is still viable, then find the root from
4686 * the txn's db table and set it as the root of the cursor's stack.
4688 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4689 DPUTS("transaction has failed, must abort");
4692 /* Make sure we're using an up-to-date root */
4693 if (*mc->mc_dbflag & DB_STALE) {
4695 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4696 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4703 MDB_node *leaf = mdb_node_search(&mc2,
4704 &mc->mc_dbx->md_name, &exact);
4706 return MDB_NOTFOUND;
4707 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4710 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4712 /* The txn may not know this DBI, or another process may
4713 * have dropped and recreated the DB with other flags.
4715 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4716 return MDB_INCOMPATIBLE;
4717 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4719 *mc->mc_dbflag &= ~DB_STALE;
4721 root = mc->mc_db->md_root;
4723 if (root == P_INVALID) { /* Tree is empty. */
4724 DPUTS("tree is empty");
4725 return MDB_NOTFOUND;
4730 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4731 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4737 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4738 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4740 if (flags & MDB_PS_MODIFY) {
4741 if ((rc = mdb_page_touch(mc)))
4745 if (flags & MDB_PS_ROOTONLY)
4748 return mdb_page_search_root(mc, key, flags);
4752 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4754 MDB_txn *txn = mc->mc_txn;
4755 pgno_t pg = mp->mp_pgno;
4756 unsigned x = 0, ovpages = mp->mp_pages;
4757 MDB_env *env = txn->mt_env;
4758 MDB_IDL sl = txn->mt_spill_pgs;
4759 MDB_ID pn = pg << 1;
4762 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4763 /* If the page is dirty or on the spill list we just acquired it,
4764 * so we should give it back to our current free list, if any.
4765 * Otherwise put it onto the list of pages we freed in this txn.
4767 * Won't create me_pghead: me_pglast must be inited along with it.
4768 * Unsupported in nested txns: They would need to hide the page
4769 * range in ancestor txns' dirty and spilled lists.
4771 if (env->me_pghead &&
4773 ((mp->mp_flags & P_DIRTY) ||
4774 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4778 MDB_ID2 *dl, ix, iy;
4779 rc = mdb_midl_need(&env->me_pghead, ovpages);
4782 if (!(mp->mp_flags & P_DIRTY)) {
4783 /* This page is no longer spilled */
4790 /* Remove from dirty list */
4791 dl = txn->mt_u.dirty_list;
4793 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4801 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4802 txn->mt_flags |= MDB_TXN_ERROR;
4803 return MDB_CORRUPTED;
4806 if (!(env->me_flags & MDB_WRITEMAP))
4807 mdb_dpage_free(env, mp);
4809 /* Insert in me_pghead */
4810 mop = env->me_pghead;
4811 j = mop[0] + ovpages;
4812 for (i = mop[0]; i && mop[i] < pg; i--)
4818 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4822 mc->mc_db->md_overflow_pages -= ovpages;
4826 /** Return the data associated with a given node.
4827 * @param[in] txn The transaction for this operation.
4828 * @param[in] leaf The node being read.
4829 * @param[out] data Updated to point to the node's data.
4830 * @return 0 on success, non-zero on failure.
4833 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4835 MDB_page *omp; /* overflow page */
4839 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4840 data->mv_size = NODEDSZ(leaf);
4841 data->mv_data = NODEDATA(leaf);
4845 /* Read overflow data.
4847 data->mv_size = NODEDSZ(leaf);
4848 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4849 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4850 DPRINTF(("read overflow page %"Z"u failed", pgno));
4853 data->mv_data = METADATA(omp);
4859 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4860 MDB_val *key, MDB_val *data)
4869 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4871 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4874 if (txn->mt_flags & MDB_TXN_ERROR)
4877 if (key->mv_size > MDB_MAXKEYSIZE) {
4878 return MDB_BAD_VALSIZE;
4881 mdb_cursor_init(&mc, txn, dbi, &mx);
4882 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4885 /** Find a sibling for a page.
4886 * Replaces the page at the top of the cursor's stack with the
4887 * specified sibling, if one exists.
4888 * @param[in] mc The cursor for this operation.
4889 * @param[in] move_right Non-zero if the right sibling is requested,
4890 * otherwise the left sibling.
4891 * @return 0 on success, non-zero on failure.
4894 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4900 if (mc->mc_snum < 2) {
4901 return MDB_NOTFOUND; /* root has no siblings */
4905 DPRINTF(("parent page is page %"Z"u, index %u",
4906 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4908 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4909 : (mc->mc_ki[mc->mc_top] == 0)) {
4910 DPRINTF(("no more keys left, moving to %s sibling",
4911 move_right ? "right" : "left"));
4912 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4913 /* undo cursor_pop before returning */
4920 mc->mc_ki[mc->mc_top]++;
4922 mc->mc_ki[mc->mc_top]--;
4923 DPRINTF(("just moving to %s index key %u",
4924 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4926 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4928 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4929 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4932 mdb_cursor_push(mc, mp);
4934 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4939 /** Move the cursor to the next data item. */
4941 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4947 if (mc->mc_flags & C_EOF) {
4948 return MDB_NOTFOUND;
4951 assert(mc->mc_flags & C_INITIALIZED);
4953 mp = mc->mc_pg[mc->mc_top];
4955 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4956 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4957 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4958 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4959 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4960 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
4961 if (rc == MDB_SUCCESS)
4962 MDB_GET_KEY(leaf, key);
4967 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4968 if (op == MDB_NEXT_DUP)
4969 return MDB_NOTFOUND;
4973 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4974 if (mc->mc_flags & C_DEL)
4977 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4978 DPUTS("=====> move to next sibling page");
4979 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4980 mc->mc_flags |= C_EOF;
4983 mp = mc->mc_pg[mc->mc_top];
4984 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
4986 mc->mc_ki[mc->mc_top]++;
4989 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
4990 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
4993 key->mv_size = mc->mc_db->md_pad;
4994 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4998 assert(IS_LEAF(mp));
4999 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5001 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5002 mdb_xcursor_init1(mc, leaf);
5005 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5008 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5009 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5010 if (rc != MDB_SUCCESS)
5015 MDB_GET_KEY(leaf, key);
5019 /** Move the cursor to the previous data item. */
5021 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5027 assert(mc->mc_flags & C_INITIALIZED);
5029 mp = mc->mc_pg[mc->mc_top];
5031 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5032 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5033 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5034 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5035 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5036 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5037 if (rc == MDB_SUCCESS)
5038 MDB_GET_KEY(leaf, key);
5042 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5043 if (op == MDB_PREV_DUP)
5044 return MDB_NOTFOUND;
5049 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5051 if (mc->mc_ki[mc->mc_top] == 0) {
5052 DPUTS("=====> move to prev sibling page");
5053 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5056 mp = mc->mc_pg[mc->mc_top];
5057 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5058 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5060 mc->mc_ki[mc->mc_top]--;
5062 mc->mc_flags &= ~C_EOF;
5064 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5065 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5068 key->mv_size = mc->mc_db->md_pad;
5069 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5073 assert(IS_LEAF(mp));
5074 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5076 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5077 mdb_xcursor_init1(mc, leaf);
5080 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5083 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5084 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5085 if (rc != MDB_SUCCESS)
5090 MDB_GET_KEY(leaf, key);
5094 /** Set the cursor on a specific data item. */
5096 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5097 MDB_cursor_op op, int *exactp)
5101 MDB_node *leaf = NULL;
5106 if (key->mv_size == 0)
5107 return MDB_BAD_VALSIZE;
5110 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5112 /* See if we're already on the right page */
5113 if (mc->mc_flags & C_INITIALIZED) {
5116 mp = mc->mc_pg[mc->mc_top];
5118 mc->mc_ki[mc->mc_top] = 0;
5119 return MDB_NOTFOUND;
5121 if (mp->mp_flags & P_LEAF2) {
5122 nodekey.mv_size = mc->mc_db->md_pad;
5123 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5125 leaf = NODEPTR(mp, 0);
5126 MDB_GET_KEY2(leaf, nodekey);
5128 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5130 /* Probably happens rarely, but first node on the page
5131 * was the one we wanted.
5133 mc->mc_ki[mc->mc_top] = 0;
5140 unsigned int nkeys = NUMKEYS(mp);
5142 if (mp->mp_flags & P_LEAF2) {
5143 nodekey.mv_data = LEAF2KEY(mp,
5144 nkeys-1, nodekey.mv_size);
5146 leaf = NODEPTR(mp, nkeys-1);
5147 MDB_GET_KEY2(leaf, nodekey);
5149 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5151 /* last node was the one we wanted */
5152 mc->mc_ki[mc->mc_top] = nkeys-1;
5158 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5159 /* This is definitely the right page, skip search_page */
5160 if (mp->mp_flags & P_LEAF2) {
5161 nodekey.mv_data = LEAF2KEY(mp,
5162 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5164 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5165 MDB_GET_KEY2(leaf, nodekey);
5167 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5169 /* current node was the one we wanted */
5179 /* If any parents have right-sibs, search.
5180 * Otherwise, there's nothing further.
5182 for (i=0; i<mc->mc_top; i++)
5184 NUMKEYS(mc->mc_pg[i])-1)
5186 if (i == mc->mc_top) {
5187 /* There are no other pages */
5188 mc->mc_ki[mc->mc_top] = nkeys;
5189 return MDB_NOTFOUND;
5193 /* There are no other pages */
5194 mc->mc_ki[mc->mc_top] = 0;
5195 if (op == MDB_SET_RANGE) {
5199 return MDB_NOTFOUND;
5203 rc = mdb_page_search(mc, key, 0);
5204 if (rc != MDB_SUCCESS)
5207 mp = mc->mc_pg[mc->mc_top];
5208 assert(IS_LEAF(mp));
5211 leaf = mdb_node_search(mc, key, exactp);
5212 if (exactp != NULL && !*exactp) {
5213 /* MDB_SET specified and not an exact match. */
5214 return MDB_NOTFOUND;
5218 DPUTS("===> inexact leaf not found, goto sibling");
5219 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5220 return rc; /* no entries matched */
5221 mp = mc->mc_pg[mc->mc_top];
5222 assert(IS_LEAF(mp));
5223 leaf = NODEPTR(mp, 0);
5227 mc->mc_flags |= C_INITIALIZED;
5228 mc->mc_flags &= ~C_EOF;
5231 key->mv_size = mc->mc_db->md_pad;
5232 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5236 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5237 mdb_xcursor_init1(mc, leaf);
5240 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5241 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5242 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5245 if (op == MDB_GET_BOTH) {
5251 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5252 if (rc != MDB_SUCCESS)
5255 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5257 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5259 rc = mc->mc_dbx->md_dcmp(data, &d2);
5261 if (op == MDB_GET_BOTH || rc > 0)
5262 return MDB_NOTFOUND;
5268 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5269 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5274 /* The key already matches in all other cases */
5275 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5276 MDB_GET_KEY(leaf, key);
5277 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5282 /** Move the cursor to the first item in the database. */
5284 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5290 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5292 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5293 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5294 if (rc != MDB_SUCCESS)
5297 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5299 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5300 mc->mc_flags |= C_INITIALIZED;
5301 mc->mc_flags &= ~C_EOF;
5303 mc->mc_ki[mc->mc_top] = 0;
5305 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5306 key->mv_size = mc->mc_db->md_pad;
5307 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5312 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5313 mdb_xcursor_init1(mc, leaf);
5314 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5318 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5322 MDB_GET_KEY(leaf, key);
5326 /** Move the cursor to the last item in the database. */
5328 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5334 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5336 if (!(mc->mc_flags & C_EOF)) {
5338 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5339 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5340 if (rc != MDB_SUCCESS)
5343 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5346 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5347 mc->mc_flags |= C_INITIALIZED|C_EOF;
5348 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5350 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5351 key->mv_size = mc->mc_db->md_pad;
5352 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5357 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5358 mdb_xcursor_init1(mc, leaf);
5359 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5363 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5368 MDB_GET_KEY(leaf, key);
5373 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5378 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5382 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5386 case MDB_GET_CURRENT:
5387 if (!(mc->mc_flags & C_INITIALIZED)) {
5390 MDB_page *mp = mc->mc_pg[mc->mc_top];
5392 mc->mc_ki[mc->mc_top] = 0;
5398 key->mv_size = mc->mc_db->md_pad;
5399 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5401 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5402 MDB_GET_KEY(leaf, key);
5404 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5405 if (mc->mc_flags & C_DEL)
5406 mdb_xcursor_init1(mc, leaf);
5407 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5409 rc = mdb_node_read(mc->mc_txn, leaf, data);
5416 case MDB_GET_BOTH_RANGE:
5421 if (mc->mc_xcursor == NULL) {
5422 rc = MDB_INCOMPATIBLE;
5431 } else if (key->mv_size > MDB_MAXKEYSIZE) {
5432 rc = MDB_BAD_VALSIZE;
5433 } else if (op == MDB_SET_RANGE)
5434 rc = mdb_cursor_set(mc, key, data, op, NULL);
5436 rc = mdb_cursor_set(mc, key, data, op, &exact);
5438 case MDB_GET_MULTIPLE:
5439 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5443 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5444 rc = MDB_INCOMPATIBLE;
5448 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5449 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5452 case MDB_NEXT_MULTIPLE:
5457 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5458 rc = MDB_INCOMPATIBLE;
5461 if (!(mc->mc_flags & C_INITIALIZED))
5462 rc = mdb_cursor_first(mc, key, data);
5464 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5465 if (rc == MDB_SUCCESS) {
5466 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5469 mx = &mc->mc_xcursor->mx_cursor;
5470 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5472 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5473 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5481 case MDB_NEXT_NODUP:
5482 if (!(mc->mc_flags & C_INITIALIZED))
5483 rc = mdb_cursor_first(mc, key, data);
5485 rc = mdb_cursor_next(mc, key, data, op);
5489 case MDB_PREV_NODUP:
5490 if (!(mc->mc_flags & C_INITIALIZED)) {
5491 rc = mdb_cursor_last(mc, key, data);
5494 mc->mc_flags |= C_INITIALIZED;
5495 mc->mc_ki[mc->mc_top]++;
5497 rc = mdb_cursor_prev(mc, key, data, op);
5500 rc = mdb_cursor_first(mc, key, data);
5503 mfunc = mdb_cursor_first;
5505 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5509 if (mc->mc_xcursor == NULL) {
5510 rc = MDB_INCOMPATIBLE;
5513 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5517 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5520 rc = mdb_cursor_last(mc, key, data);
5523 mfunc = mdb_cursor_last;
5526 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5531 if (mc->mc_flags & C_DEL)
5532 mc->mc_flags ^= C_DEL;
5537 /** Touch all the pages in the cursor stack.
5538 * Makes sure all the pages are writable, before attempting a write operation.
5539 * @param[in] mc The cursor to operate on.
5542 mdb_cursor_touch(MDB_cursor *mc)
5546 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5549 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5550 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5553 *mc->mc_dbflag |= DB_DIRTY;
5555 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5556 rc = mdb_page_touch(mc);
5560 mc->mc_top = mc->mc_snum-1;
5564 /** Do not spill pages to disk if txn is getting full, may fail instead */
5565 #define MDB_NOSPILL 0x8000
5568 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5571 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5572 MDB_node *leaf = NULL;
5573 MDB_val xdata, *rdata, dkey;
5576 int do_sub = 0, insert = 0;
5577 unsigned int mcount = 0, dcount = 0, nospill;
5581 char dbuf[MDB_MAXKEYSIZE+1];
5582 unsigned int nflags;
5585 /* Check this first so counter will always be zero on any
5588 if (flags & MDB_MULTIPLE) {
5589 dcount = data[1].mv_size;
5590 data[1].mv_size = 0;
5591 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5592 return MDB_INCOMPATIBLE;
5595 nospill = flags & MDB_NOSPILL;
5596 flags &= ~MDB_NOSPILL;
5598 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5599 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5601 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5602 return MDB_BAD_VALSIZE;
5604 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5605 return MDB_BAD_VALSIZE;
5607 #if SIZE_MAX > MAXDATASIZE
5608 if (data->mv_size > MAXDATASIZE)
5609 return MDB_BAD_VALSIZE;
5612 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5613 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5617 if (flags == MDB_CURRENT) {
5618 if (!(mc->mc_flags & C_INITIALIZED))
5621 } else if (mc->mc_db->md_root == P_INVALID) {
5622 /* new database, cursor has nothing to point to */
5624 mc->mc_flags &= ~C_INITIALIZED;
5629 if (flags & MDB_APPEND) {
5631 rc = mdb_cursor_last(mc, &k2, &d2);
5633 rc = mc->mc_dbx->md_cmp(key, &k2);
5636 mc->mc_ki[mc->mc_top]++;
5638 /* new key is <= last key */
5643 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5645 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5646 DPRINTF(("duplicate key [%s]", DKEY(key)));
5648 return MDB_KEYEXIST;
5650 if (rc && rc != MDB_NOTFOUND)
5654 if (mc->mc_flags & C_DEL)
5655 mc->mc_flags ^= C_DEL;
5657 /* Cursor is positioned, check for room in the dirty list */
5659 if (flags & MDB_MULTIPLE) {
5661 xdata.mv_size = data->mv_size * dcount;
5665 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5669 if (rc == MDB_NO_ROOT) {
5671 /* new database, write a root leaf page */
5672 DPUTS("allocating new root leaf page");
5673 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5676 mdb_cursor_push(mc, np);
5677 mc->mc_db->md_root = np->mp_pgno;
5678 mc->mc_db->md_depth++;
5679 *mc->mc_dbflag |= DB_DIRTY;
5680 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5682 np->mp_flags |= P_LEAF2;
5683 mc->mc_flags |= C_INITIALIZED;
5685 /* make sure all cursor pages are writable */
5686 rc2 = mdb_cursor_touch(mc);
5691 /* The key already exists */
5692 if (rc == MDB_SUCCESS) {
5693 /* there's only a key anyway, so this is a no-op */
5694 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5695 unsigned int ksize = mc->mc_db->md_pad;
5696 if (key->mv_size != ksize)
5697 return MDB_BAD_VALSIZE;
5698 if (flags == MDB_CURRENT) {
5699 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5700 memcpy(ptr, key->mv_data, ksize);
5705 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5708 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5709 /* Was a single item before, must convert now */
5711 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5712 /* Just overwrite the current item */
5713 if (flags == MDB_CURRENT)
5716 dkey.mv_size = NODEDSZ(leaf);
5717 dkey.mv_data = NODEDATA(leaf);
5718 #if UINT_MAX < SIZE_MAX
5719 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5720 #ifdef MISALIGNED_OK
5721 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5723 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5726 /* if data matches, skip it */
5727 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5728 if (flags & MDB_NODUPDATA)
5730 else if (flags & MDB_MULTIPLE)
5737 /* create a fake page for the dup items */
5738 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5739 dkey.mv_data = dbuf;
5740 fp = (MDB_page *)&pbuf;
5741 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5742 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5743 fp->mp_lower = PAGEHDRSZ;
5744 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5745 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5746 fp->mp_flags |= P_LEAF2;
5747 fp->mp_pad = data->mv_size;
5748 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5750 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5751 (dkey.mv_size & 1) + (data->mv_size & 1);
5753 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5756 xdata.mv_size = fp->mp_upper;
5761 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5762 /* See if we need to convert from fake page to subDB */
5764 unsigned int offset;
5768 fp = NODEDATA(leaf);
5769 if (flags == MDB_CURRENT) {
5771 fp->mp_flags |= P_DIRTY;
5772 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5773 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5777 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5778 offset = fp->mp_pad;
5779 if (SIZELEFT(fp) >= offset)
5781 offset *= 4; /* space for 4 more */
5783 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5785 offset += offset & 1;
5786 fp_flags = fp->mp_flags;
5787 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5788 offset >= mc->mc_txn->mt_env->me_nodemax) {
5789 /* yes, convert it */
5791 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5792 dummy.md_pad = fp->mp_pad;
5793 dummy.md_flags = MDB_DUPFIXED;
5794 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5795 dummy.md_flags |= MDB_INTEGERKEY;
5798 dummy.md_branch_pages = 0;
5799 dummy.md_leaf_pages = 1;
5800 dummy.md_overflow_pages = 0;
5801 dummy.md_entries = NUMKEYS(fp);
5803 xdata.mv_size = sizeof(MDB_db);
5804 xdata.mv_data = &dummy;
5805 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5807 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5808 flags |= F_DUPDATA|F_SUBDATA;
5809 dummy.md_root = mp->mp_pgno;
5810 fp_flags &= ~P_SUBP;
5812 /* no, just grow it */
5814 xdata.mv_size = NODEDSZ(leaf) + offset;
5815 xdata.mv_data = &pbuf;
5816 mp = (MDB_page *)&pbuf;
5817 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5820 mp->mp_flags = fp_flags | P_DIRTY;
5821 mp->mp_pad = fp->mp_pad;
5822 mp->mp_lower = fp->mp_lower;
5823 mp->mp_upper = fp->mp_upper + offset;
5825 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5827 nsize = NODEDSZ(leaf) - fp->mp_upper;
5828 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5829 for (i=0; i<NUMKEYS(fp); i++)
5830 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5832 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5836 /* data is on sub-DB, just store it */
5837 flags |= F_DUPDATA|F_SUBDATA;
5841 /* overflow page overwrites need special handling */
5842 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5845 unsigned psize = mc->mc_txn->mt_env->me_psize;
5846 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5848 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5849 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5851 ovpages = omp->mp_pages;
5853 /* Is the ov page large enough? */
5854 if (ovpages >= dpages) {
5855 if (!(omp->mp_flags & P_DIRTY) &&
5856 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5858 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5861 level = 0; /* dirty in this txn or clean */
5864 if (omp->mp_flags & P_DIRTY) {
5865 /* yes, overwrite it. Note in this case we don't
5866 * bother to try shrinking the page if the new data
5867 * is smaller than the overflow threshold.
5870 /* It is writable only in a parent txn */
5871 size_t sz = (size_t) psize * ovpages, off;
5872 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5878 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5879 if (!(flags & MDB_RESERVE)) {
5880 /* Copy end of page, adjusting alignment so
5881 * compiler may copy words instead of bytes.
5883 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5884 memcpy((size_t *)((char *)np + off),
5885 (size_t *)((char *)omp + off), sz - off);
5888 memcpy(np, omp, sz); /* Copy beginning of page */
5891 SETDSZ(leaf, data->mv_size);
5892 if (F_ISSET(flags, MDB_RESERVE))
5893 data->mv_data = METADATA(omp);
5895 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5899 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5901 } else if (NODEDSZ(leaf) == data->mv_size) {
5902 /* same size, just replace it. Note that we could
5903 * also reuse this node if the new data is smaller,
5904 * but instead we opt to shrink the node in that case.
5906 if (F_ISSET(flags, MDB_RESERVE))
5907 data->mv_data = NODEDATA(leaf);
5908 else if (data->mv_size)
5909 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5911 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5914 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5915 mc->mc_db->md_entries--;
5917 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5924 nflags = flags & NODE_ADD_FLAGS;
5925 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5926 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5927 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5928 nflags &= ~MDB_APPEND;
5930 nflags |= MDB_SPLIT_REPLACE;
5931 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5933 /* There is room already in this leaf page. */
5934 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5935 if (rc == 0 && !do_sub && insert) {
5936 /* Adjust other cursors pointing to mp */
5937 MDB_cursor *m2, *m3;
5938 MDB_dbi dbi = mc->mc_dbi;
5939 unsigned i = mc->mc_top;
5940 MDB_page *mp = mc->mc_pg[i];
5942 if (mc->mc_flags & C_SUB)
5945 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5946 if (mc->mc_flags & C_SUB)
5947 m3 = &m2->mc_xcursor->mx_cursor;
5950 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5951 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5958 if (rc != MDB_SUCCESS)
5959 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5961 /* Now store the actual data in the child DB. Note that we're
5962 * storing the user data in the keys field, so there are strict
5963 * size limits on dupdata. The actual data fields of the child
5964 * DB are all zero size.
5971 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5972 if (flags & MDB_CURRENT) {
5973 xflags = MDB_CURRENT|MDB_NOSPILL;
5975 mdb_xcursor_init1(mc, leaf);
5976 xflags = (flags & MDB_NODUPDATA) ?
5977 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5979 /* converted, write the original data first */
5981 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5985 /* Adjust other cursors pointing to mp */
5987 unsigned i = mc->mc_top;
5988 MDB_page *mp = mc->mc_pg[i];
5990 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5991 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5992 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5993 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5994 mdb_xcursor_init1(m2, leaf);
5998 /* we've done our job */
6001 if (flags & MDB_APPENDDUP)
6002 xflags |= MDB_APPEND;
6003 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6004 if (flags & F_SUBDATA) {
6005 void *db = NODEDATA(leaf);
6006 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6009 /* sub-writes might have failed so check rc again.
6010 * Don't increment count if we just replaced an existing item.
6012 if (!rc && !(flags & MDB_CURRENT))
6013 mc->mc_db->md_entries++;
6014 if (flags & MDB_MULTIPLE) {
6018 /* let caller know how many succeeded, if any */
6019 data[1].mv_size = mcount;
6020 if (mcount < dcount) {
6021 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6022 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6029 /* If we succeeded and the key didn't exist before, make sure
6030 * the cursor is marked valid.
6033 mc->mc_flags |= C_INITIALIZED;
6038 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6043 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6044 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6046 if (!(mc->mc_flags & C_INITIALIZED))
6049 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6052 rc = mdb_cursor_touch(mc);
6056 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6058 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6059 if (!(flags & MDB_NODUPDATA)) {
6060 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6061 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6063 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6064 /* If sub-DB still has entries, we're done */
6065 if (mc->mc_xcursor->mx_db.md_entries) {
6066 if (leaf->mn_flags & F_SUBDATA) {
6067 /* update subDB info */
6068 void *db = NODEDATA(leaf);
6069 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6072 /* shrink fake page */
6073 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6074 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6075 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6076 /* fix other sub-DB cursors pointed at this fake page */
6077 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6078 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6079 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6080 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6081 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6084 mc->mc_db->md_entries--;
6085 mc->mc_flags |= C_DEL;
6088 /* otherwise fall thru and delete the sub-DB */
6091 if (leaf->mn_flags & F_SUBDATA) {
6092 /* add all the child DB's pages to the free list */
6093 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6094 if (rc == MDB_SUCCESS) {
6095 mc->mc_db->md_entries -=
6096 mc->mc_xcursor->mx_db.md_entries;
6101 return mdb_cursor_del0(mc, leaf);
6104 /** Allocate and initialize new pages for a database.
6105 * @param[in] mc a cursor on the database being added to.
6106 * @param[in] flags flags defining what type of page is being allocated.
6107 * @param[in] num the number of pages to allocate. This is usually 1,
6108 * unless allocating overflow pages for a large record.
6109 * @param[out] mp Address of a page, or NULL on failure.
6110 * @return 0 on success, non-zero on failure.
6113 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6118 if ((rc = mdb_page_alloc(mc, num, &np)))
6120 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6121 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6122 np->mp_flags = flags | P_DIRTY;
6123 np->mp_lower = PAGEHDRSZ;
6124 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6127 mc->mc_db->md_branch_pages++;
6128 else if (IS_LEAF(np))
6129 mc->mc_db->md_leaf_pages++;
6130 else if (IS_OVERFLOW(np)) {
6131 mc->mc_db->md_overflow_pages += num;
6139 /** Calculate the size of a leaf node.
6140 * The size depends on the environment's page size; if a data item
6141 * is too large it will be put onto an overflow page and the node
6142 * size will only include the key and not the data. Sizes are always
6143 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6144 * of the #MDB_node headers.
6145 * @param[in] env The environment handle.
6146 * @param[in] key The key for the node.
6147 * @param[in] data The data for the node.
6148 * @return The number of bytes needed to store the node.
6151 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6155 sz = LEAFSIZE(key, data);
6156 if (sz >= env->me_nodemax) {
6157 /* put on overflow page */
6158 sz -= data->mv_size - sizeof(pgno_t);
6162 return sz + sizeof(indx_t);
6165 /** Calculate the size of a branch node.
6166 * The size should depend on the environment's page size but since
6167 * we currently don't support spilling large keys onto overflow
6168 * pages, it's simply the size of the #MDB_node header plus the
6169 * size of the key. Sizes are always rounded up to an even number
6170 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6171 * @param[in] env The environment handle.
6172 * @param[in] key The key for the node.
6173 * @return The number of bytes needed to store the node.
6176 mdb_branch_size(MDB_env *env, MDB_val *key)
6181 if (sz >= env->me_nodemax) {
6182 /* put on overflow page */
6183 /* not implemented */
6184 /* sz -= key->size - sizeof(pgno_t); */
6187 return sz + sizeof(indx_t);
6190 /** Add a node to the page pointed to by the cursor.
6191 * @param[in] mc The cursor for this operation.
6192 * @param[in] indx The index on the page where the new node should be added.
6193 * @param[in] key The key for the new node.
6194 * @param[in] data The data for the new node, if any.
6195 * @param[in] pgno The page number, if adding a branch node.
6196 * @param[in] flags Flags for the node.
6197 * @return 0 on success, non-zero on failure. Possible errors are:
6199 * <li>ENOMEM - failed to allocate overflow pages for the node.
6200 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6201 * should never happen since all callers already calculate the
6202 * page's free space before calling this function.
6206 mdb_node_add(MDB_cursor *mc, indx_t indx,
6207 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6210 size_t node_size = NODESIZE;
6213 MDB_page *mp = mc->mc_pg[mc->mc_top];
6214 MDB_page *ofp = NULL; /* overflow page */
6217 assert(mp->mp_upper >= mp->mp_lower);
6219 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6220 IS_LEAF(mp) ? "leaf" : "branch",
6221 IS_SUBP(mp) ? "sub-" : "",
6222 mp->mp_pgno, indx, data ? data->mv_size : 0,
6223 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6226 /* Move higher keys up one slot. */
6227 int ksize = mc->mc_db->md_pad, dif;
6228 char *ptr = LEAF2KEY(mp, indx, ksize);
6229 dif = NUMKEYS(mp) - indx;
6231 memmove(ptr+ksize, ptr, dif*ksize);
6232 /* insert new key */
6233 memcpy(ptr, key->mv_data, ksize);
6235 /* Just using these for counting */
6236 mp->mp_lower += sizeof(indx_t);
6237 mp->mp_upper -= ksize - sizeof(indx_t);
6242 node_size += key->mv_size;
6246 if (F_ISSET(flags, F_BIGDATA)) {
6247 /* Data already on overflow page. */
6248 node_size += sizeof(pgno_t);
6249 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6250 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6252 /* Put data on overflow page. */
6253 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6254 data->mv_size, node_size+data->mv_size));
6255 node_size += sizeof(pgno_t);
6256 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6258 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6261 node_size += data->mv_size;
6264 node_size += node_size & 1;
6266 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6267 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6268 mp->mp_pgno, NUMKEYS(mp)));
6269 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6270 mp->mp_upper - mp->mp_lower));
6271 DPRINTF(("node size = %"Z"u", node_size));
6272 return MDB_PAGE_FULL;
6275 /* Move higher pointers up one slot. */
6276 for (i = NUMKEYS(mp); i > indx; i--)
6277 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6279 /* Adjust free space offsets. */
6280 ofs = mp->mp_upper - node_size;
6281 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6282 mp->mp_ptrs[indx] = ofs;
6284 mp->mp_lower += sizeof(indx_t);
6286 /* Write the node data. */
6287 node = NODEPTR(mp, indx);
6288 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6289 node->mn_flags = flags;
6291 SETDSZ(node,data->mv_size);
6296 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6301 if (F_ISSET(flags, F_BIGDATA))
6302 memcpy(node->mn_data + key->mv_size, data->mv_data,
6304 else if (F_ISSET(flags, MDB_RESERVE))
6305 data->mv_data = node->mn_data + key->mv_size;
6307 memcpy(node->mn_data + key->mv_size, data->mv_data,
6310 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6312 if (F_ISSET(flags, MDB_RESERVE))
6313 data->mv_data = METADATA(ofp);
6315 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6322 /** Delete the specified node from a page.
6323 * @param[in] mp The page to operate on.
6324 * @param[in] indx The index of the node to delete.
6325 * @param[in] ksize The size of a node. Only used if the page is
6326 * part of a #MDB_DUPFIXED database.
6329 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6332 indx_t i, j, numkeys, ptr;
6339 COPY_PGNO(pgno, mp->mp_pgno);
6340 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6341 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6344 assert(indx < NUMKEYS(mp));
6347 int x = NUMKEYS(mp) - 1 - indx;
6348 base = LEAF2KEY(mp, indx, ksize);
6350 memmove(base, base + ksize, x * ksize);
6351 mp->mp_lower -= sizeof(indx_t);
6352 mp->mp_upper += ksize - sizeof(indx_t);
6356 node = NODEPTR(mp, indx);
6357 sz = NODESIZE + node->mn_ksize;
6359 if (F_ISSET(node->mn_flags, F_BIGDATA))
6360 sz += sizeof(pgno_t);
6362 sz += NODEDSZ(node);
6366 ptr = mp->mp_ptrs[indx];
6367 numkeys = NUMKEYS(mp);
6368 for (i = j = 0; i < numkeys; i++) {
6370 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6371 if (mp->mp_ptrs[i] < ptr)
6372 mp->mp_ptrs[j] += sz;
6377 base = (char *)mp + mp->mp_upper;
6378 memmove(base + sz, base, ptr - mp->mp_upper);
6380 mp->mp_lower -= sizeof(indx_t);
6384 /** Compact the main page after deleting a node on a subpage.
6385 * @param[in] mp The main page to operate on.
6386 * @param[in] indx The index of the subpage on the main page.
6389 mdb_node_shrink(MDB_page *mp, indx_t indx)
6396 indx_t i, numkeys, ptr;
6398 node = NODEPTR(mp, indx);
6399 sp = (MDB_page *)NODEDATA(node);
6400 osize = NODEDSZ(node);
6402 delta = sp->mp_upper - sp->mp_lower;
6403 SETDSZ(node, osize - delta);
6404 xp = (MDB_page *)((char *)sp + delta);
6406 /* shift subpage upward */
6408 nsize = NUMKEYS(sp) * sp->mp_pad;
6409 memmove(METADATA(xp), METADATA(sp), nsize);
6412 nsize = osize - sp->mp_upper;
6413 numkeys = NUMKEYS(sp);
6414 for (i=numkeys-1; i>=0; i--)
6415 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6417 xp->mp_upper = sp->mp_lower;
6418 xp->mp_lower = sp->mp_lower;
6419 xp->mp_flags = sp->mp_flags;
6420 xp->mp_pad = sp->mp_pad;
6421 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6423 /* shift lower nodes upward */
6424 ptr = mp->mp_ptrs[indx];
6425 numkeys = NUMKEYS(mp);
6426 for (i = 0; i < numkeys; i++) {
6427 if (mp->mp_ptrs[i] <= ptr)
6428 mp->mp_ptrs[i] += delta;
6431 base = (char *)mp + mp->mp_upper;
6432 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6433 mp->mp_upper += delta;
6436 /** Initial setup of a sorted-dups cursor.
6437 * Sorted duplicates are implemented as a sub-database for the given key.
6438 * The duplicate data items are actually keys of the sub-database.
6439 * Operations on the duplicate data items are performed using a sub-cursor
6440 * initialized when the sub-database is first accessed. This function does
6441 * the preliminary setup of the sub-cursor, filling in the fields that
6442 * depend only on the parent DB.
6443 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6446 mdb_xcursor_init0(MDB_cursor *mc)
6448 MDB_xcursor *mx = mc->mc_xcursor;
6450 mx->mx_cursor.mc_xcursor = NULL;
6451 mx->mx_cursor.mc_txn = mc->mc_txn;
6452 mx->mx_cursor.mc_db = &mx->mx_db;
6453 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6454 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6455 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6456 mx->mx_cursor.mc_snum = 0;
6457 mx->mx_cursor.mc_top = 0;
6458 mx->mx_cursor.mc_flags = C_SUB;
6459 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6460 mx->mx_dbx.md_dcmp = NULL;
6461 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6464 /** Final setup of a sorted-dups cursor.
6465 * Sets up the fields that depend on the data from the main cursor.
6466 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6467 * @param[in] node The data containing the #MDB_db record for the
6468 * sorted-dup database.
6471 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6473 MDB_xcursor *mx = mc->mc_xcursor;
6475 if (node->mn_flags & F_SUBDATA) {
6476 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6477 mx->mx_cursor.mc_pg[0] = 0;
6478 mx->mx_cursor.mc_snum = 0;
6479 mx->mx_cursor.mc_flags = C_SUB;
6481 MDB_page *fp = NODEDATA(node);
6482 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6483 mx->mx_db.md_flags = 0;
6484 mx->mx_db.md_depth = 1;
6485 mx->mx_db.md_branch_pages = 0;
6486 mx->mx_db.md_leaf_pages = 1;
6487 mx->mx_db.md_overflow_pages = 0;
6488 mx->mx_db.md_entries = NUMKEYS(fp);
6489 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6490 mx->mx_cursor.mc_snum = 1;
6491 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6492 mx->mx_cursor.mc_top = 0;
6493 mx->mx_cursor.mc_pg[0] = fp;
6494 mx->mx_cursor.mc_ki[0] = 0;
6495 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6496 mx->mx_db.md_flags = MDB_DUPFIXED;
6497 mx->mx_db.md_pad = fp->mp_pad;
6498 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6499 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6502 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6503 mx->mx_db.md_root));
6504 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6506 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6507 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6508 #if UINT_MAX < SIZE_MAX
6509 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6510 #ifdef MISALIGNED_OK
6511 mx->mx_dbx.md_cmp = mdb_cmp_long;
6513 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6518 /** Initialize a cursor for a given transaction and database. */
6520 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6523 mc->mc_backup = NULL;
6526 mc->mc_db = &txn->mt_dbs[dbi];
6527 mc->mc_dbx = &txn->mt_dbxs[dbi];
6528 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6533 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6535 mc->mc_xcursor = mx;
6536 mdb_xcursor_init0(mc);
6538 mc->mc_xcursor = NULL;
6540 if (*mc->mc_dbflag & DB_STALE) {
6541 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6546 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6549 size_t size = sizeof(MDB_cursor);
6551 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6554 if (txn->mt_flags & MDB_TXN_ERROR)
6557 /* Allow read access to the freelist */
6558 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6561 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6562 size += sizeof(MDB_xcursor);
6564 if ((mc = malloc(size)) != NULL) {
6565 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6566 if (txn->mt_cursors) {
6567 mc->mc_next = txn->mt_cursors[dbi];
6568 txn->mt_cursors[dbi] = mc;
6569 mc->mc_flags |= C_UNTRACK;
6581 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6583 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6586 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6589 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6593 /* Return the count of duplicate data items for the current key */
6595 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6599 if (mc == NULL || countp == NULL)
6602 if (mc->mc_xcursor == NULL)
6603 return MDB_INCOMPATIBLE;
6605 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6606 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6609 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6612 *countp = mc->mc_xcursor->mx_db.md_entries;
6618 mdb_cursor_close(MDB_cursor *mc)
6620 if (mc && !mc->mc_backup) {
6621 /* remove from txn, if tracked */
6622 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6623 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6624 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6626 *prev = mc->mc_next;
6633 mdb_cursor_txn(MDB_cursor *mc)
6635 if (!mc) return NULL;
6640 mdb_cursor_dbi(MDB_cursor *mc)
6646 /** Replace the key for a node with a new key.
6647 * @param[in] mc Cursor pointing to the node to operate on.
6648 * @param[in] key The new key to use.
6649 * @return 0 on success, non-zero on failure.
6652 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6659 indx_t ptr, i, numkeys, indx;
6662 indx = mc->mc_ki[mc->mc_top];
6663 mp = mc->mc_pg[mc->mc_top];
6664 node = NODEPTR(mp, indx);
6665 ptr = mp->mp_ptrs[indx];
6669 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6670 k2.mv_data = NODEKEY(node);
6671 k2.mv_size = node->mn_ksize;
6672 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6674 mdb_dkey(&k2, kbuf2),
6680 delta0 = delta = key->mv_size - node->mn_ksize;
6682 /* Must be 2-byte aligned. If new key is
6683 * shorter by 1, the shift will be skipped.
6685 delta += (delta & 1);
6687 if (delta > 0 && SIZELEFT(mp) < delta) {
6689 /* not enough space left, do a delete and split */
6690 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6691 pgno = NODEPGNO(node);
6692 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6693 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6696 numkeys = NUMKEYS(mp);
6697 for (i = 0; i < numkeys; i++) {
6698 if (mp->mp_ptrs[i] <= ptr)
6699 mp->mp_ptrs[i] -= delta;
6702 base = (char *)mp + mp->mp_upper;
6703 len = ptr - mp->mp_upper + NODESIZE;
6704 memmove(base - delta, base, len);
6705 mp->mp_upper -= delta;
6707 node = NODEPTR(mp, indx);
6710 /* But even if no shift was needed, update ksize */
6712 node->mn_ksize = key->mv_size;
6715 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6721 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6723 /** Move a node from csrc to cdst.
6726 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6733 unsigned short flags;
6737 /* Mark src and dst as dirty. */
6738 if ((rc = mdb_page_touch(csrc)) ||
6739 (rc = mdb_page_touch(cdst)))
6742 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6743 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6744 key.mv_size = csrc->mc_db->md_pad;
6745 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6747 data.mv_data = NULL;
6751 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6752 assert(!((long)srcnode&1));
6753 srcpg = NODEPGNO(srcnode);
6754 flags = srcnode->mn_flags;
6755 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6756 unsigned int snum = csrc->mc_snum;
6758 /* must find the lowest key below src */
6759 mdb_page_search_lowest(csrc);
6760 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6761 key.mv_size = csrc->mc_db->md_pad;
6762 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6764 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6765 key.mv_size = NODEKSZ(s2);
6766 key.mv_data = NODEKEY(s2);
6768 csrc->mc_snum = snum--;
6769 csrc->mc_top = snum;
6771 key.mv_size = NODEKSZ(srcnode);
6772 key.mv_data = NODEKEY(srcnode);
6774 data.mv_size = NODEDSZ(srcnode);
6775 data.mv_data = NODEDATA(srcnode);
6777 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6778 unsigned int snum = cdst->mc_snum;
6781 /* must find the lowest key below dst */
6782 mdb_page_search_lowest(cdst);
6783 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6784 bkey.mv_size = cdst->mc_db->md_pad;
6785 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6787 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6788 bkey.mv_size = NODEKSZ(s2);
6789 bkey.mv_data = NODEKEY(s2);
6791 cdst->mc_snum = snum--;
6792 cdst->mc_top = snum;
6793 mdb_cursor_copy(cdst, &mn);
6795 rc = mdb_update_key(&mn, &bkey);
6800 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6801 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6802 csrc->mc_ki[csrc->mc_top],
6804 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6805 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6807 /* Add the node to the destination page.
6809 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6810 if (rc != MDB_SUCCESS)
6813 /* Delete the node from the source page.
6815 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6818 /* Adjust other cursors pointing to mp */
6819 MDB_cursor *m2, *m3;
6820 MDB_dbi dbi = csrc->mc_dbi;
6821 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6823 if (csrc->mc_flags & C_SUB)
6826 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6827 if (csrc->mc_flags & C_SUB)
6828 m3 = &m2->mc_xcursor->mx_cursor;
6831 if (m3 == csrc) continue;
6832 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6833 csrc->mc_ki[csrc->mc_top]) {
6834 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6835 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6840 /* Update the parent separators.
6842 if (csrc->mc_ki[csrc->mc_top] == 0) {
6843 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6844 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6845 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6847 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6848 key.mv_size = NODEKSZ(srcnode);
6849 key.mv_data = NODEKEY(srcnode);
6851 DPRINTF(("update separator for source page %"Z"u to [%s]",
6852 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6853 mdb_cursor_copy(csrc, &mn);
6856 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6859 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6861 indx_t ix = csrc->mc_ki[csrc->mc_top];
6862 nullkey.mv_size = 0;
6863 csrc->mc_ki[csrc->mc_top] = 0;
6864 rc = mdb_update_key(csrc, &nullkey);
6865 csrc->mc_ki[csrc->mc_top] = ix;
6866 assert(rc == MDB_SUCCESS);
6870 if (cdst->mc_ki[cdst->mc_top] == 0) {
6871 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6872 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6873 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6875 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6876 key.mv_size = NODEKSZ(srcnode);
6877 key.mv_data = NODEKEY(srcnode);
6879 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6880 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6881 mdb_cursor_copy(cdst, &mn);
6884 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6887 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6889 indx_t ix = cdst->mc_ki[cdst->mc_top];
6890 nullkey.mv_size = 0;
6891 cdst->mc_ki[cdst->mc_top] = 0;
6892 rc = mdb_update_key(cdst, &nullkey);
6893 cdst->mc_ki[cdst->mc_top] = ix;
6894 assert(rc == MDB_SUCCESS);
6901 /** Merge one page into another.
6902 * The nodes from the page pointed to by \b csrc will
6903 * be copied to the page pointed to by \b cdst and then
6904 * the \b csrc page will be freed.
6905 * @param[in] csrc Cursor pointing to the source page.
6906 * @param[in] cdst Cursor pointing to the destination page.
6909 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6917 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6918 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6920 assert(csrc->mc_snum > 1); /* can't merge root page */
6921 assert(cdst->mc_snum > 1);
6923 /* Mark dst as dirty. */
6924 if ((rc = mdb_page_touch(cdst)))
6927 /* Move all nodes from src to dst.
6929 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6930 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6931 key.mv_size = csrc->mc_db->md_pad;
6932 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6933 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6934 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6935 if (rc != MDB_SUCCESS)
6937 key.mv_data = (char *)key.mv_data + key.mv_size;
6940 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6941 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6942 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6943 unsigned int snum = csrc->mc_snum;
6945 /* must find the lowest key below src */
6946 mdb_page_search_lowest(csrc);
6947 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6948 key.mv_size = csrc->mc_db->md_pad;
6949 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6951 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6952 key.mv_size = NODEKSZ(s2);
6953 key.mv_data = NODEKEY(s2);
6955 csrc->mc_snum = snum--;
6956 csrc->mc_top = snum;
6958 key.mv_size = srcnode->mn_ksize;
6959 key.mv_data = NODEKEY(srcnode);
6962 data.mv_size = NODEDSZ(srcnode);
6963 data.mv_data = NODEDATA(srcnode);
6964 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6965 if (rc != MDB_SUCCESS)
6970 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6971 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
6972 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
6974 /* Unlink the src page from parent and add to free list.
6976 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6977 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6980 rc = mdb_update_key(csrc, &key);
6986 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6987 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6990 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6991 csrc->mc_db->md_leaf_pages--;
6993 csrc->mc_db->md_branch_pages--;
6995 /* Adjust other cursors pointing to mp */
6996 MDB_cursor *m2, *m3;
6997 MDB_dbi dbi = csrc->mc_dbi;
6998 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7000 if (csrc->mc_flags & C_SUB)
7003 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7004 if (csrc->mc_flags & C_SUB)
7005 m3 = &m2->mc_xcursor->mx_cursor;
7008 if (m3 == csrc) continue;
7009 if (m3->mc_snum < csrc->mc_snum) continue;
7010 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7011 m3->mc_pg[csrc->mc_top] = mp;
7012 m3->mc_ki[csrc->mc_top] += nkeys;
7016 mdb_cursor_pop(csrc);
7018 return mdb_rebalance(csrc);
7021 /** Copy the contents of a cursor.
7022 * @param[in] csrc The cursor to copy from.
7023 * @param[out] cdst The cursor to copy to.
7026 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7030 cdst->mc_txn = csrc->mc_txn;
7031 cdst->mc_dbi = csrc->mc_dbi;
7032 cdst->mc_db = csrc->mc_db;
7033 cdst->mc_dbx = csrc->mc_dbx;
7034 cdst->mc_snum = csrc->mc_snum;
7035 cdst->mc_top = csrc->mc_top;
7036 cdst->mc_flags = csrc->mc_flags;
7038 for (i=0; i<csrc->mc_snum; i++) {
7039 cdst->mc_pg[i] = csrc->mc_pg[i];
7040 cdst->mc_ki[i] = csrc->mc_ki[i];
7044 /** Rebalance the tree after a delete operation.
7045 * @param[in] mc Cursor pointing to the page where rebalancing
7047 * @return 0 on success, non-zero on failure.
7050 mdb_rebalance(MDB_cursor *mc)
7054 unsigned int ptop, minkeys;
7057 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7061 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7062 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7063 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7064 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7065 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7069 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7070 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7073 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7074 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7080 if (mc->mc_snum < 2) {
7081 MDB_page *mp = mc->mc_pg[0];
7083 DPUTS("Can't rebalance a subpage, ignoring");
7086 if (NUMKEYS(mp) == 0) {
7087 DPUTS("tree is completely empty");
7088 mc->mc_db->md_root = P_INVALID;
7089 mc->mc_db->md_depth = 0;
7090 mc->mc_db->md_leaf_pages = 0;
7091 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7094 /* Adjust cursors pointing to mp */
7098 MDB_cursor *m2, *m3;
7099 MDB_dbi dbi = mc->mc_dbi;
7101 if (mc->mc_flags & C_SUB)
7104 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7105 if (mc->mc_flags & C_SUB)
7106 m3 = &m2->mc_xcursor->mx_cursor;
7109 if (m3->mc_snum < mc->mc_snum) continue;
7110 if (m3->mc_pg[0] == mp) {
7116 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7117 DPUTS("collapsing root page!");
7118 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7121 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7122 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7125 mc->mc_db->md_depth--;
7126 mc->mc_db->md_branch_pages--;
7127 mc->mc_ki[0] = mc->mc_ki[1];
7129 /* Adjust other cursors pointing to mp */
7130 MDB_cursor *m2, *m3;
7131 MDB_dbi dbi = mc->mc_dbi;
7133 if (mc->mc_flags & C_SUB)
7136 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7137 if (mc->mc_flags & C_SUB)
7138 m3 = &m2->mc_xcursor->mx_cursor;
7141 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7142 if (m3->mc_pg[0] == mp) {
7143 m3->mc_pg[0] = mc->mc_pg[0];
7146 m3->mc_ki[0] = m3->mc_ki[1];
7151 DPUTS("root page doesn't need rebalancing");
7155 /* The parent (branch page) must have at least 2 pointers,
7156 * otherwise the tree is invalid.
7158 ptop = mc->mc_top-1;
7159 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7161 /* Leaf page fill factor is below the threshold.
7162 * Try to move keys from left or right neighbor, or
7163 * merge with a neighbor page.
7168 mdb_cursor_copy(mc, &mn);
7169 mn.mc_xcursor = NULL;
7171 if (mc->mc_ki[ptop] == 0) {
7172 /* We're the leftmost leaf in our parent.
7174 DPUTS("reading right neighbor");
7176 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7177 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7180 mn.mc_ki[mn.mc_top] = 0;
7181 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7183 /* There is at least one neighbor to the left.
7185 DPUTS("reading left neighbor");
7187 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7188 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7191 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7192 mc->mc_ki[mc->mc_top] = 0;
7195 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7196 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7197 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7199 /* If the neighbor page is above threshold and has enough keys,
7200 * move one key from it. Otherwise we should try to merge them.
7201 * (A branch page must never have less than 2 keys.)
7203 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7204 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7205 return mdb_node_move(&mn, mc);
7207 if (mc->mc_ki[ptop] == 0)
7208 rc = mdb_page_merge(&mn, mc);
7210 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7211 rc = mdb_page_merge(mc, &mn);
7212 mdb_cursor_copy(&mn, mc);
7214 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7219 /** Complete a delete operation started by #mdb_cursor_del(). */
7221 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7228 mp = mc->mc_pg[mc->mc_top];
7229 ki = mc->mc_ki[mc->mc_top];
7231 /* add overflow pages to free list */
7232 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7236 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7237 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7238 (rc = mdb_ovpage_free(mc, omp)))
7241 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7242 mc->mc_db->md_entries--;
7243 rc = mdb_rebalance(mc);
7244 if (rc != MDB_SUCCESS)
7245 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7248 MDB_dbi dbi = mc->mc_dbi;
7250 mp = mc->mc_pg[mc->mc_top];
7251 nkeys = NUMKEYS(mp);
7253 /* if mc points past last node in page, find next sibling */
7254 if (mc->mc_ki[mc->mc_top] >= nkeys)
7255 mdb_cursor_sibling(mc, 1);
7257 /* Adjust other cursors pointing to mp */
7258 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7261 if (!(m2->mc_flags & C_INITIALIZED))
7263 if (m2->mc_pg[mc->mc_top] == mp) {
7264 if (m2->mc_ki[mc->mc_top] >= ki) {
7265 m2->mc_flags |= C_DEL;
7266 if (m2->mc_ki[mc->mc_top] > ki)
7267 m2->mc_ki[mc->mc_top]--;
7269 if (m2->mc_ki[mc->mc_top] >= nkeys)
7270 mdb_cursor_sibling(m2, 1);
7273 mc->mc_flags |= C_DEL;
7280 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7281 MDB_val *key, MDB_val *data)
7286 MDB_val rdata, *xdata;
7290 assert(key != NULL);
7292 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7294 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7297 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7298 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7300 if (key->mv_size > MDB_MAXKEYSIZE) {
7301 return MDB_BAD_VALSIZE;
7304 mdb_cursor_init(&mc, txn, dbi, &mx);
7307 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7308 /* must ignore any data */
7319 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7321 /* let mdb_page_split know about this cursor if needed:
7322 * delete will trigger a rebalance; if it needs to move
7323 * a node from one page to another, it will have to
7324 * update the parent's separator key(s). If the new sepkey
7325 * is larger than the current one, the parent page may
7326 * run out of space, triggering a split. We need this
7327 * cursor to be consistent until the end of the rebalance.
7329 mc.mc_flags |= C_UNTRACK;
7330 mc.mc_next = txn->mt_cursors[dbi];
7331 txn->mt_cursors[dbi] = &mc;
7332 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7333 txn->mt_cursors[dbi] = mc.mc_next;
7338 /** Split a page and insert a new node.
7339 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7340 * The cursor will be updated to point to the actual page and index where
7341 * the node got inserted after the split.
7342 * @param[in] newkey The key for the newly inserted node.
7343 * @param[in] newdata The data for the newly inserted node.
7344 * @param[in] newpgno The page number, if the new node is a branch node.
7345 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7346 * @return 0 on success, non-zero on failure.
7349 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7350 unsigned int nflags)
7353 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7356 unsigned int i, j, split_indx, nkeys, pmax;
7358 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7360 MDB_page *mp, *rp, *pp;
7365 mp = mc->mc_pg[mc->mc_top];
7366 newindx = mc->mc_ki[mc->mc_top];
7368 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7369 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7370 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7372 /* Create a right sibling. */
7373 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7375 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7377 if (mc->mc_snum < 2) {
7378 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7380 /* shift current top to make room for new parent */
7381 mc->mc_pg[1] = mc->mc_pg[0];
7382 mc->mc_ki[1] = mc->mc_ki[0];
7385 mc->mc_db->md_root = pp->mp_pgno;
7386 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7387 mc->mc_db->md_depth++;
7390 /* Add left (implicit) pointer. */
7391 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7392 /* undo the pre-push */
7393 mc->mc_pg[0] = mc->mc_pg[1];
7394 mc->mc_ki[0] = mc->mc_ki[1];
7395 mc->mc_db->md_root = mp->mp_pgno;
7396 mc->mc_db->md_depth--;
7403 ptop = mc->mc_top-1;
7404 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7407 mc->mc_flags |= C_SPLITTING;
7408 mdb_cursor_copy(mc, &mn);
7409 mn.mc_pg[mn.mc_top] = rp;
7410 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7412 if (nflags & MDB_APPEND) {
7413 mn.mc_ki[mn.mc_top] = 0;
7415 split_indx = newindx;
7420 nkeys = NUMKEYS(mp);
7421 split_indx = nkeys / 2;
7422 if (newindx < split_indx)
7428 unsigned int lsize, rsize, ksize;
7429 /* Move half of the keys to the right sibling */
7431 x = mc->mc_ki[mc->mc_top] - split_indx;
7432 ksize = mc->mc_db->md_pad;
7433 split = LEAF2KEY(mp, split_indx, ksize);
7434 rsize = (nkeys - split_indx) * ksize;
7435 lsize = (nkeys - split_indx) * sizeof(indx_t);
7436 mp->mp_lower -= lsize;
7437 rp->mp_lower += lsize;
7438 mp->mp_upper += rsize - lsize;
7439 rp->mp_upper -= rsize - lsize;
7440 sepkey.mv_size = ksize;
7441 if (newindx == split_indx) {
7442 sepkey.mv_data = newkey->mv_data;
7444 sepkey.mv_data = split;
7447 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7448 memcpy(rp->mp_ptrs, split, rsize);
7449 sepkey.mv_data = rp->mp_ptrs;
7450 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7451 memcpy(ins, newkey->mv_data, ksize);
7452 mp->mp_lower += sizeof(indx_t);
7453 mp->mp_upper -= ksize - sizeof(indx_t);
7456 memcpy(rp->mp_ptrs, split, x * ksize);
7457 ins = LEAF2KEY(rp, x, ksize);
7458 memcpy(ins, newkey->mv_data, ksize);
7459 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7460 rp->mp_lower += sizeof(indx_t);
7461 rp->mp_upper -= ksize - sizeof(indx_t);
7462 mc->mc_ki[mc->mc_top] = x;
7463 mc->mc_pg[mc->mc_top] = rp;
7468 /* For leaf pages, check the split point based on what
7469 * fits where, since otherwise mdb_node_add can fail.
7471 * This check is only needed when the data items are
7472 * relatively large, such that being off by one will
7473 * make the difference between success or failure.
7475 * It's also relevant if a page happens to be laid out
7476 * such that one half of its nodes are all "small" and
7477 * the other half of its nodes are "large." If the new
7478 * item is also "large" and falls on the half with
7479 * "large" nodes, it also may not fit.
7482 unsigned int psize, nsize;
7483 /* Maximum free space in an empty page */
7484 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7485 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7486 if ((nkeys < 20) || (nsize > pmax/16)) {
7487 if (newindx <= split_indx) {
7490 for (i=0; i<split_indx; i++) {
7491 node = NODEPTR(mp, i);
7492 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7493 if (F_ISSET(node->mn_flags, F_BIGDATA))
7494 psize += sizeof(pgno_t);
7496 psize += NODEDSZ(node);
7500 split_indx = newindx;
7511 for (i=nkeys-1; i>=split_indx; i--) {
7512 node = NODEPTR(mp, i);
7513 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7514 if (F_ISSET(node->mn_flags, F_BIGDATA))
7515 psize += sizeof(pgno_t);
7517 psize += NODEDSZ(node);
7521 split_indx = newindx;
7532 /* First find the separating key between the split pages.
7533 * The case where newindx == split_indx is ambiguous; the
7534 * new item could go to the new page or stay on the original
7535 * page. If newpos == 1 it goes to the new page.
7537 if (newindx == split_indx && newpos) {
7538 sepkey.mv_size = newkey->mv_size;
7539 sepkey.mv_data = newkey->mv_data;
7541 node = NODEPTR(mp, split_indx);
7542 sepkey.mv_size = node->mn_ksize;
7543 sepkey.mv_data = NODEKEY(node);
7547 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7549 /* Copy separator key to the parent.
7551 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7555 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7558 if (mn.mc_snum == mc->mc_snum) {
7559 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7560 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7561 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7562 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7567 /* Right page might now have changed parent.
7568 * Check if left page also changed parent.
7570 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7571 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7572 for (i=0; i<ptop; i++) {
7573 mc->mc_pg[i] = mn.mc_pg[i];
7574 mc->mc_ki[i] = mn.mc_ki[i];
7576 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7577 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7581 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7584 mc->mc_flags ^= C_SPLITTING;
7585 if (rc != MDB_SUCCESS) {
7588 if (nflags & MDB_APPEND) {
7589 mc->mc_pg[mc->mc_top] = rp;
7590 mc->mc_ki[mc->mc_top] = 0;
7591 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7594 for (i=0; i<mc->mc_top; i++)
7595 mc->mc_ki[i] = mn.mc_ki[i];
7602 /* Move half of the keys to the right sibling. */
7604 /* grab a page to hold a temporary copy */
7605 copy = mdb_page_malloc(mc->mc_txn, 1);
7609 copy->mp_pgno = mp->mp_pgno;
7610 copy->mp_flags = mp->mp_flags;
7611 copy->mp_lower = PAGEHDRSZ;
7612 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7613 mc->mc_pg[mc->mc_top] = copy;
7614 for (i = j = 0; i <= nkeys; j++) {
7615 if (i == split_indx) {
7616 /* Insert in right sibling. */
7617 /* Reset insert index for right sibling. */
7618 if (i != newindx || (newpos ^ ins_new)) {
7620 mc->mc_pg[mc->mc_top] = rp;
7624 if (i == newindx && !ins_new) {
7625 /* Insert the original entry that caused the split. */
7626 rkey.mv_data = newkey->mv_data;
7627 rkey.mv_size = newkey->mv_size;
7636 /* Update index for the new key. */
7637 mc->mc_ki[mc->mc_top] = j;
7638 } else if (i == nkeys) {
7641 node = NODEPTR(mp, i);
7642 rkey.mv_data = NODEKEY(node);
7643 rkey.mv_size = node->mn_ksize;
7645 xdata.mv_data = NODEDATA(node);
7646 xdata.mv_size = NODEDSZ(node);
7649 pgno = NODEPGNO(node);
7650 flags = node->mn_flags;
7655 if (!IS_LEAF(mp) && j == 0) {
7656 /* First branch index doesn't need key data. */
7660 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7664 nkeys = NUMKEYS(copy);
7665 for (i=0; i<nkeys; i++)
7666 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7667 mp->mp_lower = copy->mp_lower;
7668 mp->mp_upper = copy->mp_upper;
7669 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7670 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7672 /* reset back to original page */
7673 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7674 mc->mc_pg[mc->mc_top] = mp;
7675 if (nflags & MDB_RESERVE) {
7676 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7677 if (!(node->mn_flags & F_BIGDATA))
7678 newdata->mv_data = NODEDATA(node);
7682 /* Make sure mc_ki is still valid.
7684 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7685 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7686 for (i=0; i<ptop; i++) {
7687 mc->mc_pg[i] = mn.mc_pg[i];
7688 mc->mc_ki[i] = mn.mc_ki[i];
7690 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7691 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7695 /* return tmp page to freelist */
7696 mdb_page_free(mc->mc_txn->mt_env, copy);
7699 /* Adjust other cursors pointing to mp */
7700 MDB_cursor *m2, *m3;
7701 MDB_dbi dbi = mc->mc_dbi;
7702 int fixup = NUMKEYS(mp);
7704 if (mc->mc_flags & C_SUB)
7707 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7708 if (mc->mc_flags & C_SUB)
7709 m3 = &m2->mc_xcursor->mx_cursor;
7714 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7716 if (m3->mc_flags & C_SPLITTING)
7721 for (k=m3->mc_top; k>=0; k--) {
7722 m3->mc_ki[k+1] = m3->mc_ki[k];
7723 m3->mc_pg[k+1] = m3->mc_pg[k];
7725 if (m3->mc_ki[0] >= split_indx) {
7730 m3->mc_pg[0] = mc->mc_pg[0];
7734 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7735 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7736 m3->mc_ki[mc->mc_top]++;
7737 if (m3->mc_ki[mc->mc_top] >= fixup) {
7738 m3->mc_pg[mc->mc_top] = rp;
7739 m3->mc_ki[mc->mc_top] -= fixup;
7740 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7742 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7743 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7752 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7753 MDB_val *key, MDB_val *data, unsigned int flags)
7758 assert(key != NULL);
7759 assert(data != NULL);
7761 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7764 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7767 mdb_cursor_init(&mc, txn, dbi, &mx);
7768 return mdb_cursor_put(&mc, key, data, flags);
7772 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7774 if ((flag & CHANGEABLE) != flag)
7777 env->me_flags |= flag;
7779 env->me_flags &= ~flag;
7784 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7789 *arg = env->me_flags;
7794 mdb_env_get_path(MDB_env *env, const char **arg)
7799 *arg = env->me_path;
7803 /** Common code for #mdb_stat() and #mdb_env_stat().
7804 * @param[in] env the environment to operate in.
7805 * @param[in] db the #MDB_db record containing the stats to return.
7806 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7807 * @return 0, this function always succeeds.
7810 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7812 arg->ms_psize = env->me_psize;
7813 arg->ms_depth = db->md_depth;
7814 arg->ms_branch_pages = db->md_branch_pages;
7815 arg->ms_leaf_pages = db->md_leaf_pages;
7816 arg->ms_overflow_pages = db->md_overflow_pages;
7817 arg->ms_entries = db->md_entries;
7822 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7826 if (env == NULL || arg == NULL)
7829 toggle = mdb_env_pick_meta(env);
7831 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7835 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7839 if (env == NULL || arg == NULL)
7842 toggle = mdb_env_pick_meta(env);
7843 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7844 arg->me_mapsize = env->me_mapsize;
7845 arg->me_maxreaders = env->me_maxreaders;
7847 /* me_numreaders may be zero if this process never used any readers. Use
7848 * the shared numreader count if it exists.
7850 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7852 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7853 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7857 /** Set the default comparison functions for a database.
7858 * Called immediately after a database is opened to set the defaults.
7859 * The user can then override them with #mdb_set_compare() or
7860 * #mdb_set_dupsort().
7861 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7862 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7865 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7867 uint16_t f = txn->mt_dbs[dbi].md_flags;
7869 txn->mt_dbxs[dbi].md_cmp =
7870 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7871 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7873 txn->mt_dbxs[dbi].md_dcmp =
7874 !(f & MDB_DUPSORT) ? 0 :
7875 ((f & MDB_INTEGERDUP)
7876 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7877 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7880 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7885 int rc, dbflag, exact;
7886 unsigned int unused = 0;
7889 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7890 mdb_default_cmp(txn, FREE_DBI);
7893 if ((flags & VALID_FLAGS) != flags)
7895 if (txn->mt_flags & MDB_TXN_ERROR)
7901 if (flags & PERSISTENT_FLAGS) {
7902 uint16_t f2 = flags & PERSISTENT_FLAGS;
7903 /* make sure flag changes get committed */
7904 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7905 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7906 txn->mt_flags |= MDB_TXN_DIRTY;
7909 mdb_default_cmp(txn, MAIN_DBI);
7913 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7914 mdb_default_cmp(txn, MAIN_DBI);
7917 /* Is the DB already open? */
7919 for (i=2; i<txn->mt_numdbs; i++) {
7920 if (!txn->mt_dbxs[i].md_name.mv_size) {
7921 /* Remember this free slot */
7922 if (!unused) unused = i;
7925 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7926 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7932 /* If no free slot and max hit, fail */
7933 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7934 return MDB_DBS_FULL;
7936 /* Cannot mix named databases with some mainDB flags */
7937 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7938 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7940 /* Find the DB info */
7941 dbflag = DB_NEW|DB_VALID;
7944 key.mv_data = (void *)name;
7945 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7946 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7947 if (rc == MDB_SUCCESS) {
7948 /* make sure this is actually a DB */
7949 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7950 if (!(node->mn_flags & F_SUBDATA))
7951 return MDB_INCOMPATIBLE;
7952 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7953 /* Create if requested */
7955 data.mv_size = sizeof(MDB_db);
7956 data.mv_data = &dummy;
7957 memset(&dummy, 0, sizeof(dummy));
7958 dummy.md_root = P_INVALID;
7959 dummy.md_flags = flags & PERSISTENT_FLAGS;
7960 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7964 /* OK, got info, add to table */
7965 if (rc == MDB_SUCCESS) {
7966 unsigned int slot = unused ? unused : txn->mt_numdbs;
7967 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7968 txn->mt_dbxs[slot].md_name.mv_size = len;
7969 txn->mt_dbxs[slot].md_rel = NULL;
7970 txn->mt_dbflags[slot] = dbflag;
7971 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7973 mdb_default_cmp(txn, slot);
7982 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7984 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7987 if (txn->mt_dbflags[dbi] & DB_STALE) {
7990 /* Stale, must read the DB's root. cursor_init does it for us. */
7991 mdb_cursor_init(&mc, txn, dbi, &mx);
7993 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7996 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7999 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8001 ptr = env->me_dbxs[dbi].md_name.mv_data;
8002 env->me_dbxs[dbi].md_name.mv_data = NULL;
8003 env->me_dbxs[dbi].md_name.mv_size = 0;
8004 env->me_dbflags[dbi] = 0;
8008 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8010 /* We could return the flags for the FREE_DBI too but what's the point? */
8011 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8013 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8017 /** Add all the DB's pages to the free list.
8018 * @param[in] mc Cursor on the DB to free.
8019 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8020 * @return 0 on success, non-zero on failure.
8023 mdb_drop0(MDB_cursor *mc, int subs)
8027 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8028 if (rc == MDB_SUCCESS) {
8029 MDB_txn *txn = mc->mc_txn;
8034 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8035 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8038 mdb_cursor_copy(mc, &mx);
8039 while (mc->mc_snum > 0) {
8040 MDB_page *mp = mc->mc_pg[mc->mc_top];
8041 unsigned n = NUMKEYS(mp);
8043 for (i=0; i<n; i++) {
8044 ni = NODEPTR(mp, i);
8045 if (ni->mn_flags & F_BIGDATA) {
8048 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8049 rc = mdb_page_get(txn, pg, &omp, NULL);
8052 assert(IS_OVERFLOW(omp));
8053 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8057 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8058 mdb_xcursor_init1(mc, ni);
8059 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8065 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8067 for (i=0; i<n; i++) {
8069 ni = NODEPTR(mp, i);
8072 mdb_midl_xappend(txn->mt_free_pgs, pg);
8077 mc->mc_ki[mc->mc_top] = i;
8078 rc = mdb_cursor_sibling(mc, 1);
8080 /* no more siblings, go back to beginning
8081 * of previous level.
8085 for (i=1; i<mc->mc_snum; i++) {
8087 mc->mc_pg[i] = mx.mc_pg[i];
8092 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8093 } else if (rc == MDB_NOTFOUND) {
8099 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8101 MDB_cursor *mc, *m2;
8104 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8107 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8110 rc = mdb_cursor_open(txn, dbi, &mc);
8114 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8115 /* Invalidate the dropped DB's cursors */
8116 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8117 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8121 /* Can't delete the main DB */
8122 if (del && dbi > MAIN_DBI) {
8123 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8125 txn->mt_dbflags[dbi] = DB_STALE;
8126 mdb_dbi_close(txn->mt_env, dbi);
8129 /* reset the DB record, mark it dirty */
8130 txn->mt_dbflags[dbi] |= DB_DIRTY;
8131 txn->mt_dbs[dbi].md_depth = 0;
8132 txn->mt_dbs[dbi].md_branch_pages = 0;
8133 txn->mt_dbs[dbi].md_leaf_pages = 0;
8134 txn->mt_dbs[dbi].md_overflow_pages = 0;
8135 txn->mt_dbs[dbi].md_entries = 0;
8136 txn->mt_dbs[dbi].md_root = P_INVALID;
8138 txn->mt_flags |= MDB_TXN_DIRTY;
8141 mdb_cursor_close(mc);
8145 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8147 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8150 txn->mt_dbxs[dbi].md_cmp = cmp;
8154 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8156 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8159 txn->mt_dbxs[dbi].md_dcmp = cmp;
8163 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8165 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8168 txn->mt_dbxs[dbi].md_rel = rel;
8172 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8174 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8177 txn->mt_dbxs[dbi].md_relctx = ctx;
8181 int mdb_env_get_maxkeysize(MDB_env *env)
8183 return MDB_MAXKEYSIZE;
8186 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8188 unsigned int i, rdrs;
8195 if (!env->me_txns) {
8196 return func("(no reader locks)\n", ctx);
8198 rdrs = env->me_txns->mti_numreaders;
8199 mr = env->me_txns->mti_readers;
8200 for (i=0; i<rdrs; i++) {
8205 if (mr[i].mr_txnid == (txnid_t)-1) {
8206 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8208 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8212 func(" pid thread txnid\n", ctx);
8214 rc = func(buf, ctx);
8220 func("(no active readers)\n", ctx);
8225 /** Insert pid into list if not already present.
8226 * return -1 if already present.
8228 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8230 /* binary search of pid in list */
8232 unsigned cursor = 1;
8234 unsigned n = ids[0];
8237 unsigned pivot = n >> 1;
8238 cursor = base + pivot + 1;
8239 val = pid - ids[cursor];
8244 } else if ( val > 0 ) {
8249 /* found, so it's a duplicate */
8258 for (n = ids[0]; n > cursor; n--)
8264 int mdb_reader_check(MDB_env *env, int *dead)
8266 unsigned int i, j, rdrs;
8277 rdrs = env->me_txns->mti_numreaders;
8278 pids = malloc((rdrs+1) * sizeof(pid_t));
8282 mr = env->me_txns->mti_readers;
8284 for (i=0; i<rdrs; i++) {
8285 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8287 if (mdb_pid_insert(pids, pid) == 0) {
8288 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8290 /* Recheck, a new process may have reused pid */
8291 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8292 for (j=i; j<rdrs; j++)
8293 if (mr[j].mr_pid == pid) {
8294 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8295 (unsigned) pid, mr[j].mr_txnid));
8300 UNLOCK_MUTEX_R(env);