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
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 */
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 typedef struct MDB_meta {
791 /** Stamp identifying this as an MDB file. It must be set
794 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
796 void *mm_address; /**< address for fixed mapping */
797 size_t mm_mapsize; /**< size of mmap region */
798 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
799 /** The size of pages used in this DB */
800 #define mm_psize mm_dbs[0].md_pad
801 /** Any persistent environment flags. @ref mdb_env */
802 #define mm_flags mm_dbs[0].md_flags
803 pgno_t mm_last_pg; /**< last used page in file */
804 txnid_t mm_txnid; /**< txnid that committed this page */
807 /** Buffer for a stack-allocated dirty page.
808 * The members define size and alignment, and silence type
809 * aliasing warnings. They are not used directly; that could
810 * mean incorrectly using several union members in parallel.
812 typedef union MDB_pagebuf {
813 char mb_raw[MDB_PAGESIZE];
816 char mm_pad[PAGEHDRSZ];
821 /** Auxiliary DB info.
822 * The information here is mostly static/read-only. There is
823 * only a single copy of this record in the environment.
825 typedef struct MDB_dbx {
826 MDB_val md_name; /**< name of the database */
827 MDB_cmp_func *md_cmp; /**< function for comparing keys */
828 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
829 MDB_rel_func *md_rel; /**< user relocate function */
830 void *md_relctx; /**< user-provided context for md_rel */
833 /** A database transaction.
834 * Every operation requires a transaction handle.
837 MDB_txn *mt_parent; /**< parent of a nested txn */
838 MDB_txn *mt_child; /**< nested txn under this txn */
839 pgno_t mt_next_pgno; /**< next unallocated page */
840 /** The ID of this transaction. IDs are integers incrementing from 1.
841 * Only committed write transactions increment the ID. If a transaction
842 * aborts, the ID may be re-used by the next writer.
845 MDB_env *mt_env; /**< the DB environment */
846 /** The list of pages that became unused during this transaction.
849 /** The sorted list of dirty pages we temporarily wrote to disk
850 * because the dirty list was full. page numbers in here are
851 * shifted left by 1, deleted slots have the LSB set.
853 MDB_IDL mt_spill_pgs;
855 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
857 /** For read txns: This thread/txn's reader table slot, or NULL. */
860 /** Array of records for each DB known in the environment. */
862 /** Array of MDB_db records for each known DB */
864 /** @defgroup mt_dbflag Transaction DB Flags
868 #define DB_DIRTY 0x01 /**< DB was written in this txn */
869 #define DB_STALE 0x02 /**< DB record is older than txnID */
870 #define DB_NEW 0x04 /**< DB handle opened in this txn */
871 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
873 /** In write txns, array of cursors for each DB */
874 MDB_cursor **mt_cursors;
875 /** Array of flags for each DB */
876 unsigned char *mt_dbflags;
877 /** Number of DB records in use. This number only ever increments;
878 * we don't decrement it when individual DB handles are closed.
882 /** @defgroup mdb_txn Transaction Flags
886 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
887 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
888 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
889 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
891 unsigned int mt_flags; /**< @ref mdb_txn */
892 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
893 unsigned int mt_dirty_room;
894 /** Tracks which of the two meta pages was used at the start
895 * of this transaction.
897 unsigned int mt_toggle;
900 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
901 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
902 * raise this on a 64 bit machine.
904 #define CURSOR_STACK 32
908 /** Cursors are used for all DB operations */
910 /** Next cursor on this DB in this txn */
912 /** Backup of the original cursor if this cursor is a shadow */
913 MDB_cursor *mc_backup;
914 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
915 struct MDB_xcursor *mc_xcursor;
916 /** The transaction that owns this cursor */
918 /** The database handle this cursor operates on */
920 /** The database record for this cursor */
922 /** The database auxiliary record for this cursor */
924 /** The @ref mt_dbflag for this database */
925 unsigned char *mc_dbflag;
926 unsigned short mc_snum; /**< number of pushed pages */
927 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
928 /** @defgroup mdb_cursor Cursor Flags
930 * Cursor state flags.
933 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
934 #define C_EOF 0x02 /**< No more data */
935 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
936 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
937 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
939 unsigned int mc_flags; /**< @ref mdb_cursor */
940 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
941 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
944 /** Context for sorted-dup records.
945 * We could have gone to a fully recursive design, with arbitrarily
946 * deep nesting of sub-databases. But for now we only handle these
947 * levels - main DB, optional sub-DB, sorted-duplicate DB.
949 typedef struct MDB_xcursor {
950 /** A sub-cursor for traversing the Dup DB */
951 MDB_cursor mx_cursor;
952 /** The database record for this Dup DB */
954 /** The auxiliary DB record for this Dup DB */
956 /** The @ref mt_dbflag for this Dup DB */
957 unsigned char mx_dbflag;
960 /** State of FreeDB old pages, stored in the MDB_env */
961 typedef struct MDB_pgstate {
962 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
963 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
966 /** The database environment. */
968 HANDLE me_fd; /**< The main data file */
969 HANDLE me_lfd; /**< The lock file */
970 HANDLE me_mfd; /**< just for writing the meta pages */
971 /** Failed to update the meta page. Probably an I/O error. */
972 #define MDB_FATAL_ERROR 0x80000000U
973 /** Some fields are initialized. */
974 #define MDB_ENV_ACTIVE 0x20000000U
975 /** me_txkey is set */
976 #define MDB_ENV_TXKEY 0x10000000U
977 /** Have liveness lock in reader table */
978 #define MDB_LIVE_READER 0x08000000U
979 uint32_t me_flags; /**< @ref mdb_env */
980 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
981 unsigned int me_maxreaders; /**< size of the reader table */
982 unsigned int me_numreaders; /**< max numreaders set by this env */
983 MDB_dbi me_numdbs; /**< number of DBs opened */
984 MDB_dbi me_maxdbs; /**< size of the DB table */
985 pid_t me_pid; /**< process ID of this env */
986 char *me_path; /**< path to the DB files */
987 char *me_map; /**< the memory map of the data file */
988 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
989 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
990 MDB_txn *me_txn; /**< current write transaction */
991 size_t me_mapsize; /**< size of the data memory map */
992 off_t me_size; /**< current file size */
993 pgno_t me_maxpg; /**< me_mapsize / me_psize */
994 MDB_dbx *me_dbxs; /**< array of static DB info */
995 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
996 pthread_key_t me_txkey; /**< thread-key for readers */
997 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
998 # define me_pglast me_pgstate.mf_pglast
999 # define me_pghead me_pgstate.mf_pghead
1000 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1001 /** IDL of pages that became unused in a write txn */
1002 MDB_IDL me_free_pgs;
1003 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1004 MDB_ID2L me_dirty_list;
1005 /** Max number of freelist items that can fit in a single overflow page */
1007 /** Max size of a node on a page */
1008 unsigned int me_nodemax;
1010 int me_pidquery; /**< Used in OpenProcess */
1011 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1013 #elif defined(MDB_USE_POSIX_SEM)
1014 sem_t *me_rmutex; /* Shared mutexes are not supported */
1019 /** Nested transaction */
1020 typedef struct MDB_ntxn {
1021 MDB_txn mnt_txn; /* the transaction */
1022 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
1025 /** max number of pages to commit in one writev() call */
1026 #define MDB_COMMIT_PAGES 64
1027 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1028 #undef MDB_COMMIT_PAGES
1029 #define MDB_COMMIT_PAGES IOV_MAX
1032 /* max bytes to write in one call */
1033 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1035 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1036 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1037 static int mdb_page_touch(MDB_cursor *mc);
1039 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1040 static int mdb_page_search_root(MDB_cursor *mc,
1041 MDB_val *key, int modify);
1042 #define MDB_PS_MODIFY 1
1043 #define MDB_PS_ROOTONLY 2
1044 static int mdb_page_search(MDB_cursor *mc,
1045 MDB_val *key, int flags);
1046 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1048 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1049 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1050 pgno_t newpgno, unsigned int nflags);
1052 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1053 static int mdb_env_pick_meta(const MDB_env *env);
1054 static int mdb_env_write_meta(MDB_txn *txn);
1055 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1056 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1058 static void mdb_env_close0(MDB_env *env, int excl);
1060 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1061 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1062 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1063 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1064 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1065 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1066 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1067 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1068 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1070 static int mdb_rebalance(MDB_cursor *mc);
1071 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1073 static void mdb_cursor_pop(MDB_cursor *mc);
1074 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1076 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1077 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1078 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1079 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1080 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1082 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1083 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1085 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1086 static void mdb_xcursor_init0(MDB_cursor *mc);
1087 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1089 static int mdb_drop0(MDB_cursor *mc, int subs);
1090 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1093 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1097 static SECURITY_DESCRIPTOR mdb_null_sd;
1098 static SECURITY_ATTRIBUTES mdb_all_sa;
1099 static int mdb_sec_inited;
1102 /** Return the library version info. */
1104 mdb_version(int *major, int *minor, int *patch)
1106 if (major) *major = MDB_VERSION_MAJOR;
1107 if (minor) *minor = MDB_VERSION_MINOR;
1108 if (patch) *patch = MDB_VERSION_PATCH;
1109 return MDB_VERSION_STRING;
1112 /** Table of descriptions for MDB @ref errors */
1113 static char *const mdb_errstr[] = {
1114 "MDB_KEYEXIST: Key/data pair already exists",
1115 "MDB_NOTFOUND: No matching key/data pair found",
1116 "MDB_PAGE_NOTFOUND: Requested page not found",
1117 "MDB_CORRUPTED: Located page was wrong type",
1118 "MDB_PANIC: Update of meta page failed",
1119 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1120 "MDB_INVALID: File is not an MDB file",
1121 "MDB_MAP_FULL: Environment mapsize limit reached",
1122 "MDB_DBS_FULL: Environment maxdbs limit reached",
1123 "MDB_READERS_FULL: Environment maxreaders limit reached",
1124 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1125 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1126 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1127 "MDB_PAGE_FULL: Internal error - page has no more space",
1128 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1129 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1130 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1131 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1132 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1136 mdb_strerror(int err)
1140 return ("Successful return: 0");
1142 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1143 i = err - MDB_KEYEXIST;
1144 return mdb_errstr[i];
1147 return strerror(err);
1151 /** Display a key in hexadecimal and return the address of the result.
1152 * @param[in] key the key to display
1153 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1154 * @return The key in hexadecimal form.
1157 mdb_dkey(MDB_val *key, char *buf)
1160 unsigned char *c = key->mv_data;
1166 if (key->mv_size > MDB_MAXKEYSIZE)
1167 return "MDB_MAXKEYSIZE";
1168 /* may want to make this a dynamic check: if the key is mostly
1169 * printable characters, print it as-is instead of converting to hex.
1173 for (i=0; i<key->mv_size; i++)
1174 ptr += sprintf(ptr, "%02x", *c++);
1176 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1181 /** Display all the keys in the page. */
1183 mdb_page_list(MDB_page *mp)
1186 unsigned int i, nkeys, nsize;
1190 nkeys = NUMKEYS(mp);
1191 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1192 for (i=0; i<nkeys; i++) {
1193 node = NODEPTR(mp, i);
1194 key.mv_size = node->mn_ksize;
1195 key.mv_data = node->mn_data;
1196 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1197 if (IS_BRANCH(mp)) {
1198 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1201 if (F_ISSET(node->mn_flags, F_BIGDATA))
1202 nsize += sizeof(pgno_t);
1204 nsize += NODEDSZ(node);
1205 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1211 mdb_cursor_chk(MDB_cursor *mc)
1217 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1218 for (i=0; i<mc->mc_top; i++) {
1220 node = NODEPTR(mp, mc->mc_ki[i]);
1221 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1224 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1230 /** Count all the pages in each DB and in the freelist
1231 * and make sure it matches the actual number of pages
1234 static void mdb_audit(MDB_txn *txn)
1238 MDB_ID freecount, count;
1243 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1244 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1245 freecount += *(MDB_ID *)data.mv_data;
1248 for (i = 0; i<txn->mt_numdbs; i++) {
1250 mdb_cursor_init(&mc, txn, i, &mx);
1251 if (txn->mt_dbs[i].md_root == P_INVALID)
1253 count += txn->mt_dbs[i].md_branch_pages +
1254 txn->mt_dbs[i].md_leaf_pages +
1255 txn->mt_dbs[i].md_overflow_pages;
1256 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1257 mdb_page_search(&mc, NULL, 0);
1261 mp = mc.mc_pg[mc.mc_top];
1262 for (j=0; j<NUMKEYS(mp); j++) {
1263 MDB_node *leaf = NODEPTR(mp, j);
1264 if (leaf->mn_flags & F_SUBDATA) {
1266 memcpy(&db, NODEDATA(leaf), sizeof(db));
1267 count += db.md_branch_pages + db.md_leaf_pages +
1268 db.md_overflow_pages;
1272 while (mdb_cursor_sibling(&mc, 1) == 0);
1275 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1276 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1277 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1283 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1285 return txn->mt_dbxs[dbi].md_cmp(a, b);
1289 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1291 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1294 /** Allocate memory for a page.
1295 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1298 mdb_page_malloc(MDB_txn *txn, unsigned num)
1300 MDB_env *env = txn->mt_env;
1301 MDB_page *ret = env->me_dpages;
1302 size_t sz = env->me_psize;
1305 VGMEMP_ALLOC(env, ret, sz);
1306 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1307 env->me_dpages = ret->mp_next;
1313 if ((ret = malloc(sz)) != NULL) {
1314 VGMEMP_ALLOC(env, ret, sz);
1319 /** Free a single page.
1320 * Saves single pages to a list, for future reuse.
1321 * (This is not used for multi-page overflow pages.)
1324 mdb_page_free(MDB_env *env, MDB_page *mp)
1326 mp->mp_next = env->me_dpages;
1327 VGMEMP_FREE(env, mp);
1328 env->me_dpages = mp;
1331 /* Free a dirty page */
1333 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1335 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1336 mdb_page_free(env, dp);
1338 /* large pages just get freed directly */
1339 VGMEMP_FREE(env, dp);
1344 /** Return all dirty pages to dpage list */
1346 mdb_dlist_free(MDB_txn *txn)
1348 MDB_env *env = txn->mt_env;
1349 MDB_ID2L dl = txn->mt_u.dirty_list;
1350 unsigned i, n = dl[0].mid;
1352 for (i = 1; i <= n; i++) {
1353 mdb_dpage_free(env, dl[i].mptr);
1358 /* Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1359 * @param[in] mc A cursor handle for the current operation.
1360 * @param[in] pflags Flags of the pages to update:
1361 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1362 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1363 * @return 0 on success, non-zero on failure.
1366 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1368 MDB_txn *txn = mc->mc_txn;
1373 int rc = MDB_SUCCESS, level;
1375 /* Mark pages seen by cursors */
1376 if (mc->mc_flags & C_UNTRACK)
1377 mc = NULL; /* will find mc in mt_cursors */
1378 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1379 for (; mc; mc=mc->mc_next) {
1380 for (m3 = mc; m3->mc_flags & C_INITIALIZED; m3 = &mx->mx_cursor) {
1381 for (j=0; j<m3->mc_snum; j++)
1382 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1384 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1385 mx = m3->mc_xcursor;
1395 /* Mark dirty root pages */
1396 for (i=0; i<txn->mt_numdbs; i++) {
1397 if (txn->mt_dbflags[i] & DB_DIRTY) {
1398 pgno_t pgno = txn->mt_dbs[i].md_root;
1399 if (pgno == P_INVALID)
1401 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1403 if ((dp->mp_flags & (P_DIRTY|P_KEEP)) == pflags && level <= 1)
1404 dp->mp_flags ^= P_KEEP;
1412 static int mdb_page_flush(MDB_txn *txn, int keep);
1414 /** Spill pages from the dirty list back to disk.
1415 * This is intended to prevent running into #MDB_TXN_FULL situations,
1416 * but note that they may still occur in a few cases:
1417 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1418 * are too many of these dirtied in one txn, the txn may still get
1420 * 2) child txns may run out of space if their parents dirtied a
1421 * lot of pages and never spilled them. TODO: we probably should do
1422 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1423 * the parent's dirty_room is below a given threshold.
1424 * 3) our estimate of the txn size could be too small. At the
1425 * moment this seems unlikely.
1427 * Otherwise, if not using nested txns, it is expected that apps will
1428 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1429 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1430 * If the txn never references them again, they can be left alone.
1431 * If the txn only reads them, they can be used without any fuss.
1432 * If the txn writes them again, they can be dirtied immediately without
1433 * going thru all of the work of #mdb_page_touch(). Such references are
1434 * handled by #mdb_page_unspill().
1436 * Also note, we never spill DB root pages, nor pages of active cursors,
1437 * because we'll need these back again soon anyway. And in nested txns,
1438 * we can't spill a page in a child txn if it was already spilled in a
1439 * parent txn. That would alter the parent txns' data even though
1440 * the child hasn't committed yet, and we'd have no way to undo it if
1441 * the child aborted.
1443 * @param[in] m0 cursor A cursor handle identifying the transaction and
1444 * database for which we are checking space.
1445 * @param[in] key For a put operation, the key being stored.
1446 * @param[in] data For a put operation, the data being stored.
1447 * @return 0 on success, non-zero on failure.
1450 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1452 MDB_txn *txn = m0->mc_txn;
1454 MDB_ID2L dl = txn->mt_u.dirty_list;
1455 unsigned int i, j, need;
1458 if (m0->mc_flags & C_SUB)
1461 /* Estimate how much space this op will take */
1462 i = m0->mc_db->md_depth;
1463 /* Named DBs also dirty the main DB */
1464 if (m0->mc_dbi > MAIN_DBI)
1465 i += txn->mt_dbs[MAIN_DBI].md_depth;
1466 /* For puts, roughly factor in the key+data size */
1468 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1469 i += i; /* double it for good measure */
1472 if (txn->mt_dirty_room > i)
1475 if (!txn->mt_spill_pgs) {
1476 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1477 if (!txn->mt_spill_pgs)
1480 /* purge deleted slots */
1481 MDB_IDL sl = txn->mt_spill_pgs;
1482 unsigned int num = sl[0];
1484 for (i=1; i<=num; i++) {
1491 /* Preserve pages which may soon be dirtied again */
1492 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1495 /* Less aggressive spill - we originally spilled the entire dirty list,
1496 * with a few exceptions for cursor pages and DB root pages. But this
1497 * turns out to be a lot of wasted effort because in a large txn many
1498 * of those pages will need to be used again. So now we spill only 1/8th
1499 * of the dirty pages. Testing revealed this to be a good tradeoff,
1500 * better than 1/2, 1/4, or 1/10.
1502 if (need < MDB_IDL_UM_MAX / 8)
1503 need = MDB_IDL_UM_MAX / 8;
1505 /* Save the page IDs of all the pages we're flushing */
1506 /* flush from the tail forward, this saves a lot of shifting later on. */
1507 for (i=dl[0].mid; i && need; i--) {
1508 MDB_ID pn = dl[i].mid << 1;
1510 if (dp->mp_flags & P_KEEP)
1512 /* Can't spill twice, make sure it's not already in a parent's
1515 if (txn->mt_parent) {
1517 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1518 if (tx2->mt_spill_pgs) {
1519 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1520 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1521 dp->mp_flags |= P_KEEP;
1529 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1533 mdb_midl_sort(txn->mt_spill_pgs);
1535 /* Flush the spilled part of dirty list */
1536 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1539 /* Reset any dirty pages we kept that page_flush didn't see */
1540 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1544 if (txn->mt_parent) {
1545 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1546 /* dirty pages that are dirty in an ancestor don't
1547 * count against this txn's dirty_room.
1549 for (i=1; i<=dl[0].mid; i++) {
1550 pgno_t pgno = dl[i].mid;
1552 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1553 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1554 if (j <= tx2->mt_u.dirty_list[0].mid &&
1555 tx2->mt_u.dirty_list[j].mid == pgno) {
1556 txn->mt_dirty_room++;
1562 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1564 txn->mt_flags |= MDB_TXN_SPILLS;
1566 txn->mt_flags |= MDB_TXN_ERROR;
1571 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1573 mdb_find_oldest(MDB_txn *txn)
1576 txnid_t mr, oldest = txn->mt_txnid - 1;
1577 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1578 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1588 /** Add a page to the txn's dirty list */
1590 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1593 int (*insert)(MDB_ID2L, MDB_ID2 *);
1595 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1596 insert = mdb_mid2l_append;
1598 insert = mdb_mid2l_insert;
1600 mid.mid = mp->mp_pgno;
1602 insert(txn->mt_u.dirty_list, &mid);
1603 txn->mt_dirty_room--;
1606 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1607 * me_pghead and mt_next_pgno.
1609 * If there are free pages available from older transactions, they
1610 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1611 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1612 * and move me_pglast to say which records were consumed. Only this
1613 * function can create me_pghead and move me_pglast/mt_next_pgno.
1614 * @param[in] mc cursor A cursor handle identifying the transaction and
1615 * database for which we are allocating.
1616 * @param[in] num the number of pages to allocate.
1617 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1618 * will always be satisfied by a single contiguous chunk of memory.
1619 * @return 0 on success, non-zero on failure.
1622 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1624 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1625 /* Get at most <Max_retries> more freeDB records once me_pghead
1626 * has enough pages. If not enough, use new pages from the map.
1627 * If <Paranoid> and mc is updating the freeDB, only get new
1628 * records if me_pghead is empty. Then the freelist cannot play
1629 * catch-up with itself by growing while trying to save it.
1631 enum { Paranoid = 1, Max_retries = 500 };
1633 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1635 int rc, n2 = num-1, retry = Max_retries;
1636 MDB_txn *txn = mc->mc_txn;
1637 MDB_env *env = txn->mt_env;
1638 pgno_t pgno, *mop = env->me_pghead;
1639 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1641 txnid_t oldest = 0, last;
1647 /* If our dirty list is already full, we can't do anything */
1648 if (txn->mt_dirty_room == 0)
1649 return MDB_TXN_FULL;
1651 for (op = MDB_FIRST;; op = MDB_NEXT) {
1654 pgno_t *idl, old_id, new_id;
1656 /* Seek a big enough contiguous page range. Prefer
1657 * pages at the tail, just truncating the list.
1659 if (mop_len >= (unsigned)num) {
1663 if (mop[i-n2] == pgno+n2)
1665 } while (--i >= (unsigned)num);
1666 if (Max_retries < INT_MAX && --retry < 0)
1670 if (op == MDB_FIRST) { /* 1st iteration */
1671 /* Prepare to fetch more and coalesce */
1672 oldest = mdb_find_oldest(txn);
1673 last = env->me_pglast;
1674 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1677 key.mv_data = &last; /* will look up last+1 */
1678 key.mv_size = sizeof(last);
1680 if (Paranoid && mc->mc_dbi == FREE_DBI)
1683 if (Paranoid && retry < 0 && mop_len)
1687 /* Do not fetch more if the record will be too recent */
1690 rc = mdb_cursor_get(&m2, &key, NULL, op);
1692 if (rc == MDB_NOTFOUND)
1696 last = *(txnid_t*)key.mv_data;
1699 np = m2.mc_pg[m2.mc_top];
1700 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1701 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1704 idl = (MDB_ID *) data.mv_data;
1707 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1710 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1712 mop = env->me_pghead;
1714 env->me_pglast = last;
1716 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1717 last, txn->mt_dbs[FREE_DBI].md_root, i));
1719 DPRINTF(("IDL %"Z"u", idl[k]));
1721 /* Merge in descending sorted order */
1724 mop[0] = (pgno_t)-1;
1728 for (; old_id < new_id; old_id = mop[--j])
1735 /* Use new pages from the map when nothing suitable in the freeDB */
1737 pgno = txn->mt_next_pgno;
1738 if (pgno + num >= env->me_maxpg) {
1739 DPUTS("DB size maxed out");
1740 return MDB_MAP_FULL;
1744 if (env->me_flags & MDB_WRITEMAP) {
1745 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1747 if (!(np = mdb_page_malloc(txn, num)))
1751 mop[0] = mop_len -= num;
1752 /* Move any stragglers down */
1753 for (j = i-num; j < mop_len; )
1754 mop[++j] = mop[++i];
1756 txn->mt_next_pgno = pgno + num;
1759 mdb_page_dirty(txn, np);
1765 /** Copy the used portions of a non-overflow page.
1766 * @param[in] dst page to copy into
1767 * @param[in] src page to copy from
1768 * @param[in] psize size of a page
1771 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1773 enum { Align = sizeof(pgno_t) };
1774 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1776 /* If page isn't full, just copy the used portion. Adjust
1777 * alignment so memcpy may copy words instead of bytes.
1779 if ((unused &= -Align) && !IS_LEAF2(src)) {
1781 memcpy(dst, src, (lower + (Align-1)) & -Align);
1782 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1785 memcpy(dst, src, psize - unused);
1789 /** Pull a page off the txn's spill list, if present.
1790 * If a page being referenced was spilled to disk in this txn, bring
1791 * it back and make it dirty/writable again.
1792 * @param[in] tx0 the transaction handle.
1793 * @param[in] mp the page being referenced.
1794 * @param[out] ret the writable page, if any. ret is unchanged if
1795 * mp wasn't spilled.
1798 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1800 MDB_env *env = tx0->mt_env;
1803 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1805 for (txn = tx0; txn; txn=txn->mt_parent) {
1806 if (!txn->mt_spill_pgs)
1808 x = mdb_midl_search(txn->mt_spill_pgs, pn);
1809 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pn) {
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 if (txn->mt_parent) {
1842 /* If this page is also in a parent's dirty list, then
1843 * it's already accounted in dirty_room, and we need to
1844 * cancel out the decrement that mdb_page_dirty does.
1846 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1847 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1848 if (x <= tx2->mt_u.dirty_list[0].mid &&
1849 tx2->mt_u.dirty_list[x].mid == pgno) {
1850 tx0->mt_dirty_room++;
1855 mdb_page_dirty(tx0, np);
1856 np->mp_flags |= P_DIRTY;
1864 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1865 * @param[in] mc cursor pointing to the page to be touched
1866 * @return 0 on success, non-zero on failure.
1869 mdb_page_touch(MDB_cursor *mc)
1871 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1872 MDB_txn *txn = mc->mc_txn;
1873 MDB_cursor *m2, *m3;
1878 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1879 if (txn->mt_flags & MDB_TXN_SPILLS) {
1881 rc = mdb_page_unspill(txn, mp, &np);
1887 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1888 (rc = mdb_page_alloc(mc, 1, &np)))
1891 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1892 assert(mp->mp_pgno != pgno);
1893 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1894 /* Update the parent page, if any, to point to the new page */
1896 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1897 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1898 SETPGNO(node, pgno);
1900 mc->mc_db->md_root = pgno;
1902 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1903 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1905 /* If txn has a parent, make sure the page is in our
1909 unsigned x = mdb_mid2l_search(dl, pgno);
1910 if (x <= dl[0].mid && dl[x].mid == pgno) {
1911 if (mp != dl[x].mptr) { /* bad cursor? */
1912 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1913 return MDB_CORRUPTED;
1918 assert(dl[0].mid < MDB_IDL_UM_MAX);
1920 np = mdb_page_malloc(txn, 1);
1925 mdb_mid2l_insert(dl, &mid);
1930 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1932 np->mp_flags |= P_DIRTY;
1935 /* Adjust cursors pointing to mp */
1936 mc->mc_pg[mc->mc_top] = np;
1938 if (mc->mc_flags & C_SUB) {
1940 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1941 m3 = &m2->mc_xcursor->mx_cursor;
1942 if (m3->mc_snum < mc->mc_snum) continue;
1943 if (m3->mc_pg[mc->mc_top] == mp)
1944 m3->mc_pg[mc->mc_top] = np;
1947 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1948 if (m2->mc_snum < mc->mc_snum) continue;
1949 if (m2->mc_pg[mc->mc_top] == mp) {
1950 m2->mc_pg[mc->mc_top] = np;
1951 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1952 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1954 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1955 if (!(leaf->mn_flags & F_SUBDATA))
1956 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1965 mdb_env_sync(MDB_env *env, int force)
1968 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1969 if (env->me_flags & MDB_WRITEMAP) {
1970 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1971 ? MS_ASYNC : MS_SYNC;
1972 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1975 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1979 if (MDB_FDATASYNC(env->me_fd))
1986 /** Back up parent txn's cursors, then grab the originals for tracking */
1988 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1990 MDB_cursor *mc, *bk;
1995 for (i = src->mt_numdbs; --i >= 0; ) {
1996 if ((mc = src->mt_cursors[i]) != NULL) {
1997 size = sizeof(MDB_cursor);
1999 size += sizeof(MDB_xcursor);
2000 for (; mc; mc = bk->mc_next) {
2006 mc->mc_db = &dst->mt_dbs[i];
2007 /* Kill pointers into src - and dst to reduce abuse: The
2008 * user may not use mc until dst ends. Otherwise we'd...
2010 mc->mc_txn = NULL; /* ...set this to dst */
2011 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2012 if ((mx = mc->mc_xcursor) != NULL) {
2013 *(MDB_xcursor *)(bk+1) = *mx;
2014 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2016 mc->mc_next = dst->mt_cursors[i];
2017 dst->mt_cursors[i] = mc;
2024 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2025 * @param[in] txn the transaction handle.
2026 * @param[in] merge true to keep changes to parent cursors, false to revert.
2027 * @return 0 on success, non-zero on failure.
2030 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2032 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2036 for (i = txn->mt_numdbs; --i >= 0; ) {
2037 for (mc = cursors[i]; mc; mc = next) {
2039 if ((bk = mc->mc_backup) != NULL) {
2041 /* Commit changes to parent txn */
2042 mc->mc_next = bk->mc_next;
2043 mc->mc_backup = bk->mc_backup;
2044 mc->mc_txn = bk->mc_txn;
2045 mc->mc_db = bk->mc_db;
2046 mc->mc_dbflag = bk->mc_dbflag;
2047 if ((mx = mc->mc_xcursor) != NULL)
2048 mx->mx_cursor.mc_txn = bk->mc_txn;
2050 /* Abort nested txn */
2052 if ((mx = mc->mc_xcursor) != NULL)
2053 *mx = *(MDB_xcursor *)(bk+1);
2057 /* Only malloced cursors are permanently tracked. */
2065 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2068 mdb_txn_reset0(MDB_txn *txn, const char *act);
2070 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2076 Pidset = F_SETLK, Pidcheck = F_GETLK
2080 /** Set or check a pid lock. Set returns 0 on success.
2081 * Check returns 0 if the process is certainly dead, nonzero if it may
2082 * be alive (the lock exists or an error happened so we do not know).
2084 * On Windows Pidset is a no-op, we merely check for the existence
2085 * of the process with the given pid. On POSIX we use a single byte
2086 * lock on the lockfile, set at an offset equal to the pid.
2089 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2091 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2094 if (op == Pidcheck) {
2095 h = OpenProcess(env->me_pidquery, FALSE, pid);
2096 /* No documented "no such process" code, but other program use this: */
2098 return ErrCode() != ERROR_INVALID_PARAMETER;
2099 /* A process exists until all handles to it close. Has it exited? */
2100 ret = WaitForSingleObject(h, 0) != 0;
2107 struct flock lock_info;
2108 memset(&lock_info, 0, sizeof(lock_info));
2109 lock_info.l_type = F_WRLCK;
2110 lock_info.l_whence = SEEK_SET;
2111 lock_info.l_start = pid;
2112 lock_info.l_len = 1;
2113 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2114 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2116 } else if ((rc = ErrCode()) == EINTR) {
2124 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2125 * @param[in] txn the transaction handle to initialize
2126 * @return 0 on success, non-zero on failure.
2129 mdb_txn_renew0(MDB_txn *txn)
2131 MDB_env *env = txn->mt_env;
2134 int rc, new_notls = 0;
2137 txn->mt_numdbs = env->me_numdbs;
2138 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2140 if (txn->mt_flags & MDB_TXN_RDONLY) {
2141 if (!env->me_txns) {
2142 i = mdb_env_pick_meta(env);
2143 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2144 txn->mt_u.reader = NULL;
2146 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2147 pthread_getspecific(env->me_txkey);
2149 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2150 return MDB_BAD_RSLOT;
2152 pid_t pid = env->me_pid;
2153 pthread_t tid = pthread_self();
2155 if (!(env->me_flags & MDB_LIVE_READER)) {
2156 rc = mdb_reader_pid(env, Pidset, pid);
2158 UNLOCK_MUTEX_R(env);
2161 env->me_flags |= MDB_LIVE_READER;
2165 for (i=0; i<env->me_txns->mti_numreaders; i++)
2166 if (env->me_txns->mti_readers[i].mr_pid == 0)
2168 if (i == env->me_maxreaders) {
2169 UNLOCK_MUTEX_R(env);
2170 return MDB_READERS_FULL;
2172 env->me_txns->mti_readers[i].mr_pid = pid;
2173 env->me_txns->mti_readers[i].mr_tid = tid;
2174 if (i >= env->me_txns->mti_numreaders)
2175 env->me_txns->mti_numreaders = i+1;
2176 /* Save numreaders for un-mutexed mdb_env_close() */
2177 env->me_numreaders = env->me_txns->mti_numreaders;
2178 UNLOCK_MUTEX_R(env);
2179 r = &env->me_txns->mti_readers[i];
2180 new_notls = (env->me_flags & MDB_NOTLS);
2181 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2186 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2187 txn->mt_u.reader = r;
2189 txn->mt_toggle = txn->mt_txnid & 1;
2193 txn->mt_txnid = env->me_txns->mti_txnid;
2194 txn->mt_toggle = txn->mt_txnid & 1;
2197 if (txn->mt_txnid == mdb_debug_start)
2200 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2201 txn->mt_u.dirty_list = env->me_dirty_list;
2202 txn->mt_u.dirty_list[0].mid = 0;
2203 txn->mt_free_pgs = env->me_free_pgs;
2204 txn->mt_free_pgs[0] = 0;
2205 txn->mt_spill_pgs = NULL;
2209 /* Copy the DB info and flags */
2210 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2212 /* Moved to here to avoid a data race in read TXNs */
2213 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2215 for (i=2; i<txn->mt_numdbs; i++) {
2216 x = env->me_dbflags[i];
2217 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2218 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2220 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2222 if (env->me_maxpg < txn->mt_next_pgno) {
2223 mdb_txn_reset0(txn, "renew0-mapfail");
2225 txn->mt_u.reader->mr_pid = 0;
2226 txn->mt_u.reader = NULL;
2228 return MDB_MAP_RESIZED;
2235 mdb_txn_renew(MDB_txn *txn)
2239 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2242 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2243 DPUTS("environment had fatal error, must shutdown!");
2247 rc = mdb_txn_renew0(txn);
2248 if (rc == MDB_SUCCESS) {
2249 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2250 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2251 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2257 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2261 int rc, size, tsize = sizeof(MDB_txn);
2263 if (env->me_flags & MDB_FATAL_ERROR) {
2264 DPUTS("environment had fatal error, must shutdown!");
2267 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2270 /* Nested transactions: Max 1 child, write txns only, no writemap */
2271 if (parent->mt_child ||
2272 (flags & MDB_RDONLY) ||
2273 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2274 (env->me_flags & MDB_WRITEMAP))
2276 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2278 tsize = sizeof(MDB_ntxn);
2280 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2281 if (!(flags & MDB_RDONLY))
2282 size += env->me_maxdbs * sizeof(MDB_cursor *);
2284 if ((txn = calloc(1, size)) == NULL) {
2285 DPRINTF(("calloc: %s", strerror(ErrCode())));
2288 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2289 if (flags & MDB_RDONLY) {
2290 txn->mt_flags |= MDB_TXN_RDONLY;
2291 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2293 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2294 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2300 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2301 if (!txn->mt_u.dirty_list ||
2302 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2304 free(txn->mt_u.dirty_list);
2308 txn->mt_txnid = parent->mt_txnid;
2309 txn->mt_toggle = parent->mt_toggle;
2310 txn->mt_dirty_room = parent->mt_dirty_room;
2311 txn->mt_u.dirty_list[0].mid = 0;
2312 txn->mt_spill_pgs = NULL;
2313 txn->mt_next_pgno = parent->mt_next_pgno;
2314 parent->mt_child = txn;
2315 txn->mt_parent = parent;
2316 txn->mt_numdbs = parent->mt_numdbs;
2317 txn->mt_flags = parent->mt_flags;
2318 txn->mt_dbxs = parent->mt_dbxs;
2319 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2320 /* Copy parent's mt_dbflags, but clear DB_NEW */
2321 for (i=0; i<txn->mt_numdbs; i++)
2322 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2324 ntxn = (MDB_ntxn *)txn;
2325 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2326 if (env->me_pghead) {
2327 size = MDB_IDL_SIZEOF(env->me_pghead);
2328 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2330 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2335 rc = mdb_cursor_shadow(parent, txn);
2337 mdb_txn_reset0(txn, "beginchild-fail");
2339 rc = mdb_txn_renew0(txn);
2345 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2346 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2347 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2354 mdb_txn_env(MDB_txn *txn)
2356 if(!txn) return NULL;
2360 /** Export or close DBI handles opened in this txn. */
2362 mdb_dbis_update(MDB_txn *txn, int keep)
2365 MDB_dbi n = txn->mt_numdbs;
2366 MDB_env *env = txn->mt_env;
2367 unsigned char *tdbflags = txn->mt_dbflags;
2369 for (i = n; --i >= 2;) {
2370 if (tdbflags[i] & DB_NEW) {
2372 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2374 char *ptr = env->me_dbxs[i].md_name.mv_data;
2375 env->me_dbxs[i].md_name.mv_data = NULL;
2376 env->me_dbxs[i].md_name.mv_size = 0;
2377 env->me_dbflags[i] = 0;
2382 if (keep && env->me_numdbs < n)
2386 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2387 * May be called twice for readonly txns: First reset it, then abort.
2388 * @param[in] txn the transaction handle to reset
2389 * @param[in] act why the transaction is being reset
2392 mdb_txn_reset0(MDB_txn *txn, const char *act)
2394 MDB_env *env = txn->mt_env;
2396 /* Close any DBI handles opened in this txn */
2397 mdb_dbis_update(txn, 0);
2399 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2400 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2401 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2403 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2404 if (txn->mt_u.reader) {
2405 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2406 if (!(env->me_flags & MDB_NOTLS))
2407 txn->mt_u.reader = NULL; /* txn does not own reader */
2409 txn->mt_numdbs = 0; /* close nothing if called again */
2410 txn->mt_dbxs = NULL; /* mark txn as reset */
2412 mdb_cursors_close(txn, 0);
2414 if (!(env->me_flags & MDB_WRITEMAP)) {
2415 mdb_dlist_free(txn);
2417 mdb_midl_free(env->me_pghead);
2419 if (txn->mt_parent) {
2420 txn->mt_parent->mt_child = NULL;
2421 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2422 mdb_midl_free(txn->mt_free_pgs);
2423 mdb_midl_free(txn->mt_spill_pgs);
2424 free(txn->mt_u.dirty_list);
2428 if (mdb_midl_shrink(&txn->mt_free_pgs))
2429 env->me_free_pgs = txn->mt_free_pgs;
2430 env->me_pghead = NULL;
2434 /* The writer mutex was locked in mdb_txn_begin. */
2435 UNLOCK_MUTEX_W(env);
2440 mdb_txn_reset(MDB_txn *txn)
2445 /* This call is only valid for read-only txns */
2446 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2449 mdb_txn_reset0(txn, "reset");
2453 mdb_txn_abort(MDB_txn *txn)
2459 mdb_txn_abort(txn->mt_child);
2461 mdb_txn_reset0(txn, "abort");
2462 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2463 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2464 txn->mt_u.reader->mr_pid = 0;
2469 /** Save the freelist as of this transaction to the freeDB.
2470 * This changes the freelist. Keep trying until it stabilizes.
2473 mdb_freelist_save(MDB_txn *txn)
2475 /* env->me_pghead[] can grow and shrink during this call.
2476 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2477 * Page numbers cannot disappear from txn->mt_free_pgs[].
2480 MDB_env *env = txn->mt_env;
2481 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2482 txnid_t pglast = 0, head_id = 0;
2483 pgno_t freecnt = 0, *free_pgs, *mop;
2484 ssize_t head_room = 0, total_room = 0, mop_len;
2486 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2488 if (env->me_pghead) {
2489 /* Make sure first page of freeDB is touched and on freelist */
2490 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2491 if (rc && rc != MDB_NOTFOUND)
2496 /* Come back here after each Put() in case freelist changed */
2499 /* If using records from freeDB which we have not yet
2500 * deleted, delete them and any we reserved for me_pghead.
2502 while (pglast < env->me_pglast) {
2503 rc = mdb_cursor_first(&mc, &key, NULL);
2506 pglast = head_id = *(txnid_t *)key.mv_data;
2507 total_room = head_room = 0;
2508 assert(pglast <= env->me_pglast);
2509 rc = mdb_cursor_del(&mc, 0);
2514 /* Save the IDL of pages freed by this txn, to a single record */
2515 if (freecnt < txn->mt_free_pgs[0]) {
2517 /* Make sure last page of freeDB is touched and on freelist */
2518 key.mv_size = MDB_MAXKEYSIZE+1;
2520 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2521 if (rc && rc != MDB_NOTFOUND)
2524 free_pgs = txn->mt_free_pgs;
2525 /* Write to last page of freeDB */
2526 key.mv_size = sizeof(txn->mt_txnid);
2527 key.mv_data = &txn->mt_txnid;
2529 freecnt = free_pgs[0];
2530 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2531 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2534 /* Retry if mt_free_pgs[] grew during the Put() */
2535 free_pgs = txn->mt_free_pgs;
2536 } while (freecnt < free_pgs[0]);
2537 mdb_midl_sort(free_pgs);
2538 memcpy(data.mv_data, free_pgs, data.mv_size);
2541 unsigned int i = free_pgs[0];
2542 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2543 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2545 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2551 mop = env->me_pghead;
2552 mop_len = mop ? mop[0] : 0;
2554 /* Reserve records for me_pghead[]. Split it if multi-page,
2555 * to avoid searching freeDB for a page range. Use keys in
2556 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2558 if (total_room >= mop_len) {
2559 if (total_room == mop_len || --more < 0)
2561 } else if (head_room >= maxfree_1pg && head_id > 1) {
2562 /* Keep current record (overflow page), add a new one */
2566 /* (Re)write {key = head_id, IDL length = head_room} */
2567 total_room -= head_room;
2568 head_room = mop_len - total_room;
2569 if (head_room > maxfree_1pg && head_id > 1) {
2570 /* Overflow multi-page for part of me_pghead */
2571 head_room /= head_id; /* amortize page sizes */
2572 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2573 } else if (head_room < 0) {
2574 /* Rare case, not bothering to delete this record */
2577 key.mv_size = sizeof(head_id);
2578 key.mv_data = &head_id;
2579 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2580 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2583 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2584 total_room += head_room;
2587 /* Fill in the reserved, touched me_pghead records */
2593 rc = mdb_cursor_first(&mc, &key, &data);
2594 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2595 unsigned flags = MDB_CURRENT;
2596 txnid_t id = *(txnid_t *)key.mv_data;
2597 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2600 assert(len >= 0 && id <= env->me_pglast);
2602 if (len > mop_len) {
2604 data.mv_size = (len + 1) * sizeof(MDB_ID);
2607 data.mv_data = mop -= len;
2610 rc = mdb_cursor_put(&mc, &key, &data, flags);
2612 if (rc || !(mop_len -= len))
2619 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2620 * @param[in] txn the transaction that's being committed
2621 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2622 * @return 0 on success, non-zero on failure.
2625 mdb_page_flush(MDB_txn *txn, int keep)
2627 MDB_env *env = txn->mt_env;
2628 MDB_ID2L dl = txn->mt_u.dirty_list;
2629 unsigned psize = env->me_psize, j;
2630 int i, pagecount = dl[0].mid, rc;
2631 size_t size = 0, pos = 0;
2633 MDB_page *dp = NULL;
2637 struct iovec iov[MDB_COMMIT_PAGES];
2638 ssize_t wpos = 0, wsize = 0, wres;
2639 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2645 if (env->me_flags & MDB_WRITEMAP) {
2646 /* Clear dirty flags */
2647 while (++i <= pagecount) {
2649 /* Don't flush this page yet */
2650 if (dp->mp_flags & P_KEEP) {
2651 dp->mp_flags ^= P_KEEP;
2655 dp->mp_flags &= ~P_DIRTY;
2661 /* Write the pages */
2663 if (++i <= pagecount) {
2665 /* Don't flush this page yet */
2666 if (dp->mp_flags & P_KEEP) {
2667 dp->mp_flags ^= P_KEEP;
2672 /* clear dirty flag */
2673 dp->mp_flags &= ~P_DIRTY;
2676 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2681 /* Windows actually supports scatter/gather I/O, but only on
2682 * unbuffered file handles. Since we're relying on the OS page
2683 * cache for all our data, that's self-defeating. So we just
2684 * write pages one at a time. We use the ov structure to set
2685 * the write offset, to at least save the overhead of a Seek
2688 DPRINTF(("committing page %"Z"u", pgno));
2689 memset(&ov, 0, sizeof(ov));
2690 ov.Offset = pos & 0xffffffff;
2691 ov.OffsetHigh = pos >> 16 >> 16;
2692 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2694 DPRINTF(("WriteFile: %d", rc));
2698 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2699 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2701 /* Write previous page(s) */
2702 #ifdef MDB_USE_PWRITEV
2703 wres = pwritev(env->me_fd, iov, n, wpos);
2706 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2708 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2710 DPRINTF(("lseek: %s", strerror(rc)));
2713 wres = writev(env->me_fd, iov, n);
2716 if (wres != wsize) {
2719 DPRINTF(("Write error: %s", strerror(rc)));
2721 rc = EIO; /* TODO: Use which error code? */
2722 DPUTS("short write, filesystem full?");
2733 DPRINTF(("committing page %"Z"u", pgno));
2734 next_pos = pos + size;
2735 iov[n].iov_len = size;
2736 iov[n].iov_base = (char *)dp;
2742 for (i = keep; ++i <= pagecount; ) {
2744 /* This is a page we skipped above */
2747 dl[j].mid = dp->mp_pgno;
2750 mdb_dpage_free(env, dp);
2758 mdb_txn_commit(MDB_txn *txn)
2764 assert(txn != NULL);
2765 assert(txn->mt_env != NULL);
2767 if (txn->mt_child) {
2768 rc = mdb_txn_commit(txn->mt_child);
2769 txn->mt_child = NULL;
2776 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2777 mdb_dbis_update(txn, 1);
2778 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2783 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2784 DPUTS("error flag is set, can't commit");
2786 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2791 if (txn->mt_parent) {
2792 MDB_txn *parent = txn->mt_parent;
2796 /* Append our free list to parent's */
2797 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2800 mdb_midl_free(txn->mt_free_pgs);
2802 parent->mt_next_pgno = txn->mt_next_pgno;
2803 parent->mt_flags = txn->mt_flags;
2805 /* Merge our cursors into parent's and close them */
2806 mdb_cursors_close(txn, 1);
2808 /* Update parent's DB table. */
2809 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2810 parent->mt_numdbs = txn->mt_numdbs;
2811 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2812 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2813 for (i=2; i<txn->mt_numdbs; i++) {
2814 /* preserve parent's DB_NEW status */
2815 x = parent->mt_dbflags[i] & DB_NEW;
2816 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2819 dst = parent->mt_u.dirty_list;
2820 src = txn->mt_u.dirty_list;
2821 /* Remove anything in our dirty list from parent's spill list */
2822 if (parent->mt_spill_pgs) {
2823 x = parent->mt_spill_pgs[0];
2825 /* zero out our dirty pages in parent spill list */
2826 for (i=1; i<=src[0].mid; i++) {
2827 MDB_ID pn = src[i].mid << 1;
2828 if (pn < parent->mt_spill_pgs[x])
2830 if (pn > parent->mt_spill_pgs[x]) {
2836 parent->mt_spill_pgs[x] = 0;
2839 /* OK, we had a few hits, squash zeros from the spill list */
2840 if (len < parent->mt_spill_pgs[0]) {
2842 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2843 if (parent->mt_spill_pgs[y]) {
2845 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2850 parent->mt_spill_pgs[0] = len;
2853 /* Find len = length of merging our dirty list with parent's */
2855 dst[0].mid = 0; /* simplify loops */
2856 if (parent->mt_parent) {
2857 len = x + src[0].mid;
2858 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2859 for (i = x; y && i; y--) {
2860 pgno_t yp = src[y].mid;
2861 while (yp < dst[i].mid)
2863 if (yp == dst[i].mid) {
2868 } else { /* Simplify the above for single-ancestor case */
2869 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2871 /* Merge our dirty list with parent's */
2873 for (i = len; y; dst[i--] = src[y--]) {
2874 pgno_t yp = src[y].mid;
2875 while (yp < dst[x].mid)
2876 dst[i--] = dst[x--];
2877 if (yp == dst[x].mid)
2878 free(dst[x--].mptr);
2882 free(txn->mt_u.dirty_list);
2883 parent->mt_dirty_room = txn->mt_dirty_room;
2884 if (txn->mt_spill_pgs) {
2885 if (parent->mt_spill_pgs) {
2886 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2887 mdb_midl_free(txn->mt_spill_pgs);
2888 mdb_midl_sort(parent->mt_spill_pgs);
2890 parent->mt_spill_pgs = txn->mt_spill_pgs;
2894 parent->mt_child = NULL;
2895 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2900 if (txn != env->me_txn) {
2901 DPUTS("attempt to commit unknown transaction");
2906 mdb_cursors_close(txn, 0);
2908 if (!txn->mt_u.dirty_list[0].mid &&
2909 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2912 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2913 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2915 /* Update DB root pointers */
2916 if (txn->mt_numdbs > 2) {
2920 data.mv_size = sizeof(MDB_db);
2922 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2923 for (i = 2; i < txn->mt_numdbs; i++) {
2924 if (txn->mt_dbflags[i] & DB_DIRTY) {
2925 data.mv_data = &txn->mt_dbs[i];
2926 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2933 rc = mdb_freelist_save(txn);
2937 mdb_midl_free(env->me_pghead);
2938 env->me_pghead = NULL;
2939 if (mdb_midl_shrink(&txn->mt_free_pgs))
2940 env->me_free_pgs = txn->mt_free_pgs;
2946 if ((rc = mdb_page_flush(txn, 0)) ||
2947 (rc = mdb_env_sync(env, 0)) ||
2948 (rc = mdb_env_write_meta(txn)))
2954 mdb_dbis_update(txn, 1);
2956 UNLOCK_MUTEX_W(env);
2966 /** Read the environment parameters of a DB environment before
2967 * mapping it into memory.
2968 * @param[in] env the environment handle
2969 * @param[out] meta address of where to store the meta information
2970 * @return 0 on success, non-zero on failure.
2973 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2980 /* We don't know the page size yet, so use a minimum value.
2981 * Read both meta pages so we can use the latest one.
2984 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2988 memset(&ov, 0, sizeof(ov));
2990 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2991 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2994 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2996 if (rc != MDB_PAGESIZE) {
2997 if (rc == 0 && off == 0)
2999 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3000 DPRINTF(("read: %s", mdb_strerror(rc)));
3004 p = (MDB_page *)&pbuf;
3006 if (!F_ISSET(p->mp_flags, P_META)) {
3007 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3012 if (m->mm_magic != MDB_MAGIC) {
3013 DPUTS("meta has invalid magic");
3017 if (m->mm_version != MDB_DATA_VERSION) {
3018 DPRINTF(("database is version %u, expected version %u",
3019 m->mm_version, MDB_DATA_VERSION));
3020 return MDB_VERSION_MISMATCH;
3023 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3029 /** Write the environment parameters of a freshly created DB environment.
3030 * @param[in] env the environment handle
3031 * @param[out] meta address of where to store the meta information
3032 * @return 0 on success, non-zero on failure.
3035 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3043 memset(&ov, 0, sizeof(ov));
3044 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3046 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3049 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3050 len = pwrite(fd, ptr, size, pos); \
3051 rc = (len >= 0); } while(0)
3054 DPUTS("writing new meta page");
3056 psize = env->me_psize;
3058 meta->mm_magic = MDB_MAGIC;
3059 meta->mm_version = MDB_DATA_VERSION;
3060 meta->mm_mapsize = env->me_mapsize;
3061 meta->mm_psize = psize;
3062 meta->mm_last_pg = 1;
3063 meta->mm_flags = env->me_flags & 0xffff;
3064 meta->mm_flags |= MDB_INTEGERKEY;
3065 meta->mm_dbs[0].md_root = P_INVALID;
3066 meta->mm_dbs[1].md_root = P_INVALID;
3068 p = calloc(2, psize);
3070 p->mp_flags = P_META;
3071 *(MDB_meta *)METADATA(p) = *meta;
3073 q = (MDB_page *)((char *)p + psize);
3075 q->mp_flags = P_META;
3076 *(MDB_meta *)METADATA(q) = *meta;
3078 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3081 else if ((unsigned) len == psize * 2)
3089 /** Update the environment info to commit a transaction.
3090 * @param[in] txn the transaction that's being committed
3091 * @return 0 on success, non-zero on failure.
3094 mdb_env_write_meta(MDB_txn *txn)
3097 MDB_meta meta, metab, *mp;
3099 int rc, len, toggle;
3108 assert(txn != NULL);
3109 assert(txn->mt_env != NULL);
3111 toggle = !txn->mt_toggle;
3112 DPRINTF(("writing meta page %d for root page %"Z"u",
3113 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3116 mp = env->me_metas[toggle];
3118 if (env->me_flags & MDB_WRITEMAP) {
3119 /* Persist any increases of mapsize config */
3120 if (env->me_mapsize > mp->mm_mapsize)
3121 mp->mm_mapsize = env->me_mapsize;
3122 mp->mm_dbs[0] = txn->mt_dbs[0];
3123 mp->mm_dbs[1] = txn->mt_dbs[1];
3124 mp->mm_last_pg = txn->mt_next_pgno - 1;
3125 mp->mm_txnid = txn->mt_txnid;
3126 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3127 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3130 ptr += env->me_psize;
3131 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3138 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3139 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3141 ptr = (char *)&meta;
3142 if (env->me_mapsize > mp->mm_mapsize) {
3143 /* Persist any increases of mapsize config */
3144 meta.mm_mapsize = env->me_mapsize;
3145 off = offsetof(MDB_meta, mm_mapsize);
3147 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3149 len = sizeof(MDB_meta) - off;
3152 meta.mm_dbs[0] = txn->mt_dbs[0];
3153 meta.mm_dbs[1] = txn->mt_dbs[1];
3154 meta.mm_last_pg = txn->mt_next_pgno - 1;
3155 meta.mm_txnid = txn->mt_txnid;
3158 off += env->me_psize;
3161 /* Write to the SYNC fd */
3162 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3163 env->me_fd : env->me_mfd;
3166 memset(&ov, 0, sizeof(ov));
3168 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3172 rc = pwrite(mfd, ptr, len, off);
3175 rc = rc < 0 ? ErrCode() : EIO;
3176 DPUTS("write failed, disk error?");
3177 /* On a failure, the pagecache still contains the new data.
3178 * Write some old data back, to prevent it from being used.
3179 * Use the non-SYNC fd; we know it will fail anyway.
3181 meta.mm_last_pg = metab.mm_last_pg;
3182 meta.mm_txnid = metab.mm_txnid;
3184 memset(&ov, 0, sizeof(ov));
3186 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3188 r2 = pwrite(env->me_fd, ptr, len, off);
3189 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3192 env->me_flags |= MDB_FATAL_ERROR;
3196 /* Memory ordering issues are irrelevant; since the entire writer
3197 * is wrapped by wmutex, all of these changes will become visible
3198 * after the wmutex is unlocked. Since the DB is multi-version,
3199 * readers will get consistent data regardless of how fresh or
3200 * how stale their view of these values is.
3202 env->me_txns->mti_txnid = txn->mt_txnid;
3207 /** Check both meta pages to see which one is newer.
3208 * @param[in] env the environment handle
3209 * @return meta toggle (0 or 1).
3212 mdb_env_pick_meta(const MDB_env *env)
3214 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3218 mdb_env_create(MDB_env **env)
3222 e = calloc(1, sizeof(MDB_env));
3226 e->me_maxreaders = DEFAULT_READERS;
3227 e->me_maxdbs = e->me_numdbs = 2;
3228 e->me_fd = INVALID_HANDLE_VALUE;
3229 e->me_lfd = INVALID_HANDLE_VALUE;
3230 e->me_mfd = INVALID_HANDLE_VALUE;
3231 #ifdef MDB_USE_POSIX_SEM
3232 e->me_rmutex = SEM_FAILED;
3233 e->me_wmutex = SEM_FAILED;
3235 e->me_pid = getpid();
3236 VGMEMP_CREATE(e,0,0);
3242 mdb_env_map(MDB_env *env, void *addr, int newsize)
3245 unsigned int flags = env->me_flags;
3249 LONG sizelo, sizehi;
3250 sizelo = env->me_mapsize & 0xffffffff;
3251 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3253 /* Windows won't create mappings for zero length files.
3254 * Just allocate the maxsize right now.
3257 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3258 || !SetEndOfFile(env->me_fd)
3259 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3262 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3263 PAGE_READWRITE : PAGE_READONLY,
3264 sizehi, sizelo, NULL);
3267 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3268 FILE_MAP_WRITE : FILE_MAP_READ,
3269 0, 0, env->me_mapsize, addr);
3270 rc = env->me_map ? 0 : ErrCode();
3275 int prot = PROT_READ;
3276 if (flags & MDB_WRITEMAP) {
3278 if (newsize && ftruncate(env->me_fd, env->me_mapsize) < 0)
3281 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3283 if (env->me_map == MAP_FAILED) {
3287 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3289 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3291 #ifdef POSIX_MADV_RANDOM
3292 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3293 #endif /* POSIX_MADV_RANDOM */
3294 #endif /* MADV_RANDOM */
3297 /* Can happen because the address argument to mmap() is just a
3298 * hint. mmap() can pick another, e.g. if the range is in use.
3299 * The MAP_FIXED flag would prevent that, but then mmap could
3300 * instead unmap existing pages to make room for the new map.
3302 if (addr && env->me_map != addr)
3303 return EBUSY; /* TODO: Make a new MDB_* error code? */
3305 p = (MDB_page *)env->me_map;
3306 env->me_metas[0] = METADATA(p);
3307 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3313 mdb_env_set_mapsize(MDB_env *env, size_t size)
3315 /* If env is already open, caller is responsible for making
3316 * sure there are no active txns.
3324 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3325 munmap(env->me_map, env->me_mapsize);
3326 env->me_mapsize = size;
3327 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3328 rc = mdb_env_map(env, old, 1);
3332 env->me_mapsize = size;
3334 env->me_maxpg = env->me_mapsize / env->me_psize;
3339 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3343 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3348 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3350 if (env->me_map || readers < 1)
3352 env->me_maxreaders = readers;
3357 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3359 if (!env || !readers)
3361 *readers = env->me_maxreaders;
3365 /** Further setup required for opening an MDB environment
3368 mdb_env_open2(MDB_env *env)
3370 unsigned int flags = env->me_flags;
3371 int i, newenv = 0, rc;
3375 /* See if we should use QueryLimited */
3377 if ((rc & 0xff) > 5)
3378 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3380 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3383 memset(&meta, 0, sizeof(meta));
3385 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3388 DPUTS("new mdbenv");
3390 GET_PAGESIZE(env->me_psize);
3392 env->me_psize = meta.mm_psize;
3395 /* Was a mapsize configured? */
3396 if (!env->me_mapsize) {
3397 /* If this is a new environment, take the default,
3398 * else use the size recorded in the existing env.
3400 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3401 } else if (env->me_mapsize < meta.mm_mapsize) {
3402 /* If the configured size is smaller, make sure it's
3403 * still big enough. Silently round up to minimum if not.
3405 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3406 if (env->me_mapsize < minsize)
3407 env->me_mapsize = minsize;
3410 rc = mdb_env_map(env, meta.mm_address, newenv);
3415 if (flags & MDB_FIXEDMAP)
3416 meta.mm_address = env->me_map;
3417 i = mdb_env_init_meta(env, &meta);
3418 if (i != MDB_SUCCESS) {
3422 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3423 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3425 env->me_maxpg = env->me_mapsize / env->me_psize;
3428 int toggle = mdb_env_pick_meta(env);
3429 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3431 DPRINTF(("opened database version %u, pagesize %u",
3432 env->me_metas[0]->mm_version, env->me_psize));
3433 DPRINTF(("using meta page %d", toggle));
3434 DPRINTF(("depth: %u", db->md_depth));
3435 DPRINTF(("entries: %"Z"u", db->md_entries));
3436 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3437 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3438 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3439 DPRINTF(("root: %"Z"u", db->md_root));
3447 /** Release a reader thread's slot in the reader lock table.
3448 * This function is called automatically when a thread exits.
3449 * @param[in] ptr This points to the slot in the reader lock table.
3452 mdb_env_reader_dest(void *ptr)
3454 MDB_reader *reader = ptr;
3460 /** Junk for arranging thread-specific callbacks on Windows. This is
3461 * necessarily platform and compiler-specific. Windows supports up
3462 * to 1088 keys. Let's assume nobody opens more than 64 environments
3463 * in a single process, for now. They can override this if needed.
3465 #ifndef MAX_TLS_KEYS
3466 #define MAX_TLS_KEYS 64
3468 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3469 static int mdb_tls_nkeys;
3471 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3475 case DLL_PROCESS_ATTACH: break;
3476 case DLL_THREAD_ATTACH: break;
3477 case DLL_THREAD_DETACH:
3478 for (i=0; i<mdb_tls_nkeys; i++) {
3479 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3480 mdb_env_reader_dest(r);
3483 case DLL_PROCESS_DETACH: break;
3488 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3490 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3494 /* Force some symbol references.
3495 * _tls_used forces the linker to create the TLS directory if not already done
3496 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3498 #pragma comment(linker, "/INCLUDE:_tls_used")
3499 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3500 #pragma const_seg(".CRT$XLB")
3501 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3502 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3505 #pragma comment(linker, "/INCLUDE:__tls_used")
3506 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3507 #pragma data_seg(".CRT$XLB")
3508 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3510 #endif /* WIN 32/64 */
3511 #endif /* !__GNUC__ */
3514 /** Downgrade the exclusive lock on the region back to shared */
3516 mdb_env_share_locks(MDB_env *env, int *excl)
3518 int rc = 0, toggle = mdb_env_pick_meta(env);
3520 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3525 /* First acquire a shared lock. The Unlock will
3526 * then release the existing exclusive lock.
3528 memset(&ov, 0, sizeof(ov));
3529 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3532 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3538 struct flock lock_info;
3539 /* The shared lock replaces the existing lock */
3540 memset((void *)&lock_info, 0, sizeof(lock_info));
3541 lock_info.l_type = F_RDLCK;
3542 lock_info.l_whence = SEEK_SET;
3543 lock_info.l_start = 0;
3544 lock_info.l_len = 1;
3545 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3546 (rc = ErrCode()) == EINTR) ;
3547 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3554 /** Try to get exlusive lock, otherwise shared.
3555 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3558 mdb_env_excl_lock(MDB_env *env, int *excl)
3562 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3566 memset(&ov, 0, sizeof(ov));
3567 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3574 struct flock lock_info;
3575 memset((void *)&lock_info, 0, sizeof(lock_info));
3576 lock_info.l_type = F_WRLCK;
3577 lock_info.l_whence = SEEK_SET;
3578 lock_info.l_start = 0;
3579 lock_info.l_len = 1;
3580 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3581 (rc = ErrCode()) == EINTR) ;
3585 # ifdef MDB_USE_POSIX_SEM
3586 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3589 lock_info.l_type = F_RDLCK;
3590 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3591 (rc = ErrCode()) == EINTR) ;
3599 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3601 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3603 * @(#) $Revision: 5.1 $
3604 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3605 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3607 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3611 * Please do not copyright this code. This code is in the public domain.
3613 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3614 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3615 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3616 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3617 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3618 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3619 * PERFORMANCE OF THIS SOFTWARE.
3622 * chongo <Landon Curt Noll> /\oo/\
3623 * http://www.isthe.com/chongo/
3625 * Share and Enjoy! :-)
3628 typedef unsigned long long mdb_hash_t;
3629 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3631 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3632 * @param[in] val value to hash
3633 * @param[in] hval initial value for hash
3634 * @return 64 bit hash
3636 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3637 * hval arg on the first call.
3640 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3642 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3643 unsigned char *end = s + val->mv_size;
3645 * FNV-1a hash each octet of the string
3648 /* xor the bottom with the current octet */
3649 hval ^= (mdb_hash_t)*s++;
3651 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3652 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3653 (hval << 7) + (hval << 8) + (hval << 40);
3655 /* return our new hash value */
3659 /** Hash the string and output the encoded hash.
3660 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3661 * very short name limits. We don't care about the encoding being reversible,
3662 * we just want to preserve as many bits of the input as possible in a
3663 * small printable string.
3664 * @param[in] str string to hash
3665 * @param[out] encbuf an array of 11 chars to hold the hash
3667 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3670 mdb_pack85(unsigned long l, char *out)
3674 for (i=0; i<5; i++) {
3675 *out++ = mdb_a85[l % 85];
3681 mdb_hash_enc(MDB_val *val, char *encbuf)
3683 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3685 mdb_pack85(h, encbuf);
3686 mdb_pack85(h>>32, encbuf+5);
3691 /** Open and/or initialize the lock region for the environment.
3692 * @param[in] env The MDB environment.
3693 * @param[in] lpath The pathname of the file used for the lock region.
3694 * @param[in] mode The Unix permissions for the file, if we create it.
3695 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3696 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3697 * @return 0 on success, non-zero on failure.
3700 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3703 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3705 # define MDB_ERRCODE_ROFS EROFS
3706 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3707 # define MDB_CLOEXEC O_CLOEXEC
3710 # define MDB_CLOEXEC 0
3717 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3718 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3719 FILE_ATTRIBUTE_NORMAL, NULL);
3721 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3723 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3725 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3730 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3731 /* Lose record locks when exec*() */
3732 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3733 fcntl(env->me_lfd, F_SETFD, fdflags);
3736 if (!(env->me_flags & MDB_NOTLS)) {
3737 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3740 env->me_flags |= MDB_ENV_TXKEY;
3742 /* Windows TLS callbacks need help finding their TLS info. */
3743 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3747 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3751 /* Try to get exclusive lock. If we succeed, then
3752 * nobody is using the lock region and we should initialize it.
3754 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3757 size = GetFileSize(env->me_lfd, NULL);
3759 size = lseek(env->me_lfd, 0, SEEK_END);
3760 if (size == -1) goto fail_errno;
3762 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3763 if (size < rsize && *excl > 0) {
3765 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3766 || !SetEndOfFile(env->me_lfd))
3769 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3773 size = rsize - sizeof(MDB_txninfo);
3774 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3779 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3781 if (!mh) goto fail_errno;
3782 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3784 if (!env->me_txns) goto fail_errno;
3786 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3788 if (m == MAP_FAILED) goto fail_errno;
3794 BY_HANDLE_FILE_INFORMATION stbuf;
3803 if (!mdb_sec_inited) {
3804 InitializeSecurityDescriptor(&mdb_null_sd,
3805 SECURITY_DESCRIPTOR_REVISION);
3806 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3807 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3808 mdb_all_sa.bInheritHandle = FALSE;
3809 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3812 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3813 idbuf.volume = stbuf.dwVolumeSerialNumber;
3814 idbuf.nhigh = stbuf.nFileIndexHigh;
3815 idbuf.nlow = stbuf.nFileIndexLow;
3816 val.mv_data = &idbuf;
3817 val.mv_size = sizeof(idbuf);
3818 mdb_hash_enc(&val, encbuf);
3819 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3820 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3821 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3822 if (!env->me_rmutex) goto fail_errno;
3823 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3824 if (!env->me_wmutex) goto fail_errno;
3825 #elif defined(MDB_USE_POSIX_SEM)
3834 #if defined(__NetBSD__)
3835 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3837 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3838 idbuf.dev = stbuf.st_dev;
3839 idbuf.ino = stbuf.st_ino;
3840 val.mv_data = &idbuf;
3841 val.mv_size = sizeof(idbuf);
3842 mdb_hash_enc(&val, encbuf);
3843 #ifdef MDB_SHORT_SEMNAMES
3844 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3846 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3847 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3848 /* Clean up after a previous run, if needed: Try to
3849 * remove both semaphores before doing anything else.
3851 sem_unlink(env->me_txns->mti_rmname);
3852 sem_unlink(env->me_txns->mti_wmname);
3853 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3854 O_CREAT|O_EXCL, mode, 1);
3855 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3856 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3857 O_CREAT|O_EXCL, mode, 1);
3858 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3859 #else /* MDB_USE_POSIX_SEM */
3860 pthread_mutexattr_t mattr;
3862 if ((rc = pthread_mutexattr_init(&mattr))
3863 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3864 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3865 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3867 pthread_mutexattr_destroy(&mattr);
3868 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3870 env->me_txns->mti_magic = MDB_MAGIC;
3871 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3872 env->me_txns->mti_txnid = 0;
3873 env->me_txns->mti_numreaders = 0;
3876 if (env->me_txns->mti_magic != MDB_MAGIC) {
3877 DPUTS("lock region has invalid magic");
3881 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3882 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3883 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3884 rc = MDB_VERSION_MISMATCH;
3888 if (rc && rc != EACCES && rc != EAGAIN) {
3892 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3893 if (!env->me_rmutex) goto fail_errno;
3894 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3895 if (!env->me_wmutex) goto fail_errno;
3896 #elif defined(MDB_USE_POSIX_SEM)
3897 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3898 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3899 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3900 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3911 /** The name of the lock file in the DB environment */
3912 #define LOCKNAME "/lock.mdb"
3913 /** The name of the data file in the DB environment */
3914 #define DATANAME "/data.mdb"
3915 /** The suffix of the lock file when no subdir is used */
3916 #define LOCKSUFF "-lock"
3917 /** Only a subset of the @ref mdb_env flags can be changed
3918 * at runtime. Changing other flags requires closing the
3919 * environment and re-opening it with the new flags.
3921 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3922 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3925 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3927 int oflags, rc, len, excl = -1;
3928 char *lpath, *dpath;
3930 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3934 if (flags & MDB_NOSUBDIR) {
3935 rc = len + sizeof(LOCKSUFF) + len + 1;
3937 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3942 if (flags & MDB_NOSUBDIR) {
3943 dpath = lpath + len + sizeof(LOCKSUFF);
3944 sprintf(lpath, "%s" LOCKSUFF, path);
3945 strcpy(dpath, path);
3947 dpath = lpath + len + sizeof(LOCKNAME);
3948 sprintf(lpath, "%s" LOCKNAME, path);
3949 sprintf(dpath, "%s" DATANAME, path);
3953 flags |= env->me_flags;
3954 if (flags & MDB_RDONLY) {
3955 /* silently ignore WRITEMAP when we're only getting read access */
3956 flags &= ~MDB_WRITEMAP;
3958 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3959 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3962 env->me_flags = flags |= MDB_ENV_ACTIVE;
3966 env->me_path = strdup(path);
3967 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3968 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3969 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3974 /* For RDONLY, get lockfile after we know datafile exists */
3975 if (!F_ISSET(flags, MDB_RDONLY)) {
3976 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3982 if (F_ISSET(flags, MDB_RDONLY)) {
3983 oflags = GENERIC_READ;
3984 len = OPEN_EXISTING;
3986 oflags = GENERIC_READ|GENERIC_WRITE;
3989 mode = FILE_ATTRIBUTE_NORMAL;
3990 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3991 NULL, len, mode, NULL);
3993 if (F_ISSET(flags, MDB_RDONLY))
3996 oflags = O_RDWR | O_CREAT;
3998 env->me_fd = open(dpath, oflags, mode);
4000 if (env->me_fd == INVALID_HANDLE_VALUE) {
4005 if (F_ISSET(flags, MDB_RDONLY)) {
4006 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4011 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4012 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4013 env->me_mfd = env->me_fd;
4015 /* Synchronous fd for meta writes. Needed even with
4016 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4019 len = OPEN_EXISTING;
4020 env->me_mfd = CreateFile(dpath, oflags,
4021 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4022 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4025 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4027 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4032 DPRINTF(("opened dbenv %p", (void *) env));
4034 rc = mdb_env_share_locks(env, &excl);
4040 mdb_env_close0(env, excl);
4046 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4048 mdb_env_close0(MDB_env *env, int excl)
4052 if (!(env->me_flags & MDB_ENV_ACTIVE))
4055 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4056 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4057 free(env->me_dbxs[i].md_name.mv_data);
4059 free(env->me_dbflags);
4062 free(env->me_dirty_list);
4063 mdb_midl_free(env->me_free_pgs);
4065 if (env->me_flags & MDB_ENV_TXKEY) {
4066 pthread_key_delete(env->me_txkey);
4068 /* Delete our key from the global list */
4069 for (i=0; i<mdb_tls_nkeys; i++)
4070 if (mdb_tls_keys[i] == env->me_txkey) {
4071 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4079 munmap(env->me_map, env->me_mapsize);
4081 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4082 (void) close(env->me_mfd);
4083 if (env->me_fd != INVALID_HANDLE_VALUE)
4084 (void) close(env->me_fd);
4086 pid_t pid = env->me_pid;
4087 /* Clearing readers is done in this function because
4088 * me_txkey with its destructor must be disabled first.
4090 for (i = env->me_numreaders; --i >= 0; )
4091 if (env->me_txns->mti_readers[i].mr_pid == pid)
4092 env->me_txns->mti_readers[i].mr_pid = 0;
4094 if (env->me_rmutex) {
4095 CloseHandle(env->me_rmutex);
4096 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4098 /* Windows automatically destroys the mutexes when
4099 * the last handle closes.
4101 #elif defined(MDB_USE_POSIX_SEM)
4102 if (env->me_rmutex != SEM_FAILED) {
4103 sem_close(env->me_rmutex);
4104 if (env->me_wmutex != SEM_FAILED)
4105 sem_close(env->me_wmutex);
4106 /* If we have the filelock: If we are the
4107 * only remaining user, clean up semaphores.
4110 mdb_env_excl_lock(env, &excl);
4112 sem_unlink(env->me_txns->mti_rmname);
4113 sem_unlink(env->me_txns->mti_wmname);
4117 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4119 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4122 /* Unlock the lockfile. Windows would have unlocked it
4123 * after closing anyway, but not necessarily at once.
4125 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4128 (void) close(env->me_lfd);
4131 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4135 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4137 MDB_txn *txn = NULL;
4143 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4147 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4150 /* Do the lock/unlock of the reader mutex before starting the
4151 * write txn. Otherwise other read txns could block writers.
4153 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4158 /* We must start the actual read txn after blocking writers */
4159 mdb_txn_reset0(txn, "reset-stage1");
4161 /* Temporarily block writers until we snapshot the meta pages */
4164 rc = mdb_txn_renew0(txn);
4166 UNLOCK_MUTEX_W(env);
4171 wsize = env->me_psize * 2;
4175 DO_WRITE(rc, fd, ptr, w2, len);
4179 } else if (len > 0) {
4185 /* Non-blocking or async handles are not supported */
4191 UNLOCK_MUTEX_W(env);
4196 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4198 if (wsize > MAX_WRITE)
4202 DO_WRITE(rc, fd, ptr, w2, len);
4206 } else if (len > 0) {
4223 mdb_env_copy(MDB_env *env, const char *path)
4227 HANDLE newfd = INVALID_HANDLE_VALUE;
4229 if (env->me_flags & MDB_NOSUBDIR) {
4230 lpath = (char *)path;
4233 len += sizeof(DATANAME);
4234 lpath = malloc(len);
4237 sprintf(lpath, "%s" DATANAME, path);
4240 /* The destination path must exist, but the destination file must not.
4241 * We don't want the OS to cache the writes, since the source data is
4242 * already in the OS cache.
4245 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4246 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4248 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4254 if (newfd == INVALID_HANDLE_VALUE) {
4259 #ifdef F_NOCACHE /* __APPLE__ */
4260 rc = fcntl(newfd, F_NOCACHE, 1);
4267 rc = mdb_env_copyfd(env, newfd);
4270 if (!(env->me_flags & MDB_NOSUBDIR))
4272 if (newfd != INVALID_HANDLE_VALUE)
4273 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4280 mdb_env_close(MDB_env *env)
4287 VGMEMP_DESTROY(env);
4288 while ((dp = env->me_dpages) != NULL) {
4289 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4290 env->me_dpages = dp->mp_next;
4294 mdb_env_close0(env, 0);
4298 /** Compare two items pointing at aligned size_t's */
4300 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4302 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4303 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4306 /** Compare two items pointing at aligned int's */
4308 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4310 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4311 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4314 /** Compare two items pointing at ints of unknown alignment.
4315 * Nodes and keys are guaranteed to be 2-byte aligned.
4318 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4320 #if BYTE_ORDER == LITTLE_ENDIAN
4321 unsigned short *u, *c;
4324 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4325 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4328 } while(!x && u > (unsigned short *)a->mv_data);
4331 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4335 /** Compare two items lexically */
4337 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4344 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4350 diff = memcmp(a->mv_data, b->mv_data, len);
4351 return diff ? diff : len_diff<0 ? -1 : len_diff;
4354 /** Compare two items in reverse byte order */
4356 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4358 const unsigned char *p1, *p2, *p1_lim;
4362 p1_lim = (const unsigned char *)a->mv_data;
4363 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4364 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4366 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4372 while (p1 > p1_lim) {
4373 diff = *--p1 - *--p2;
4377 return len_diff<0 ? -1 : len_diff;
4380 /** Search for key within a page, using binary search.
4381 * Returns the smallest entry larger or equal to the key.
4382 * If exactp is non-null, stores whether the found entry was an exact match
4383 * in *exactp (1 or 0).
4384 * Updates the cursor index with the index of the found entry.
4385 * If no entry larger or equal to the key is found, returns NULL.
4388 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4390 unsigned int i = 0, nkeys;
4393 MDB_page *mp = mc->mc_pg[mc->mc_top];
4394 MDB_node *node = NULL;
4399 nkeys = NUMKEYS(mp);
4404 COPY_PGNO(pgno, mp->mp_pgno);
4405 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4406 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4413 low = IS_LEAF(mp) ? 0 : 1;
4415 cmp = mc->mc_dbx->md_cmp;
4417 /* Branch pages have no data, so if using integer keys,
4418 * alignment is guaranteed. Use faster mdb_cmp_int.
4420 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4421 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4428 nodekey.mv_size = mc->mc_db->md_pad;
4429 node = NODEPTR(mp, 0); /* fake */
4430 while (low <= high) {
4431 i = (low + high) >> 1;
4432 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4433 rc = cmp(key, &nodekey);
4434 DPRINTF(("found leaf index %u [%s], rc = %i",
4435 i, DKEY(&nodekey), rc));
4444 while (low <= high) {
4445 i = (low + high) >> 1;
4447 node = NODEPTR(mp, i);
4448 nodekey.mv_size = NODEKSZ(node);
4449 nodekey.mv_data = NODEKEY(node);
4451 rc = cmp(key, &nodekey);
4454 DPRINTF(("found leaf index %u [%s], rc = %i",
4455 i, DKEY(&nodekey), rc));
4457 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4458 i, DKEY(&nodekey), NODEPGNO(node), rc));
4469 if (rc > 0) { /* Found entry is less than the key. */
4470 i++; /* Skip to get the smallest entry larger than key. */
4472 node = NODEPTR(mp, i);
4475 *exactp = (rc == 0);
4476 /* store the key index */
4477 mc->mc_ki[mc->mc_top] = i;
4479 /* There is no entry larger or equal to the key. */
4482 /* nodeptr is fake for LEAF2 */
4488 mdb_cursor_adjust(MDB_cursor *mc, func)
4492 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4493 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4500 /** Pop a page off the top of the cursor's stack. */
4502 mdb_cursor_pop(MDB_cursor *mc)
4506 MDB_page *top = mc->mc_pg[mc->mc_top];
4512 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4513 mc->mc_dbi, (void *) mc));
4517 /** Push a page onto the top of the cursor's stack. */
4519 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4521 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4522 mc->mc_dbi, (void *) mc));
4524 if (mc->mc_snum >= CURSOR_STACK) {
4525 assert(mc->mc_snum < CURSOR_STACK);
4526 return MDB_CURSOR_FULL;
4529 mc->mc_top = mc->mc_snum++;
4530 mc->mc_pg[mc->mc_top] = mp;
4531 mc->mc_ki[mc->mc_top] = 0;
4536 /** Find the address of the page corresponding to a given page number.
4537 * @param[in] txn the transaction for this access.
4538 * @param[in] pgno the page number for the page to retrieve.
4539 * @param[out] ret address of a pointer where the page's address will be stored.
4540 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4541 * @return 0 on success, non-zero on failure.
4544 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4546 MDB_env *env = txn->mt_env;
4550 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4554 MDB_ID2L dl = tx2->mt_u.dirty_list;
4556 /* Spilled pages were dirtied in this txn and flushed
4557 * because the dirty list got full. Bring this page
4558 * back in from the map (but don't unspill it here,
4559 * leave that unless page_touch happens again).
4561 if (tx2->mt_spill_pgs) {
4562 MDB_ID pn = pgno << 1;
4563 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4564 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4565 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4570 unsigned x = mdb_mid2l_search(dl, pgno);
4571 if (x <= dl[0].mid && dl[x].mid == pgno) {
4577 } while ((tx2 = tx2->mt_parent) != NULL);
4580 if (pgno < txn->mt_next_pgno) {
4582 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4584 DPRINTF(("page %"Z"u not found", pgno));
4586 return MDB_PAGE_NOTFOUND;
4596 /** Search for the page a given key should be in.
4597 * Pushes parent pages on the cursor stack. This function continues a
4598 * search on a cursor that has already been initialized. (Usually by
4599 * #mdb_page_search() but also by #mdb_node_move().)
4600 * @param[in,out] mc the cursor for this operation.
4601 * @param[in] key the key to search for. If NULL, search for the lowest
4602 * page. (This is used by #mdb_cursor_first().)
4603 * @param[in] modify If true, visited pages are updated with new page numbers.
4604 * @return 0 on success, non-zero on failure.
4607 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4609 MDB_page *mp = mc->mc_pg[mc->mc_top];
4613 while (IS_BRANCH(mp)) {
4617 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4618 assert(NUMKEYS(mp) > 1);
4619 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4621 if (key == NULL) /* Initialize cursor to first page. */
4623 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4624 /* cursor to last page */
4628 node = mdb_node_search(mc, key, &exact);
4630 i = NUMKEYS(mp) - 1;
4632 i = mc->mc_ki[mc->mc_top];
4641 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4642 assert(i < NUMKEYS(mp));
4643 node = NODEPTR(mp, i);
4645 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4648 mc->mc_ki[mc->mc_top] = i;
4649 if ((rc = mdb_cursor_push(mc, mp)))
4653 if ((rc = mdb_page_touch(mc)) != 0)
4655 mp = mc->mc_pg[mc->mc_top];
4660 DPRINTF(("internal error, index points to a %02X page!?",
4662 return MDB_CORRUPTED;
4665 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4666 key ? DKEY(key) : NULL));
4667 mc->mc_flags |= C_INITIALIZED;
4668 mc->mc_flags &= ~C_EOF;
4673 /** Search for the lowest key under the current branch page.
4674 * This just bypasses a NUMKEYS check in the current page
4675 * before calling mdb_page_search_root(), because the callers
4676 * are all in situations where the current page is known to
4680 mdb_page_search_lowest(MDB_cursor *mc)
4682 MDB_page *mp = mc->mc_pg[mc->mc_top];
4683 MDB_node *node = NODEPTR(mp, 0);
4686 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4689 mc->mc_ki[mc->mc_top] = 0;
4690 if ((rc = mdb_cursor_push(mc, mp)))
4692 return mdb_page_search_root(mc, NULL, 0);
4695 /** Search for the page a given key should be in.
4696 * Pushes parent pages on the cursor stack. This function just sets up
4697 * the search; it finds the root page for \b mc's database and sets this
4698 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4699 * called to complete the search.
4700 * @param[in,out] mc the cursor for this operation.
4701 * @param[in] key the key to search for. If NULL, search for the lowest
4702 * page. (This is used by #mdb_cursor_first().)
4703 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4704 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4705 * @return 0 on success, non-zero on failure.
4708 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4713 /* Make sure the txn is still viable, then find the root from
4714 * the txn's db table.
4716 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4717 DPUTS("transaction has failed, must abort");
4720 /* Make sure we're using an up-to-date root */
4721 if (mc->mc_dbi > MAIN_DBI) {
4722 if ((*mc->mc_dbflag & DB_STALE) ||
4723 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4725 unsigned char dbflag = 0;
4726 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4727 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4730 if (*mc->mc_dbflag & DB_STALE) {
4734 MDB_node *leaf = mdb_node_search(&mc2,
4735 &mc->mc_dbx->md_name, &exact);
4737 return MDB_NOTFOUND;
4738 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4741 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4743 /* The txn may not know this DBI, or another process may
4744 * have dropped and recreated the DB with other flags.
4746 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4747 return MDB_INCOMPATIBLE;
4748 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4750 if (flags & MDB_PS_MODIFY)
4752 *mc->mc_dbflag &= ~DB_STALE;
4753 *mc->mc_dbflag |= dbflag;
4756 root = mc->mc_db->md_root;
4758 if (root == P_INVALID) { /* Tree is empty. */
4759 DPUTS("tree is empty");
4760 return MDB_NOTFOUND;
4765 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4766 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4772 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4773 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4775 if (flags & MDB_PS_MODIFY) {
4776 if ((rc = mdb_page_touch(mc)))
4780 if (flags & MDB_PS_ROOTONLY)
4783 return mdb_page_search_root(mc, key, flags);
4787 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4789 MDB_txn *txn = mc->mc_txn;
4790 pgno_t pg = mp->mp_pgno;
4791 unsigned x = 0, ovpages = mp->mp_pages;
4792 MDB_env *env = txn->mt_env;
4793 MDB_IDL sl = txn->mt_spill_pgs;
4794 MDB_ID pn = pg << 1;
4797 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4798 /* If the page is dirty or on the spill list we just acquired it,
4799 * so we should give it back to our current free list, if any.
4800 * Otherwise put it onto the list of pages we freed in this txn.
4802 * Won't create me_pghead: me_pglast must be inited along with it.
4803 * Unsupported in nested txns: They would need to hide the page
4804 * range in ancestor txns' dirty and spilled lists.
4806 if (env->me_pghead &&
4808 ((mp->mp_flags & P_DIRTY) ||
4809 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4813 MDB_ID2 *dl, ix, iy;
4814 rc = mdb_midl_need(&env->me_pghead, ovpages);
4817 if (!(mp->mp_flags & P_DIRTY)) {
4818 /* This page is no longer spilled */
4825 /* Remove from dirty list */
4826 dl = txn->mt_u.dirty_list;
4828 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4836 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4837 txn->mt_flags |= MDB_TXN_ERROR;
4838 return MDB_CORRUPTED;
4841 if (!(env->me_flags & MDB_WRITEMAP))
4842 mdb_dpage_free(env, mp);
4844 /* Insert in me_pghead */
4845 mop = env->me_pghead;
4846 j = mop[0] + ovpages;
4847 for (i = mop[0]; i && mop[i] < pg; i--)
4853 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4857 mc->mc_db->md_overflow_pages -= ovpages;
4861 /** Return the data associated with a given node.
4862 * @param[in] txn The transaction for this operation.
4863 * @param[in] leaf The node being read.
4864 * @param[out] data Updated to point to the node's data.
4865 * @return 0 on success, non-zero on failure.
4868 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4870 MDB_page *omp; /* overflow page */
4874 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4875 data->mv_size = NODEDSZ(leaf);
4876 data->mv_data = NODEDATA(leaf);
4880 /* Read overflow data.
4882 data->mv_size = NODEDSZ(leaf);
4883 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4884 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4885 DPRINTF(("read overflow page %"Z"u failed", pgno));
4888 data->mv_data = METADATA(omp);
4894 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4895 MDB_val *key, MDB_val *data)
4904 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4906 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4909 if (txn->mt_flags & MDB_TXN_ERROR)
4912 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4913 return MDB_BAD_VALSIZE;
4916 mdb_cursor_init(&mc, txn, dbi, &mx);
4917 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4920 /** Find a sibling for a page.
4921 * Replaces the page at the top of the cursor's stack with the
4922 * specified sibling, if one exists.
4923 * @param[in] mc The cursor for this operation.
4924 * @param[in] move_right Non-zero if the right sibling is requested,
4925 * otherwise the left sibling.
4926 * @return 0 on success, non-zero on failure.
4929 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4935 if (mc->mc_snum < 2) {
4936 return MDB_NOTFOUND; /* root has no siblings */
4940 DPRINTF(("parent page is page %"Z"u, index %u",
4941 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4943 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4944 : (mc->mc_ki[mc->mc_top] == 0)) {
4945 DPRINTF(("no more keys left, moving to %s sibling",
4946 move_right ? "right" : "left"));
4947 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4948 /* undo cursor_pop before returning */
4955 mc->mc_ki[mc->mc_top]++;
4957 mc->mc_ki[mc->mc_top]--;
4958 DPRINTF(("just moving to %s index key %u",
4959 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4961 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4963 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4964 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4967 mdb_cursor_push(mc, mp);
4969 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4974 /** Move the cursor to the next data item. */
4976 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4982 if (mc->mc_flags & C_EOF) {
4983 return MDB_NOTFOUND;
4986 assert(mc->mc_flags & C_INITIALIZED);
4988 mp = mc->mc_pg[mc->mc_top];
4990 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4991 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4992 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4993 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4994 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4995 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
4996 if (rc == MDB_SUCCESS)
4997 MDB_GET_KEY(leaf, key);
5002 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5003 if (op == MDB_NEXT_DUP)
5004 return MDB_NOTFOUND;
5008 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5010 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5011 DPUTS("=====> move to next sibling page");
5012 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5013 mc->mc_flags |= C_EOF;
5016 mp = mc->mc_pg[mc->mc_top];
5017 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5019 mc->mc_ki[mc->mc_top]++;
5021 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5022 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5025 key->mv_size = mc->mc_db->md_pad;
5026 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5030 assert(IS_LEAF(mp));
5031 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5033 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5034 mdb_xcursor_init1(mc, leaf);
5037 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5040 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5041 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5042 if (rc != MDB_SUCCESS)
5047 MDB_GET_KEY(leaf, key);
5051 /** Move the cursor to the previous data item. */
5053 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5059 assert(mc->mc_flags & C_INITIALIZED);
5061 mp = mc->mc_pg[mc->mc_top];
5063 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5064 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5065 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5066 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5067 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5068 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5069 if (rc == MDB_SUCCESS)
5070 MDB_GET_KEY(leaf, key);
5074 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5075 if (op == MDB_PREV_DUP)
5076 return MDB_NOTFOUND;
5081 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5083 if (mc->mc_ki[mc->mc_top] == 0) {
5084 DPUTS("=====> move to prev sibling page");
5085 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5088 mp = mc->mc_pg[mc->mc_top];
5089 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5090 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5092 mc->mc_ki[mc->mc_top]--;
5094 mc->mc_flags &= ~C_EOF;
5096 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5097 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5100 key->mv_size = mc->mc_db->md_pad;
5101 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5105 assert(IS_LEAF(mp));
5106 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5108 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5109 mdb_xcursor_init1(mc, leaf);
5112 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5115 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5116 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5117 if (rc != MDB_SUCCESS)
5122 MDB_GET_KEY(leaf, key);
5126 /** Set the cursor on a specific data item. */
5128 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5129 MDB_cursor_op op, int *exactp)
5133 MDB_node *leaf = NULL;
5138 assert(key->mv_size > 0);
5141 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5143 /* See if we're already on the right page */
5144 if (mc->mc_flags & C_INITIALIZED) {
5147 mp = mc->mc_pg[mc->mc_top];
5149 mc->mc_ki[mc->mc_top] = 0;
5150 return MDB_NOTFOUND;
5152 if (mp->mp_flags & P_LEAF2) {
5153 nodekey.mv_size = mc->mc_db->md_pad;
5154 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5156 leaf = NODEPTR(mp, 0);
5157 MDB_GET_KEY2(leaf, nodekey);
5159 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5161 /* Probably happens rarely, but first node on the page
5162 * was the one we wanted.
5164 mc->mc_ki[mc->mc_top] = 0;
5171 unsigned int nkeys = NUMKEYS(mp);
5173 if (mp->mp_flags & P_LEAF2) {
5174 nodekey.mv_data = LEAF2KEY(mp,
5175 nkeys-1, nodekey.mv_size);
5177 leaf = NODEPTR(mp, nkeys-1);
5178 MDB_GET_KEY2(leaf, nodekey);
5180 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5182 /* last node was the one we wanted */
5183 mc->mc_ki[mc->mc_top] = nkeys-1;
5189 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5190 /* This is definitely the right page, skip search_page */
5191 if (mp->mp_flags & P_LEAF2) {
5192 nodekey.mv_data = LEAF2KEY(mp,
5193 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5195 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5196 MDB_GET_KEY2(leaf, nodekey);
5198 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5200 /* current node was the one we wanted */
5210 /* If any parents have right-sibs, search.
5211 * Otherwise, there's nothing further.
5213 for (i=0; i<mc->mc_top; i++)
5215 NUMKEYS(mc->mc_pg[i])-1)
5217 if (i == mc->mc_top) {
5218 /* There are no other pages */
5219 mc->mc_ki[mc->mc_top] = nkeys;
5220 return MDB_NOTFOUND;
5224 /* There are no other pages */
5225 mc->mc_ki[mc->mc_top] = 0;
5226 return MDB_NOTFOUND;
5230 rc = mdb_page_search(mc, key, 0);
5231 if (rc != MDB_SUCCESS)
5234 mp = mc->mc_pg[mc->mc_top];
5235 assert(IS_LEAF(mp));
5238 leaf = mdb_node_search(mc, key, exactp);
5239 if (exactp != NULL && !*exactp) {
5240 /* MDB_SET specified and not an exact match. */
5241 return MDB_NOTFOUND;
5245 DPUTS("===> inexact leaf not found, goto sibling");
5246 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5247 return rc; /* no entries matched */
5248 mp = mc->mc_pg[mc->mc_top];
5249 assert(IS_LEAF(mp));
5250 leaf = NODEPTR(mp, 0);
5254 mc->mc_flags |= C_INITIALIZED;
5255 mc->mc_flags &= ~C_EOF;
5258 key->mv_size = mc->mc_db->md_pad;
5259 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5263 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5264 mdb_xcursor_init1(mc, leaf);
5267 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5268 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5269 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5272 if (op == MDB_GET_BOTH) {
5278 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5279 if (rc != MDB_SUCCESS)
5282 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5284 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5286 rc = mc->mc_dbx->md_dcmp(data, &d2);
5288 if (op == MDB_GET_BOTH || rc > 0)
5289 return MDB_NOTFOUND;
5294 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5295 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5300 /* The key already matches in all other cases */
5301 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5302 MDB_GET_KEY(leaf, key);
5303 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5308 /** Move the cursor to the first item in the database. */
5310 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5316 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5318 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5319 rc = mdb_page_search(mc, NULL, 0);
5320 if (rc != MDB_SUCCESS)
5323 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5325 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5326 mc->mc_flags |= C_INITIALIZED;
5327 mc->mc_flags &= ~C_EOF;
5329 mc->mc_ki[mc->mc_top] = 0;
5331 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5332 key->mv_size = mc->mc_db->md_pad;
5333 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5338 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5339 mdb_xcursor_init1(mc, leaf);
5340 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5344 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5348 MDB_GET_KEY(leaf, key);
5352 /** Move the cursor to the last item in the database. */
5354 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5360 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5362 if (!(mc->mc_flags & C_EOF)) {
5364 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5367 lkey.mv_size = MDB_MAXKEYSIZE+1;
5368 lkey.mv_data = NULL;
5369 rc = mdb_page_search(mc, &lkey, 0);
5370 if (rc != MDB_SUCCESS)
5373 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5376 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5377 mc->mc_flags |= C_INITIALIZED|C_EOF;
5378 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5380 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5381 key->mv_size = mc->mc_db->md_pad;
5382 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5387 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5388 mdb_xcursor_init1(mc, leaf);
5389 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5393 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5398 MDB_GET_KEY(leaf, key);
5403 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5408 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5412 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5416 case MDB_GET_CURRENT:
5417 if (!(mc->mc_flags & C_INITIALIZED)) {
5420 MDB_page *mp = mc->mc_pg[mc->mc_top];
5422 mc->mc_ki[mc->mc_top] = 0;
5428 key->mv_size = mc->mc_db->md_pad;
5429 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5431 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5432 MDB_GET_KEY(leaf, key);
5434 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5435 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5437 rc = mdb_node_read(mc->mc_txn, leaf, data);
5444 case MDB_GET_BOTH_RANGE:
5449 if (mc->mc_xcursor == NULL) {
5450 rc = MDB_INCOMPATIBLE;
5459 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5460 rc = MDB_BAD_VALSIZE;
5461 } else if (op == MDB_SET_RANGE)
5462 rc = mdb_cursor_set(mc, key, data, op, NULL);
5464 rc = mdb_cursor_set(mc, key, data, op, &exact);
5466 case MDB_GET_MULTIPLE:
5467 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5471 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5472 rc = MDB_INCOMPATIBLE;
5476 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5477 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5480 case MDB_NEXT_MULTIPLE:
5485 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5486 rc = MDB_INCOMPATIBLE;
5489 if (!(mc->mc_flags & C_INITIALIZED))
5490 rc = mdb_cursor_first(mc, key, data);
5492 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5493 if (rc == MDB_SUCCESS) {
5494 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5497 mx = &mc->mc_xcursor->mx_cursor;
5498 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5500 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5501 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5509 case MDB_NEXT_NODUP:
5510 if (!(mc->mc_flags & C_INITIALIZED))
5511 rc = mdb_cursor_first(mc, key, data);
5513 rc = mdb_cursor_next(mc, key, data, op);
5517 case MDB_PREV_NODUP:
5518 if (!(mc->mc_flags & C_INITIALIZED)) {
5519 rc = mdb_cursor_last(mc, key, data);
5522 mc->mc_flags |= C_INITIALIZED;
5523 mc->mc_ki[mc->mc_top]++;
5525 rc = mdb_cursor_prev(mc, key, data, op);
5528 rc = mdb_cursor_first(mc, key, data);
5531 mfunc = mdb_cursor_first;
5533 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5537 if (mc->mc_xcursor == NULL) {
5538 rc = MDB_INCOMPATIBLE;
5541 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5545 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5548 rc = mdb_cursor_last(mc, key, data);
5551 mfunc = mdb_cursor_last;
5554 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5562 /** Touch all the pages in the cursor stack.
5563 * Makes sure all the pages are writable, before attempting a write operation.
5564 * @param[in] mc The cursor to operate on.
5567 mdb_cursor_touch(MDB_cursor *mc)
5571 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5574 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5575 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5578 *mc->mc_dbflag |= DB_DIRTY;
5580 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5581 rc = mdb_page_touch(mc);
5585 mc->mc_top = mc->mc_snum-1;
5589 /** Do not spill pages to disk if txn is getting full, may fail instead */
5590 #define MDB_NOSPILL 0x8000
5593 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5596 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5597 MDB_node *leaf = NULL;
5598 MDB_val xdata, *rdata, dkey;
5601 int do_sub = 0, insert = 0;
5602 unsigned int mcount = 0, dcount = 0, nospill;
5606 char dbuf[MDB_MAXKEYSIZE+1];
5607 unsigned int nflags;
5610 /* Check this first so counter will always be zero on any
5613 if (flags & MDB_MULTIPLE) {
5614 dcount = data[1].mv_size;
5615 data[1].mv_size = 0;
5616 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5617 return MDB_INCOMPATIBLE;
5620 nospill = flags & MDB_NOSPILL;
5621 flags &= ~MDB_NOSPILL;
5623 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5624 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5626 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5627 return MDB_BAD_VALSIZE;
5629 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5630 return MDB_BAD_VALSIZE;
5632 #if SIZE_MAX > MAXDATASIZE
5633 if (data->mv_size > MAXDATASIZE)
5634 return MDB_BAD_VALSIZE;
5637 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5638 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5642 if (flags == MDB_CURRENT) {
5643 if (!(mc->mc_flags & C_INITIALIZED))
5646 } else if (mc->mc_db->md_root == P_INVALID) {
5647 /* new database, cursor has nothing to point to */
5649 mc->mc_flags &= ~C_INITIALIZED;
5654 if (flags & MDB_APPEND) {
5656 rc = mdb_cursor_last(mc, &k2, &d2);
5658 rc = mc->mc_dbx->md_cmp(key, &k2);
5661 mc->mc_ki[mc->mc_top]++;
5663 /* new key is <= last key */
5668 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5670 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5671 DPRINTF(("duplicate key [%s]", DKEY(key)));
5673 return MDB_KEYEXIST;
5675 if (rc && rc != MDB_NOTFOUND)
5679 /* Cursor is positioned, check for room in the dirty list */
5681 if (flags & MDB_MULTIPLE) {
5683 xdata.mv_size = data->mv_size * dcount;
5687 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5691 if (rc == MDB_NO_ROOT) {
5693 /* new database, write a root leaf page */
5694 DPUTS("allocating new root leaf page");
5695 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5698 mdb_cursor_push(mc, np);
5699 mc->mc_db->md_root = np->mp_pgno;
5700 mc->mc_db->md_depth++;
5701 *mc->mc_dbflag |= DB_DIRTY;
5702 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5704 np->mp_flags |= P_LEAF2;
5705 mc->mc_flags |= C_INITIALIZED;
5707 /* make sure all cursor pages are writable */
5708 rc2 = mdb_cursor_touch(mc);
5713 /* The key already exists */
5714 if (rc == MDB_SUCCESS) {
5715 /* there's only a key anyway, so this is a no-op */
5716 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5717 unsigned int ksize = mc->mc_db->md_pad;
5718 if (key->mv_size != ksize)
5719 return MDB_BAD_VALSIZE;
5720 if (flags == MDB_CURRENT) {
5721 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5722 memcpy(ptr, key->mv_data, ksize);
5727 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5730 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5731 /* Was a single item before, must convert now */
5733 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5734 /* Just overwrite the current item */
5735 if (flags == MDB_CURRENT)
5738 dkey.mv_size = NODEDSZ(leaf);
5739 dkey.mv_data = NODEDATA(leaf);
5740 #if UINT_MAX < SIZE_MAX
5741 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5742 #ifdef MISALIGNED_OK
5743 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5745 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5748 /* if data matches, skip it */
5749 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5750 if (flags & MDB_NODUPDATA)
5752 else if (flags & MDB_MULTIPLE)
5759 /* create a fake page for the dup items */
5760 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5761 dkey.mv_data = dbuf;
5762 fp = (MDB_page *)&pbuf;
5763 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5764 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5765 fp->mp_lower = PAGEHDRSZ;
5766 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5767 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5768 fp->mp_flags |= P_LEAF2;
5769 fp->mp_pad = data->mv_size;
5770 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5772 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5773 (dkey.mv_size & 1) + (data->mv_size & 1);
5775 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5778 xdata.mv_size = fp->mp_upper;
5783 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5784 /* See if we need to convert from fake page to subDB */
5786 unsigned int offset;
5790 fp = NODEDATA(leaf);
5791 if (flags == MDB_CURRENT) {
5793 fp->mp_flags |= P_DIRTY;
5794 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5795 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5799 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5800 offset = fp->mp_pad;
5801 if (SIZELEFT(fp) >= offset)
5803 offset *= 4; /* space for 4 more */
5805 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5807 offset += offset & 1;
5808 fp_flags = fp->mp_flags;
5809 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5810 offset >= mc->mc_txn->mt_env->me_nodemax) {
5811 /* yes, convert it */
5813 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5814 dummy.md_pad = fp->mp_pad;
5815 dummy.md_flags = MDB_DUPFIXED;
5816 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5817 dummy.md_flags |= MDB_INTEGERKEY;
5820 dummy.md_branch_pages = 0;
5821 dummy.md_leaf_pages = 1;
5822 dummy.md_overflow_pages = 0;
5823 dummy.md_entries = NUMKEYS(fp);
5825 xdata.mv_size = sizeof(MDB_db);
5826 xdata.mv_data = &dummy;
5827 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5829 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5830 flags |= F_DUPDATA|F_SUBDATA;
5831 dummy.md_root = mp->mp_pgno;
5832 fp_flags &= ~P_SUBP;
5834 /* no, just grow it */
5836 xdata.mv_size = NODEDSZ(leaf) + offset;
5837 xdata.mv_data = &pbuf;
5838 mp = (MDB_page *)&pbuf;
5839 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5842 mp->mp_flags = fp_flags | P_DIRTY;
5843 mp->mp_pad = fp->mp_pad;
5844 mp->mp_lower = fp->mp_lower;
5845 mp->mp_upper = fp->mp_upper + offset;
5847 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5849 nsize = NODEDSZ(leaf) - fp->mp_upper;
5850 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5851 for (i=0; i<NUMKEYS(fp); i++)
5852 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5854 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5858 /* data is on sub-DB, just store it */
5859 flags |= F_DUPDATA|F_SUBDATA;
5863 /* overflow page overwrites need special handling */
5864 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5867 unsigned psize = mc->mc_txn->mt_env->me_psize;
5868 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5870 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5871 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5873 ovpages = omp->mp_pages;
5875 /* Is the ov page large enough? */
5876 if (ovpages >= dpages) {
5877 if (!(omp->mp_flags & P_DIRTY) &&
5878 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5880 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5883 level = 0; /* dirty in this txn or clean */
5886 if (omp->mp_flags & P_DIRTY) {
5887 /* yes, overwrite it. Note in this case we don't
5888 * bother to try shrinking the page if the new data
5889 * is smaller than the overflow threshold.
5892 /* It is writable only in a parent txn */
5893 size_t sz = (size_t) psize * ovpages, off;
5894 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5900 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5901 if (!(flags & MDB_RESERVE)) {
5902 /* Copy end of page, adjusting alignment so
5903 * compiler may copy words instead of bytes.
5905 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5906 memcpy((size_t *)((char *)np + off),
5907 (size_t *)((char *)omp + off), sz - off);
5910 memcpy(np, omp, sz); /* Copy beginning of page */
5913 SETDSZ(leaf, data->mv_size);
5914 if (F_ISSET(flags, MDB_RESERVE))
5915 data->mv_data = METADATA(omp);
5917 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5921 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5923 } else if (NODEDSZ(leaf) == data->mv_size) {
5924 /* same size, just replace it. Note that we could
5925 * also reuse this node if the new data is smaller,
5926 * but instead we opt to shrink the node in that case.
5928 if (F_ISSET(flags, MDB_RESERVE))
5929 data->mv_data = NODEDATA(leaf);
5930 else if (data->mv_size)
5931 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5933 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5936 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5937 mc->mc_db->md_entries--;
5939 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5946 nflags = flags & NODE_ADD_FLAGS;
5947 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5948 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5949 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5950 nflags &= ~MDB_APPEND;
5952 nflags |= MDB_SPLIT_REPLACE;
5953 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5955 /* There is room already in this leaf page. */
5956 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5957 if (rc == 0 && !do_sub && insert) {
5958 /* Adjust other cursors pointing to mp */
5959 MDB_cursor *m2, *m3;
5960 MDB_dbi dbi = mc->mc_dbi;
5961 unsigned i = mc->mc_top;
5962 MDB_page *mp = mc->mc_pg[i];
5964 if (mc->mc_flags & C_SUB)
5967 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5968 if (mc->mc_flags & C_SUB)
5969 m3 = &m2->mc_xcursor->mx_cursor;
5972 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5973 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5980 if (rc != MDB_SUCCESS)
5981 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5983 /* Now store the actual data in the child DB. Note that we're
5984 * storing the user data in the keys field, so there are strict
5985 * size limits on dupdata. The actual data fields of the child
5986 * DB are all zero size.
5993 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5994 if (flags & MDB_CURRENT) {
5995 xflags = MDB_CURRENT|MDB_NOSPILL;
5997 mdb_xcursor_init1(mc, leaf);
5998 xflags = (flags & MDB_NODUPDATA) ?
5999 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6001 /* converted, write the original data first */
6003 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6007 /* Adjust other cursors pointing to mp */
6009 unsigned i = mc->mc_top;
6010 MDB_page *mp = mc->mc_pg[i];
6012 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6013 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6014 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6015 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6016 mdb_xcursor_init1(m2, leaf);
6020 /* we've done our job */
6023 if (flags & MDB_APPENDDUP)
6024 xflags |= MDB_APPEND;
6025 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6026 if (flags & F_SUBDATA) {
6027 void *db = NODEDATA(leaf);
6028 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6031 /* sub-writes might have failed so check rc again.
6032 * Don't increment count if we just replaced an existing item.
6034 if (!rc && !(flags & MDB_CURRENT))
6035 mc->mc_db->md_entries++;
6036 if (flags & MDB_MULTIPLE) {
6040 /* let caller know how many succeeded, if any */
6041 data[1].mv_size = mcount;
6042 if (mcount < dcount) {
6043 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6044 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6051 /* If we succeeded and the key didn't exist before, make sure
6052 * the cursor is marked valid.
6055 mc->mc_flags |= C_INITIALIZED;
6060 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6065 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6066 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6068 if (!(mc->mc_flags & C_INITIALIZED))
6071 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6073 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
6075 rc = mdb_cursor_touch(mc);
6079 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6081 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6082 if (!(flags & MDB_NODUPDATA)) {
6083 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6084 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6086 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6087 /* If sub-DB still has entries, we're done */
6088 if (mc->mc_xcursor->mx_db.md_entries) {
6089 if (leaf->mn_flags & F_SUBDATA) {
6090 /* update subDB info */
6091 void *db = NODEDATA(leaf);
6092 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6095 /* shrink fake page */
6096 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6097 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6098 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6099 /* fix other sub-DB cursors pointed at this fake page */
6100 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6101 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6102 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6103 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6104 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6107 mc->mc_db->md_entries--;
6110 /* otherwise fall thru and delete the sub-DB */
6113 if (leaf->mn_flags & F_SUBDATA) {
6114 /* add all the child DB's pages to the free list */
6115 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6116 if (rc == MDB_SUCCESS) {
6117 mc->mc_db->md_entries -=
6118 mc->mc_xcursor->mx_db.md_entries;
6123 return mdb_cursor_del0(mc, leaf);
6126 /** Allocate and initialize new pages for a database.
6127 * @param[in] mc a cursor on the database being added to.
6128 * @param[in] flags flags defining what type of page is being allocated.
6129 * @param[in] num the number of pages to allocate. This is usually 1,
6130 * unless allocating overflow pages for a large record.
6131 * @param[out] mp Address of a page, or NULL on failure.
6132 * @return 0 on success, non-zero on failure.
6135 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6140 if ((rc = mdb_page_alloc(mc, num, &np)))
6142 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6143 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6144 np->mp_flags = flags | P_DIRTY;
6145 np->mp_lower = PAGEHDRSZ;
6146 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6149 mc->mc_db->md_branch_pages++;
6150 else if (IS_LEAF(np))
6151 mc->mc_db->md_leaf_pages++;
6152 else if (IS_OVERFLOW(np)) {
6153 mc->mc_db->md_overflow_pages += num;
6161 /** Calculate the size of a leaf node.
6162 * The size depends on the environment's page size; if a data item
6163 * is too large it will be put onto an overflow page and the node
6164 * size will only include the key and not the data. Sizes are always
6165 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6166 * of the #MDB_node headers.
6167 * @param[in] env The environment handle.
6168 * @param[in] key The key for the node.
6169 * @param[in] data The data for the node.
6170 * @return The number of bytes needed to store the node.
6173 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6177 sz = LEAFSIZE(key, data);
6178 if (sz >= env->me_nodemax) {
6179 /* put on overflow page */
6180 sz -= data->mv_size - sizeof(pgno_t);
6184 return sz + sizeof(indx_t);
6187 /** Calculate the size of a branch node.
6188 * The size should depend on the environment's page size but since
6189 * we currently don't support spilling large keys onto overflow
6190 * pages, it's simply the size of the #MDB_node header plus the
6191 * size of the key. Sizes are always rounded up to an even number
6192 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6193 * @param[in] env The environment handle.
6194 * @param[in] key The key for the node.
6195 * @return The number of bytes needed to store the node.
6198 mdb_branch_size(MDB_env *env, MDB_val *key)
6203 if (sz >= env->me_nodemax) {
6204 /* put on overflow page */
6205 /* not implemented */
6206 /* sz -= key->size - sizeof(pgno_t); */
6209 return sz + sizeof(indx_t);
6212 /** Add a node to the page pointed to by the cursor.
6213 * @param[in] mc The cursor for this operation.
6214 * @param[in] indx The index on the page where the new node should be added.
6215 * @param[in] key The key for the new node.
6216 * @param[in] data The data for the new node, if any.
6217 * @param[in] pgno The page number, if adding a branch node.
6218 * @param[in] flags Flags for the node.
6219 * @return 0 on success, non-zero on failure. Possible errors are:
6221 * <li>ENOMEM - failed to allocate overflow pages for the node.
6222 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6223 * should never happen since all callers already calculate the
6224 * page's free space before calling this function.
6228 mdb_node_add(MDB_cursor *mc, indx_t indx,
6229 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6232 size_t node_size = NODESIZE;
6235 MDB_page *mp = mc->mc_pg[mc->mc_top];
6236 MDB_page *ofp = NULL; /* overflow page */
6239 assert(mp->mp_upper >= mp->mp_lower);
6241 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6242 IS_LEAF(mp) ? "leaf" : "branch",
6243 IS_SUBP(mp) ? "sub-" : "",
6244 mp->mp_pgno, indx, data ? data->mv_size : 0,
6245 key ? key->mv_size : 0, key ? DKEY(key) : NULL));
6248 /* Move higher keys up one slot. */
6249 int ksize = mc->mc_db->md_pad, dif;
6250 char *ptr = LEAF2KEY(mp, indx, ksize);
6251 dif = NUMKEYS(mp) - indx;
6253 memmove(ptr+ksize, ptr, dif*ksize);
6254 /* insert new key */
6255 memcpy(ptr, key->mv_data, ksize);
6257 /* Just using these for counting */
6258 mp->mp_lower += sizeof(indx_t);
6259 mp->mp_upper -= ksize - sizeof(indx_t);
6264 node_size += key->mv_size;
6268 if (F_ISSET(flags, F_BIGDATA)) {
6269 /* Data already on overflow page. */
6270 node_size += sizeof(pgno_t);
6271 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6272 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6274 /* Put data on overflow page. */
6275 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6276 data->mv_size, node_size+data->mv_size));
6277 node_size += sizeof(pgno_t);
6278 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6280 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6283 node_size += data->mv_size;
6286 node_size += node_size & 1;
6288 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6289 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6290 mp->mp_pgno, NUMKEYS(mp)));
6291 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6292 mp->mp_upper - mp->mp_lower));
6293 DPRINTF(("node size = %"Z"u", node_size));
6294 return MDB_PAGE_FULL;
6297 /* Move higher pointers up one slot. */
6298 for (i = NUMKEYS(mp); i > indx; i--)
6299 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6301 /* Adjust free space offsets. */
6302 ofs = mp->mp_upper - node_size;
6303 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6304 mp->mp_ptrs[indx] = ofs;
6306 mp->mp_lower += sizeof(indx_t);
6308 /* Write the node data. */
6309 node = NODEPTR(mp, indx);
6310 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6311 node->mn_flags = flags;
6313 SETDSZ(node,data->mv_size);
6318 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6323 if (F_ISSET(flags, F_BIGDATA))
6324 memcpy(node->mn_data + key->mv_size, data->mv_data,
6326 else if (F_ISSET(flags, MDB_RESERVE))
6327 data->mv_data = node->mn_data + key->mv_size;
6329 memcpy(node->mn_data + key->mv_size, data->mv_data,
6332 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6334 if (F_ISSET(flags, MDB_RESERVE))
6335 data->mv_data = METADATA(ofp);
6337 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6344 /** Delete the specified node from a page.
6345 * @param[in] mp The page to operate on.
6346 * @param[in] indx The index of the node to delete.
6347 * @param[in] ksize The size of a node. Only used if the page is
6348 * part of a #MDB_DUPFIXED database.
6351 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6354 indx_t i, j, numkeys, ptr;
6361 COPY_PGNO(pgno, mp->mp_pgno);
6362 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6363 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6366 assert(indx < NUMKEYS(mp));
6369 int x = NUMKEYS(mp) - 1 - indx;
6370 base = LEAF2KEY(mp, indx, ksize);
6372 memmove(base, base + ksize, x * ksize);
6373 mp->mp_lower -= sizeof(indx_t);
6374 mp->mp_upper += ksize - sizeof(indx_t);
6378 node = NODEPTR(mp, indx);
6379 sz = NODESIZE + node->mn_ksize;
6381 if (F_ISSET(node->mn_flags, F_BIGDATA))
6382 sz += sizeof(pgno_t);
6384 sz += NODEDSZ(node);
6388 ptr = mp->mp_ptrs[indx];
6389 numkeys = NUMKEYS(mp);
6390 for (i = j = 0; i < numkeys; i++) {
6392 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6393 if (mp->mp_ptrs[i] < ptr)
6394 mp->mp_ptrs[j] += sz;
6399 base = (char *)mp + mp->mp_upper;
6400 memmove(base + sz, base, ptr - mp->mp_upper);
6402 mp->mp_lower -= sizeof(indx_t);
6406 /** Compact the main page after deleting a node on a subpage.
6407 * @param[in] mp The main page to operate on.
6408 * @param[in] indx The index of the subpage on the main page.
6411 mdb_node_shrink(MDB_page *mp, indx_t indx)
6418 indx_t i, numkeys, ptr;
6420 node = NODEPTR(mp, indx);
6421 sp = (MDB_page *)NODEDATA(node);
6422 osize = NODEDSZ(node);
6424 delta = sp->mp_upper - sp->mp_lower;
6425 SETDSZ(node, osize - delta);
6426 xp = (MDB_page *)((char *)sp + delta);
6428 /* shift subpage upward */
6430 nsize = NUMKEYS(sp) * sp->mp_pad;
6431 memmove(METADATA(xp), METADATA(sp), nsize);
6434 nsize = osize - sp->mp_upper;
6435 numkeys = NUMKEYS(sp);
6436 for (i=numkeys-1; i>=0; i--)
6437 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6439 xp->mp_upper = sp->mp_lower;
6440 xp->mp_lower = sp->mp_lower;
6441 xp->mp_flags = sp->mp_flags;
6442 xp->mp_pad = sp->mp_pad;
6443 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6445 /* shift lower nodes upward */
6446 ptr = mp->mp_ptrs[indx];
6447 numkeys = NUMKEYS(mp);
6448 for (i = 0; i < numkeys; i++) {
6449 if (mp->mp_ptrs[i] <= ptr)
6450 mp->mp_ptrs[i] += delta;
6453 base = (char *)mp + mp->mp_upper;
6454 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6455 mp->mp_upper += delta;
6458 /** Initial setup of a sorted-dups cursor.
6459 * Sorted duplicates are implemented as a sub-database for the given key.
6460 * The duplicate data items are actually keys of the sub-database.
6461 * Operations on the duplicate data items are performed using a sub-cursor
6462 * initialized when the sub-database is first accessed. This function does
6463 * the preliminary setup of the sub-cursor, filling in the fields that
6464 * depend only on the parent DB.
6465 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6468 mdb_xcursor_init0(MDB_cursor *mc)
6470 MDB_xcursor *mx = mc->mc_xcursor;
6472 mx->mx_cursor.mc_xcursor = NULL;
6473 mx->mx_cursor.mc_txn = mc->mc_txn;
6474 mx->mx_cursor.mc_db = &mx->mx_db;
6475 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6476 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6477 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6478 mx->mx_cursor.mc_snum = 0;
6479 mx->mx_cursor.mc_top = 0;
6480 mx->mx_cursor.mc_flags = C_SUB;
6481 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6482 mx->mx_dbx.md_dcmp = NULL;
6483 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6486 /** Final setup of a sorted-dups cursor.
6487 * Sets up the fields that depend on the data from the main cursor.
6488 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6489 * @param[in] node The data containing the #MDB_db record for the
6490 * sorted-dup database.
6493 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6495 MDB_xcursor *mx = mc->mc_xcursor;
6497 if (node->mn_flags & F_SUBDATA) {
6498 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6499 mx->mx_cursor.mc_pg[0] = 0;
6500 mx->mx_cursor.mc_snum = 0;
6501 mx->mx_cursor.mc_flags = C_SUB;
6503 MDB_page *fp = NODEDATA(node);
6504 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6505 mx->mx_db.md_flags = 0;
6506 mx->mx_db.md_depth = 1;
6507 mx->mx_db.md_branch_pages = 0;
6508 mx->mx_db.md_leaf_pages = 1;
6509 mx->mx_db.md_overflow_pages = 0;
6510 mx->mx_db.md_entries = NUMKEYS(fp);
6511 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6512 mx->mx_cursor.mc_snum = 1;
6513 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6514 mx->mx_cursor.mc_top = 0;
6515 mx->mx_cursor.mc_pg[0] = fp;
6516 mx->mx_cursor.mc_ki[0] = 0;
6517 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6518 mx->mx_db.md_flags = MDB_DUPFIXED;
6519 mx->mx_db.md_pad = fp->mp_pad;
6520 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6521 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6524 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6525 mx->mx_db.md_root));
6526 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6528 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6529 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6530 #if UINT_MAX < SIZE_MAX
6531 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6532 #ifdef MISALIGNED_OK
6533 mx->mx_dbx.md_cmp = mdb_cmp_long;
6535 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6540 /** Initialize a cursor for a given transaction and database. */
6542 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6545 mc->mc_backup = NULL;
6548 mc->mc_db = &txn->mt_dbs[dbi];
6549 mc->mc_dbx = &txn->mt_dbxs[dbi];
6550 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6555 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6557 mc->mc_xcursor = mx;
6558 mdb_xcursor_init0(mc);
6560 mc->mc_xcursor = NULL;
6562 if (*mc->mc_dbflag & DB_STALE) {
6563 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6568 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6571 size_t size = sizeof(MDB_cursor);
6573 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6576 if (txn->mt_flags & MDB_TXN_ERROR)
6579 /* Allow read access to the freelist */
6580 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6583 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6584 size += sizeof(MDB_xcursor);
6586 if ((mc = malloc(size)) != NULL) {
6587 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6588 if (txn->mt_cursors) {
6589 mc->mc_next = txn->mt_cursors[dbi];
6590 txn->mt_cursors[dbi] = mc;
6591 mc->mc_flags |= C_UNTRACK;
6603 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6605 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6608 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6611 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6615 /* Return the count of duplicate data items for the current key */
6617 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6621 if (mc == NULL || countp == NULL)
6624 if (mc->mc_xcursor == NULL)
6625 return MDB_INCOMPATIBLE;
6627 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6628 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6631 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6634 *countp = mc->mc_xcursor->mx_db.md_entries;
6640 mdb_cursor_close(MDB_cursor *mc)
6642 if (mc && !mc->mc_backup) {
6643 /* remove from txn, if tracked */
6644 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6645 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6646 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6648 *prev = mc->mc_next;
6655 mdb_cursor_txn(MDB_cursor *mc)
6657 if (!mc) return NULL;
6662 mdb_cursor_dbi(MDB_cursor *mc)
6668 /** Replace the key for a node with a new key.
6669 * @param[in] mc Cursor pointing to the node to operate on.
6670 * @param[in] key The new key to use.
6671 * @return 0 on success, non-zero on failure.
6674 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6681 indx_t ptr, i, numkeys, indx;
6684 indx = mc->mc_ki[mc->mc_top];
6685 mp = mc->mc_pg[mc->mc_top];
6686 node = NODEPTR(mp, indx);
6687 ptr = mp->mp_ptrs[indx];
6691 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6692 k2.mv_data = NODEKEY(node);
6693 k2.mv_size = node->mn_ksize;
6694 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6696 mdb_dkey(&k2, kbuf2),
6702 delta0 = delta = key->mv_size - node->mn_ksize;
6704 /* Must be 2-byte aligned. If new key is
6705 * shorter by 1, the shift will be skipped.
6707 delta += (delta & 1);
6709 if (delta > 0 && SIZELEFT(mp) < delta) {
6711 /* not enough space left, do a delete and split */
6712 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6713 pgno = NODEPGNO(node);
6714 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6715 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6718 numkeys = NUMKEYS(mp);
6719 for (i = 0; i < numkeys; i++) {
6720 if (mp->mp_ptrs[i] <= ptr)
6721 mp->mp_ptrs[i] -= delta;
6724 base = (char *)mp + mp->mp_upper;
6725 len = ptr - mp->mp_upper + NODESIZE;
6726 memmove(base - delta, base, len);
6727 mp->mp_upper -= delta;
6729 node = NODEPTR(mp, indx);
6732 /* But even if no shift was needed, update ksize */
6734 node->mn_ksize = key->mv_size;
6737 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6743 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6745 /** Move a node from csrc to cdst.
6748 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6755 unsigned short flags;
6759 /* Mark src and dst as dirty. */
6760 if ((rc = mdb_page_touch(csrc)) ||
6761 (rc = mdb_page_touch(cdst)))
6764 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6765 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6766 key.mv_size = csrc->mc_db->md_pad;
6767 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6769 data.mv_data = NULL;
6773 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6774 assert(!((long)srcnode&1));
6775 srcpg = NODEPGNO(srcnode);
6776 flags = srcnode->mn_flags;
6777 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6778 unsigned int snum = csrc->mc_snum;
6780 /* must find the lowest key below src */
6781 mdb_page_search_lowest(csrc);
6782 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6783 key.mv_size = csrc->mc_db->md_pad;
6784 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6786 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6787 key.mv_size = NODEKSZ(s2);
6788 key.mv_data = NODEKEY(s2);
6790 csrc->mc_snum = snum--;
6791 csrc->mc_top = snum;
6793 key.mv_size = NODEKSZ(srcnode);
6794 key.mv_data = NODEKEY(srcnode);
6796 data.mv_size = NODEDSZ(srcnode);
6797 data.mv_data = NODEDATA(srcnode);
6799 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6800 unsigned int snum = cdst->mc_snum;
6803 /* must find the lowest key below dst */
6804 mdb_page_search_lowest(cdst);
6805 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6806 bkey.mv_size = cdst->mc_db->md_pad;
6807 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6809 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6810 bkey.mv_size = NODEKSZ(s2);
6811 bkey.mv_data = NODEKEY(s2);
6813 cdst->mc_snum = snum--;
6814 cdst->mc_top = snum;
6815 mdb_cursor_copy(cdst, &mn);
6817 rc = mdb_update_key(&mn, &bkey);
6822 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6823 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6824 csrc->mc_ki[csrc->mc_top],
6826 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6827 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6829 /* Add the node to the destination page.
6831 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6832 if (rc != MDB_SUCCESS)
6835 /* Delete the node from the source page.
6837 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6840 /* Adjust other cursors pointing to mp */
6841 MDB_cursor *m2, *m3;
6842 MDB_dbi dbi = csrc->mc_dbi;
6843 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6845 if (csrc->mc_flags & C_SUB)
6848 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6849 if (csrc->mc_flags & C_SUB)
6850 m3 = &m2->mc_xcursor->mx_cursor;
6853 if (m3 == csrc) continue;
6854 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6855 csrc->mc_ki[csrc->mc_top]) {
6856 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6857 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6862 /* Update the parent separators.
6864 if (csrc->mc_ki[csrc->mc_top] == 0) {
6865 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6866 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6867 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6869 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6870 key.mv_size = NODEKSZ(srcnode);
6871 key.mv_data = NODEKEY(srcnode);
6873 DPRINTF(("update separator for source page %"Z"u to [%s]",
6874 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6875 mdb_cursor_copy(csrc, &mn);
6878 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6881 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6883 indx_t ix = csrc->mc_ki[csrc->mc_top];
6884 nullkey.mv_size = 0;
6885 csrc->mc_ki[csrc->mc_top] = 0;
6886 rc = mdb_update_key(csrc, &nullkey);
6887 csrc->mc_ki[csrc->mc_top] = ix;
6888 assert(rc == MDB_SUCCESS);
6892 if (cdst->mc_ki[cdst->mc_top] == 0) {
6893 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6894 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6895 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6897 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6898 key.mv_size = NODEKSZ(srcnode);
6899 key.mv_data = NODEKEY(srcnode);
6901 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6902 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6903 mdb_cursor_copy(cdst, &mn);
6906 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6909 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6911 indx_t ix = cdst->mc_ki[cdst->mc_top];
6912 nullkey.mv_size = 0;
6913 cdst->mc_ki[cdst->mc_top] = 0;
6914 rc = mdb_update_key(cdst, &nullkey);
6915 cdst->mc_ki[cdst->mc_top] = ix;
6916 assert(rc == MDB_SUCCESS);
6923 /** Merge one page into another.
6924 * The nodes from the page pointed to by \b csrc will
6925 * be copied to the page pointed to by \b cdst and then
6926 * the \b csrc page will be freed.
6927 * @param[in] csrc Cursor pointing to the source page.
6928 * @param[in] cdst Cursor pointing to the destination page.
6931 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6939 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6940 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6942 assert(csrc->mc_snum > 1); /* can't merge root page */
6943 assert(cdst->mc_snum > 1);
6945 /* Mark dst as dirty. */
6946 if ((rc = mdb_page_touch(cdst)))
6949 /* Move all nodes from src to dst.
6951 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6952 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6953 key.mv_size = csrc->mc_db->md_pad;
6954 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6955 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6956 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6957 if (rc != MDB_SUCCESS)
6959 key.mv_data = (char *)key.mv_data + key.mv_size;
6962 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6963 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6964 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6965 unsigned int snum = csrc->mc_snum;
6967 /* must find the lowest key below src */
6968 mdb_page_search_lowest(csrc);
6969 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6970 key.mv_size = csrc->mc_db->md_pad;
6971 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6973 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6974 key.mv_size = NODEKSZ(s2);
6975 key.mv_data = NODEKEY(s2);
6977 csrc->mc_snum = snum--;
6978 csrc->mc_top = snum;
6980 key.mv_size = srcnode->mn_ksize;
6981 key.mv_data = NODEKEY(srcnode);
6984 data.mv_size = NODEDSZ(srcnode);
6985 data.mv_data = NODEDATA(srcnode);
6986 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6987 if (rc != MDB_SUCCESS)
6992 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6993 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
6994 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
6996 /* Unlink the src page from parent and add to free list.
6998 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6999 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7002 rc = mdb_update_key(csrc, &key);
7008 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7009 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7012 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7013 csrc->mc_db->md_leaf_pages--;
7015 csrc->mc_db->md_branch_pages--;
7017 /* Adjust other cursors pointing to mp */
7018 MDB_cursor *m2, *m3;
7019 MDB_dbi dbi = csrc->mc_dbi;
7020 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7022 if (csrc->mc_flags & C_SUB)
7025 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7026 if (csrc->mc_flags & C_SUB)
7027 m3 = &m2->mc_xcursor->mx_cursor;
7030 if (m3 == csrc) continue;
7031 if (m3->mc_snum < csrc->mc_snum) continue;
7032 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7033 m3->mc_pg[csrc->mc_top] = mp;
7034 m3->mc_ki[csrc->mc_top] += nkeys;
7038 mdb_cursor_pop(csrc);
7040 return mdb_rebalance(csrc);
7043 /** Copy the contents of a cursor.
7044 * @param[in] csrc The cursor to copy from.
7045 * @param[out] cdst The cursor to copy to.
7048 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7052 cdst->mc_txn = csrc->mc_txn;
7053 cdst->mc_dbi = csrc->mc_dbi;
7054 cdst->mc_db = csrc->mc_db;
7055 cdst->mc_dbx = csrc->mc_dbx;
7056 cdst->mc_snum = csrc->mc_snum;
7057 cdst->mc_top = csrc->mc_top;
7058 cdst->mc_flags = csrc->mc_flags;
7060 for (i=0; i<csrc->mc_snum; i++) {
7061 cdst->mc_pg[i] = csrc->mc_pg[i];
7062 cdst->mc_ki[i] = csrc->mc_ki[i];
7066 /** Rebalance the tree after a delete operation.
7067 * @param[in] mc Cursor pointing to the page where rebalancing
7069 * @return 0 on success, non-zero on failure.
7072 mdb_rebalance(MDB_cursor *mc)
7076 unsigned int ptop, minkeys;
7079 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7083 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7084 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7085 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7086 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7087 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7091 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7092 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7095 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7096 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7102 if (mc->mc_snum < 2) {
7103 MDB_page *mp = mc->mc_pg[0];
7105 DPUTS("Can't rebalance a subpage, ignoring");
7108 if (NUMKEYS(mp) == 0) {
7109 DPUTS("tree is completely empty");
7110 mc->mc_db->md_root = P_INVALID;
7111 mc->mc_db->md_depth = 0;
7112 mc->mc_db->md_leaf_pages = 0;
7113 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7116 /* Adjust cursors pointing to mp */
7120 MDB_cursor *m2, *m3;
7121 MDB_dbi dbi = mc->mc_dbi;
7123 if (mc->mc_flags & C_SUB)
7126 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7127 if (mc->mc_flags & C_SUB)
7128 m3 = &m2->mc_xcursor->mx_cursor;
7131 if (m3->mc_snum < mc->mc_snum) continue;
7132 if (m3->mc_pg[0] == mp) {
7138 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7139 DPUTS("collapsing root page!");
7140 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7143 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7144 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7147 mc->mc_db->md_depth--;
7148 mc->mc_db->md_branch_pages--;
7149 mc->mc_ki[0] = mc->mc_ki[1];
7151 /* Adjust other cursors pointing to mp */
7152 MDB_cursor *m2, *m3;
7153 MDB_dbi dbi = mc->mc_dbi;
7155 if (mc->mc_flags & C_SUB)
7158 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7159 if (mc->mc_flags & C_SUB)
7160 m3 = &m2->mc_xcursor->mx_cursor;
7163 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7164 if (m3->mc_pg[0] == mp) {
7165 m3->mc_pg[0] = mc->mc_pg[0];
7168 m3->mc_ki[0] = m3->mc_ki[1];
7173 DPUTS("root page doesn't need rebalancing");
7177 /* The parent (branch page) must have at least 2 pointers,
7178 * otherwise the tree is invalid.
7180 ptop = mc->mc_top-1;
7181 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7183 /* Leaf page fill factor is below the threshold.
7184 * Try to move keys from left or right neighbor, or
7185 * merge with a neighbor page.
7190 mdb_cursor_copy(mc, &mn);
7191 mn.mc_xcursor = NULL;
7193 if (mc->mc_ki[ptop] == 0) {
7194 /* We're the leftmost leaf in our parent.
7196 DPUTS("reading right neighbor");
7198 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7199 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7202 mn.mc_ki[mn.mc_top] = 0;
7203 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7205 /* There is at least one neighbor to the left.
7207 DPUTS("reading left neighbor");
7209 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7210 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7213 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7214 mc->mc_ki[mc->mc_top] = 0;
7217 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7218 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7219 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7221 /* If the neighbor page is above threshold and has enough keys,
7222 * move one key from it. Otherwise we should try to merge them.
7223 * (A branch page must never have less than 2 keys.)
7225 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7226 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7227 return mdb_node_move(&mn, mc);
7229 if (mc->mc_ki[ptop] == 0)
7230 rc = mdb_page_merge(&mn, mc);
7232 rc = mdb_page_merge(mc, &mn);
7233 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7238 /** Complete a delete operation started by #mdb_cursor_del(). */
7240 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7247 mp = mc->mc_pg[mc->mc_top];
7248 ki = mc->mc_ki[mc->mc_top];
7250 /* add overflow pages to free list */
7251 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7255 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7256 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7257 (rc = mdb_ovpage_free(mc, omp)))
7260 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7261 mc->mc_db->md_entries--;
7262 rc = mdb_rebalance(mc);
7263 if (rc != MDB_SUCCESS)
7264 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7267 MDB_dbi dbi = mc->mc_dbi;
7269 mp = mc->mc_pg[mc->mc_top];
7270 nkeys = NUMKEYS(mp);
7272 /* if mc points past last node in page, find next sibling */
7273 if (mc->mc_ki[mc->mc_top] >= nkeys)
7274 mdb_cursor_sibling(mc, 1);
7276 /* Adjust other cursors pointing to mp */
7277 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7280 if (!(m2->mc_flags & C_INITIALIZED))
7282 if (m2->mc_pg[mc->mc_top] == mp) {
7283 if (m2->mc_ki[mc->mc_top] > ki)
7284 m2->mc_ki[mc->mc_top]--;
7285 if (m2->mc_ki[mc->mc_top] >= nkeys)
7286 mdb_cursor_sibling(m2, 1);
7295 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7296 MDB_val *key, MDB_val *data)
7301 MDB_val rdata, *xdata;
7305 assert(key != NULL);
7307 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7309 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7312 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7313 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7315 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7316 return MDB_BAD_VALSIZE;
7319 mdb_cursor_init(&mc, txn, dbi, &mx);
7322 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7323 /* must ignore any data */
7334 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7336 /* let mdb_page_split know about this cursor if needed:
7337 * delete will trigger a rebalance; if it needs to move
7338 * a node from one page to another, it will have to
7339 * update the parent's separator key(s). If the new sepkey
7340 * is larger than the current one, the parent page may
7341 * run out of space, triggering a split. We need this
7342 * cursor to be consistent until the end of the rebalance.
7344 mc.mc_flags |= C_UNTRACK;
7345 mc.mc_next = txn->mt_cursors[dbi];
7346 txn->mt_cursors[dbi] = &mc;
7347 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7348 txn->mt_cursors[dbi] = mc.mc_next;
7353 /** Split a page and insert a new node.
7354 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7355 * The cursor will be updated to point to the actual page and index where
7356 * the node got inserted after the split.
7357 * @param[in] newkey The key for the newly inserted node.
7358 * @param[in] newdata The data for the newly inserted node.
7359 * @param[in] newpgno The page number, if the new node is a branch node.
7360 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7361 * @return 0 on success, non-zero on failure.
7364 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7365 unsigned int nflags)
7368 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7371 unsigned int i, j, split_indx, nkeys, pmax;
7373 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7375 MDB_page *mp, *rp, *pp;
7380 mp = mc->mc_pg[mc->mc_top];
7381 newindx = mc->mc_ki[mc->mc_top];
7383 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7384 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7385 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7387 /* Create a right sibling. */
7388 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7390 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7392 if (mc->mc_snum < 2) {
7393 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7395 /* shift current top to make room for new parent */
7396 mc->mc_pg[1] = mc->mc_pg[0];
7397 mc->mc_ki[1] = mc->mc_ki[0];
7400 mc->mc_db->md_root = pp->mp_pgno;
7401 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7402 mc->mc_db->md_depth++;
7405 /* Add left (implicit) pointer. */
7406 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7407 /* undo the pre-push */
7408 mc->mc_pg[0] = mc->mc_pg[1];
7409 mc->mc_ki[0] = mc->mc_ki[1];
7410 mc->mc_db->md_root = mp->mp_pgno;
7411 mc->mc_db->md_depth--;
7418 ptop = mc->mc_top-1;
7419 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7422 mc->mc_flags |= C_SPLITTING;
7423 mdb_cursor_copy(mc, &mn);
7424 mn.mc_pg[mn.mc_top] = rp;
7425 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7427 if (nflags & MDB_APPEND) {
7428 mn.mc_ki[mn.mc_top] = 0;
7430 split_indx = newindx;
7435 nkeys = NUMKEYS(mp);
7436 split_indx = nkeys / 2;
7437 if (newindx < split_indx)
7443 unsigned int lsize, rsize, ksize;
7444 /* Move half of the keys to the right sibling */
7446 x = mc->mc_ki[mc->mc_top] - split_indx;
7447 ksize = mc->mc_db->md_pad;
7448 split = LEAF2KEY(mp, split_indx, ksize);
7449 rsize = (nkeys - split_indx) * ksize;
7450 lsize = (nkeys - split_indx) * sizeof(indx_t);
7451 mp->mp_lower -= lsize;
7452 rp->mp_lower += lsize;
7453 mp->mp_upper += rsize - lsize;
7454 rp->mp_upper -= rsize - lsize;
7455 sepkey.mv_size = ksize;
7456 if (newindx == split_indx) {
7457 sepkey.mv_data = newkey->mv_data;
7459 sepkey.mv_data = split;
7462 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7463 memcpy(rp->mp_ptrs, split, rsize);
7464 sepkey.mv_data = rp->mp_ptrs;
7465 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7466 memcpy(ins, newkey->mv_data, ksize);
7467 mp->mp_lower += sizeof(indx_t);
7468 mp->mp_upper -= ksize - sizeof(indx_t);
7471 memcpy(rp->mp_ptrs, split, x * ksize);
7472 ins = LEAF2KEY(rp, x, ksize);
7473 memcpy(ins, newkey->mv_data, ksize);
7474 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7475 rp->mp_lower += sizeof(indx_t);
7476 rp->mp_upper -= ksize - sizeof(indx_t);
7477 mc->mc_ki[mc->mc_top] = x;
7478 mc->mc_pg[mc->mc_top] = rp;
7483 /* For leaf pages, check the split point based on what
7484 * fits where, since otherwise mdb_node_add can fail.
7486 * This check is only needed when the data items are
7487 * relatively large, such that being off by one will
7488 * make the difference between success or failure.
7490 * It's also relevant if a page happens to be laid out
7491 * such that one half of its nodes are all "small" and
7492 * the other half of its nodes are "large." If the new
7493 * item is also "large" and falls on the half with
7494 * "large" nodes, it also may not fit.
7497 unsigned int psize, nsize;
7498 /* Maximum free space in an empty page */
7499 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7500 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7501 if ((nkeys < 20) || (nsize > pmax/16)) {
7502 if (newindx <= split_indx) {
7505 for (i=0; i<split_indx; i++) {
7506 node = NODEPTR(mp, i);
7507 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7508 if (F_ISSET(node->mn_flags, F_BIGDATA))
7509 psize += sizeof(pgno_t);
7511 psize += NODEDSZ(node);
7515 split_indx = newindx;
7526 for (i=nkeys-1; i>=split_indx; i--) {
7527 node = NODEPTR(mp, i);
7528 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7529 if (F_ISSET(node->mn_flags, F_BIGDATA))
7530 psize += sizeof(pgno_t);
7532 psize += NODEDSZ(node);
7536 split_indx = newindx;
7547 /* First find the separating key between the split pages.
7548 * The case where newindx == split_indx is ambiguous; the
7549 * new item could go to the new page or stay on the original
7550 * page. If newpos == 1 it goes to the new page.
7552 if (newindx == split_indx && newpos) {
7553 sepkey.mv_size = newkey->mv_size;
7554 sepkey.mv_data = newkey->mv_data;
7556 node = NODEPTR(mp, split_indx);
7557 sepkey.mv_size = node->mn_ksize;
7558 sepkey.mv_data = NODEKEY(node);
7562 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7564 /* Copy separator key to the parent.
7566 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7570 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7573 if (mn.mc_snum == mc->mc_snum) {
7574 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7575 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7576 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7577 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7582 /* Right page might now have changed parent.
7583 * Check if left page also changed parent.
7585 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7586 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7587 for (i=0; i<ptop; i++) {
7588 mc->mc_pg[i] = mn.mc_pg[i];
7589 mc->mc_ki[i] = mn.mc_ki[i];
7591 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7592 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7596 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7599 mc->mc_flags ^= C_SPLITTING;
7600 if (rc != MDB_SUCCESS) {
7603 if (nflags & MDB_APPEND) {
7604 mc->mc_pg[mc->mc_top] = rp;
7605 mc->mc_ki[mc->mc_top] = 0;
7606 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7609 for (i=0; i<mc->mc_top; i++)
7610 mc->mc_ki[i] = mn.mc_ki[i];
7617 /* Move half of the keys to the right sibling. */
7619 /* grab a page to hold a temporary copy */
7620 copy = mdb_page_malloc(mc->mc_txn, 1);
7624 copy->mp_pgno = mp->mp_pgno;
7625 copy->mp_flags = mp->mp_flags;
7626 copy->mp_lower = PAGEHDRSZ;
7627 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7628 mc->mc_pg[mc->mc_top] = copy;
7629 for (i = j = 0; i <= nkeys; j++) {
7630 if (i == split_indx) {
7631 /* Insert in right sibling. */
7632 /* Reset insert index for right sibling. */
7633 if (i != newindx || (newpos ^ ins_new)) {
7635 mc->mc_pg[mc->mc_top] = rp;
7639 if (i == newindx && !ins_new) {
7640 /* Insert the original entry that caused the split. */
7641 rkey.mv_data = newkey->mv_data;
7642 rkey.mv_size = newkey->mv_size;
7651 /* Update index for the new key. */
7652 mc->mc_ki[mc->mc_top] = j;
7653 } else if (i == nkeys) {
7656 node = NODEPTR(mp, i);
7657 rkey.mv_data = NODEKEY(node);
7658 rkey.mv_size = node->mn_ksize;
7660 xdata.mv_data = NODEDATA(node);
7661 xdata.mv_size = NODEDSZ(node);
7664 pgno = NODEPGNO(node);
7665 flags = node->mn_flags;
7670 if (!IS_LEAF(mp) && j == 0) {
7671 /* First branch index doesn't need key data. */
7675 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7679 nkeys = NUMKEYS(copy);
7680 for (i=0; i<nkeys; i++)
7681 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7682 mp->mp_lower = copy->mp_lower;
7683 mp->mp_upper = copy->mp_upper;
7684 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7685 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7687 /* reset back to original page */
7688 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7689 mc->mc_pg[mc->mc_top] = mp;
7690 if (nflags & MDB_RESERVE) {
7691 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7692 if (!(node->mn_flags & F_BIGDATA))
7693 newdata->mv_data = NODEDATA(node);
7697 /* Make sure mc_ki is still valid.
7699 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7700 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7701 for (i=0; i<ptop; i++) {
7702 mc->mc_pg[i] = mn.mc_pg[i];
7703 mc->mc_ki[i] = mn.mc_ki[i];
7705 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7706 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7710 /* return tmp page to freelist */
7711 mdb_page_free(mc->mc_txn->mt_env, copy);
7714 /* Adjust other cursors pointing to mp */
7715 MDB_cursor *m2, *m3;
7716 MDB_dbi dbi = mc->mc_dbi;
7717 int fixup = NUMKEYS(mp);
7719 if (mc->mc_flags & C_SUB)
7722 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7723 if (mc->mc_flags & C_SUB)
7724 m3 = &m2->mc_xcursor->mx_cursor;
7729 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7731 if (m3->mc_flags & C_SPLITTING)
7736 for (k=m3->mc_top; k>=0; k--) {
7737 m3->mc_ki[k+1] = m3->mc_ki[k];
7738 m3->mc_pg[k+1] = m3->mc_pg[k];
7740 if (m3->mc_ki[0] >= split_indx) {
7745 m3->mc_pg[0] = mc->mc_pg[0];
7749 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7750 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7751 m3->mc_ki[mc->mc_top]++;
7752 if (m3->mc_ki[mc->mc_top] >= fixup) {
7753 m3->mc_pg[mc->mc_top] = rp;
7754 m3->mc_ki[mc->mc_top] -= fixup;
7755 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7757 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7758 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7767 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7768 MDB_val *key, MDB_val *data, unsigned int flags)
7773 assert(key != NULL);
7774 assert(data != NULL);
7776 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7779 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7780 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7782 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7783 return MDB_BAD_VALSIZE;
7786 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7789 mdb_cursor_init(&mc, txn, dbi, &mx);
7790 return mdb_cursor_put(&mc, key, data, flags);
7794 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7796 if ((flag & CHANGEABLE) != flag)
7799 env->me_flags |= flag;
7801 env->me_flags &= ~flag;
7806 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7811 *arg = env->me_flags;
7816 mdb_env_get_path(MDB_env *env, const char **arg)
7821 *arg = env->me_path;
7825 /** Common code for #mdb_stat() and #mdb_env_stat().
7826 * @param[in] env the environment to operate in.
7827 * @param[in] db the #MDB_db record containing the stats to return.
7828 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7829 * @return 0, this function always succeeds.
7832 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7834 arg->ms_psize = env->me_psize;
7835 arg->ms_depth = db->md_depth;
7836 arg->ms_branch_pages = db->md_branch_pages;
7837 arg->ms_leaf_pages = db->md_leaf_pages;
7838 arg->ms_overflow_pages = db->md_overflow_pages;
7839 arg->ms_entries = db->md_entries;
7844 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7848 if (env == NULL || arg == NULL)
7851 toggle = mdb_env_pick_meta(env);
7853 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7857 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7861 if (env == NULL || arg == NULL)
7864 toggle = mdb_env_pick_meta(env);
7865 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7866 arg->me_mapsize = env->me_mapsize;
7867 arg->me_maxreaders = env->me_maxreaders;
7869 /* me_numreaders may be zero if this process never used any readers. Use
7870 * the shared numreader count if it exists.
7872 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7874 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7875 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7879 /** Set the default comparison functions for a database.
7880 * Called immediately after a database is opened to set the defaults.
7881 * The user can then override them with #mdb_set_compare() or
7882 * #mdb_set_dupsort().
7883 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7884 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7887 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7889 uint16_t f = txn->mt_dbs[dbi].md_flags;
7891 txn->mt_dbxs[dbi].md_cmp =
7892 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7893 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7895 txn->mt_dbxs[dbi].md_dcmp =
7896 !(f & MDB_DUPSORT) ? 0 :
7897 ((f & MDB_INTEGERDUP)
7898 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7899 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7902 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7907 int rc, dbflag, exact;
7908 unsigned int unused = 0;
7911 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7912 mdb_default_cmp(txn, FREE_DBI);
7915 if ((flags & VALID_FLAGS) != flags)
7917 if (txn->mt_flags & MDB_TXN_ERROR)
7923 if (flags & PERSISTENT_FLAGS) {
7924 uint16_t f2 = flags & PERSISTENT_FLAGS;
7925 /* make sure flag changes get committed */
7926 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7927 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7928 txn->mt_flags |= MDB_TXN_DIRTY;
7931 mdb_default_cmp(txn, MAIN_DBI);
7935 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7936 mdb_default_cmp(txn, MAIN_DBI);
7939 /* Is the DB already open? */
7941 for (i=2; i<txn->mt_numdbs; i++) {
7942 if (!txn->mt_dbxs[i].md_name.mv_size) {
7943 /* Remember this free slot */
7944 if (!unused) unused = i;
7947 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7948 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7954 /* If no free slot and max hit, fail */
7955 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7956 return MDB_DBS_FULL;
7958 /* Cannot mix named databases with some mainDB flags */
7959 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7960 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7962 /* Find the DB info */
7963 dbflag = DB_NEW|DB_VALID;
7966 key.mv_data = (void *)name;
7967 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7968 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7969 if (rc == MDB_SUCCESS) {
7970 /* make sure this is actually a DB */
7971 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7972 if (!(node->mn_flags & F_SUBDATA))
7973 return MDB_INCOMPATIBLE;
7974 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7975 /* Create if requested */
7977 data.mv_size = sizeof(MDB_db);
7978 data.mv_data = &dummy;
7979 memset(&dummy, 0, sizeof(dummy));
7980 dummy.md_root = P_INVALID;
7981 dummy.md_flags = flags & PERSISTENT_FLAGS;
7982 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7986 /* OK, got info, add to table */
7987 if (rc == MDB_SUCCESS) {
7988 unsigned int slot = unused ? unused : txn->mt_numdbs;
7989 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7990 txn->mt_dbxs[slot].md_name.mv_size = len;
7991 txn->mt_dbxs[slot].md_rel = NULL;
7992 txn->mt_dbflags[slot] = dbflag;
7993 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7995 mdb_default_cmp(txn, slot);
8004 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8006 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8009 if (txn->mt_dbflags[dbi] & DB_STALE) {
8012 /* Stale, must read the DB's root. cursor_init does it for us. */
8013 mdb_cursor_init(&mc, txn, dbi, &mx);
8015 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8018 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8021 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8023 ptr = env->me_dbxs[dbi].md_name.mv_data;
8024 env->me_dbxs[dbi].md_name.mv_data = NULL;
8025 env->me_dbxs[dbi].md_name.mv_size = 0;
8026 env->me_dbflags[dbi] = 0;
8030 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8032 /* We could return the flags for the FREE_DBI too but what's the point? */
8033 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8035 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8039 /** Add all the DB's pages to the free list.
8040 * @param[in] mc Cursor on the DB to free.
8041 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8042 * @return 0 on success, non-zero on failure.
8045 mdb_drop0(MDB_cursor *mc, int subs)
8049 rc = mdb_page_search(mc, NULL, 0);
8050 if (rc == MDB_SUCCESS) {
8051 MDB_txn *txn = mc->mc_txn;
8056 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8057 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8060 mdb_cursor_copy(mc, &mx);
8061 while (mc->mc_snum > 0) {
8062 MDB_page *mp = mc->mc_pg[mc->mc_top];
8063 unsigned n = NUMKEYS(mp);
8065 for (i=0; i<n; i++) {
8066 ni = NODEPTR(mp, i);
8067 if (ni->mn_flags & F_BIGDATA) {
8070 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8071 rc = mdb_page_get(txn, pg, &omp, NULL);
8074 assert(IS_OVERFLOW(omp));
8075 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8079 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8080 mdb_xcursor_init1(mc, ni);
8081 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8087 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8089 for (i=0; i<n; i++) {
8091 ni = NODEPTR(mp, i);
8094 mdb_midl_xappend(txn->mt_free_pgs, pg);
8099 mc->mc_ki[mc->mc_top] = i;
8100 rc = mdb_cursor_sibling(mc, 1);
8102 /* no more siblings, go back to beginning
8103 * of previous level.
8107 for (i=1; i<mc->mc_snum; i++) {
8109 mc->mc_pg[i] = mx.mc_pg[i];
8114 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8115 } else if (rc == MDB_NOTFOUND) {
8121 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8123 MDB_cursor *mc, *m2;
8126 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8129 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8132 rc = mdb_cursor_open(txn, dbi, &mc);
8136 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8137 /* Invalidate the dropped DB's cursors */
8138 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8139 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8143 /* Can't delete the main DB */
8144 if (del && dbi > MAIN_DBI) {
8145 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8147 txn->mt_dbflags[dbi] = DB_STALE;
8148 mdb_dbi_close(txn->mt_env, dbi);
8151 /* reset the DB record, mark it dirty */
8152 txn->mt_dbflags[dbi] |= DB_DIRTY;
8153 txn->mt_dbs[dbi].md_depth = 0;
8154 txn->mt_dbs[dbi].md_branch_pages = 0;
8155 txn->mt_dbs[dbi].md_leaf_pages = 0;
8156 txn->mt_dbs[dbi].md_overflow_pages = 0;
8157 txn->mt_dbs[dbi].md_entries = 0;
8158 txn->mt_dbs[dbi].md_root = P_INVALID;
8160 txn->mt_flags |= MDB_TXN_DIRTY;
8163 mdb_cursor_close(mc);
8167 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8169 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8172 txn->mt_dbxs[dbi].md_cmp = cmp;
8176 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8178 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8181 txn->mt_dbxs[dbi].md_dcmp = cmp;
8185 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8187 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8190 txn->mt_dbxs[dbi].md_rel = rel;
8194 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8196 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8199 txn->mt_dbxs[dbi].md_relctx = ctx;
8203 int mdb_env_get_maxkeysize(MDB_env *env)
8205 return MDB_MAXKEYSIZE;
8208 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8210 unsigned int i, rdrs;
8217 if (!env->me_txns) {
8218 return func("(no reader locks)\n", ctx);
8220 rdrs = env->me_txns->mti_numreaders;
8221 mr = env->me_txns->mti_readers;
8222 for (i=0; i<rdrs; i++) {
8227 if (mr[i].mr_txnid == (txnid_t)-1) {
8228 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8230 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8234 func(" pid thread txnid\n", ctx);
8236 rc = func(buf, ctx);
8242 func("(no active readers)\n", ctx);
8247 /* insert pid into list if not already present.
8248 * return -1 if already present.
8250 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8252 /* binary search of pid in list */
8254 unsigned cursor = 1;
8256 unsigned n = ids[0];
8259 unsigned pivot = n >> 1;
8260 cursor = base + pivot + 1;
8261 val = pid - ids[cursor];
8266 } else if ( val > 0 ) {
8271 /* found, so it's a duplicate */
8280 for (n = ids[0]; n > cursor; n--)
8286 int mdb_reader_check(MDB_env *env, int *dead)
8288 unsigned int i, j, rdrs;
8299 rdrs = env->me_txns->mti_numreaders;
8300 pids = malloc((rdrs+1) * sizeof(pid_t));
8304 mr = env->me_txns->mti_readers;
8306 for (i=0; i<rdrs; i++) {
8307 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8309 if (mdb_pid_insert(pids, pid) == 0) {
8310 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8312 /* Recheck, a new process may have reused pid */
8313 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8314 for (j=i; j<rdrs; j++)
8315 if (mr[j].mr_pid == pid) {
8320 UNLOCK_MUTEX_R(env);
projects / openldap / blob