2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define MDB_PIDLOCK 0
144 #define pthread_t DWORD
145 #define pthread_mutex_t HANDLE
146 #define pthread_key_t DWORD
147 #define pthread_self() GetCurrentThreadId()
148 #define pthread_key_create(x,y) \
149 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
150 #define pthread_key_delete(x) TlsFree(x)
151 #define pthread_getspecific(x) TlsGetValue(x)
152 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
153 #define pthread_mutex_unlock(x) ReleaseMutex(x)
154 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
155 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
156 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
157 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
158 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
159 #define getpid() GetCurrentProcessId()
160 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
161 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
162 #define ErrCode() GetLastError()
163 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
164 #define close(fd) (CloseHandle(fd) ? 0 : -1)
165 #define munmap(ptr,len) UnmapViewOfFile(ptr)
166 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
167 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
169 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
174 #define Z "z" /**< printf format modifier for size_t */
176 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
177 #define MDB_PIDLOCK 1
179 #ifdef MDB_USE_POSIX_SEM
181 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
182 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
183 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
184 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
187 mdb_sem_wait(sem_t *sem)
190 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
195 /** Lock the reader mutex.
197 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
198 /** Unlock the reader mutex.
200 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
202 /** Lock the writer mutex.
203 * Only a single write transaction is allowed at a time. Other writers
204 * will block waiting for this mutex.
206 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
207 /** Unlock the writer mutex.
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
210 #endif /* MDB_USE_POSIX_SEM */
212 /** Get the error code for the last failed system function.
214 #define ErrCode() errno
216 /** An abstraction for a file handle.
217 * On POSIX systems file handles are small integers. On Windows
218 * they're opaque pointers.
222 /** A value for an invalid file handle.
223 * Mainly used to initialize file variables and signify that they are
226 #define INVALID_HANDLE_VALUE (-1)
228 /** Get the size of a memory page for the system.
229 * This is the basic size that the platform's memory manager uses, and is
230 * fundamental to the use of memory-mapped files.
232 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
235 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
238 #define MNAME_LEN (sizeof(pthread_mutex_t))
244 /** A flag for opening a file and requesting synchronous data writes.
245 * This is only used when writing a meta page. It's not strictly needed;
246 * we could just do a normal write and then immediately perform a flush.
247 * But if this flag is available it saves us an extra system call.
249 * @note If O_DSYNC is undefined but exists in /usr/include,
250 * preferably set some compiler flag to get the definition.
251 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
254 # define MDB_DSYNC O_DSYNC
258 /** Function for flushing the data of a file. Define this to fsync
259 * if fdatasync() is not supported.
261 #ifndef MDB_FDATASYNC
262 # define MDB_FDATASYNC fdatasync
266 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
277 /** A page number in the database.
278 * Note that 64 bit page numbers are overkill, since pages themselves
279 * already represent 12-13 bits of addressable memory, and the OS will
280 * always limit applications to a maximum of 63 bits of address space.
282 * @note In the #MDB_node structure, we only store 48 bits of this value,
283 * which thus limits us to only 60 bits of addressable data.
285 typedef MDB_ID pgno_t;
287 /** A transaction ID.
288 * See struct MDB_txn.mt_txnid for details.
290 typedef MDB_ID txnid_t;
292 /** @defgroup debug Debug Macros
296 /** Enable debug output. Needs variable argument macros (a C99 feature).
297 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
298 * read from and written to the database (used for free space management).
304 static int mdb_debug;
305 static txnid_t mdb_debug_start;
307 /** Print a debug message with printf formatting.
308 * Requires double parenthesis around 2 or more args.
310 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
311 # define DPRINTF0(fmt, ...) \
312 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
314 # define DPRINTF(args) ((void) 0)
316 /** Print a debug string.
317 * The string is printed literally, with no format processing.
319 #define DPUTS(arg) DPRINTF(("%s", arg))
322 /** A default memory page size.
323 * The actual size is platform-dependent, but we use this for
324 * boot-strapping. We probably should not be using this any more.
325 * The #GET_PAGESIZE() macro is used to get the actual size.
327 * Note that we don't currently support Huge pages. On Linux,
328 * regular data files cannot use Huge pages, and in general
329 * Huge pages aren't actually pageable. We rely on the OS
330 * demand-pager to read our data and page it out when memory
331 * pressure from other processes is high. So until OSs have
332 * actual paging support for Huge pages, they're not viable.
334 #define MDB_PAGESIZE 4096
336 /** The minimum number of keys required in a database page.
337 * Setting this to a larger value will place a smaller bound on the
338 * maximum size of a data item. Data items larger than this size will
339 * be pushed into overflow pages instead of being stored directly in
340 * the B-tree node. This value used to default to 4. With a page size
341 * of 4096 bytes that meant that any item larger than 1024 bytes would
342 * go into an overflow page. That also meant that on average 2-3KB of
343 * each overflow page was wasted space. The value cannot be lower than
344 * 2 because then there would no longer be a tree structure. With this
345 * value, items larger than 2KB will go into overflow pages, and on
346 * average only 1KB will be wasted.
348 #define MDB_MINKEYS 2
350 /** A stamp that identifies a file as an MDB file.
351 * There's nothing special about this value other than that it is easily
352 * recognizable, and it will reflect any byte order mismatches.
354 #define MDB_MAGIC 0xBEEFC0DE
356 /** The version number for a database's datafile format. */
357 #define MDB_DATA_VERSION 1
358 /** The version number for a database's lockfile format. */
359 #define MDB_LOCK_VERSION 1
361 /** @brief The maximum size of a key in the database.
363 * The library rejects bigger keys, and cannot deal with records
364 * with bigger keys stored by a library with bigger max keysize.
366 * We require that keys all fit onto a regular page. This limit
367 * could be raised a bit further if needed; to something just
368 * under #MDB_PAGESIZE / #MDB_MINKEYS.
370 * Note that data items in an #MDB_DUPSORT database are actually keys
371 * of a subDB, so they're also limited to this size.
373 #ifndef MDB_MAXKEYSIZE
374 #define MDB_MAXKEYSIZE 511
377 /** @brief The maximum size of a data item.
379 * We only store a 32 bit value for node sizes.
381 #define MAXDATASIZE 0xffffffffUL
386 * This is used for printing a hex dump of a key's contents.
388 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
389 /** Display a key in hex.
391 * Invoke a function to display a key in hex.
393 #define DKEY(x) mdb_dkey(x, kbuf)
399 /** An invalid page number.
400 * Mainly used to denote an empty tree.
402 #define P_INVALID (~(pgno_t)0)
404 /** Test if the flags \b f are set in a flag word \b w. */
405 #define F_ISSET(w, f) (((w) & (f)) == (f))
407 /** Used for offsets within a single page.
408 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
411 typedef uint16_t indx_t;
413 /** Default size of memory map.
414 * This is certainly too small for any actual applications. Apps should always set
415 * the size explicitly using #mdb_env_set_mapsize().
417 #define DEFAULT_MAPSIZE 1048576
419 /** @defgroup readers Reader Lock Table
420 * Readers don't acquire any locks for their data access. Instead, they
421 * simply record their transaction ID in the reader table. The reader
422 * mutex is needed just to find an empty slot in the reader table. The
423 * slot's address is saved in thread-specific data so that subsequent read
424 * transactions started by the same thread need no further locking to proceed.
426 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
428 * No reader table is used if the database is on a read-only filesystem.
430 * Since the database uses multi-version concurrency control, readers don't
431 * actually need any locking. This table is used to keep track of which
432 * readers are using data from which old transactions, so that we'll know
433 * when a particular old transaction is no longer in use. Old transactions
434 * that have discarded any data pages can then have those pages reclaimed
435 * for use by a later write transaction.
437 * The lock table is constructed such that reader slots are aligned with the
438 * processor's cache line size. Any slot is only ever used by one thread.
439 * This alignment guarantees that there will be no contention or cache
440 * thrashing as threads update their own slot info, and also eliminates
441 * any need for locking when accessing a slot.
443 * A writer thread will scan every slot in the table to determine the oldest
444 * outstanding reader transaction. Any freed pages older than this will be
445 * reclaimed by the writer. The writer doesn't use any locks when scanning
446 * this table. This means that there's no guarantee that the writer will
447 * see the most up-to-date reader info, but that's not required for correct
448 * operation - all we need is to know the upper bound on the oldest reader,
449 * we don't care at all about the newest reader. So the only consequence of
450 * reading stale information here is that old pages might hang around a
451 * while longer before being reclaimed. That's actually good anyway, because
452 * the longer we delay reclaiming old pages, the more likely it is that a
453 * string of contiguous pages can be found after coalescing old pages from
454 * many old transactions together.
457 /** Number of slots in the reader table.
458 * This value was chosen somewhat arbitrarily. 126 readers plus a
459 * couple mutexes fit exactly into 8KB on my development machine.
460 * Applications should set the table size using #mdb_env_set_maxreaders().
462 #define DEFAULT_READERS 126
464 /** The size of a CPU cache line in bytes. We want our lock structures
465 * aligned to this size to avoid false cache line sharing in the
467 * This value works for most CPUs. For Itanium this should be 128.
473 /** The information we store in a single slot of the reader table.
474 * In addition to a transaction ID, we also record the process and
475 * thread ID that owns a slot, so that we can detect stale information,
476 * e.g. threads or processes that went away without cleaning up.
477 * @note We currently don't check for stale records. We simply re-init
478 * the table when we know that we're the only process opening the
481 typedef struct MDB_rxbody {
482 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
483 * Multiple readers that start at the same time will probably have the
484 * same ID here. Again, it's not important to exclude them from
485 * anything; all we need to know is which version of the DB they
486 * started from so we can avoid overwriting any data used in that
487 * particular version.
490 /** The process ID of the process owning this reader txn. */
492 /** The thread ID of the thread owning this txn. */
496 /** The actual reader record, with cacheline padding. */
497 typedef struct MDB_reader {
500 /** shorthand for mrb_txnid */
501 #define mr_txnid mru.mrx.mrb_txnid
502 #define mr_pid mru.mrx.mrb_pid
503 #define mr_tid mru.mrx.mrb_tid
504 /** cache line alignment */
505 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
509 /** The header for the reader table.
510 * The table resides in a memory-mapped file. (This is a different file
511 * than is used for the main database.)
513 * For POSIX the actual mutexes reside in the shared memory of this
514 * mapped file. On Windows, mutexes are named objects allocated by the
515 * kernel; we store the mutex names in this mapped file so that other
516 * processes can grab them. This same approach is also used on
517 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
518 * process-shared POSIX mutexes. For these cases where a named object
519 * is used, the object name is derived from a 64 bit FNV hash of the
520 * environment pathname. As such, naming collisions are extremely
521 * unlikely. If a collision occurs, the results are unpredictable.
523 typedef struct MDB_txbody {
524 /** Stamp identifying this as an MDB file. It must be set
527 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
529 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
530 char mtb_rmname[MNAME_LEN];
532 /** Mutex protecting access to this table.
533 * This is the reader lock that #LOCK_MUTEX_R acquires.
535 pthread_mutex_t mtb_mutex;
537 /** The ID of the last transaction committed to the database.
538 * This is recorded here only for convenience; the value can always
539 * be determined by reading the main database meta pages.
542 /** The number of slots that have been used in the reader table.
543 * This always records the maximum count, it is not decremented
544 * when readers release their slots.
546 unsigned mtb_numreaders;
549 /** The actual reader table definition. */
550 typedef struct MDB_txninfo {
553 #define mti_magic mt1.mtb.mtb_magic
554 #define mti_format mt1.mtb.mtb_format
555 #define mti_mutex mt1.mtb.mtb_mutex
556 #define mti_rmname mt1.mtb.mtb_rmname
557 #define mti_txnid mt1.mtb.mtb_txnid
558 #define mti_numreaders mt1.mtb.mtb_numreaders
559 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
562 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
563 char mt2_wmname[MNAME_LEN];
564 #define mti_wmname mt2.mt2_wmname
566 pthread_mutex_t mt2_wmutex;
567 #define mti_wmutex mt2.mt2_wmutex
569 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
571 MDB_reader mti_readers[1];
574 /** Lockfile format signature: version, features and field layout */
575 #define MDB_LOCK_FORMAT \
577 ((MDB_LOCK_VERSION) \
578 /* Flags which describe functionality */ \
579 + (((MDB_PIDLOCK) != 0) << 16)))
582 /** Common header for all page types.
583 * Overflow records occupy a number of contiguous pages with no
584 * headers on any page after the first.
586 typedef struct MDB_page {
587 #define mp_pgno mp_p.p_pgno
588 #define mp_next mp_p.p_next
590 pgno_t p_pgno; /**< page number */
591 void * p_next; /**< for in-memory list of freed structs */
594 /** @defgroup mdb_page Page Flags
596 * Flags for the page headers.
599 #define P_BRANCH 0x01 /**< branch page */
600 #define P_LEAF 0x02 /**< leaf page */
601 #define P_OVERFLOW 0x04 /**< overflow page */
602 #define P_META 0x08 /**< meta page */
603 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
604 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
605 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
606 #define P_KEEP 0x8000 /**< leave this page alone during spill */
608 uint16_t mp_flags; /**< @ref mdb_page */
609 #define mp_lower mp_pb.pb.pb_lower
610 #define mp_upper mp_pb.pb.pb_upper
611 #define mp_pages mp_pb.pb_pages
614 indx_t pb_lower; /**< lower bound of free space */
615 indx_t pb_upper; /**< upper bound of free space */
617 uint32_t pb_pages; /**< number of overflow pages */
619 indx_t mp_ptrs[1]; /**< dynamic size */
622 /** Size of the page header, excluding dynamic data at the end */
623 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
625 /** Address of first usable data byte in a page, after the header */
626 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
628 /** Number of nodes on a page */
629 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
631 /** The amount of space remaining in the page */
632 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
634 /** The percentage of space used in the page, in tenths of a percent. */
635 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
636 ((env)->me_psize - PAGEHDRSZ))
637 /** The minimum page fill factor, in tenths of a percent.
638 * Pages emptier than this are candidates for merging.
640 #define FILL_THRESHOLD 250
642 /** Test if a page is a leaf page */
643 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
644 /** Test if a page is a LEAF2 page */
645 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
646 /** Test if a page is a branch page */
647 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
648 /** Test if a page is an overflow page */
649 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
650 /** Test if a page is a sub page */
651 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
653 /** The number of overflow pages needed to store the given size. */
654 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
656 /** Header for a single key/data pair within a page.
657 * We guarantee 2-byte alignment for nodes.
659 typedef struct MDB_node {
660 /** lo and hi are used for data size on leaf nodes and for
661 * child pgno on branch nodes. On 64 bit platforms, flags
662 * is also used for pgno. (Branch nodes have no flags).
663 * They are in host byte order in case that lets some
664 * accesses be optimized into a 32-bit word access.
666 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
667 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
668 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
669 /** @defgroup mdb_node Node Flags
671 * Flags for node headers.
674 #define F_BIGDATA 0x01 /**< data put on overflow page */
675 #define F_SUBDATA 0x02 /**< data is a sub-database */
676 #define F_DUPDATA 0x04 /**< data has duplicates */
678 /** valid flags for #mdb_node_add() */
679 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
682 unsigned short mn_flags; /**< @ref mdb_node */
683 unsigned short mn_ksize; /**< key size */
684 char mn_data[1]; /**< key and data are appended here */
687 /** Size of the node header, excluding dynamic data at the end */
688 #define NODESIZE offsetof(MDB_node, mn_data)
690 /** Bit position of top word in page number, for shifting mn_flags */
691 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
693 /** Size of a node in a branch page with a given key.
694 * This is just the node header plus the key, there is no data.
696 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
698 /** Size of a node in a leaf page with a given key and data.
699 * This is node header plus key plus data size.
701 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
703 /** Address of node \b i in page \b p */
704 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
706 /** Address of the key for the node */
707 #define NODEKEY(node) (void *)((node)->mn_data)
709 /** Address of the data for a node */
710 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
712 /** Get the page number pointed to by a branch node */
713 #define NODEPGNO(node) \
714 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
715 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
716 /** Set the page number in a branch node */
717 #define SETPGNO(node,pgno) do { \
718 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
719 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
721 /** Get the size of the data in a leaf node */
722 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
723 /** Set the size of the data for a leaf node */
724 #define SETDSZ(node,size) do { \
725 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
726 /** The size of a key in a node */
727 #define NODEKSZ(node) ((node)->mn_ksize)
729 /** Copy a page number from src to dst */
731 #define COPY_PGNO(dst,src) dst = src
733 #if SIZE_MAX > 4294967295UL
734 #define COPY_PGNO(dst,src) do { \
735 unsigned short *s, *d; \
736 s = (unsigned short *)&(src); \
737 d = (unsigned short *)&(dst); \
744 #define COPY_PGNO(dst,src) do { \
745 unsigned short *s, *d; \
746 s = (unsigned short *)&(src); \
747 d = (unsigned short *)&(dst); \
753 /** The address of a key in a LEAF2 page.
754 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
755 * There are no node headers, keys are stored contiguously.
757 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
759 /** Set the \b node's key into \b keyptr, if requested. */
760 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
761 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
763 /** Set the \b node's key into \b key. */
764 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
766 /** Information about a single database in the environment. */
767 typedef struct MDB_db {
768 uint32_t md_pad; /**< also ksize for LEAF2 pages */
769 uint16_t md_flags; /**< @ref mdb_dbi_open */
770 uint16_t md_depth; /**< depth of this tree */
771 pgno_t md_branch_pages; /**< number of internal pages */
772 pgno_t md_leaf_pages; /**< number of leaf pages */
773 pgno_t md_overflow_pages; /**< number of overflow pages */
774 size_t md_entries; /**< number of data items */
775 pgno_t md_root; /**< the root page of this tree */
778 /** mdb_dbi_open flags */
779 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
780 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
781 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
782 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
784 /** Handle for the DB used to track free pages. */
786 /** Handle for the default DB. */
789 /** Meta page content.
790 * A meta page is the start point for accessing a database snapshot.
791 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
793 typedef struct MDB_meta {
794 /** Stamp identifying this as an MDB file. It must be set
797 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
799 void *mm_address; /**< address for fixed mapping */
800 size_t mm_mapsize; /**< size of mmap region */
801 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
802 /** The size of pages used in this DB */
803 #define mm_psize mm_dbs[0].md_pad
804 /** Any persistent environment flags. @ref mdb_env */
805 #define mm_flags mm_dbs[0].md_flags
806 pgno_t mm_last_pg; /**< last used page in file */
807 txnid_t mm_txnid; /**< txnid that committed this page */
810 /** Buffer for a stack-allocated dirty page.
811 * The members define size and alignment, and silence type
812 * aliasing warnings. They are not used directly; that could
813 * mean incorrectly using several union members in parallel.
815 typedef union MDB_pagebuf {
816 char mb_raw[MDB_PAGESIZE];
819 char mm_pad[PAGEHDRSZ];
824 /** Auxiliary DB info.
825 * The information here is mostly static/read-only. There is
826 * only a single copy of this record in the environment.
828 typedef struct MDB_dbx {
829 MDB_val md_name; /**< name of the database */
830 MDB_cmp_func *md_cmp; /**< function for comparing keys */
831 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
832 MDB_rel_func *md_rel; /**< user relocate function */
833 void *md_relctx; /**< user-provided context for md_rel */
836 /** A database transaction.
837 * Every operation requires a transaction handle.
840 MDB_txn *mt_parent; /**< parent of a nested txn */
841 MDB_txn *mt_child; /**< nested txn under this txn */
842 pgno_t mt_next_pgno; /**< next unallocated page */
843 /** The ID of this transaction. IDs are integers incrementing from 1.
844 * Only committed write transactions increment the ID. If a transaction
845 * aborts, the ID may be re-used by the next writer.
848 MDB_env *mt_env; /**< the DB environment */
849 /** The list of pages that became unused during this transaction.
852 /** The sorted list of dirty pages we temporarily wrote to disk
853 * because the dirty list was full. page numbers in here are
854 * shifted left by 1, deleted slots have the LSB set.
856 MDB_IDL mt_spill_pgs;
858 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
860 /** For read txns: This thread/txn's reader table slot, or NULL. */
863 /** Array of records for each DB known in the environment. */
865 /** Array of MDB_db records for each known DB */
867 /** @defgroup mt_dbflag Transaction DB Flags
871 #define DB_DIRTY 0x01 /**< DB was written in this txn */
872 #define DB_STALE 0x02 /**< DB record is older than txnID */
873 #define DB_NEW 0x04 /**< DB handle opened in this txn */
874 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
876 /** In write txns, array of cursors for each DB */
877 MDB_cursor **mt_cursors;
878 /** Array of flags for each DB */
879 unsigned char *mt_dbflags;
880 /** Number of DB records in use. This number only ever increments;
881 * we don't decrement it when individual DB handles are closed.
885 /** @defgroup mdb_txn Transaction Flags
889 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
890 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
891 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
892 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
894 unsigned int mt_flags; /**< @ref mdb_txn */
895 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
896 unsigned int mt_dirty_room;
897 /** Tracks which of the two meta pages was used at the start
898 * of this transaction.
900 unsigned int mt_toggle;
903 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
904 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
905 * raise this on a 64 bit machine.
907 #define CURSOR_STACK 32
911 /** Cursors are used for all DB operations.
912 * A cursor holds a path of (page pointer, key index) from the DB
913 * root to a position in the DB, plus other state. #MDB_DUPSORT
914 * cursors include an xcursor to the current data item. Write txns
915 * track their cursors and keep them up to date when data moves.
916 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
917 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
920 /** Next cursor on this DB in this txn */
922 /** Backup of the original cursor if this cursor is a shadow */
923 MDB_cursor *mc_backup;
924 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
925 struct MDB_xcursor *mc_xcursor;
926 /** The transaction that owns this cursor */
928 /** The database handle this cursor operates on */
930 /** The database record for this cursor */
932 /** The database auxiliary record for this cursor */
934 /** The @ref mt_dbflag for this database */
935 unsigned char *mc_dbflag;
936 unsigned short mc_snum; /**< number of pushed pages */
937 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
938 /** @defgroup mdb_cursor Cursor Flags
940 * Cursor state flags.
943 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
944 #define C_EOF 0x02 /**< No more data */
945 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
946 #define C_DEL 0x08 /**< last op was a cursor_del */
947 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
948 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
950 unsigned int mc_flags; /**< @ref mdb_cursor */
951 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
952 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
955 /** Context for sorted-dup records.
956 * We could have gone to a fully recursive design, with arbitrarily
957 * deep nesting of sub-databases. But for now we only handle these
958 * levels - main DB, optional sub-DB, sorted-duplicate DB.
960 typedef struct MDB_xcursor {
961 /** A sub-cursor for traversing the Dup DB */
962 MDB_cursor mx_cursor;
963 /** The database record for this Dup DB */
965 /** The auxiliary DB record for this Dup DB */
967 /** The @ref mt_dbflag for this Dup DB */
968 unsigned char mx_dbflag;
971 /** State of FreeDB old pages, stored in the MDB_env */
972 typedef struct MDB_pgstate {
973 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
974 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
977 /** The database environment. */
979 HANDLE me_fd; /**< The main data file */
980 HANDLE me_lfd; /**< The lock file */
981 HANDLE me_mfd; /**< just for writing the meta pages */
982 /** Failed to update the meta page. Probably an I/O error. */
983 #define MDB_FATAL_ERROR 0x80000000U
984 /** Some fields are initialized. */
985 #define MDB_ENV_ACTIVE 0x20000000U
986 /** me_txkey is set */
987 #define MDB_ENV_TXKEY 0x10000000U
988 /** Have liveness lock in reader table */
989 #define MDB_LIVE_READER 0x08000000U
990 uint32_t me_flags; /**< @ref mdb_env */
991 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
992 unsigned int me_maxreaders; /**< size of the reader table */
993 unsigned int me_numreaders; /**< max numreaders set by this env */
994 MDB_dbi me_numdbs; /**< number of DBs opened */
995 MDB_dbi me_maxdbs; /**< size of the DB table */
996 pid_t me_pid; /**< process ID of this env */
997 char *me_path; /**< path to the DB files */
998 char *me_map; /**< the memory map of the data file */
999 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1000 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1001 MDB_txn *me_txn; /**< current write transaction */
1002 size_t me_mapsize; /**< size of the data memory map */
1003 off_t me_size; /**< current file size */
1004 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1005 MDB_dbx *me_dbxs; /**< array of static DB info */
1006 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1007 pthread_key_t me_txkey; /**< thread-key for readers */
1008 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1009 # define me_pglast me_pgstate.mf_pglast
1010 # define me_pghead me_pgstate.mf_pghead
1011 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1012 /** IDL of pages that became unused in a write txn */
1013 MDB_IDL me_free_pgs;
1014 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1015 MDB_ID2L me_dirty_list;
1016 /** Max number of freelist items that can fit in a single overflow page */
1018 /** Max size of a node on a page */
1019 unsigned int me_nodemax;
1021 int me_pidquery; /**< Used in OpenProcess */
1022 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1024 #elif defined(MDB_USE_POSIX_SEM)
1025 sem_t *me_rmutex; /* Shared mutexes are not supported */
1030 /** Nested transaction */
1031 typedef struct MDB_ntxn {
1032 MDB_txn mnt_txn; /**< the transaction */
1033 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1036 /** max number of pages to commit in one writev() call */
1037 #define MDB_COMMIT_PAGES 64
1038 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1039 #undef MDB_COMMIT_PAGES
1040 #define MDB_COMMIT_PAGES IOV_MAX
1043 /* max bytes to write in one call */
1044 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1046 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1047 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1048 static int mdb_page_touch(MDB_cursor *mc);
1050 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1051 static int mdb_page_search_root(MDB_cursor *mc,
1052 MDB_val *key, int modify);
1053 #define MDB_PS_MODIFY 1
1054 #define MDB_PS_ROOTONLY 2
1055 static int mdb_page_search(MDB_cursor *mc,
1056 MDB_val *key, int flags);
1057 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1059 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1060 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1061 pgno_t newpgno, unsigned int nflags);
1063 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1064 static int mdb_env_pick_meta(const MDB_env *env);
1065 static int mdb_env_write_meta(MDB_txn *txn);
1066 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1067 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1069 static void mdb_env_close0(MDB_env *env, int excl);
1071 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1072 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1073 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1074 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1075 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1076 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1077 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1078 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1079 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1081 static int mdb_rebalance(MDB_cursor *mc);
1082 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1084 static void mdb_cursor_pop(MDB_cursor *mc);
1085 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1087 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1088 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1089 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1090 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1091 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1093 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1094 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1096 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1097 static void mdb_xcursor_init0(MDB_cursor *mc);
1098 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1100 static int mdb_drop0(MDB_cursor *mc, int subs);
1101 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1104 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1108 static SECURITY_DESCRIPTOR mdb_null_sd;
1109 static SECURITY_ATTRIBUTES mdb_all_sa;
1110 static int mdb_sec_inited;
1113 /** Return the library version info. */
1115 mdb_version(int *major, int *minor, int *patch)
1117 if (major) *major = MDB_VERSION_MAJOR;
1118 if (minor) *minor = MDB_VERSION_MINOR;
1119 if (patch) *patch = MDB_VERSION_PATCH;
1120 return MDB_VERSION_STRING;
1123 /** Table of descriptions for MDB @ref errors */
1124 static char *const mdb_errstr[] = {
1125 "MDB_KEYEXIST: Key/data pair already exists",
1126 "MDB_NOTFOUND: No matching key/data pair found",
1127 "MDB_PAGE_NOTFOUND: Requested page not found",
1128 "MDB_CORRUPTED: Located page was wrong type",
1129 "MDB_PANIC: Update of meta page failed",
1130 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1131 "MDB_INVALID: File is not an MDB file",
1132 "MDB_MAP_FULL: Environment mapsize limit reached",
1133 "MDB_DBS_FULL: Environment maxdbs limit reached",
1134 "MDB_READERS_FULL: Environment maxreaders limit reached",
1135 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1136 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1137 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1138 "MDB_PAGE_FULL: Internal error - page has no more space",
1139 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1140 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1141 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1142 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1143 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1147 mdb_strerror(int err)
1151 return ("Successful return: 0");
1153 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1154 i = err - MDB_KEYEXIST;
1155 return mdb_errstr[i];
1158 return strerror(err);
1162 /** Display a key in hexadecimal and return the address of the result.
1163 * @param[in] key the key to display
1164 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1165 * @return The key in hexadecimal form.
1168 mdb_dkey(MDB_val *key, char *buf)
1171 unsigned char *c = key->mv_data;
1177 if (key->mv_size > MDB_MAXKEYSIZE)
1178 return "MDB_MAXKEYSIZE";
1179 /* may want to make this a dynamic check: if the key is mostly
1180 * printable characters, print it as-is instead of converting to hex.
1184 for (i=0; i<key->mv_size; i++)
1185 ptr += sprintf(ptr, "%02x", *c++);
1187 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1192 /** Display all the keys in the page. */
1194 mdb_page_list(MDB_page *mp)
1197 unsigned int i, nkeys, nsize;
1201 nkeys = NUMKEYS(mp);
1202 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1203 for (i=0; i<nkeys; i++) {
1204 node = NODEPTR(mp, i);
1205 key.mv_size = node->mn_ksize;
1206 key.mv_data = node->mn_data;
1207 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1208 if (IS_BRANCH(mp)) {
1209 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1212 if (F_ISSET(node->mn_flags, F_BIGDATA))
1213 nsize += sizeof(pgno_t);
1215 nsize += NODEDSZ(node);
1216 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1222 mdb_cursor_chk(MDB_cursor *mc)
1228 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1229 for (i=0; i<mc->mc_top; i++) {
1231 node = NODEPTR(mp, mc->mc_ki[i]);
1232 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1235 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1241 /** Count all the pages in each DB and in the freelist
1242 * and make sure it matches the actual number of pages
1245 static void mdb_audit(MDB_txn *txn)
1249 MDB_ID freecount, count;
1254 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1255 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1256 freecount += *(MDB_ID *)data.mv_data;
1259 for (i = 0; i<txn->mt_numdbs; i++) {
1261 mdb_cursor_init(&mc, txn, i, &mx);
1262 if (txn->mt_dbs[i].md_root == P_INVALID)
1264 count += txn->mt_dbs[i].md_branch_pages +
1265 txn->mt_dbs[i].md_leaf_pages +
1266 txn->mt_dbs[i].md_overflow_pages;
1267 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1268 mdb_page_search(&mc, NULL, 0);
1272 mp = mc.mc_pg[mc.mc_top];
1273 for (j=0; j<NUMKEYS(mp); j++) {
1274 MDB_node *leaf = NODEPTR(mp, j);
1275 if (leaf->mn_flags & F_SUBDATA) {
1277 memcpy(&db, NODEDATA(leaf), sizeof(db));
1278 count += db.md_branch_pages + db.md_leaf_pages +
1279 db.md_overflow_pages;
1283 while (mdb_cursor_sibling(&mc, 1) == 0);
1286 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1287 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1288 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1294 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1296 return txn->mt_dbxs[dbi].md_cmp(a, b);
1300 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1302 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1305 /** Allocate memory for a page.
1306 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1309 mdb_page_malloc(MDB_txn *txn, unsigned num)
1311 MDB_env *env = txn->mt_env;
1312 MDB_page *ret = env->me_dpages;
1313 size_t sz = env->me_psize;
1316 VGMEMP_ALLOC(env, ret, sz);
1317 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1318 env->me_dpages = ret->mp_next;
1324 if ((ret = malloc(sz)) != NULL) {
1325 VGMEMP_ALLOC(env, ret, sz);
1330 /** Free a single page.
1331 * Saves single pages to a list, for future reuse.
1332 * (This is not used for multi-page overflow pages.)
1335 mdb_page_free(MDB_env *env, MDB_page *mp)
1337 mp->mp_next = env->me_dpages;
1338 VGMEMP_FREE(env, mp);
1339 env->me_dpages = mp;
1342 /** Free a dirty page */
1344 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1346 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1347 mdb_page_free(env, dp);
1349 /* large pages just get freed directly */
1350 VGMEMP_FREE(env, dp);
1355 /** Return all dirty pages to dpage list */
1357 mdb_dlist_free(MDB_txn *txn)
1359 MDB_env *env = txn->mt_env;
1360 MDB_ID2L dl = txn->mt_u.dirty_list;
1361 unsigned i, n = dl[0].mid;
1363 for (i = 1; i <= n; i++) {
1364 mdb_dpage_free(env, dl[i].mptr);
1369 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1370 * @param[in] mc A cursor handle for the current operation.
1371 * @param[in] pflags Flags of the pages to update:
1372 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1373 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1374 * @return 0 on success, non-zero on failure.
1377 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1379 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1380 MDB_txn *txn = mc->mc_txn;
1386 int rc = MDB_SUCCESS, level;
1388 /* Mark pages seen by cursors */
1389 if (mc->mc_flags & C_UNTRACK)
1390 mc = NULL; /* will find mc in mt_cursors */
1391 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1392 for (; mc; mc=mc->mc_next) {
1393 if (!(mc->mc_flags & C_INITIALIZED))
1395 for (m3 = mc;; m3 = &mx->mx_cursor) {
1397 for (j=0; j<m3->mc_snum; j++) {
1399 if ((mp->mp_flags & Mask) == pflags)
1400 mp->mp_flags ^= P_KEEP;
1402 mx = m3->mc_xcursor;
1403 /* Proceed to mx if it is at a sub-database */
1404 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1406 if (! (mp && (mp->mp_flags & P_LEAF)))
1408 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1409 if (!(leaf->mn_flags & F_SUBDATA))
1418 /* Mark dirty root pages */
1419 for (i=0; i<txn->mt_numdbs; i++) {
1420 if (txn->mt_dbflags[i] & DB_DIRTY) {
1421 pgno_t pgno = txn->mt_dbs[i].md_root;
1422 if (pgno == P_INVALID)
1424 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1426 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1427 dp->mp_flags ^= P_KEEP;
1435 static int mdb_page_flush(MDB_txn *txn, int keep);
1437 /** Spill pages from the dirty list back to disk.
1438 * This is intended to prevent running into #MDB_TXN_FULL situations,
1439 * but note that they may still occur in a few cases:
1440 * 1) our estimate of the txn size could be too small. Currently this
1441 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1442 * 2) child txns may run out of space if their parents dirtied a
1443 * lot of pages and never spilled them. TODO: we probably should do
1444 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1445 * the parent's dirty_room is below a given threshold.
1447 * Otherwise, if not using nested txns, it is expected that apps will
1448 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1449 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1450 * If the txn never references them again, they can be left alone.
1451 * If the txn only reads them, they can be used without any fuss.
1452 * If the txn writes them again, they can be dirtied immediately without
1453 * going thru all of the work of #mdb_page_touch(). Such references are
1454 * handled by #mdb_page_unspill().
1456 * Also note, we never spill DB root pages, nor pages of active cursors,
1457 * because we'll need these back again soon anyway. And in nested txns,
1458 * we can't spill a page in a child txn if it was already spilled in a
1459 * parent txn. That would alter the parent txns' data even though
1460 * the child hasn't committed yet, and we'd have no way to undo it if
1461 * the child aborted.
1463 * @param[in] m0 cursor A cursor handle identifying the transaction and
1464 * database for which we are checking space.
1465 * @param[in] key For a put operation, the key being stored.
1466 * @param[in] data For a put operation, the data being stored.
1467 * @return 0 on success, non-zero on failure.
1470 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1472 MDB_txn *txn = m0->mc_txn;
1474 MDB_ID2L dl = txn->mt_u.dirty_list;
1475 unsigned int i, j, need;
1478 if (m0->mc_flags & C_SUB)
1481 /* Estimate how much space this op will take */
1482 i = m0->mc_db->md_depth;
1483 /* Named DBs also dirty the main DB */
1484 if (m0->mc_dbi > MAIN_DBI)
1485 i += txn->mt_dbs[MAIN_DBI].md_depth;
1486 /* For puts, roughly factor in the key+data size */
1488 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1489 i += i; /* double it for good measure */
1492 if (txn->mt_dirty_room > i)
1495 if (!txn->mt_spill_pgs) {
1496 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1497 if (!txn->mt_spill_pgs)
1500 /* purge deleted slots */
1501 MDB_IDL sl = txn->mt_spill_pgs;
1502 unsigned int num = sl[0];
1504 for (i=1; i<=num; i++) {
1511 /* Preserve pages which may soon be dirtied again */
1512 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1515 /* Less aggressive spill - we originally spilled the entire dirty list,
1516 * with a few exceptions for cursor pages and DB root pages. But this
1517 * turns out to be a lot of wasted effort because in a large txn many
1518 * of those pages will need to be used again. So now we spill only 1/8th
1519 * of the dirty pages. Testing revealed this to be a good tradeoff,
1520 * better than 1/2, 1/4, or 1/10.
1522 if (need < MDB_IDL_UM_MAX / 8)
1523 need = MDB_IDL_UM_MAX / 8;
1525 /* Save the page IDs of all the pages we're flushing */
1526 /* flush from the tail forward, this saves a lot of shifting later on. */
1527 for (i=dl[0].mid; i && need; i--) {
1528 MDB_ID pn = dl[i].mid << 1;
1530 if (dp->mp_flags & P_KEEP)
1532 /* Can't spill twice, make sure it's not already in a parent's
1535 if (txn->mt_parent) {
1537 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1538 if (tx2->mt_spill_pgs) {
1539 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1540 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1541 dp->mp_flags |= P_KEEP;
1549 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1553 mdb_midl_sort(txn->mt_spill_pgs);
1555 /* Flush the spilled part of dirty list */
1556 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1559 /* Reset any dirty pages we kept that page_flush didn't see */
1560 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1564 if (txn->mt_parent) {
1565 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1566 /* dirty pages that are dirty in an ancestor don't
1567 * count against this txn's dirty_room.
1569 for (i=1; i<=dl[0].mid; i++) {
1570 pgno_t pgno = dl[i].mid;
1572 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1573 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1574 if (j <= tx2->mt_u.dirty_list[0].mid &&
1575 tx2->mt_u.dirty_list[j].mid == pgno) {
1576 txn->mt_dirty_room++;
1582 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1584 txn->mt_flags |= MDB_TXN_SPILLS;
1586 txn->mt_flags |= MDB_TXN_ERROR;
1591 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1593 mdb_find_oldest(MDB_txn *txn)
1596 txnid_t mr, oldest = txn->mt_txnid - 1;
1597 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1598 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1608 /** Add a page to the txn's dirty list */
1610 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1613 int (*insert)(MDB_ID2L, MDB_ID2 *);
1615 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1616 insert = mdb_mid2l_append;
1618 insert = mdb_mid2l_insert;
1620 mid.mid = mp->mp_pgno;
1622 insert(txn->mt_u.dirty_list, &mid);
1623 txn->mt_dirty_room--;
1626 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1627 * me_pghead and mt_next_pgno.
1629 * If there are free pages available from older transactions, they
1630 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1631 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1632 * and move me_pglast to say which records were consumed. Only this
1633 * function can create me_pghead and move me_pglast/mt_next_pgno.
1634 * @param[in] mc cursor A cursor handle identifying the transaction and
1635 * database for which we are allocating.
1636 * @param[in] num the number of pages to allocate.
1637 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1638 * will always be satisfied by a single contiguous chunk of memory.
1639 * @return 0 on success, non-zero on failure.
1642 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1644 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1645 /* Get at most <Max_retries> more freeDB records once me_pghead
1646 * has enough pages. If not enough, use new pages from the map.
1647 * If <Paranoid> and mc is updating the freeDB, only get new
1648 * records if me_pghead is empty. Then the freelist cannot play
1649 * catch-up with itself by growing while trying to save it.
1651 enum { Paranoid = 1, Max_retries = 500 };
1653 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1655 int rc, n2 = num-1, retry = Max_retries;
1656 MDB_txn *txn = mc->mc_txn;
1657 MDB_env *env = txn->mt_env;
1658 pgno_t pgno, *mop = env->me_pghead;
1659 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1661 txnid_t oldest = 0, last;
1667 /* If our dirty list is already full, we can't do anything */
1668 if (txn->mt_dirty_room == 0)
1669 return MDB_TXN_FULL;
1671 for (op = MDB_FIRST;; op = MDB_NEXT) {
1674 pgno_t *idl, old_id, new_id;
1676 /* Seek a big enough contiguous page range. Prefer
1677 * pages at the tail, just truncating the list.
1679 if (mop_len >= (unsigned)num) {
1683 if (mop[i-n2] == pgno+n2)
1685 } while (--i >= (unsigned)num);
1686 if (Max_retries < INT_MAX && --retry < 0)
1690 if (op == MDB_FIRST) { /* 1st iteration */
1691 /* Prepare to fetch more and coalesce */
1692 oldest = mdb_find_oldest(txn);
1693 last = env->me_pglast;
1694 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1697 key.mv_data = &last; /* will look up last+1 */
1698 key.mv_size = sizeof(last);
1700 if (Paranoid && mc->mc_dbi == FREE_DBI)
1703 if (Paranoid && retry < 0 && mop_len)
1707 /* Do not fetch more if the record will be too recent */
1710 rc = mdb_cursor_get(&m2, &key, NULL, op);
1712 if (rc == MDB_NOTFOUND)
1716 last = *(txnid_t*)key.mv_data;
1719 np = m2.mc_pg[m2.mc_top];
1720 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1721 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1724 idl = (MDB_ID *) data.mv_data;
1727 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1730 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1732 mop = env->me_pghead;
1734 env->me_pglast = last;
1736 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1737 last, txn->mt_dbs[FREE_DBI].md_root, i));
1739 DPRINTF(("IDL %"Z"u", idl[k]));
1741 /* Merge in descending sorted order */
1744 mop[0] = (pgno_t)-1;
1748 for (; old_id < new_id; old_id = mop[--j])
1755 /* Use new pages from the map when nothing suitable in the freeDB */
1757 pgno = txn->mt_next_pgno;
1758 if (pgno + num >= env->me_maxpg) {
1759 DPUTS("DB size maxed out");
1760 return MDB_MAP_FULL;
1764 if (env->me_flags & MDB_WRITEMAP) {
1765 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1767 if (!(np = mdb_page_malloc(txn, num)))
1771 mop[0] = mop_len -= num;
1772 /* Move any stragglers down */
1773 for (j = i-num; j < mop_len; )
1774 mop[++j] = mop[++i];
1776 txn->mt_next_pgno = pgno + num;
1779 mdb_page_dirty(txn, np);
1785 /** Copy the used portions of a non-overflow page.
1786 * @param[in] dst page to copy into
1787 * @param[in] src page to copy from
1788 * @param[in] psize size of a page
1791 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1793 enum { Align = sizeof(pgno_t) };
1794 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1796 /* If page isn't full, just copy the used portion. Adjust
1797 * alignment so memcpy may copy words instead of bytes.
1799 if ((unused &= -Align) && !IS_LEAF2(src)) {
1801 memcpy(dst, src, (lower + (Align-1)) & -Align);
1802 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1805 memcpy(dst, src, psize - unused);
1809 /** Pull a page off the txn's spill list, if present.
1810 * If a page being referenced was spilled to disk in this txn, bring
1811 * it back and make it dirty/writable again.
1812 * @param[in] tx0 the transaction handle.
1813 * @param[in] mp the page being referenced.
1814 * @param[out] ret the writable page, if any. ret is unchanged if
1815 * mp wasn't spilled.
1818 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1820 MDB_env *env = tx0->mt_env;
1823 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1825 for (txn = tx0; txn; txn=txn->mt_parent) {
1826 if (!txn->mt_spill_pgs)
1828 x = mdb_midl_search(txn->mt_spill_pgs, pn);
1829 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pn) {
1832 if (IS_OVERFLOW(mp))
1836 if (env->me_flags & MDB_WRITEMAP) {
1839 np = mdb_page_malloc(txn, num);
1843 memcpy(np, mp, num * env->me_psize);
1845 mdb_page_copy(np, mp, env->me_psize);
1848 /* If in current txn, this page is no longer spilled.
1849 * If it happens to be the last page, truncate the spill list.
1850 * Otherwise mark it as deleted by setting the LSB.
1852 if (x == txn->mt_spill_pgs[0])
1853 txn->mt_spill_pgs[0]--;
1855 txn->mt_spill_pgs[x] |= 1;
1856 } /* otherwise, if belonging to a parent txn, the
1857 * page remains spilled until child commits
1860 if (txn->mt_parent) {
1862 /* If this page is also in a parent's dirty list, then
1863 * it's already accounted in dirty_room, and we need to
1864 * cancel out the decrement that mdb_page_dirty does.
1866 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1867 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1868 if (x <= tx2->mt_u.dirty_list[0].mid &&
1869 tx2->mt_u.dirty_list[x].mid == pgno) {
1870 tx0->mt_dirty_room++;
1875 mdb_page_dirty(tx0, np);
1876 np->mp_flags |= P_DIRTY;
1884 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1885 * @param[in] mc cursor pointing to the page to be touched
1886 * @return 0 on success, non-zero on failure.
1889 mdb_page_touch(MDB_cursor *mc)
1891 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1892 MDB_txn *txn = mc->mc_txn;
1893 MDB_cursor *m2, *m3;
1898 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1899 if (txn->mt_flags & MDB_TXN_SPILLS) {
1901 rc = mdb_page_unspill(txn, mp, &np);
1907 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1908 (rc = mdb_page_alloc(mc, 1, &np)))
1911 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1912 assert(mp->mp_pgno != pgno);
1913 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1914 /* Update the parent page, if any, to point to the new page */
1916 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1917 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1918 SETPGNO(node, pgno);
1920 mc->mc_db->md_root = pgno;
1922 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1923 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1925 /* If txn has a parent, make sure the page is in our
1929 unsigned x = mdb_mid2l_search(dl, pgno);
1930 if (x <= dl[0].mid && dl[x].mid == pgno) {
1931 if (mp != dl[x].mptr) { /* bad cursor? */
1932 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1933 return MDB_CORRUPTED;
1938 assert(dl[0].mid < MDB_IDL_UM_MAX);
1940 np = mdb_page_malloc(txn, 1);
1945 mdb_mid2l_insert(dl, &mid);
1950 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1952 np->mp_flags |= P_DIRTY;
1955 /* Adjust cursors pointing to mp */
1956 mc->mc_pg[mc->mc_top] = np;
1958 if (mc->mc_flags & C_SUB) {
1960 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1961 m3 = &m2->mc_xcursor->mx_cursor;
1962 if (m3->mc_snum < mc->mc_snum) continue;
1963 if (m3->mc_pg[mc->mc_top] == mp)
1964 m3->mc_pg[mc->mc_top] = np;
1967 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1968 if (m2->mc_snum < mc->mc_snum) continue;
1969 if (m2->mc_pg[mc->mc_top] == mp) {
1970 m2->mc_pg[mc->mc_top] = np;
1971 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1972 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1974 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1975 if (!(leaf->mn_flags & F_SUBDATA))
1976 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1985 mdb_env_sync(MDB_env *env, int force)
1988 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1989 if (env->me_flags & MDB_WRITEMAP) {
1990 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1991 ? MS_ASYNC : MS_SYNC;
1992 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1995 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1999 if (MDB_FDATASYNC(env->me_fd))
2006 /** Back up parent txn's cursors, then grab the originals for tracking */
2008 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2010 MDB_cursor *mc, *bk;
2015 for (i = src->mt_numdbs; --i >= 0; ) {
2016 if ((mc = src->mt_cursors[i]) != NULL) {
2017 size = sizeof(MDB_cursor);
2019 size += sizeof(MDB_xcursor);
2020 for (; mc; mc = bk->mc_next) {
2026 mc->mc_db = &dst->mt_dbs[i];
2027 /* Kill pointers into src - and dst to reduce abuse: The
2028 * user may not use mc until dst ends. Otherwise we'd...
2030 mc->mc_txn = NULL; /* ...set this to dst */
2031 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2032 if ((mx = mc->mc_xcursor) != NULL) {
2033 *(MDB_xcursor *)(bk+1) = *mx;
2034 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2036 mc->mc_next = dst->mt_cursors[i];
2037 dst->mt_cursors[i] = mc;
2044 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2045 * @param[in] txn the transaction handle.
2046 * @param[in] merge true to keep changes to parent cursors, false to revert.
2047 * @return 0 on success, non-zero on failure.
2050 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2052 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2056 for (i = txn->mt_numdbs; --i >= 0; ) {
2057 for (mc = cursors[i]; mc; mc = next) {
2059 if ((bk = mc->mc_backup) != NULL) {
2061 /* Commit changes to parent txn */
2062 mc->mc_next = bk->mc_next;
2063 mc->mc_backup = bk->mc_backup;
2064 mc->mc_txn = bk->mc_txn;
2065 mc->mc_db = bk->mc_db;
2066 mc->mc_dbflag = bk->mc_dbflag;
2067 if ((mx = mc->mc_xcursor) != NULL)
2068 mx->mx_cursor.mc_txn = bk->mc_txn;
2070 /* Abort nested txn */
2072 if ((mx = mc->mc_xcursor) != NULL)
2073 *mx = *(MDB_xcursor *)(bk+1);
2077 /* Only malloced cursors are permanently tracked. */
2085 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2088 mdb_txn_reset0(MDB_txn *txn, const char *act);
2090 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2096 Pidset = F_SETLK, Pidcheck = F_GETLK
2100 /** Set or check a pid lock. Set returns 0 on success.
2101 * Check returns 0 if the process is certainly dead, nonzero if it may
2102 * be alive (the lock exists or an error happened so we do not know).
2104 * On Windows Pidset is a no-op, we merely check for the existence
2105 * of the process with the given pid. On POSIX we use a single byte
2106 * lock on the lockfile, set at an offset equal to the pid.
2109 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2111 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2114 if (op == Pidcheck) {
2115 h = OpenProcess(env->me_pidquery, FALSE, pid);
2116 /* No documented "no such process" code, but other program use this: */
2118 return ErrCode() != ERROR_INVALID_PARAMETER;
2119 /* A process exists until all handles to it close. Has it exited? */
2120 ret = WaitForSingleObject(h, 0) != 0;
2127 struct flock lock_info;
2128 memset(&lock_info, 0, sizeof(lock_info));
2129 lock_info.l_type = F_WRLCK;
2130 lock_info.l_whence = SEEK_SET;
2131 lock_info.l_start = pid;
2132 lock_info.l_len = 1;
2133 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2134 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2136 } else if ((rc = ErrCode()) == EINTR) {
2144 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2145 * @param[in] txn the transaction handle to initialize
2146 * @return 0 on success, non-zero on failure.
2149 mdb_txn_renew0(MDB_txn *txn)
2151 MDB_env *env = txn->mt_env;
2154 int rc, new_notls = 0;
2157 txn->mt_numdbs = env->me_numdbs;
2158 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2160 if (txn->mt_flags & MDB_TXN_RDONLY) {
2161 if (!env->me_txns) {
2162 i = mdb_env_pick_meta(env);
2163 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2164 txn->mt_u.reader = NULL;
2166 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2167 pthread_getspecific(env->me_txkey);
2169 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2170 return MDB_BAD_RSLOT;
2172 pid_t pid = env->me_pid;
2173 pthread_t tid = pthread_self();
2175 if (!(env->me_flags & MDB_LIVE_READER)) {
2176 rc = mdb_reader_pid(env, Pidset, pid);
2178 UNLOCK_MUTEX_R(env);
2181 env->me_flags |= MDB_LIVE_READER;
2185 for (i=0; i<env->me_txns->mti_numreaders; i++)
2186 if (env->me_txns->mti_readers[i].mr_pid == 0)
2188 if (i == env->me_maxreaders) {
2189 UNLOCK_MUTEX_R(env);
2190 return MDB_READERS_FULL;
2192 env->me_txns->mti_readers[i].mr_pid = pid;
2193 env->me_txns->mti_readers[i].mr_tid = tid;
2194 if (i >= env->me_txns->mti_numreaders)
2195 env->me_txns->mti_numreaders = i+1;
2196 /* Save numreaders for un-mutexed mdb_env_close() */
2197 env->me_numreaders = env->me_txns->mti_numreaders;
2198 UNLOCK_MUTEX_R(env);
2199 r = &env->me_txns->mti_readers[i];
2200 new_notls = (env->me_flags & MDB_NOTLS);
2201 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2206 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2207 txn->mt_u.reader = r;
2209 txn->mt_toggle = txn->mt_txnid & 1;
2213 txn->mt_txnid = env->me_txns->mti_txnid;
2214 txn->mt_toggle = txn->mt_txnid & 1;
2217 if (txn->mt_txnid == mdb_debug_start)
2220 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2221 txn->mt_u.dirty_list = env->me_dirty_list;
2222 txn->mt_u.dirty_list[0].mid = 0;
2223 txn->mt_free_pgs = env->me_free_pgs;
2224 txn->mt_free_pgs[0] = 0;
2225 txn->mt_spill_pgs = NULL;
2229 /* Copy the DB info and flags */
2230 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2232 /* Moved to here to avoid a data race in read TXNs */
2233 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2235 for (i=2; i<txn->mt_numdbs; i++) {
2236 x = env->me_dbflags[i];
2237 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2238 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2240 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2242 if (env->me_maxpg < txn->mt_next_pgno) {
2243 mdb_txn_reset0(txn, "renew0-mapfail");
2245 txn->mt_u.reader->mr_pid = 0;
2246 txn->mt_u.reader = NULL;
2248 return MDB_MAP_RESIZED;
2255 mdb_txn_renew(MDB_txn *txn)
2259 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2262 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2263 DPUTS("environment had fatal error, must shutdown!");
2267 rc = mdb_txn_renew0(txn);
2268 if (rc == MDB_SUCCESS) {
2269 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2270 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2271 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2277 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2281 int rc, size, tsize = sizeof(MDB_txn);
2283 if (env->me_flags & MDB_FATAL_ERROR) {
2284 DPUTS("environment had fatal error, must shutdown!");
2287 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2290 /* Nested transactions: Max 1 child, write txns only, no writemap */
2291 if (parent->mt_child ||
2292 (flags & MDB_RDONLY) ||
2293 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2294 (env->me_flags & MDB_WRITEMAP))
2296 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2298 tsize = sizeof(MDB_ntxn);
2300 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2301 if (!(flags & MDB_RDONLY))
2302 size += env->me_maxdbs * sizeof(MDB_cursor *);
2304 if ((txn = calloc(1, size)) == NULL) {
2305 DPRINTF(("calloc: %s", strerror(ErrCode())));
2308 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2309 if (flags & MDB_RDONLY) {
2310 txn->mt_flags |= MDB_TXN_RDONLY;
2311 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2313 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2314 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2320 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2321 if (!txn->mt_u.dirty_list ||
2322 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2324 free(txn->mt_u.dirty_list);
2328 txn->mt_txnid = parent->mt_txnid;
2329 txn->mt_toggle = parent->mt_toggle;
2330 txn->mt_dirty_room = parent->mt_dirty_room;
2331 txn->mt_u.dirty_list[0].mid = 0;
2332 txn->mt_spill_pgs = NULL;
2333 txn->mt_next_pgno = parent->mt_next_pgno;
2334 parent->mt_child = txn;
2335 txn->mt_parent = parent;
2336 txn->mt_numdbs = parent->mt_numdbs;
2337 txn->mt_flags = parent->mt_flags;
2338 txn->mt_dbxs = parent->mt_dbxs;
2339 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2340 /* Copy parent's mt_dbflags, but clear DB_NEW */
2341 for (i=0; i<txn->mt_numdbs; i++)
2342 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2344 ntxn = (MDB_ntxn *)txn;
2345 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2346 if (env->me_pghead) {
2347 size = MDB_IDL_SIZEOF(env->me_pghead);
2348 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2350 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2355 rc = mdb_cursor_shadow(parent, txn);
2357 mdb_txn_reset0(txn, "beginchild-fail");
2359 rc = mdb_txn_renew0(txn);
2365 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2366 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2367 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2374 mdb_txn_env(MDB_txn *txn)
2376 if(!txn) return NULL;
2380 /** Export or close DBI handles opened in this txn. */
2382 mdb_dbis_update(MDB_txn *txn, int keep)
2385 MDB_dbi n = txn->mt_numdbs;
2386 MDB_env *env = txn->mt_env;
2387 unsigned char *tdbflags = txn->mt_dbflags;
2389 for (i = n; --i >= 2;) {
2390 if (tdbflags[i] & DB_NEW) {
2392 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2394 char *ptr = env->me_dbxs[i].md_name.mv_data;
2395 env->me_dbxs[i].md_name.mv_data = NULL;
2396 env->me_dbxs[i].md_name.mv_size = 0;
2397 env->me_dbflags[i] = 0;
2402 if (keep && env->me_numdbs < n)
2406 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2407 * May be called twice for readonly txns: First reset it, then abort.
2408 * @param[in] txn the transaction handle to reset
2409 * @param[in] act why the transaction is being reset
2412 mdb_txn_reset0(MDB_txn *txn, const char *act)
2414 MDB_env *env = txn->mt_env;
2416 /* Close any DBI handles opened in this txn */
2417 mdb_dbis_update(txn, 0);
2419 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2420 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2421 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2423 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2424 if (txn->mt_u.reader) {
2425 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2426 if (!(env->me_flags & MDB_NOTLS))
2427 txn->mt_u.reader = NULL; /* txn does not own reader */
2429 txn->mt_numdbs = 0; /* close nothing if called again */
2430 txn->mt_dbxs = NULL; /* mark txn as reset */
2432 mdb_cursors_close(txn, 0);
2434 if (!(env->me_flags & MDB_WRITEMAP)) {
2435 mdb_dlist_free(txn);
2437 mdb_midl_free(env->me_pghead);
2439 if (txn->mt_parent) {
2440 txn->mt_parent->mt_child = NULL;
2441 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2442 mdb_midl_free(txn->mt_free_pgs);
2443 mdb_midl_free(txn->mt_spill_pgs);
2444 free(txn->mt_u.dirty_list);
2448 if (mdb_midl_shrink(&txn->mt_free_pgs))
2449 env->me_free_pgs = txn->mt_free_pgs;
2450 env->me_pghead = NULL;
2454 /* The writer mutex was locked in mdb_txn_begin. */
2455 UNLOCK_MUTEX_W(env);
2460 mdb_txn_reset(MDB_txn *txn)
2465 /* This call is only valid for read-only txns */
2466 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2469 mdb_txn_reset0(txn, "reset");
2473 mdb_txn_abort(MDB_txn *txn)
2479 mdb_txn_abort(txn->mt_child);
2481 mdb_txn_reset0(txn, "abort");
2482 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2483 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2484 txn->mt_u.reader->mr_pid = 0;
2489 /** Save the freelist as of this transaction to the freeDB.
2490 * This changes the freelist. Keep trying until it stabilizes.
2493 mdb_freelist_save(MDB_txn *txn)
2495 /* env->me_pghead[] can grow and shrink during this call.
2496 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2497 * Page numbers cannot disappear from txn->mt_free_pgs[].
2500 MDB_env *env = txn->mt_env;
2501 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2502 txnid_t pglast = 0, head_id = 0;
2503 pgno_t freecnt = 0, *free_pgs, *mop;
2504 ssize_t head_room = 0, total_room = 0, mop_len;
2506 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2508 if (env->me_pghead) {
2509 /* Make sure first page of freeDB is touched and on freelist */
2510 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2511 if (rc && rc != MDB_NOTFOUND)
2516 /* Come back here after each Put() in case freelist changed */
2519 /* If using records from freeDB which we have not yet
2520 * deleted, delete them and any we reserved for me_pghead.
2522 while (pglast < env->me_pglast) {
2523 rc = mdb_cursor_first(&mc, &key, NULL);
2526 pglast = head_id = *(txnid_t *)key.mv_data;
2527 total_room = head_room = 0;
2528 assert(pglast <= env->me_pglast);
2529 rc = mdb_cursor_del(&mc, 0);
2534 /* Save the IDL of pages freed by this txn, to a single record */
2535 if (freecnt < txn->mt_free_pgs[0]) {
2537 /* Make sure last page of freeDB is touched and on freelist */
2538 key.mv_size = MDB_MAXKEYSIZE+1;
2540 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2541 if (rc && rc != MDB_NOTFOUND)
2544 free_pgs = txn->mt_free_pgs;
2545 /* Write to last page of freeDB */
2546 key.mv_size = sizeof(txn->mt_txnid);
2547 key.mv_data = &txn->mt_txnid;
2549 freecnt = free_pgs[0];
2550 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2551 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2554 /* Retry if mt_free_pgs[] grew during the Put() */
2555 free_pgs = txn->mt_free_pgs;
2556 } while (freecnt < free_pgs[0]);
2557 mdb_midl_sort(free_pgs);
2558 memcpy(data.mv_data, free_pgs, data.mv_size);
2561 unsigned int i = free_pgs[0];
2562 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2563 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2565 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2571 mop = env->me_pghead;
2572 mop_len = mop ? mop[0] : 0;
2574 /* Reserve records for me_pghead[]. Split it if multi-page,
2575 * to avoid searching freeDB for a page range. Use keys in
2576 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2578 if (total_room >= mop_len) {
2579 if (total_room == mop_len || --more < 0)
2581 } else if (head_room >= maxfree_1pg && head_id > 1) {
2582 /* Keep current record (overflow page), add a new one */
2586 /* (Re)write {key = head_id, IDL length = head_room} */
2587 total_room -= head_room;
2588 head_room = mop_len - total_room;
2589 if (head_room > maxfree_1pg && head_id > 1) {
2590 /* Overflow multi-page for part of me_pghead */
2591 head_room /= head_id; /* amortize page sizes */
2592 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2593 } else if (head_room < 0) {
2594 /* Rare case, not bothering to delete this record */
2597 key.mv_size = sizeof(head_id);
2598 key.mv_data = &head_id;
2599 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2600 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2603 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2604 total_room += head_room;
2607 /* Fill in the reserved me_pghead records */
2613 rc = mdb_cursor_first(&mc, &key, &data);
2614 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2615 unsigned flags = MDB_CURRENT;
2616 txnid_t id = *(txnid_t *)key.mv_data;
2617 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2620 assert(len >= 0 && id <= env->me_pglast);
2622 if (len > mop_len) {
2624 data.mv_size = (len + 1) * sizeof(MDB_ID);
2627 data.mv_data = mop -= len;
2630 rc = mdb_cursor_put(&mc, &key, &data, flags);
2632 if (rc || !(mop_len -= len))
2639 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2640 * @param[in] txn the transaction that's being committed
2641 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2642 * @return 0 on success, non-zero on failure.
2645 mdb_page_flush(MDB_txn *txn, int keep)
2647 MDB_env *env = txn->mt_env;
2648 MDB_ID2L dl = txn->mt_u.dirty_list;
2649 unsigned psize = env->me_psize, j;
2650 int i, pagecount = dl[0].mid, rc;
2651 size_t size = 0, pos = 0;
2653 MDB_page *dp = NULL;
2657 struct iovec iov[MDB_COMMIT_PAGES];
2658 ssize_t wpos = 0, wsize = 0, wres;
2659 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2665 if (env->me_flags & MDB_WRITEMAP) {
2666 /* Clear dirty flags */
2667 while (++i <= pagecount) {
2669 /* Don't flush this page yet */
2670 if (dp->mp_flags & P_KEEP) {
2671 dp->mp_flags ^= P_KEEP;
2675 dp->mp_flags &= ~P_DIRTY;
2681 /* Write the pages */
2683 if (++i <= pagecount) {
2685 /* Don't flush this page yet */
2686 if (dp->mp_flags & P_KEEP) {
2687 dp->mp_flags ^= P_KEEP;
2692 /* clear dirty flag */
2693 dp->mp_flags &= ~P_DIRTY;
2696 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2701 /* Windows actually supports scatter/gather I/O, but only on
2702 * unbuffered file handles. Since we're relying on the OS page
2703 * cache for all our data, that's self-defeating. So we just
2704 * write pages one at a time. We use the ov structure to set
2705 * the write offset, to at least save the overhead of a Seek
2708 DPRINTF(("committing page %"Z"u", pgno));
2709 memset(&ov, 0, sizeof(ov));
2710 ov.Offset = pos & 0xffffffff;
2711 ov.OffsetHigh = pos >> 16 >> 16;
2712 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2714 DPRINTF(("WriteFile: %d", rc));
2718 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2719 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2721 /* Write previous page(s) */
2722 #ifdef MDB_USE_PWRITEV
2723 wres = pwritev(env->me_fd, iov, n, wpos);
2726 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2728 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2730 DPRINTF(("lseek: %s", strerror(rc)));
2733 wres = writev(env->me_fd, iov, n);
2736 if (wres != wsize) {
2739 DPRINTF(("Write error: %s", strerror(rc)));
2741 rc = EIO; /* TODO: Use which error code? */
2742 DPUTS("short write, filesystem full?");
2753 DPRINTF(("committing page %"Z"u", pgno));
2754 next_pos = pos + size;
2755 iov[n].iov_len = size;
2756 iov[n].iov_base = (char *)dp;
2762 for (i = keep; ++i <= pagecount; ) {
2764 /* This is a page we skipped above */
2767 dl[j].mid = dp->mp_pgno;
2770 mdb_dpage_free(env, dp);
2778 mdb_txn_commit(MDB_txn *txn)
2784 assert(txn != NULL);
2785 assert(txn->mt_env != NULL);
2787 if (txn->mt_child) {
2788 rc = mdb_txn_commit(txn->mt_child);
2789 txn->mt_child = NULL;
2796 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2797 mdb_dbis_update(txn, 1);
2798 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2803 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2804 DPUTS("error flag is set, can't commit");
2806 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2811 if (txn->mt_parent) {
2812 MDB_txn *parent = txn->mt_parent;
2816 /* Append our free list to parent's */
2817 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2820 mdb_midl_free(txn->mt_free_pgs);
2822 parent->mt_next_pgno = txn->mt_next_pgno;
2823 parent->mt_flags = txn->mt_flags;
2825 /* Merge our cursors into parent's and close them */
2826 mdb_cursors_close(txn, 1);
2828 /* Update parent's DB table. */
2829 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2830 parent->mt_numdbs = txn->mt_numdbs;
2831 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2832 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2833 for (i=2; i<txn->mt_numdbs; i++) {
2834 /* preserve parent's DB_NEW status */
2835 x = parent->mt_dbflags[i] & DB_NEW;
2836 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2839 dst = parent->mt_u.dirty_list;
2840 src = txn->mt_u.dirty_list;
2841 /* Remove anything in our dirty list from parent's spill list */
2842 if (parent->mt_spill_pgs) {
2843 x = parent->mt_spill_pgs[0];
2845 /* zero out our dirty pages in parent spill list */
2846 for (i=1; i<=src[0].mid; i++) {
2847 MDB_ID pn = src[i].mid << 1;
2848 if (pn < parent->mt_spill_pgs[x])
2850 if (pn > parent->mt_spill_pgs[x]) {
2856 parent->mt_spill_pgs[x] = 0;
2859 /* OK, we had a few hits, squash zeros from the spill list */
2860 if (len < parent->mt_spill_pgs[0]) {
2862 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2863 if (parent->mt_spill_pgs[y]) {
2865 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2870 parent->mt_spill_pgs[0] = len;
2873 /* Find len = length of merging our dirty list with parent's */
2875 dst[0].mid = 0; /* simplify loops */
2876 if (parent->mt_parent) {
2877 len = x + src[0].mid;
2878 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2879 for (i = x; y && i; y--) {
2880 pgno_t yp = src[y].mid;
2881 while (yp < dst[i].mid)
2883 if (yp == dst[i].mid) {
2888 } else { /* Simplify the above for single-ancestor case */
2889 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2891 /* Merge our dirty list with parent's */
2893 for (i = len; y; dst[i--] = src[y--]) {
2894 pgno_t yp = src[y].mid;
2895 while (yp < dst[x].mid)
2896 dst[i--] = dst[x--];
2897 if (yp == dst[x].mid)
2898 free(dst[x--].mptr);
2902 free(txn->mt_u.dirty_list);
2903 parent->mt_dirty_room = txn->mt_dirty_room;
2904 if (txn->mt_spill_pgs) {
2905 if (parent->mt_spill_pgs) {
2906 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2907 mdb_midl_free(txn->mt_spill_pgs);
2908 mdb_midl_sort(parent->mt_spill_pgs);
2910 parent->mt_spill_pgs = txn->mt_spill_pgs;
2914 parent->mt_child = NULL;
2915 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2920 if (txn != env->me_txn) {
2921 DPUTS("attempt to commit unknown transaction");
2926 mdb_cursors_close(txn, 0);
2928 if (!txn->mt_u.dirty_list[0].mid &&
2929 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2932 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2933 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2935 /* Update DB root pointers */
2936 if (txn->mt_numdbs > 2) {
2940 data.mv_size = sizeof(MDB_db);
2942 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2943 for (i = 2; i < txn->mt_numdbs; i++) {
2944 if (txn->mt_dbflags[i] & DB_DIRTY) {
2945 data.mv_data = &txn->mt_dbs[i];
2946 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2953 rc = mdb_freelist_save(txn);
2957 mdb_midl_free(env->me_pghead);
2958 env->me_pghead = NULL;
2959 if (mdb_midl_shrink(&txn->mt_free_pgs))
2960 env->me_free_pgs = txn->mt_free_pgs;
2966 if ((rc = mdb_page_flush(txn, 0)) ||
2967 (rc = mdb_env_sync(env, 0)) ||
2968 (rc = mdb_env_write_meta(txn)))
2974 mdb_dbis_update(txn, 1);
2976 UNLOCK_MUTEX_W(env);
2986 /** Read the environment parameters of a DB environment before
2987 * mapping it into memory.
2988 * @param[in] env the environment handle
2989 * @param[out] meta address of where to store the meta information
2990 * @return 0 on success, non-zero on failure.
2993 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3000 /* We don't know the page size yet, so use a minimum value.
3001 * Read both meta pages so we can use the latest one.
3004 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3008 memset(&ov, 0, sizeof(ov));
3010 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
3011 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3014 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
3016 if (rc != MDB_PAGESIZE) {
3017 if (rc == 0 && off == 0)
3019 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3020 DPRINTF(("read: %s", mdb_strerror(rc)));
3024 p = (MDB_page *)&pbuf;
3026 if (!F_ISSET(p->mp_flags, P_META)) {
3027 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3032 if (m->mm_magic != MDB_MAGIC) {
3033 DPUTS("meta has invalid magic");
3037 if (m->mm_version != MDB_DATA_VERSION) {
3038 DPRINTF(("database is version %u, expected version %u",
3039 m->mm_version, MDB_DATA_VERSION));
3040 return MDB_VERSION_MISMATCH;
3043 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3049 /** Write the environment parameters of a freshly created DB environment.
3050 * @param[in] env the environment handle
3051 * @param[out] meta address of where to store the meta information
3052 * @return 0 on success, non-zero on failure.
3055 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3063 memset(&ov, 0, sizeof(ov));
3064 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3066 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3069 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3070 len = pwrite(fd, ptr, size, pos); \
3071 rc = (len >= 0); } while(0)
3074 DPUTS("writing new meta page");
3076 psize = env->me_psize;
3078 meta->mm_magic = MDB_MAGIC;
3079 meta->mm_version = MDB_DATA_VERSION;
3080 meta->mm_mapsize = env->me_mapsize;
3081 meta->mm_psize = psize;
3082 meta->mm_last_pg = 1;
3083 meta->mm_flags = env->me_flags & 0xffff;
3084 meta->mm_flags |= MDB_INTEGERKEY;
3085 meta->mm_dbs[0].md_root = P_INVALID;
3086 meta->mm_dbs[1].md_root = P_INVALID;
3088 p = calloc(2, psize);
3090 p->mp_flags = P_META;
3091 *(MDB_meta *)METADATA(p) = *meta;
3093 q = (MDB_page *)((char *)p + psize);
3095 q->mp_flags = P_META;
3096 *(MDB_meta *)METADATA(q) = *meta;
3098 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3101 else if ((unsigned) len == psize * 2)
3109 /** Update the environment info to commit a transaction.
3110 * @param[in] txn the transaction that's being committed
3111 * @return 0 on success, non-zero on failure.
3114 mdb_env_write_meta(MDB_txn *txn)
3117 MDB_meta meta, metab, *mp;
3119 int rc, len, toggle;
3128 assert(txn != NULL);
3129 assert(txn->mt_env != NULL);
3131 toggle = !txn->mt_toggle;
3132 DPRINTF(("writing meta page %d for root page %"Z"u",
3133 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3136 mp = env->me_metas[toggle];
3138 if (env->me_flags & MDB_WRITEMAP) {
3139 /* Persist any increases of mapsize config */
3140 if (env->me_mapsize > mp->mm_mapsize)
3141 mp->mm_mapsize = env->me_mapsize;
3142 mp->mm_dbs[0] = txn->mt_dbs[0];
3143 mp->mm_dbs[1] = txn->mt_dbs[1];
3144 mp->mm_last_pg = txn->mt_next_pgno - 1;
3145 mp->mm_txnid = txn->mt_txnid;
3146 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3147 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3150 ptr += env->me_psize;
3151 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3158 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3159 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3161 ptr = (char *)&meta;
3162 if (env->me_mapsize > mp->mm_mapsize) {
3163 /* Persist any increases of mapsize config */
3164 meta.mm_mapsize = env->me_mapsize;
3165 off = offsetof(MDB_meta, mm_mapsize);
3167 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3169 len = sizeof(MDB_meta) - off;
3172 meta.mm_dbs[0] = txn->mt_dbs[0];
3173 meta.mm_dbs[1] = txn->mt_dbs[1];
3174 meta.mm_last_pg = txn->mt_next_pgno - 1;
3175 meta.mm_txnid = txn->mt_txnid;
3178 off += env->me_psize;
3181 /* Write to the SYNC fd */
3182 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3183 env->me_fd : env->me_mfd;
3186 memset(&ov, 0, sizeof(ov));
3188 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3192 rc = pwrite(mfd, ptr, len, off);
3195 rc = rc < 0 ? ErrCode() : EIO;
3196 DPUTS("write failed, disk error?");
3197 /* On a failure, the pagecache still contains the new data.
3198 * Write some old data back, to prevent it from being used.
3199 * Use the non-SYNC fd; we know it will fail anyway.
3201 meta.mm_last_pg = metab.mm_last_pg;
3202 meta.mm_txnid = metab.mm_txnid;
3204 memset(&ov, 0, sizeof(ov));
3206 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3208 r2 = pwrite(env->me_fd, ptr, len, off);
3209 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3212 env->me_flags |= MDB_FATAL_ERROR;
3216 /* Memory ordering issues are irrelevant; since the entire writer
3217 * is wrapped by wmutex, all of these changes will become visible
3218 * after the wmutex is unlocked. Since the DB is multi-version,
3219 * readers will get consistent data regardless of how fresh or
3220 * how stale their view of these values is.
3222 env->me_txns->mti_txnid = txn->mt_txnid;
3227 /** Check both meta pages to see which one is newer.
3228 * @param[in] env the environment handle
3229 * @return meta toggle (0 or 1).
3232 mdb_env_pick_meta(const MDB_env *env)
3234 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3238 mdb_env_create(MDB_env **env)
3242 e = calloc(1, sizeof(MDB_env));
3246 e->me_maxreaders = DEFAULT_READERS;
3247 e->me_maxdbs = e->me_numdbs = 2;
3248 e->me_fd = INVALID_HANDLE_VALUE;
3249 e->me_lfd = INVALID_HANDLE_VALUE;
3250 e->me_mfd = INVALID_HANDLE_VALUE;
3251 #ifdef MDB_USE_POSIX_SEM
3252 e->me_rmutex = SEM_FAILED;
3253 e->me_wmutex = SEM_FAILED;
3255 e->me_pid = getpid();
3256 VGMEMP_CREATE(e,0,0);
3262 mdb_env_map(MDB_env *env, void *addr, int newsize)
3265 unsigned int flags = env->me_flags;
3269 LONG sizelo, sizehi;
3270 sizelo = env->me_mapsize & 0xffffffff;
3271 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3273 /* Windows won't create mappings for zero length files.
3274 * Just allocate the maxsize right now.
3277 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3278 || !SetEndOfFile(env->me_fd)
3279 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3282 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3283 PAGE_READWRITE : PAGE_READONLY,
3284 sizehi, sizelo, NULL);
3287 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3288 FILE_MAP_WRITE : FILE_MAP_READ,
3289 0, 0, env->me_mapsize, addr);
3290 rc = env->me_map ? 0 : ErrCode();
3295 int prot = PROT_READ;
3296 if (flags & MDB_WRITEMAP) {
3298 if (newsize && ftruncate(env->me_fd, env->me_mapsize) < 0)
3301 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3303 if (env->me_map == MAP_FAILED) {
3307 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3309 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3311 #ifdef POSIX_MADV_RANDOM
3312 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3313 #endif /* POSIX_MADV_RANDOM */
3314 #endif /* MADV_RANDOM */
3317 /* Can happen because the address argument to mmap() is just a
3318 * hint. mmap() can pick another, e.g. if the range is in use.
3319 * The MAP_FIXED flag would prevent that, but then mmap could
3320 * instead unmap existing pages to make room for the new map.
3322 if (addr && env->me_map != addr)
3323 return EBUSY; /* TODO: Make a new MDB_* error code? */
3325 p = (MDB_page *)env->me_map;
3326 env->me_metas[0] = METADATA(p);
3327 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3333 mdb_env_set_mapsize(MDB_env *env, size_t size)
3335 /* If env is already open, caller is responsible for making
3336 * sure there are no active txns.
3344 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3345 munmap(env->me_map, env->me_mapsize);
3346 env->me_mapsize = size;
3347 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3348 rc = mdb_env_map(env, old, 1);
3352 env->me_mapsize = size;
3354 env->me_maxpg = env->me_mapsize / env->me_psize;
3359 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3363 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3368 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3370 if (env->me_map || readers < 1)
3372 env->me_maxreaders = readers;
3377 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3379 if (!env || !readers)
3381 *readers = env->me_maxreaders;
3385 /** Further setup required for opening an MDB environment
3388 mdb_env_open2(MDB_env *env)
3390 unsigned int flags = env->me_flags;
3391 int i, newenv = 0, rc;
3395 /* See if we should use QueryLimited */
3397 if ((rc & 0xff) > 5)
3398 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3400 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3403 memset(&meta, 0, sizeof(meta));
3405 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3408 DPUTS("new mdbenv");
3410 GET_PAGESIZE(env->me_psize);
3412 env->me_psize = meta.mm_psize;
3415 /* Was a mapsize configured? */
3416 if (!env->me_mapsize) {
3417 /* If this is a new environment, take the default,
3418 * else use the size recorded in the existing env.
3420 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3421 } else if (env->me_mapsize < meta.mm_mapsize) {
3422 /* If the configured size is smaller, make sure it's
3423 * still big enough. Silently round up to minimum if not.
3425 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3426 if (env->me_mapsize < minsize)
3427 env->me_mapsize = minsize;
3430 rc = mdb_env_map(env, meta.mm_address, newenv);
3435 if (flags & MDB_FIXEDMAP)
3436 meta.mm_address = env->me_map;
3437 i = mdb_env_init_meta(env, &meta);
3438 if (i != MDB_SUCCESS) {
3442 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3443 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3445 env->me_maxpg = env->me_mapsize / env->me_psize;
3448 int toggle = mdb_env_pick_meta(env);
3449 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3451 DPRINTF(("opened database version %u, pagesize %u",
3452 env->me_metas[0]->mm_version, env->me_psize));
3453 DPRINTF(("using meta page %d", toggle));
3454 DPRINTF(("depth: %u", db->md_depth));
3455 DPRINTF(("entries: %"Z"u", db->md_entries));
3456 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3457 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3458 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3459 DPRINTF(("root: %"Z"u", db->md_root));
3467 /** Release a reader thread's slot in the reader lock table.
3468 * This function is called automatically when a thread exits.
3469 * @param[in] ptr This points to the slot in the reader lock table.
3472 mdb_env_reader_dest(void *ptr)
3474 MDB_reader *reader = ptr;
3480 /** Junk for arranging thread-specific callbacks on Windows. This is
3481 * necessarily platform and compiler-specific. Windows supports up
3482 * to 1088 keys. Let's assume nobody opens more than 64 environments
3483 * in a single process, for now. They can override this if needed.
3485 #ifndef MAX_TLS_KEYS
3486 #define MAX_TLS_KEYS 64
3488 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3489 static int mdb_tls_nkeys;
3491 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3495 case DLL_PROCESS_ATTACH: break;
3496 case DLL_THREAD_ATTACH: break;
3497 case DLL_THREAD_DETACH:
3498 for (i=0; i<mdb_tls_nkeys; i++) {
3499 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3500 mdb_env_reader_dest(r);
3503 case DLL_PROCESS_DETACH: break;
3508 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3510 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3514 /* Force some symbol references.
3515 * _tls_used forces the linker to create the TLS directory if not already done
3516 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3518 #pragma comment(linker, "/INCLUDE:_tls_used")
3519 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3520 #pragma const_seg(".CRT$XLB")
3521 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3522 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3525 #pragma comment(linker, "/INCLUDE:__tls_used")
3526 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3527 #pragma data_seg(".CRT$XLB")
3528 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3530 #endif /* WIN 32/64 */
3531 #endif /* !__GNUC__ */
3534 /** Downgrade the exclusive lock on the region back to shared */
3536 mdb_env_share_locks(MDB_env *env, int *excl)
3538 int rc = 0, toggle = mdb_env_pick_meta(env);
3540 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3545 /* First acquire a shared lock. The Unlock will
3546 * then release the existing exclusive lock.
3548 memset(&ov, 0, sizeof(ov));
3549 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3552 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3558 struct flock lock_info;
3559 /* The shared lock replaces the existing lock */
3560 memset((void *)&lock_info, 0, sizeof(lock_info));
3561 lock_info.l_type = F_RDLCK;
3562 lock_info.l_whence = SEEK_SET;
3563 lock_info.l_start = 0;
3564 lock_info.l_len = 1;
3565 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3566 (rc = ErrCode()) == EINTR) ;
3567 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3574 /** Try to get exlusive lock, otherwise shared.
3575 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3578 mdb_env_excl_lock(MDB_env *env, int *excl)
3582 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3586 memset(&ov, 0, sizeof(ov));
3587 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3594 struct flock lock_info;
3595 memset((void *)&lock_info, 0, sizeof(lock_info));
3596 lock_info.l_type = F_WRLCK;
3597 lock_info.l_whence = SEEK_SET;
3598 lock_info.l_start = 0;
3599 lock_info.l_len = 1;
3600 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3601 (rc = ErrCode()) == EINTR) ;
3605 # ifdef MDB_USE_POSIX_SEM
3606 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3609 lock_info.l_type = F_RDLCK;
3610 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3611 (rc = ErrCode()) == EINTR) ;
3619 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3621 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3623 * @(#) $Revision: 5.1 $
3624 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3625 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3627 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3631 * Please do not copyright this code. This code is in the public domain.
3633 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3634 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3635 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3636 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3637 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3638 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3639 * PERFORMANCE OF THIS SOFTWARE.
3642 * chongo <Landon Curt Noll> /\oo/\
3643 * http://www.isthe.com/chongo/
3645 * Share and Enjoy! :-)
3648 typedef unsigned long long mdb_hash_t;
3649 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3651 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3652 * @param[in] val value to hash
3653 * @param[in] hval initial value for hash
3654 * @return 64 bit hash
3656 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3657 * hval arg on the first call.
3660 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3662 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3663 unsigned char *end = s + val->mv_size;
3665 * FNV-1a hash each octet of the string
3668 /* xor the bottom with the current octet */
3669 hval ^= (mdb_hash_t)*s++;
3671 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3672 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3673 (hval << 7) + (hval << 8) + (hval << 40);
3675 /* return our new hash value */
3679 /** Hash the string and output the encoded hash.
3680 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3681 * very short name limits. We don't care about the encoding being reversible,
3682 * we just want to preserve as many bits of the input as possible in a
3683 * small printable string.
3684 * @param[in] str string to hash
3685 * @param[out] encbuf an array of 11 chars to hold the hash
3687 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3690 mdb_pack85(unsigned long l, char *out)
3694 for (i=0; i<5; i++) {
3695 *out++ = mdb_a85[l % 85];
3701 mdb_hash_enc(MDB_val *val, char *encbuf)
3703 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3705 mdb_pack85(h, encbuf);
3706 mdb_pack85(h>>32, encbuf+5);
3711 /** Open and/or initialize the lock region for the environment.
3712 * @param[in] env The MDB environment.
3713 * @param[in] lpath The pathname of the file used for the lock region.
3714 * @param[in] mode The Unix permissions for the file, if we create it.
3715 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3716 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3717 * @return 0 on success, non-zero on failure.
3720 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3723 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3725 # define MDB_ERRCODE_ROFS EROFS
3726 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3727 # define MDB_CLOEXEC O_CLOEXEC
3730 # define MDB_CLOEXEC 0
3737 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3738 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3739 FILE_ATTRIBUTE_NORMAL, NULL);
3741 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3743 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3745 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3750 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3751 /* Lose record locks when exec*() */
3752 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3753 fcntl(env->me_lfd, F_SETFD, fdflags);
3756 if (!(env->me_flags & MDB_NOTLS)) {
3757 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3760 env->me_flags |= MDB_ENV_TXKEY;
3762 /* Windows TLS callbacks need help finding their TLS info. */
3763 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3767 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3771 /* Try to get exclusive lock. If we succeed, then
3772 * nobody is using the lock region and we should initialize it.
3774 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3777 size = GetFileSize(env->me_lfd, NULL);
3779 size = lseek(env->me_lfd, 0, SEEK_END);
3780 if (size == -1) goto fail_errno;
3782 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3783 if (size < rsize && *excl > 0) {
3785 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3786 || !SetEndOfFile(env->me_lfd))
3789 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3793 size = rsize - sizeof(MDB_txninfo);
3794 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3799 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3801 if (!mh) goto fail_errno;
3802 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3804 if (!env->me_txns) goto fail_errno;
3806 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3808 if (m == MAP_FAILED) goto fail_errno;
3814 BY_HANDLE_FILE_INFORMATION stbuf;
3823 if (!mdb_sec_inited) {
3824 InitializeSecurityDescriptor(&mdb_null_sd,
3825 SECURITY_DESCRIPTOR_REVISION);
3826 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3827 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3828 mdb_all_sa.bInheritHandle = FALSE;
3829 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3832 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3833 idbuf.volume = stbuf.dwVolumeSerialNumber;
3834 idbuf.nhigh = stbuf.nFileIndexHigh;
3835 idbuf.nlow = stbuf.nFileIndexLow;
3836 val.mv_data = &idbuf;
3837 val.mv_size = sizeof(idbuf);
3838 mdb_hash_enc(&val, encbuf);
3839 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3840 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3841 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3842 if (!env->me_rmutex) goto fail_errno;
3843 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3844 if (!env->me_wmutex) goto fail_errno;
3845 #elif defined(MDB_USE_POSIX_SEM)
3854 #if defined(__NetBSD__)
3855 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3857 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3858 idbuf.dev = stbuf.st_dev;
3859 idbuf.ino = stbuf.st_ino;
3860 val.mv_data = &idbuf;
3861 val.mv_size = sizeof(idbuf);
3862 mdb_hash_enc(&val, encbuf);
3863 #ifdef MDB_SHORT_SEMNAMES
3864 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3866 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3867 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3868 /* Clean up after a previous run, if needed: Try to
3869 * remove both semaphores before doing anything else.
3871 sem_unlink(env->me_txns->mti_rmname);
3872 sem_unlink(env->me_txns->mti_wmname);
3873 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3874 O_CREAT|O_EXCL, mode, 1);
3875 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3876 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3877 O_CREAT|O_EXCL, mode, 1);
3878 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3879 #else /* MDB_USE_POSIX_SEM */
3880 pthread_mutexattr_t mattr;
3882 if ((rc = pthread_mutexattr_init(&mattr))
3883 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3884 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3885 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3887 pthread_mutexattr_destroy(&mattr);
3888 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3890 env->me_txns->mti_magic = MDB_MAGIC;
3891 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3892 env->me_txns->mti_txnid = 0;
3893 env->me_txns->mti_numreaders = 0;
3896 if (env->me_txns->mti_magic != MDB_MAGIC) {
3897 DPUTS("lock region has invalid magic");
3901 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3902 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3903 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3904 rc = MDB_VERSION_MISMATCH;
3908 if (rc && rc != EACCES && rc != EAGAIN) {
3912 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3913 if (!env->me_rmutex) goto fail_errno;
3914 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3915 if (!env->me_wmutex) goto fail_errno;
3916 #elif defined(MDB_USE_POSIX_SEM)
3917 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3918 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3919 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3920 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3931 /** The name of the lock file in the DB environment */
3932 #define LOCKNAME "/lock.mdb"
3933 /** The name of the data file in the DB environment */
3934 #define DATANAME "/data.mdb"
3935 /** The suffix of the lock file when no subdir is used */
3936 #define LOCKSUFF "-lock"
3937 /** Only a subset of the @ref mdb_env flags can be changed
3938 * at runtime. Changing other flags requires closing the
3939 * environment and re-opening it with the new flags.
3941 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3942 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3945 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3947 int oflags, rc, len, excl = -1;
3948 char *lpath, *dpath;
3950 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3954 if (flags & MDB_NOSUBDIR) {
3955 rc = len + sizeof(LOCKSUFF) + len + 1;
3957 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3962 if (flags & MDB_NOSUBDIR) {
3963 dpath = lpath + len + sizeof(LOCKSUFF);
3964 sprintf(lpath, "%s" LOCKSUFF, path);
3965 strcpy(dpath, path);
3967 dpath = lpath + len + sizeof(LOCKNAME);
3968 sprintf(lpath, "%s" LOCKNAME, path);
3969 sprintf(dpath, "%s" DATANAME, path);
3973 flags |= env->me_flags;
3974 if (flags & MDB_RDONLY) {
3975 /* silently ignore WRITEMAP when we're only getting read access */
3976 flags &= ~MDB_WRITEMAP;
3978 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3979 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3982 env->me_flags = flags |= MDB_ENV_ACTIVE;
3986 env->me_path = strdup(path);
3987 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3988 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3989 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3994 /* For RDONLY, get lockfile after we know datafile exists */
3995 if (!F_ISSET(flags, MDB_RDONLY)) {
3996 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4002 if (F_ISSET(flags, MDB_RDONLY)) {
4003 oflags = GENERIC_READ;
4004 len = OPEN_EXISTING;
4006 oflags = GENERIC_READ|GENERIC_WRITE;
4009 mode = FILE_ATTRIBUTE_NORMAL;
4010 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4011 NULL, len, mode, NULL);
4013 if (F_ISSET(flags, MDB_RDONLY))
4016 oflags = O_RDWR | O_CREAT;
4018 env->me_fd = open(dpath, oflags, mode);
4020 if (env->me_fd == INVALID_HANDLE_VALUE) {
4025 if (F_ISSET(flags, MDB_RDONLY)) {
4026 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4031 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4032 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4033 env->me_mfd = env->me_fd;
4035 /* Synchronous fd for meta writes. Needed even with
4036 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4039 len = OPEN_EXISTING;
4040 env->me_mfd = CreateFile(dpath, oflags,
4041 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4042 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4045 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4047 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4052 DPRINTF(("opened dbenv %p", (void *) env));
4054 rc = mdb_env_share_locks(env, &excl);
4060 mdb_env_close0(env, excl);
4066 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4068 mdb_env_close0(MDB_env *env, int excl)
4072 if (!(env->me_flags & MDB_ENV_ACTIVE))
4075 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4076 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4077 free(env->me_dbxs[i].md_name.mv_data);
4079 free(env->me_dbflags);
4082 free(env->me_dirty_list);
4083 mdb_midl_free(env->me_free_pgs);
4085 if (env->me_flags & MDB_ENV_TXKEY) {
4086 pthread_key_delete(env->me_txkey);
4088 /* Delete our key from the global list */
4089 for (i=0; i<mdb_tls_nkeys; i++)
4090 if (mdb_tls_keys[i] == env->me_txkey) {
4091 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4099 munmap(env->me_map, env->me_mapsize);
4101 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4102 (void) close(env->me_mfd);
4103 if (env->me_fd != INVALID_HANDLE_VALUE)
4104 (void) close(env->me_fd);
4106 pid_t pid = env->me_pid;
4107 /* Clearing readers is done in this function because
4108 * me_txkey with its destructor must be disabled first.
4110 for (i = env->me_numreaders; --i >= 0; )
4111 if (env->me_txns->mti_readers[i].mr_pid == pid)
4112 env->me_txns->mti_readers[i].mr_pid = 0;
4114 if (env->me_rmutex) {
4115 CloseHandle(env->me_rmutex);
4116 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4118 /* Windows automatically destroys the mutexes when
4119 * the last handle closes.
4121 #elif defined(MDB_USE_POSIX_SEM)
4122 if (env->me_rmutex != SEM_FAILED) {
4123 sem_close(env->me_rmutex);
4124 if (env->me_wmutex != SEM_FAILED)
4125 sem_close(env->me_wmutex);
4126 /* If we have the filelock: If we are the
4127 * only remaining user, clean up semaphores.
4130 mdb_env_excl_lock(env, &excl);
4132 sem_unlink(env->me_txns->mti_rmname);
4133 sem_unlink(env->me_txns->mti_wmname);
4137 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4139 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4142 /* Unlock the lockfile. Windows would have unlocked it
4143 * after closing anyway, but not necessarily at once.
4145 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4148 (void) close(env->me_lfd);
4151 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4155 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4157 MDB_txn *txn = NULL;
4163 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4167 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4170 /* Do the lock/unlock of the reader mutex before starting the
4171 * write txn. Otherwise other read txns could block writers.
4173 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4178 /* We must start the actual read txn after blocking writers */
4179 mdb_txn_reset0(txn, "reset-stage1");
4181 /* Temporarily block writers until we snapshot the meta pages */
4184 rc = mdb_txn_renew0(txn);
4186 UNLOCK_MUTEX_W(env);
4191 wsize = env->me_psize * 2;
4195 DO_WRITE(rc, fd, ptr, w2, len);
4199 } else if (len > 0) {
4205 /* Non-blocking or async handles are not supported */
4211 UNLOCK_MUTEX_W(env);
4216 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4218 if (wsize > MAX_WRITE)
4222 DO_WRITE(rc, fd, ptr, w2, len);
4226 } else if (len > 0) {
4243 mdb_env_copy(MDB_env *env, const char *path)
4247 HANDLE newfd = INVALID_HANDLE_VALUE;
4249 if (env->me_flags & MDB_NOSUBDIR) {
4250 lpath = (char *)path;
4253 len += sizeof(DATANAME);
4254 lpath = malloc(len);
4257 sprintf(lpath, "%s" DATANAME, path);
4260 /* The destination path must exist, but the destination file must not.
4261 * We don't want the OS to cache the writes, since the source data is
4262 * already in the OS cache.
4265 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4266 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4268 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4270 if (newfd == INVALID_HANDLE_VALUE) {
4276 /* Set O_DIRECT if the file system supports it */
4277 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4278 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4280 #ifdef F_NOCACHE /* __APPLE__ */
4281 rc = fcntl(newfd, F_NOCACHE, 1);
4288 rc = mdb_env_copyfd(env, newfd);
4291 if (!(env->me_flags & MDB_NOSUBDIR))
4293 if (newfd != INVALID_HANDLE_VALUE)
4294 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4301 mdb_env_close(MDB_env *env)
4308 VGMEMP_DESTROY(env);
4309 while ((dp = env->me_dpages) != NULL) {
4310 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4311 env->me_dpages = dp->mp_next;
4315 mdb_env_close0(env, 0);
4319 /** Compare two items pointing at aligned size_t's */
4321 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4323 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4324 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4327 /** Compare two items pointing at aligned unsigned int's */
4329 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4331 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4332 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4335 /** Compare two items pointing at unsigned ints of unknown alignment.
4336 * Nodes and keys are guaranteed to be 2-byte aligned.
4339 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4341 #if BYTE_ORDER == LITTLE_ENDIAN
4342 unsigned short *u, *c;
4345 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4346 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4349 } while(!x && u > (unsigned short *)a->mv_data);
4352 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4356 /** Compare two items lexically */
4358 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4365 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4371 diff = memcmp(a->mv_data, b->mv_data, len);
4372 return diff ? diff : len_diff<0 ? -1 : len_diff;
4375 /** Compare two items in reverse byte order */
4377 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4379 const unsigned char *p1, *p2, *p1_lim;
4383 p1_lim = (const unsigned char *)a->mv_data;
4384 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4385 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4387 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4393 while (p1 > p1_lim) {
4394 diff = *--p1 - *--p2;
4398 return len_diff<0 ? -1 : len_diff;
4401 /** Search for key within a page, using binary search.
4402 * Returns the smallest entry larger or equal to the key.
4403 * If exactp is non-null, stores whether the found entry was an exact match
4404 * in *exactp (1 or 0).
4405 * Updates the cursor index with the index of the found entry.
4406 * If no entry larger or equal to the key is found, returns NULL.
4409 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4411 unsigned int i = 0, nkeys;
4414 MDB_page *mp = mc->mc_pg[mc->mc_top];
4415 MDB_node *node = NULL;
4420 nkeys = NUMKEYS(mp);
4425 COPY_PGNO(pgno, mp->mp_pgno);
4426 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4427 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4434 low = IS_LEAF(mp) ? 0 : 1;
4436 cmp = mc->mc_dbx->md_cmp;
4438 /* Branch pages have no data, so if using integer keys,
4439 * alignment is guaranteed. Use faster mdb_cmp_int.
4441 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4442 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4449 nodekey.mv_size = mc->mc_db->md_pad;
4450 node = NODEPTR(mp, 0); /* fake */
4451 while (low <= high) {
4452 i = (low + high) >> 1;
4453 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4454 rc = cmp(key, &nodekey);
4455 DPRINTF(("found leaf index %u [%s], rc = %i",
4456 i, DKEY(&nodekey), rc));
4465 while (low <= high) {
4466 i = (low + high) >> 1;
4468 node = NODEPTR(mp, i);
4469 nodekey.mv_size = NODEKSZ(node);
4470 nodekey.mv_data = NODEKEY(node);
4472 rc = cmp(key, &nodekey);
4475 DPRINTF(("found leaf index %u [%s], rc = %i",
4476 i, DKEY(&nodekey), rc));
4478 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4479 i, DKEY(&nodekey), NODEPGNO(node), rc));
4490 if (rc > 0) { /* Found entry is less than the key. */
4491 i++; /* Skip to get the smallest entry larger than key. */
4493 node = NODEPTR(mp, i);
4496 *exactp = (rc == 0);
4497 /* store the key index */
4498 mc->mc_ki[mc->mc_top] = i;
4500 /* There is no entry larger or equal to the key. */
4503 /* nodeptr is fake for LEAF2 */
4509 mdb_cursor_adjust(MDB_cursor *mc, func)
4513 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4514 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4521 /** Pop a page off the top of the cursor's stack. */
4523 mdb_cursor_pop(MDB_cursor *mc)
4527 MDB_page *top = mc->mc_pg[mc->mc_top];
4533 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4534 mc->mc_dbi, (void *) mc));
4538 /** Push a page onto the top of the cursor's stack. */
4540 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4542 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4543 mc->mc_dbi, (void *) mc));
4545 if (mc->mc_snum >= CURSOR_STACK) {
4546 assert(mc->mc_snum < CURSOR_STACK);
4547 return MDB_CURSOR_FULL;
4550 mc->mc_top = mc->mc_snum++;
4551 mc->mc_pg[mc->mc_top] = mp;
4552 mc->mc_ki[mc->mc_top] = 0;
4557 /** Find the address of the page corresponding to a given page number.
4558 * @param[in] txn the transaction for this access.
4559 * @param[in] pgno the page number for the page to retrieve.
4560 * @param[out] ret address of a pointer where the page's address will be stored.
4561 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4562 * @return 0 on success, non-zero on failure.
4565 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4567 MDB_env *env = txn->mt_env;
4571 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4575 MDB_ID2L dl = tx2->mt_u.dirty_list;
4577 /* Spilled pages were dirtied in this txn and flushed
4578 * because the dirty list got full. Bring this page
4579 * back in from the map (but don't unspill it here,
4580 * leave that unless page_touch happens again).
4582 if (tx2->mt_spill_pgs) {
4583 MDB_ID pn = pgno << 1;
4584 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4585 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4586 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4591 unsigned x = mdb_mid2l_search(dl, pgno);
4592 if (x <= dl[0].mid && dl[x].mid == pgno) {
4598 } while ((tx2 = tx2->mt_parent) != NULL);
4601 if (pgno < txn->mt_next_pgno) {
4603 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4605 DPRINTF(("page %"Z"u not found", pgno));
4607 return MDB_PAGE_NOTFOUND;
4617 /** Search for the page a given key should be in.
4618 * Pushes parent pages on the cursor stack. This function continues a
4619 * search on a cursor that has already been initialized. (Usually by
4620 * #mdb_page_search() but also by #mdb_node_move().)
4621 * @param[in,out] mc the cursor for this operation.
4622 * @param[in] key the key to search for. If NULL, search for the lowest
4623 * page. (This is used by #mdb_cursor_first().)
4624 * @param[in] modify If true, visited pages are updated with new page numbers.
4625 * @return 0 on success, non-zero on failure.
4628 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4630 MDB_page *mp = mc->mc_pg[mc->mc_top];
4634 while (IS_BRANCH(mp)) {
4638 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4639 assert(NUMKEYS(mp) > 1);
4640 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4642 if (key == NULL) /* Initialize cursor to first page. */
4644 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4645 /* cursor to last page */
4649 node = mdb_node_search(mc, key, &exact);
4651 i = NUMKEYS(mp) - 1;
4653 i = mc->mc_ki[mc->mc_top];
4662 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4663 assert(i < NUMKEYS(mp));
4664 node = NODEPTR(mp, i);
4666 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4669 mc->mc_ki[mc->mc_top] = i;
4670 if ((rc = mdb_cursor_push(mc, mp)))
4674 if ((rc = mdb_page_touch(mc)) != 0)
4676 mp = mc->mc_pg[mc->mc_top];
4681 DPRINTF(("internal error, index points to a %02X page!?",
4683 return MDB_CORRUPTED;
4686 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4687 key ? DKEY(key) : NULL));
4688 mc->mc_flags |= C_INITIALIZED;
4689 mc->mc_flags &= ~C_EOF;
4694 /** Search for the lowest key under the current branch page.
4695 * This just bypasses a NUMKEYS check in the current page
4696 * before calling mdb_page_search_root(), because the callers
4697 * are all in situations where the current page is known to
4701 mdb_page_search_lowest(MDB_cursor *mc)
4703 MDB_page *mp = mc->mc_pg[mc->mc_top];
4704 MDB_node *node = NODEPTR(mp, 0);
4707 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4710 mc->mc_ki[mc->mc_top] = 0;
4711 if ((rc = mdb_cursor_push(mc, mp)))
4713 return mdb_page_search_root(mc, NULL, 0);
4716 /** Search for the page a given key should be in.
4717 * Pushes parent pages on the cursor stack. This function just sets up
4718 * the search; it finds the root page for \b mc's database and sets this
4719 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4720 * called to complete the search.
4721 * @param[in,out] mc the cursor for this operation.
4722 * @param[in] key the key to search for. If NULL, search for the lowest
4723 * page. (This is used by #mdb_cursor_first().)
4724 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4725 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4726 * @return 0 on success, non-zero on failure.
4729 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4734 /* Make sure the txn is still viable, then find the root from
4735 * the txn's db table.
4737 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4738 DPUTS("transaction has failed, must abort");
4741 /* Make sure we're using an up-to-date root */
4742 if (mc->mc_dbi > MAIN_DBI) {
4743 if ((*mc->mc_dbflag & DB_STALE) ||
4744 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4746 unsigned char dbflag = 0;
4747 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4748 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4751 if (*mc->mc_dbflag & DB_STALE) {
4755 MDB_node *leaf = mdb_node_search(&mc2,
4756 &mc->mc_dbx->md_name, &exact);
4758 return MDB_NOTFOUND;
4759 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4762 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4764 /* The txn may not know this DBI, or another process may
4765 * have dropped and recreated the DB with other flags.
4767 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4768 return MDB_INCOMPATIBLE;
4769 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4771 if (flags & MDB_PS_MODIFY)
4773 *mc->mc_dbflag &= ~DB_STALE;
4774 *mc->mc_dbflag |= dbflag;
4777 root = mc->mc_db->md_root;
4779 if (root == P_INVALID) { /* Tree is empty. */
4780 DPUTS("tree is empty");
4781 return MDB_NOTFOUND;
4786 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4787 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4793 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4794 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4796 if (flags & MDB_PS_MODIFY) {
4797 if ((rc = mdb_page_touch(mc)))
4801 if (flags & MDB_PS_ROOTONLY)
4804 return mdb_page_search_root(mc, key, flags);
4808 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4810 MDB_txn *txn = mc->mc_txn;
4811 pgno_t pg = mp->mp_pgno;
4812 unsigned x = 0, ovpages = mp->mp_pages;
4813 MDB_env *env = txn->mt_env;
4814 MDB_IDL sl = txn->mt_spill_pgs;
4815 MDB_ID pn = pg << 1;
4818 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4819 /* If the page is dirty or on the spill list we just acquired it,
4820 * so we should give it back to our current free list, if any.
4821 * Otherwise put it onto the list of pages we freed in this txn.
4823 * Won't create me_pghead: me_pglast must be inited along with it.
4824 * Unsupported in nested txns: They would need to hide the page
4825 * range in ancestor txns' dirty and spilled lists.
4827 if (env->me_pghead &&
4829 ((mp->mp_flags & P_DIRTY) ||
4830 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4834 MDB_ID2 *dl, ix, iy;
4835 rc = mdb_midl_need(&env->me_pghead, ovpages);
4838 if (!(mp->mp_flags & P_DIRTY)) {
4839 /* This page is no longer spilled */
4846 /* Remove from dirty list */
4847 dl = txn->mt_u.dirty_list;
4849 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4857 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4858 txn->mt_flags |= MDB_TXN_ERROR;
4859 return MDB_CORRUPTED;
4862 if (!(env->me_flags & MDB_WRITEMAP))
4863 mdb_dpage_free(env, mp);
4865 /* Insert in me_pghead */
4866 mop = env->me_pghead;
4867 j = mop[0] + ovpages;
4868 for (i = mop[0]; i && mop[i] < pg; i--)
4874 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4878 mc->mc_db->md_overflow_pages -= ovpages;
4882 /** Return the data associated with a given node.
4883 * @param[in] txn The transaction for this operation.
4884 * @param[in] leaf The node being read.
4885 * @param[out] data Updated to point to the node's data.
4886 * @return 0 on success, non-zero on failure.
4889 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4891 MDB_page *omp; /* overflow page */
4895 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4896 data->mv_size = NODEDSZ(leaf);
4897 data->mv_data = NODEDATA(leaf);
4901 /* Read overflow data.
4903 data->mv_size = NODEDSZ(leaf);
4904 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4905 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4906 DPRINTF(("read overflow page %"Z"u failed", pgno));
4909 data->mv_data = METADATA(omp);
4915 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4916 MDB_val *key, MDB_val *data)
4925 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4927 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4930 if (txn->mt_flags & MDB_TXN_ERROR)
4933 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4934 return MDB_BAD_VALSIZE;
4937 mdb_cursor_init(&mc, txn, dbi, &mx);
4938 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4941 /** Find a sibling for a page.
4942 * Replaces the page at the top of the cursor's stack with the
4943 * specified sibling, if one exists.
4944 * @param[in] mc The cursor for this operation.
4945 * @param[in] move_right Non-zero if the right sibling is requested,
4946 * otherwise the left sibling.
4947 * @return 0 on success, non-zero on failure.
4950 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4956 if (mc->mc_snum < 2) {
4957 return MDB_NOTFOUND; /* root has no siblings */
4961 DPRINTF(("parent page is page %"Z"u, index %u",
4962 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4964 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4965 : (mc->mc_ki[mc->mc_top] == 0)) {
4966 DPRINTF(("no more keys left, moving to %s sibling",
4967 move_right ? "right" : "left"));
4968 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4969 /* undo cursor_pop before returning */
4976 mc->mc_ki[mc->mc_top]++;
4978 mc->mc_ki[mc->mc_top]--;
4979 DPRINTF(("just moving to %s index key %u",
4980 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4982 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4984 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4985 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4988 mdb_cursor_push(mc, mp);
4990 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4995 /** Move the cursor to the next data item. */
4997 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5003 if (mc->mc_flags & C_EOF) {
5004 return MDB_NOTFOUND;
5007 assert(mc->mc_flags & C_INITIALIZED);
5009 mp = mc->mc_pg[mc->mc_top];
5011 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5012 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5013 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5014 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5015 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5016 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5017 if (rc == MDB_SUCCESS)
5018 MDB_GET_KEY(leaf, key);
5023 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5024 if (op == MDB_NEXT_DUP)
5025 return MDB_NOTFOUND;
5029 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5030 if (mc->mc_flags & C_DEL)
5033 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5034 DPUTS("=====> move to next sibling page");
5035 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5036 mc->mc_flags |= C_EOF;
5039 mp = mc->mc_pg[mc->mc_top];
5040 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5042 mc->mc_ki[mc->mc_top]++;
5045 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5046 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5049 key->mv_size = mc->mc_db->md_pad;
5050 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5054 assert(IS_LEAF(mp));
5055 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5057 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5058 mdb_xcursor_init1(mc, leaf);
5061 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5064 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5065 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5066 if (rc != MDB_SUCCESS)
5071 MDB_GET_KEY(leaf, key);
5075 /** Move the cursor to the previous data item. */
5077 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5083 assert(mc->mc_flags & C_INITIALIZED);
5085 mp = mc->mc_pg[mc->mc_top];
5087 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5088 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5089 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5090 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5091 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5092 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5093 if (rc == MDB_SUCCESS)
5094 MDB_GET_KEY(leaf, key);
5098 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5099 if (op == MDB_PREV_DUP)
5100 return MDB_NOTFOUND;
5105 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5107 if (mc->mc_ki[mc->mc_top] == 0) {
5108 DPUTS("=====> move to prev sibling page");
5109 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5112 mp = mc->mc_pg[mc->mc_top];
5113 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5114 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5116 mc->mc_ki[mc->mc_top]--;
5118 mc->mc_flags &= ~C_EOF;
5120 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5121 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5124 key->mv_size = mc->mc_db->md_pad;
5125 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5129 assert(IS_LEAF(mp));
5130 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5132 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5133 mdb_xcursor_init1(mc, leaf);
5136 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5139 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5140 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5141 if (rc != MDB_SUCCESS)
5146 MDB_GET_KEY(leaf, key);
5150 /** Set the cursor on a specific data item. */
5152 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5153 MDB_cursor_op op, int *exactp)
5157 MDB_node *leaf = NULL;
5162 assert(key->mv_size > 0);
5165 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5167 /* See if we're already on the right page */
5168 if (mc->mc_flags & C_INITIALIZED) {
5171 mp = mc->mc_pg[mc->mc_top];
5173 mc->mc_ki[mc->mc_top] = 0;
5174 return MDB_NOTFOUND;
5176 if (mp->mp_flags & P_LEAF2) {
5177 nodekey.mv_size = mc->mc_db->md_pad;
5178 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5180 leaf = NODEPTR(mp, 0);
5181 MDB_GET_KEY2(leaf, nodekey);
5183 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5185 /* Probably happens rarely, but first node on the page
5186 * was the one we wanted.
5188 mc->mc_ki[mc->mc_top] = 0;
5195 unsigned int nkeys = NUMKEYS(mp);
5197 if (mp->mp_flags & P_LEAF2) {
5198 nodekey.mv_data = LEAF2KEY(mp,
5199 nkeys-1, nodekey.mv_size);
5201 leaf = NODEPTR(mp, nkeys-1);
5202 MDB_GET_KEY2(leaf, nodekey);
5204 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5206 /* last node was the one we wanted */
5207 mc->mc_ki[mc->mc_top] = nkeys-1;
5213 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5214 /* This is definitely the right page, skip search_page */
5215 if (mp->mp_flags & P_LEAF2) {
5216 nodekey.mv_data = LEAF2KEY(mp,
5217 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5219 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5220 MDB_GET_KEY2(leaf, nodekey);
5222 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5224 /* current node was the one we wanted */
5234 /* If any parents have right-sibs, search.
5235 * Otherwise, there's nothing further.
5237 for (i=0; i<mc->mc_top; i++)
5239 NUMKEYS(mc->mc_pg[i])-1)
5241 if (i == mc->mc_top) {
5242 /* There are no other pages */
5243 mc->mc_ki[mc->mc_top] = nkeys;
5244 return MDB_NOTFOUND;
5248 /* There are no other pages */
5249 mc->mc_ki[mc->mc_top] = 0;
5250 if (op == MDB_SET_RANGE) {
5254 return MDB_NOTFOUND;
5258 rc = mdb_page_search(mc, key, 0);
5259 if (rc != MDB_SUCCESS)
5262 mp = mc->mc_pg[mc->mc_top];
5263 assert(IS_LEAF(mp));
5266 leaf = mdb_node_search(mc, key, exactp);
5267 if (exactp != NULL && !*exactp) {
5268 /* MDB_SET specified and not an exact match. */
5269 return MDB_NOTFOUND;
5273 DPUTS("===> inexact leaf not found, goto sibling");
5274 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5275 return rc; /* no entries matched */
5276 mp = mc->mc_pg[mc->mc_top];
5277 assert(IS_LEAF(mp));
5278 leaf = NODEPTR(mp, 0);
5282 mc->mc_flags |= C_INITIALIZED;
5283 mc->mc_flags &= ~C_EOF;
5286 key->mv_size = mc->mc_db->md_pad;
5287 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5291 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5292 mdb_xcursor_init1(mc, leaf);
5295 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5296 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5297 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5300 if (op == MDB_GET_BOTH) {
5306 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5307 if (rc != MDB_SUCCESS)
5310 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5312 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5314 rc = mc->mc_dbx->md_dcmp(data, &d2);
5316 if (op == MDB_GET_BOTH || rc > 0)
5317 return MDB_NOTFOUND;
5323 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5324 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5329 /* The key already matches in all other cases */
5330 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5331 MDB_GET_KEY(leaf, key);
5332 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5337 /** Move the cursor to the first item in the database. */
5339 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5345 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5347 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5348 rc = mdb_page_search(mc, NULL, 0);
5349 if (rc != MDB_SUCCESS)
5352 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5354 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5355 mc->mc_flags |= C_INITIALIZED;
5356 mc->mc_flags &= ~C_EOF;
5358 mc->mc_ki[mc->mc_top] = 0;
5360 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5361 key->mv_size = mc->mc_db->md_pad;
5362 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5367 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5368 mdb_xcursor_init1(mc, leaf);
5369 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5373 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5377 MDB_GET_KEY(leaf, key);
5381 /** Move the cursor to the last item in the database. */
5383 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5389 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5391 if (!(mc->mc_flags & C_EOF)) {
5393 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5396 lkey.mv_size = MDB_MAXKEYSIZE+1;
5397 lkey.mv_data = NULL;
5398 rc = mdb_page_search(mc, &lkey, 0);
5399 if (rc != MDB_SUCCESS)
5402 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5405 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5406 mc->mc_flags |= C_INITIALIZED|C_EOF;
5407 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5409 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5410 key->mv_size = mc->mc_db->md_pad;
5411 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5416 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5417 mdb_xcursor_init1(mc, leaf);
5418 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5422 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5427 MDB_GET_KEY(leaf, key);
5432 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5437 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5441 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5445 case MDB_GET_CURRENT:
5446 if (!(mc->mc_flags & C_INITIALIZED)) {
5449 MDB_page *mp = mc->mc_pg[mc->mc_top];
5451 mc->mc_ki[mc->mc_top] = 0;
5457 key->mv_size = mc->mc_db->md_pad;
5458 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5460 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5461 MDB_GET_KEY(leaf, key);
5463 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5464 if (mc->mc_flags & C_DEL)
5465 mdb_xcursor_init1(mc, leaf);
5466 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5468 rc = mdb_node_read(mc->mc_txn, leaf, data);
5475 case MDB_GET_BOTH_RANGE:
5480 if (mc->mc_xcursor == NULL) {
5481 rc = MDB_INCOMPATIBLE;
5490 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5491 rc = MDB_BAD_VALSIZE;
5492 } else if (op == MDB_SET_RANGE)
5493 rc = mdb_cursor_set(mc, key, data, op, NULL);
5495 rc = mdb_cursor_set(mc, key, data, op, &exact);
5497 case MDB_GET_MULTIPLE:
5498 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5502 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5503 rc = MDB_INCOMPATIBLE;
5507 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5508 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5511 case MDB_NEXT_MULTIPLE:
5516 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5517 rc = MDB_INCOMPATIBLE;
5520 if (!(mc->mc_flags & C_INITIALIZED))
5521 rc = mdb_cursor_first(mc, key, data);
5523 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5524 if (rc == MDB_SUCCESS) {
5525 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5528 mx = &mc->mc_xcursor->mx_cursor;
5529 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5531 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5532 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5540 case MDB_NEXT_NODUP:
5541 if (!(mc->mc_flags & C_INITIALIZED))
5542 rc = mdb_cursor_first(mc, key, data);
5544 rc = mdb_cursor_next(mc, key, data, op);
5548 case MDB_PREV_NODUP:
5549 if (!(mc->mc_flags & C_INITIALIZED)) {
5550 rc = mdb_cursor_last(mc, key, data);
5553 mc->mc_flags |= C_INITIALIZED;
5554 mc->mc_ki[mc->mc_top]++;
5556 rc = mdb_cursor_prev(mc, key, data, op);
5559 rc = mdb_cursor_first(mc, key, data);
5562 mfunc = mdb_cursor_first;
5564 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5568 if (mc->mc_xcursor == NULL) {
5569 rc = MDB_INCOMPATIBLE;
5572 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5576 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5579 rc = mdb_cursor_last(mc, key, data);
5582 mfunc = mdb_cursor_last;
5585 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5590 if (mc->mc_flags & C_DEL)
5591 mc->mc_flags ^= C_DEL;
5596 /** Touch all the pages in the cursor stack.
5597 * Makes sure all the pages are writable, before attempting a write operation.
5598 * @param[in] mc The cursor to operate on.
5601 mdb_cursor_touch(MDB_cursor *mc)
5605 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5608 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5609 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5612 *mc->mc_dbflag |= DB_DIRTY;
5614 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5615 rc = mdb_page_touch(mc);
5619 mc->mc_top = mc->mc_snum-1;
5623 /** Do not spill pages to disk if txn is getting full, may fail instead */
5624 #define MDB_NOSPILL 0x8000
5627 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5630 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5631 MDB_node *leaf = NULL;
5632 MDB_val xdata, *rdata, dkey;
5635 int do_sub = 0, insert = 0;
5636 unsigned int mcount = 0, dcount = 0, nospill;
5640 char dbuf[MDB_MAXKEYSIZE+1];
5641 unsigned int nflags;
5644 /* Check this first so counter will always be zero on any
5647 if (flags & MDB_MULTIPLE) {
5648 dcount = data[1].mv_size;
5649 data[1].mv_size = 0;
5650 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5651 return MDB_INCOMPATIBLE;
5654 nospill = flags & MDB_NOSPILL;
5655 flags &= ~MDB_NOSPILL;
5657 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5658 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5660 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5661 return MDB_BAD_VALSIZE;
5663 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5664 return MDB_BAD_VALSIZE;
5666 #if SIZE_MAX > MAXDATASIZE
5667 if (data->mv_size > MAXDATASIZE)
5668 return MDB_BAD_VALSIZE;
5671 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5672 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5676 if (flags == MDB_CURRENT) {
5677 if (!(mc->mc_flags & C_INITIALIZED))
5680 } else if (mc->mc_db->md_root == P_INVALID) {
5681 /* new database, cursor has nothing to point to */
5683 mc->mc_flags &= ~C_INITIALIZED;
5688 if (flags & MDB_APPEND) {
5690 rc = mdb_cursor_last(mc, &k2, &d2);
5692 rc = mc->mc_dbx->md_cmp(key, &k2);
5695 mc->mc_ki[mc->mc_top]++;
5697 /* new key is <= last key */
5702 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5704 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5705 DPRINTF(("duplicate key [%s]", DKEY(key)));
5707 return MDB_KEYEXIST;
5709 if (rc && rc != MDB_NOTFOUND)
5713 if (mc->mc_flags & C_DEL)
5714 mc->mc_flags ^= C_DEL;
5716 /* Cursor is positioned, check for room in the dirty list */
5718 if (flags & MDB_MULTIPLE) {
5720 xdata.mv_size = data->mv_size * dcount;
5724 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5728 if (rc == MDB_NO_ROOT) {
5730 /* new database, write a root leaf page */
5731 DPUTS("allocating new root leaf page");
5732 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5735 mdb_cursor_push(mc, np);
5736 mc->mc_db->md_root = np->mp_pgno;
5737 mc->mc_db->md_depth++;
5738 *mc->mc_dbflag |= DB_DIRTY;
5739 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5741 np->mp_flags |= P_LEAF2;
5742 mc->mc_flags |= C_INITIALIZED;
5744 /* make sure all cursor pages are writable */
5745 rc2 = mdb_cursor_touch(mc);
5750 /* The key already exists */
5751 if (rc == MDB_SUCCESS) {
5752 /* there's only a key anyway, so this is a no-op */
5753 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5754 unsigned int ksize = mc->mc_db->md_pad;
5755 if (key->mv_size != ksize)
5756 return MDB_BAD_VALSIZE;
5757 if (flags == MDB_CURRENT) {
5758 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5759 memcpy(ptr, key->mv_data, ksize);
5764 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5767 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5768 /* Was a single item before, must convert now */
5770 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5771 /* Just overwrite the current item */
5772 if (flags == MDB_CURRENT)
5775 dkey.mv_size = NODEDSZ(leaf);
5776 dkey.mv_data = NODEDATA(leaf);
5777 #if UINT_MAX < SIZE_MAX
5778 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5779 #ifdef MISALIGNED_OK
5780 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5782 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5785 /* if data matches, skip it */
5786 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5787 if (flags & MDB_NODUPDATA)
5789 else if (flags & MDB_MULTIPLE)
5796 /* create a fake page for the dup items */
5797 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5798 dkey.mv_data = dbuf;
5799 fp = (MDB_page *)&pbuf;
5800 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5801 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5802 fp->mp_lower = PAGEHDRSZ;
5803 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5804 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5805 fp->mp_flags |= P_LEAF2;
5806 fp->mp_pad = data->mv_size;
5807 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5809 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5810 (dkey.mv_size & 1) + (data->mv_size & 1);
5812 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5815 xdata.mv_size = fp->mp_upper;
5820 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5821 /* See if we need to convert from fake page to subDB */
5823 unsigned int offset;
5827 fp = NODEDATA(leaf);
5828 if (flags == MDB_CURRENT) {
5830 fp->mp_flags |= P_DIRTY;
5831 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5832 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5836 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5837 offset = fp->mp_pad;
5838 if (SIZELEFT(fp) >= offset)
5840 offset *= 4; /* space for 4 more */
5842 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5844 offset += offset & 1;
5845 fp_flags = fp->mp_flags;
5846 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5847 offset >= mc->mc_txn->mt_env->me_nodemax) {
5848 /* yes, convert it */
5850 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5851 dummy.md_pad = fp->mp_pad;
5852 dummy.md_flags = MDB_DUPFIXED;
5853 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5854 dummy.md_flags |= MDB_INTEGERKEY;
5857 dummy.md_branch_pages = 0;
5858 dummy.md_leaf_pages = 1;
5859 dummy.md_overflow_pages = 0;
5860 dummy.md_entries = NUMKEYS(fp);
5862 xdata.mv_size = sizeof(MDB_db);
5863 xdata.mv_data = &dummy;
5864 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5866 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5867 flags |= F_DUPDATA|F_SUBDATA;
5868 dummy.md_root = mp->mp_pgno;
5869 fp_flags &= ~P_SUBP;
5871 /* no, just grow it */
5873 xdata.mv_size = NODEDSZ(leaf) + offset;
5874 xdata.mv_data = &pbuf;
5875 mp = (MDB_page *)&pbuf;
5876 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5879 mp->mp_flags = fp_flags | P_DIRTY;
5880 mp->mp_pad = fp->mp_pad;
5881 mp->mp_lower = fp->mp_lower;
5882 mp->mp_upper = fp->mp_upper + offset;
5884 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5886 nsize = NODEDSZ(leaf) - fp->mp_upper;
5887 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5888 for (i=0; i<NUMKEYS(fp); i++)
5889 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5891 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5895 /* data is on sub-DB, just store it */
5896 flags |= F_DUPDATA|F_SUBDATA;
5900 /* overflow page overwrites need special handling */
5901 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5904 unsigned psize = mc->mc_txn->mt_env->me_psize;
5905 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5907 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5908 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5910 ovpages = omp->mp_pages;
5912 /* Is the ov page large enough? */
5913 if (ovpages >= dpages) {
5914 if (!(omp->mp_flags & P_DIRTY) &&
5915 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5917 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5920 level = 0; /* dirty in this txn or clean */
5923 if (omp->mp_flags & P_DIRTY) {
5924 /* yes, overwrite it. Note in this case we don't
5925 * bother to try shrinking the page if the new data
5926 * is smaller than the overflow threshold.
5929 /* It is writable only in a parent txn */
5930 size_t sz = (size_t) psize * ovpages, off;
5931 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5937 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5938 if (!(flags & MDB_RESERVE)) {
5939 /* Copy end of page, adjusting alignment so
5940 * compiler may copy words instead of bytes.
5942 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5943 memcpy((size_t *)((char *)np + off),
5944 (size_t *)((char *)omp + off), sz - off);
5947 memcpy(np, omp, sz); /* Copy beginning of page */
5950 SETDSZ(leaf, data->mv_size);
5951 if (F_ISSET(flags, MDB_RESERVE))
5952 data->mv_data = METADATA(omp);
5954 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5958 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5960 } else if (NODEDSZ(leaf) == data->mv_size) {
5961 /* same size, just replace it. Note that we could
5962 * also reuse this node if the new data is smaller,
5963 * but instead we opt to shrink the node in that case.
5965 if (F_ISSET(flags, MDB_RESERVE))
5966 data->mv_data = NODEDATA(leaf);
5967 else if (data->mv_size)
5968 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5970 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5973 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5974 mc->mc_db->md_entries--;
5976 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5983 nflags = flags & NODE_ADD_FLAGS;
5984 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5985 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5986 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5987 nflags &= ~MDB_APPEND;
5989 nflags |= MDB_SPLIT_REPLACE;
5990 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5992 /* There is room already in this leaf page. */
5993 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5994 if (rc == 0 && !do_sub && insert) {
5995 /* Adjust other cursors pointing to mp */
5996 MDB_cursor *m2, *m3;
5997 MDB_dbi dbi = mc->mc_dbi;
5998 unsigned i = mc->mc_top;
5999 MDB_page *mp = mc->mc_pg[i];
6001 if (mc->mc_flags & C_SUB)
6004 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6005 if (mc->mc_flags & C_SUB)
6006 m3 = &m2->mc_xcursor->mx_cursor;
6009 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6010 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6017 if (rc != MDB_SUCCESS)
6018 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6020 /* Now store the actual data in the child DB. Note that we're
6021 * storing the user data in the keys field, so there are strict
6022 * size limits on dupdata. The actual data fields of the child
6023 * DB are all zero size.
6030 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6031 if (flags & MDB_CURRENT) {
6032 xflags = MDB_CURRENT|MDB_NOSPILL;
6034 mdb_xcursor_init1(mc, leaf);
6035 xflags = (flags & MDB_NODUPDATA) ?
6036 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6038 /* converted, write the original data first */
6040 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6044 /* Adjust other cursors pointing to mp */
6046 unsigned i = mc->mc_top;
6047 MDB_page *mp = mc->mc_pg[i];
6049 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6050 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6051 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6052 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6053 mdb_xcursor_init1(m2, leaf);
6057 /* we've done our job */
6060 if (flags & MDB_APPENDDUP)
6061 xflags |= MDB_APPEND;
6062 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6063 if (flags & F_SUBDATA) {
6064 void *db = NODEDATA(leaf);
6065 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6068 /* sub-writes might have failed so check rc again.
6069 * Don't increment count if we just replaced an existing item.
6071 if (!rc && !(flags & MDB_CURRENT))
6072 mc->mc_db->md_entries++;
6073 if (flags & MDB_MULTIPLE) {
6077 /* let caller know how many succeeded, if any */
6078 data[1].mv_size = mcount;
6079 if (mcount < dcount) {
6080 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6081 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6088 /* If we succeeded and the key didn't exist before, make sure
6089 * the cursor is marked valid.
6092 mc->mc_flags |= C_INITIALIZED;
6097 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6102 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6103 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6105 if (!(mc->mc_flags & C_INITIALIZED))
6108 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6110 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
6112 rc = mdb_cursor_touch(mc);
6116 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6118 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6119 if (!(flags & MDB_NODUPDATA)) {
6120 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6121 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6123 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6124 /* If sub-DB still has entries, we're done */
6125 if (mc->mc_xcursor->mx_db.md_entries) {
6126 if (leaf->mn_flags & F_SUBDATA) {
6127 /* update subDB info */
6128 void *db = NODEDATA(leaf);
6129 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6132 /* shrink fake page */
6133 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6134 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6135 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6136 /* fix other sub-DB cursors pointed at this fake page */
6137 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6138 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6139 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6140 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6141 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6144 mc->mc_db->md_entries--;
6145 mc->mc_flags |= C_DEL;
6148 /* otherwise fall thru and delete the sub-DB */
6151 if (leaf->mn_flags & F_SUBDATA) {
6152 /* add all the child DB's pages to the free list */
6153 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6154 if (rc == MDB_SUCCESS) {
6155 mc->mc_db->md_entries -=
6156 mc->mc_xcursor->mx_db.md_entries;
6161 return mdb_cursor_del0(mc, leaf);
6164 /** Allocate and initialize new pages for a database.
6165 * @param[in] mc a cursor on the database being added to.
6166 * @param[in] flags flags defining what type of page is being allocated.
6167 * @param[in] num the number of pages to allocate. This is usually 1,
6168 * unless allocating overflow pages for a large record.
6169 * @param[out] mp Address of a page, or NULL on failure.
6170 * @return 0 on success, non-zero on failure.
6173 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6178 if ((rc = mdb_page_alloc(mc, num, &np)))
6180 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6181 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6182 np->mp_flags = flags | P_DIRTY;
6183 np->mp_lower = PAGEHDRSZ;
6184 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6187 mc->mc_db->md_branch_pages++;
6188 else if (IS_LEAF(np))
6189 mc->mc_db->md_leaf_pages++;
6190 else if (IS_OVERFLOW(np)) {
6191 mc->mc_db->md_overflow_pages += num;
6199 /** Calculate the size of a leaf node.
6200 * The size depends on the environment's page size; if a data item
6201 * is too large it will be put onto an overflow page and the node
6202 * size will only include the key and not the data. Sizes are always
6203 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6204 * of the #MDB_node headers.
6205 * @param[in] env The environment handle.
6206 * @param[in] key The key for the node.
6207 * @param[in] data The data for the node.
6208 * @return The number of bytes needed to store the node.
6211 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6215 sz = LEAFSIZE(key, data);
6216 if (sz >= env->me_nodemax) {
6217 /* put on overflow page */
6218 sz -= data->mv_size - sizeof(pgno_t);
6222 return sz + sizeof(indx_t);
6225 /** Calculate the size of a branch node.
6226 * The size should depend on the environment's page size but since
6227 * we currently don't support spilling large keys onto overflow
6228 * pages, it's simply the size of the #MDB_node header plus the
6229 * size of the key. Sizes are always rounded up to an even number
6230 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6231 * @param[in] env The environment handle.
6232 * @param[in] key The key for the node.
6233 * @return The number of bytes needed to store the node.
6236 mdb_branch_size(MDB_env *env, MDB_val *key)
6241 if (sz >= env->me_nodemax) {
6242 /* put on overflow page */
6243 /* not implemented */
6244 /* sz -= key->size - sizeof(pgno_t); */
6247 return sz + sizeof(indx_t);
6250 /** Add a node to the page pointed to by the cursor.
6251 * @param[in] mc The cursor for this operation.
6252 * @param[in] indx The index on the page where the new node should be added.
6253 * @param[in] key The key for the new node.
6254 * @param[in] data The data for the new node, if any.
6255 * @param[in] pgno The page number, if adding a branch node.
6256 * @param[in] flags Flags for the node.
6257 * @return 0 on success, non-zero on failure. Possible errors are:
6259 * <li>ENOMEM - failed to allocate overflow pages for the node.
6260 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6261 * should never happen since all callers already calculate the
6262 * page's free space before calling this function.
6266 mdb_node_add(MDB_cursor *mc, indx_t indx,
6267 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6270 size_t node_size = NODESIZE;
6273 MDB_page *mp = mc->mc_pg[mc->mc_top];
6274 MDB_page *ofp = NULL; /* overflow page */
6277 assert(mp->mp_upper >= mp->mp_lower);
6279 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6280 IS_LEAF(mp) ? "leaf" : "branch",
6281 IS_SUBP(mp) ? "sub-" : "",
6282 mp->mp_pgno, indx, data ? data->mv_size : 0,
6283 key ? key->mv_size : 0, key ? DKEY(key) : NULL));
6286 /* Move higher keys up one slot. */
6287 int ksize = mc->mc_db->md_pad, dif;
6288 char *ptr = LEAF2KEY(mp, indx, ksize);
6289 dif = NUMKEYS(mp) - indx;
6291 memmove(ptr+ksize, ptr, dif*ksize);
6292 /* insert new key */
6293 memcpy(ptr, key->mv_data, ksize);
6295 /* Just using these for counting */
6296 mp->mp_lower += sizeof(indx_t);
6297 mp->mp_upper -= ksize - sizeof(indx_t);
6302 node_size += key->mv_size;
6306 if (F_ISSET(flags, F_BIGDATA)) {
6307 /* Data already on overflow page. */
6308 node_size += sizeof(pgno_t);
6309 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6310 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6312 /* Put data on overflow page. */
6313 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6314 data->mv_size, node_size+data->mv_size));
6315 node_size += sizeof(pgno_t);
6316 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6318 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6321 node_size += data->mv_size;
6324 node_size += node_size & 1;
6326 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6327 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6328 mp->mp_pgno, NUMKEYS(mp)));
6329 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6330 mp->mp_upper - mp->mp_lower));
6331 DPRINTF(("node size = %"Z"u", node_size));
6332 return MDB_PAGE_FULL;
6335 /* Move higher pointers up one slot. */
6336 for (i = NUMKEYS(mp); i > indx; i--)
6337 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6339 /* Adjust free space offsets. */
6340 ofs = mp->mp_upper - node_size;
6341 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6342 mp->mp_ptrs[indx] = ofs;
6344 mp->mp_lower += sizeof(indx_t);
6346 /* Write the node data. */
6347 node = NODEPTR(mp, indx);
6348 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6349 node->mn_flags = flags;
6351 SETDSZ(node,data->mv_size);
6356 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6361 if (F_ISSET(flags, F_BIGDATA))
6362 memcpy(node->mn_data + key->mv_size, data->mv_data,
6364 else if (F_ISSET(flags, MDB_RESERVE))
6365 data->mv_data = node->mn_data + key->mv_size;
6367 memcpy(node->mn_data + key->mv_size, data->mv_data,
6370 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6372 if (F_ISSET(flags, MDB_RESERVE))
6373 data->mv_data = METADATA(ofp);
6375 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6382 /** Delete the specified node from a page.
6383 * @param[in] mp The page to operate on.
6384 * @param[in] indx The index of the node to delete.
6385 * @param[in] ksize The size of a node. Only used if the page is
6386 * part of a #MDB_DUPFIXED database.
6389 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6392 indx_t i, j, numkeys, ptr;
6399 COPY_PGNO(pgno, mp->mp_pgno);
6400 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6401 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6404 assert(indx < NUMKEYS(mp));
6407 int x = NUMKEYS(mp) - 1 - indx;
6408 base = LEAF2KEY(mp, indx, ksize);
6410 memmove(base, base + ksize, x * ksize);
6411 mp->mp_lower -= sizeof(indx_t);
6412 mp->mp_upper += ksize - sizeof(indx_t);
6416 node = NODEPTR(mp, indx);
6417 sz = NODESIZE + node->mn_ksize;
6419 if (F_ISSET(node->mn_flags, F_BIGDATA))
6420 sz += sizeof(pgno_t);
6422 sz += NODEDSZ(node);
6426 ptr = mp->mp_ptrs[indx];
6427 numkeys = NUMKEYS(mp);
6428 for (i = j = 0; i < numkeys; i++) {
6430 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6431 if (mp->mp_ptrs[i] < ptr)
6432 mp->mp_ptrs[j] += sz;
6437 base = (char *)mp + mp->mp_upper;
6438 memmove(base + sz, base, ptr - mp->mp_upper);
6440 mp->mp_lower -= sizeof(indx_t);
6444 /** Compact the main page after deleting a node on a subpage.
6445 * @param[in] mp The main page to operate on.
6446 * @param[in] indx The index of the subpage on the main page.
6449 mdb_node_shrink(MDB_page *mp, indx_t indx)
6456 indx_t i, numkeys, ptr;
6458 node = NODEPTR(mp, indx);
6459 sp = (MDB_page *)NODEDATA(node);
6460 osize = NODEDSZ(node);
6462 delta = sp->mp_upper - sp->mp_lower;
6463 SETDSZ(node, osize - delta);
6464 xp = (MDB_page *)((char *)sp + delta);
6466 /* shift subpage upward */
6468 nsize = NUMKEYS(sp) * sp->mp_pad;
6469 memmove(METADATA(xp), METADATA(sp), nsize);
6472 nsize = osize - sp->mp_upper;
6473 numkeys = NUMKEYS(sp);
6474 for (i=numkeys-1; i>=0; i--)
6475 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6477 xp->mp_upper = sp->mp_lower;
6478 xp->mp_lower = sp->mp_lower;
6479 xp->mp_flags = sp->mp_flags;
6480 xp->mp_pad = sp->mp_pad;
6481 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6483 /* shift lower nodes upward */
6484 ptr = mp->mp_ptrs[indx];
6485 numkeys = NUMKEYS(mp);
6486 for (i = 0; i < numkeys; i++) {
6487 if (mp->mp_ptrs[i] <= ptr)
6488 mp->mp_ptrs[i] += delta;
6491 base = (char *)mp + mp->mp_upper;
6492 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6493 mp->mp_upper += delta;
6496 /** Initial setup of a sorted-dups cursor.
6497 * Sorted duplicates are implemented as a sub-database for the given key.
6498 * The duplicate data items are actually keys of the sub-database.
6499 * Operations on the duplicate data items are performed using a sub-cursor
6500 * initialized when the sub-database is first accessed. This function does
6501 * the preliminary setup of the sub-cursor, filling in the fields that
6502 * depend only on the parent DB.
6503 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6506 mdb_xcursor_init0(MDB_cursor *mc)
6508 MDB_xcursor *mx = mc->mc_xcursor;
6510 mx->mx_cursor.mc_xcursor = NULL;
6511 mx->mx_cursor.mc_txn = mc->mc_txn;
6512 mx->mx_cursor.mc_db = &mx->mx_db;
6513 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6514 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6515 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6516 mx->mx_cursor.mc_snum = 0;
6517 mx->mx_cursor.mc_top = 0;
6518 mx->mx_cursor.mc_flags = C_SUB;
6519 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6520 mx->mx_dbx.md_dcmp = NULL;
6521 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6524 /** Final setup of a sorted-dups cursor.
6525 * Sets up the fields that depend on the data from the main cursor.
6526 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6527 * @param[in] node The data containing the #MDB_db record for the
6528 * sorted-dup database.
6531 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6533 MDB_xcursor *mx = mc->mc_xcursor;
6535 if (node->mn_flags & F_SUBDATA) {
6536 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6537 mx->mx_cursor.mc_pg[0] = 0;
6538 mx->mx_cursor.mc_snum = 0;
6539 mx->mx_cursor.mc_flags = C_SUB;
6541 MDB_page *fp = NODEDATA(node);
6542 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6543 mx->mx_db.md_flags = 0;
6544 mx->mx_db.md_depth = 1;
6545 mx->mx_db.md_branch_pages = 0;
6546 mx->mx_db.md_leaf_pages = 1;
6547 mx->mx_db.md_overflow_pages = 0;
6548 mx->mx_db.md_entries = NUMKEYS(fp);
6549 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6550 mx->mx_cursor.mc_snum = 1;
6551 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6552 mx->mx_cursor.mc_top = 0;
6553 mx->mx_cursor.mc_pg[0] = fp;
6554 mx->mx_cursor.mc_ki[0] = 0;
6555 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6556 mx->mx_db.md_flags = MDB_DUPFIXED;
6557 mx->mx_db.md_pad = fp->mp_pad;
6558 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6559 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6562 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6563 mx->mx_db.md_root));
6564 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6566 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6567 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6568 #if UINT_MAX < SIZE_MAX
6569 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6570 #ifdef MISALIGNED_OK
6571 mx->mx_dbx.md_cmp = mdb_cmp_long;
6573 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6578 /** Initialize a cursor for a given transaction and database. */
6580 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6583 mc->mc_backup = NULL;
6586 mc->mc_db = &txn->mt_dbs[dbi];
6587 mc->mc_dbx = &txn->mt_dbxs[dbi];
6588 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6593 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6595 mc->mc_xcursor = mx;
6596 mdb_xcursor_init0(mc);
6598 mc->mc_xcursor = NULL;
6600 if (*mc->mc_dbflag & DB_STALE) {
6601 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6606 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6609 size_t size = sizeof(MDB_cursor);
6611 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6614 if (txn->mt_flags & MDB_TXN_ERROR)
6617 /* Allow read access to the freelist */
6618 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6621 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6622 size += sizeof(MDB_xcursor);
6624 if ((mc = malloc(size)) != NULL) {
6625 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6626 if (txn->mt_cursors) {
6627 mc->mc_next = txn->mt_cursors[dbi];
6628 txn->mt_cursors[dbi] = mc;
6629 mc->mc_flags |= C_UNTRACK;
6641 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6643 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6646 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6649 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6653 /* Return the count of duplicate data items for the current key */
6655 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6659 if (mc == NULL || countp == NULL)
6662 if (mc->mc_xcursor == NULL)
6663 return MDB_INCOMPATIBLE;
6665 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6666 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6669 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6672 *countp = mc->mc_xcursor->mx_db.md_entries;
6678 mdb_cursor_close(MDB_cursor *mc)
6680 if (mc && !mc->mc_backup) {
6681 /* remove from txn, if tracked */
6682 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6683 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6684 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6686 *prev = mc->mc_next;
6693 mdb_cursor_txn(MDB_cursor *mc)
6695 if (!mc) return NULL;
6700 mdb_cursor_dbi(MDB_cursor *mc)
6706 /** Replace the key for a node with a new key.
6707 * @param[in] mc Cursor pointing to the node to operate on.
6708 * @param[in] key The new key to use.
6709 * @return 0 on success, non-zero on failure.
6712 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6719 indx_t ptr, i, numkeys, indx;
6722 indx = mc->mc_ki[mc->mc_top];
6723 mp = mc->mc_pg[mc->mc_top];
6724 node = NODEPTR(mp, indx);
6725 ptr = mp->mp_ptrs[indx];
6729 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6730 k2.mv_data = NODEKEY(node);
6731 k2.mv_size = node->mn_ksize;
6732 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6734 mdb_dkey(&k2, kbuf2),
6740 delta0 = delta = key->mv_size - node->mn_ksize;
6742 /* Must be 2-byte aligned. If new key is
6743 * shorter by 1, the shift will be skipped.
6745 delta += (delta & 1);
6747 if (delta > 0 && SIZELEFT(mp) < delta) {
6749 /* not enough space left, do a delete and split */
6750 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6751 pgno = NODEPGNO(node);
6752 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6753 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6756 numkeys = NUMKEYS(mp);
6757 for (i = 0; i < numkeys; i++) {
6758 if (mp->mp_ptrs[i] <= ptr)
6759 mp->mp_ptrs[i] -= delta;
6762 base = (char *)mp + mp->mp_upper;
6763 len = ptr - mp->mp_upper + NODESIZE;
6764 memmove(base - delta, base, len);
6765 mp->mp_upper -= delta;
6767 node = NODEPTR(mp, indx);
6770 /* But even if no shift was needed, update ksize */
6772 node->mn_ksize = key->mv_size;
6775 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6781 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6783 /** Move a node from csrc to cdst.
6786 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6793 unsigned short flags;
6797 /* Mark src and dst as dirty. */
6798 if ((rc = mdb_page_touch(csrc)) ||
6799 (rc = mdb_page_touch(cdst)))
6802 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6803 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6804 key.mv_size = csrc->mc_db->md_pad;
6805 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6807 data.mv_data = NULL;
6811 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6812 assert(!((long)srcnode&1));
6813 srcpg = NODEPGNO(srcnode);
6814 flags = srcnode->mn_flags;
6815 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6816 unsigned int snum = csrc->mc_snum;
6818 /* must find the lowest key below src */
6819 mdb_page_search_lowest(csrc);
6820 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6821 key.mv_size = csrc->mc_db->md_pad;
6822 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6824 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6825 key.mv_size = NODEKSZ(s2);
6826 key.mv_data = NODEKEY(s2);
6828 csrc->mc_snum = snum--;
6829 csrc->mc_top = snum;
6831 key.mv_size = NODEKSZ(srcnode);
6832 key.mv_data = NODEKEY(srcnode);
6834 data.mv_size = NODEDSZ(srcnode);
6835 data.mv_data = NODEDATA(srcnode);
6837 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6838 unsigned int snum = cdst->mc_snum;
6841 /* must find the lowest key below dst */
6842 mdb_page_search_lowest(cdst);
6843 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6844 bkey.mv_size = cdst->mc_db->md_pad;
6845 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6847 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6848 bkey.mv_size = NODEKSZ(s2);
6849 bkey.mv_data = NODEKEY(s2);
6851 cdst->mc_snum = snum--;
6852 cdst->mc_top = snum;
6853 mdb_cursor_copy(cdst, &mn);
6855 rc = mdb_update_key(&mn, &bkey);
6860 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6861 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6862 csrc->mc_ki[csrc->mc_top],
6864 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6865 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6867 /* Add the node to the destination page.
6869 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6870 if (rc != MDB_SUCCESS)
6873 /* Delete the node from the source page.
6875 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6878 /* Adjust other cursors pointing to mp */
6879 MDB_cursor *m2, *m3;
6880 MDB_dbi dbi = csrc->mc_dbi;
6881 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6883 if (csrc->mc_flags & C_SUB)
6886 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6887 if (csrc->mc_flags & C_SUB)
6888 m3 = &m2->mc_xcursor->mx_cursor;
6891 if (m3 == csrc) continue;
6892 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6893 csrc->mc_ki[csrc->mc_top]) {
6894 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6895 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6900 /* Update the parent separators.
6902 if (csrc->mc_ki[csrc->mc_top] == 0) {
6903 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6904 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6905 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6907 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6908 key.mv_size = NODEKSZ(srcnode);
6909 key.mv_data = NODEKEY(srcnode);
6911 DPRINTF(("update separator for source page %"Z"u to [%s]",
6912 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6913 mdb_cursor_copy(csrc, &mn);
6916 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6919 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6921 indx_t ix = csrc->mc_ki[csrc->mc_top];
6922 nullkey.mv_size = 0;
6923 csrc->mc_ki[csrc->mc_top] = 0;
6924 rc = mdb_update_key(csrc, &nullkey);
6925 csrc->mc_ki[csrc->mc_top] = ix;
6926 assert(rc == MDB_SUCCESS);
6930 if (cdst->mc_ki[cdst->mc_top] == 0) {
6931 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6932 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6933 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6935 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6936 key.mv_size = NODEKSZ(srcnode);
6937 key.mv_data = NODEKEY(srcnode);
6939 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6940 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6941 mdb_cursor_copy(cdst, &mn);
6944 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6947 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6949 indx_t ix = cdst->mc_ki[cdst->mc_top];
6950 nullkey.mv_size = 0;
6951 cdst->mc_ki[cdst->mc_top] = 0;
6952 rc = mdb_update_key(cdst, &nullkey);
6953 cdst->mc_ki[cdst->mc_top] = ix;
6954 assert(rc == MDB_SUCCESS);
6961 /** Merge one page into another.
6962 * The nodes from the page pointed to by \b csrc will
6963 * be copied to the page pointed to by \b cdst and then
6964 * the \b csrc page will be freed.
6965 * @param[in] csrc Cursor pointing to the source page.
6966 * @param[in] cdst Cursor pointing to the destination page.
6969 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6977 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6978 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6980 assert(csrc->mc_snum > 1); /* can't merge root page */
6981 assert(cdst->mc_snum > 1);
6983 /* Mark dst as dirty. */
6984 if ((rc = mdb_page_touch(cdst)))
6987 /* Move all nodes from src to dst.
6989 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6990 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6991 key.mv_size = csrc->mc_db->md_pad;
6992 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6993 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6994 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6995 if (rc != MDB_SUCCESS)
6997 key.mv_data = (char *)key.mv_data + key.mv_size;
7000 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7001 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7002 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7003 unsigned int snum = csrc->mc_snum;
7005 /* must find the lowest key below src */
7006 mdb_page_search_lowest(csrc);
7007 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7008 key.mv_size = csrc->mc_db->md_pad;
7009 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7011 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7012 key.mv_size = NODEKSZ(s2);
7013 key.mv_data = NODEKEY(s2);
7015 csrc->mc_snum = snum--;
7016 csrc->mc_top = snum;
7018 key.mv_size = srcnode->mn_ksize;
7019 key.mv_data = NODEKEY(srcnode);
7022 data.mv_size = NODEDSZ(srcnode);
7023 data.mv_data = NODEDATA(srcnode);
7024 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7025 if (rc != MDB_SUCCESS)
7030 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7031 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7032 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7034 /* Unlink the src page from parent and add to free list.
7036 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7037 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7040 rc = mdb_update_key(csrc, &key);
7046 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7047 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7050 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7051 csrc->mc_db->md_leaf_pages--;
7053 csrc->mc_db->md_branch_pages--;
7055 /* Adjust other cursors pointing to mp */
7056 MDB_cursor *m2, *m3;
7057 MDB_dbi dbi = csrc->mc_dbi;
7058 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7060 if (csrc->mc_flags & C_SUB)
7063 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7064 if (csrc->mc_flags & C_SUB)
7065 m3 = &m2->mc_xcursor->mx_cursor;
7068 if (m3 == csrc) continue;
7069 if (m3->mc_snum < csrc->mc_snum) continue;
7070 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7071 m3->mc_pg[csrc->mc_top] = mp;
7072 m3->mc_ki[csrc->mc_top] += nkeys;
7076 mdb_cursor_pop(csrc);
7078 return mdb_rebalance(csrc);
7081 /** Copy the contents of a cursor.
7082 * @param[in] csrc The cursor to copy from.
7083 * @param[out] cdst The cursor to copy to.
7086 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7090 cdst->mc_txn = csrc->mc_txn;
7091 cdst->mc_dbi = csrc->mc_dbi;
7092 cdst->mc_db = csrc->mc_db;
7093 cdst->mc_dbx = csrc->mc_dbx;
7094 cdst->mc_snum = csrc->mc_snum;
7095 cdst->mc_top = csrc->mc_top;
7096 cdst->mc_flags = csrc->mc_flags;
7098 for (i=0; i<csrc->mc_snum; i++) {
7099 cdst->mc_pg[i] = csrc->mc_pg[i];
7100 cdst->mc_ki[i] = csrc->mc_ki[i];
7104 /** Rebalance the tree after a delete operation.
7105 * @param[in] mc Cursor pointing to the page where rebalancing
7107 * @return 0 on success, non-zero on failure.
7110 mdb_rebalance(MDB_cursor *mc)
7114 unsigned int ptop, minkeys;
7117 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7121 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7122 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7123 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7124 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7125 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7129 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7130 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7133 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7134 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7140 if (mc->mc_snum < 2) {
7141 MDB_page *mp = mc->mc_pg[0];
7143 DPUTS("Can't rebalance a subpage, ignoring");
7146 if (NUMKEYS(mp) == 0) {
7147 DPUTS("tree is completely empty");
7148 mc->mc_db->md_root = P_INVALID;
7149 mc->mc_db->md_depth = 0;
7150 mc->mc_db->md_leaf_pages = 0;
7151 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7154 /* Adjust cursors pointing to mp */
7158 MDB_cursor *m2, *m3;
7159 MDB_dbi dbi = mc->mc_dbi;
7161 if (mc->mc_flags & C_SUB)
7164 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7165 if (mc->mc_flags & C_SUB)
7166 m3 = &m2->mc_xcursor->mx_cursor;
7169 if (m3->mc_snum < mc->mc_snum) continue;
7170 if (m3->mc_pg[0] == mp) {
7176 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7177 DPUTS("collapsing root page!");
7178 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7181 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7182 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7185 mc->mc_db->md_depth--;
7186 mc->mc_db->md_branch_pages--;
7187 mc->mc_ki[0] = mc->mc_ki[1];
7189 /* Adjust other cursors pointing to mp */
7190 MDB_cursor *m2, *m3;
7191 MDB_dbi dbi = mc->mc_dbi;
7193 if (mc->mc_flags & C_SUB)
7196 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7197 if (mc->mc_flags & C_SUB)
7198 m3 = &m2->mc_xcursor->mx_cursor;
7201 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7202 if (m3->mc_pg[0] == mp) {
7203 m3->mc_pg[0] = mc->mc_pg[0];
7206 m3->mc_ki[0] = m3->mc_ki[1];
7211 DPUTS("root page doesn't need rebalancing");
7215 /* The parent (branch page) must have at least 2 pointers,
7216 * otherwise the tree is invalid.
7218 ptop = mc->mc_top-1;
7219 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7221 /* Leaf page fill factor is below the threshold.
7222 * Try to move keys from left or right neighbor, or
7223 * merge with a neighbor page.
7228 mdb_cursor_copy(mc, &mn);
7229 mn.mc_xcursor = NULL;
7231 if (mc->mc_ki[ptop] == 0) {
7232 /* We're the leftmost leaf in our parent.
7234 DPUTS("reading right neighbor");
7236 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7237 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7240 mn.mc_ki[mn.mc_top] = 0;
7241 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7243 /* There is at least one neighbor to the left.
7245 DPUTS("reading left neighbor");
7247 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7248 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7251 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7252 mc->mc_ki[mc->mc_top] = 0;
7255 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7256 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7257 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7259 /* If the neighbor page is above threshold and has enough keys,
7260 * move one key from it. Otherwise we should try to merge them.
7261 * (A branch page must never have less than 2 keys.)
7263 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7264 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7265 return mdb_node_move(&mn, mc);
7267 if (mc->mc_ki[ptop] == 0)
7268 rc = mdb_page_merge(&mn, mc);
7270 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7271 rc = mdb_page_merge(mc, &mn);
7272 mdb_cursor_copy(&mn, mc);
7274 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7279 /** Complete a delete operation started by #mdb_cursor_del(). */
7281 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7288 mp = mc->mc_pg[mc->mc_top];
7289 ki = mc->mc_ki[mc->mc_top];
7291 /* add overflow pages to free list */
7292 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7296 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7297 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7298 (rc = mdb_ovpage_free(mc, omp)))
7301 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7302 mc->mc_db->md_entries--;
7303 rc = mdb_rebalance(mc);
7304 if (rc != MDB_SUCCESS)
7305 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7308 MDB_dbi dbi = mc->mc_dbi;
7310 mp = mc->mc_pg[mc->mc_top];
7311 nkeys = NUMKEYS(mp);
7313 /* if mc points past last node in page, find next sibling */
7314 if (mc->mc_ki[mc->mc_top] >= nkeys)
7315 mdb_cursor_sibling(mc, 1);
7317 /* Adjust other cursors pointing to mp */
7318 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7321 if (!(m2->mc_flags & C_INITIALIZED))
7323 if (m2->mc_pg[mc->mc_top] == mp) {
7324 if (m2->mc_ki[mc->mc_top] >= ki) {
7325 m2->mc_flags |= C_DEL;
7326 if (m2->mc_ki[mc->mc_top] > ki)
7327 m2->mc_ki[mc->mc_top]--;
7329 if (m2->mc_ki[mc->mc_top] >= nkeys)
7330 mdb_cursor_sibling(m2, 1);
7333 mc->mc_flags |= C_DEL;
7340 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7341 MDB_val *key, MDB_val *data)
7346 MDB_val rdata, *xdata;
7350 assert(key != NULL);
7352 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7354 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7357 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7358 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7360 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7361 return MDB_BAD_VALSIZE;
7364 mdb_cursor_init(&mc, txn, dbi, &mx);
7367 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7368 /* must ignore any data */
7379 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7381 /* let mdb_page_split know about this cursor if needed:
7382 * delete will trigger a rebalance; if it needs to move
7383 * a node from one page to another, it will have to
7384 * update the parent's separator key(s). If the new sepkey
7385 * is larger than the current one, the parent page may
7386 * run out of space, triggering a split. We need this
7387 * cursor to be consistent until the end of the rebalance.
7389 mc.mc_flags |= C_UNTRACK;
7390 mc.mc_next = txn->mt_cursors[dbi];
7391 txn->mt_cursors[dbi] = &mc;
7392 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7393 txn->mt_cursors[dbi] = mc.mc_next;
7398 /** Split a page and insert a new node.
7399 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7400 * The cursor will be updated to point to the actual page and index where
7401 * the node got inserted after the split.
7402 * @param[in] newkey The key for the newly inserted node.
7403 * @param[in] newdata The data for the newly inserted node.
7404 * @param[in] newpgno The page number, if the new node is a branch node.
7405 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7406 * @return 0 on success, non-zero on failure.
7409 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7410 unsigned int nflags)
7413 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7416 unsigned int i, j, split_indx, nkeys, pmax;
7418 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7420 MDB_page *mp, *rp, *pp;
7425 mp = mc->mc_pg[mc->mc_top];
7426 newindx = mc->mc_ki[mc->mc_top];
7428 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7429 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7430 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7432 /* Create a right sibling. */
7433 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7435 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7437 if (mc->mc_snum < 2) {
7438 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7440 /* shift current top to make room for new parent */
7441 mc->mc_pg[1] = mc->mc_pg[0];
7442 mc->mc_ki[1] = mc->mc_ki[0];
7445 mc->mc_db->md_root = pp->mp_pgno;
7446 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7447 mc->mc_db->md_depth++;
7450 /* Add left (implicit) pointer. */
7451 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7452 /* undo the pre-push */
7453 mc->mc_pg[0] = mc->mc_pg[1];
7454 mc->mc_ki[0] = mc->mc_ki[1];
7455 mc->mc_db->md_root = mp->mp_pgno;
7456 mc->mc_db->md_depth--;
7463 ptop = mc->mc_top-1;
7464 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7467 mc->mc_flags |= C_SPLITTING;
7468 mdb_cursor_copy(mc, &mn);
7469 mn.mc_pg[mn.mc_top] = rp;
7470 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7472 if (nflags & MDB_APPEND) {
7473 mn.mc_ki[mn.mc_top] = 0;
7475 split_indx = newindx;
7480 nkeys = NUMKEYS(mp);
7481 split_indx = nkeys / 2;
7482 if (newindx < split_indx)
7488 unsigned int lsize, rsize, ksize;
7489 /* Move half of the keys to the right sibling */
7491 x = mc->mc_ki[mc->mc_top] - split_indx;
7492 ksize = mc->mc_db->md_pad;
7493 split = LEAF2KEY(mp, split_indx, ksize);
7494 rsize = (nkeys - split_indx) * ksize;
7495 lsize = (nkeys - split_indx) * sizeof(indx_t);
7496 mp->mp_lower -= lsize;
7497 rp->mp_lower += lsize;
7498 mp->mp_upper += rsize - lsize;
7499 rp->mp_upper -= rsize - lsize;
7500 sepkey.mv_size = ksize;
7501 if (newindx == split_indx) {
7502 sepkey.mv_data = newkey->mv_data;
7504 sepkey.mv_data = split;
7507 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7508 memcpy(rp->mp_ptrs, split, rsize);
7509 sepkey.mv_data = rp->mp_ptrs;
7510 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7511 memcpy(ins, newkey->mv_data, ksize);
7512 mp->mp_lower += sizeof(indx_t);
7513 mp->mp_upper -= ksize - sizeof(indx_t);
7516 memcpy(rp->mp_ptrs, split, x * ksize);
7517 ins = LEAF2KEY(rp, x, ksize);
7518 memcpy(ins, newkey->mv_data, ksize);
7519 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7520 rp->mp_lower += sizeof(indx_t);
7521 rp->mp_upper -= ksize - sizeof(indx_t);
7522 mc->mc_ki[mc->mc_top] = x;
7523 mc->mc_pg[mc->mc_top] = rp;
7528 /* For leaf pages, check the split point based on what
7529 * fits where, since otherwise mdb_node_add can fail.
7531 * This check is only needed when the data items are
7532 * relatively large, such that being off by one will
7533 * make the difference between success or failure.
7535 * It's also relevant if a page happens to be laid out
7536 * such that one half of its nodes are all "small" and
7537 * the other half of its nodes are "large." If the new
7538 * item is also "large" and falls on the half with
7539 * "large" nodes, it also may not fit.
7542 unsigned int psize, nsize;
7543 /* Maximum free space in an empty page */
7544 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7545 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7546 if ((nkeys < 20) || (nsize > pmax/16)) {
7547 if (newindx <= split_indx) {
7550 for (i=0; i<split_indx; i++) {
7551 node = NODEPTR(mp, i);
7552 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7553 if (F_ISSET(node->mn_flags, F_BIGDATA))
7554 psize += sizeof(pgno_t);
7556 psize += NODEDSZ(node);
7560 split_indx = newindx;
7571 for (i=nkeys-1; i>=split_indx; i--) {
7572 node = NODEPTR(mp, i);
7573 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7574 if (F_ISSET(node->mn_flags, F_BIGDATA))
7575 psize += sizeof(pgno_t);
7577 psize += NODEDSZ(node);
7581 split_indx = newindx;
7592 /* First find the separating key between the split pages.
7593 * The case where newindx == split_indx is ambiguous; the
7594 * new item could go to the new page or stay on the original
7595 * page. If newpos == 1 it goes to the new page.
7597 if (newindx == split_indx && newpos) {
7598 sepkey.mv_size = newkey->mv_size;
7599 sepkey.mv_data = newkey->mv_data;
7601 node = NODEPTR(mp, split_indx);
7602 sepkey.mv_size = node->mn_ksize;
7603 sepkey.mv_data = NODEKEY(node);
7607 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7609 /* Copy separator key to the parent.
7611 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7615 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7618 if (mn.mc_snum == mc->mc_snum) {
7619 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7620 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7621 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7622 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7627 /* Right page might now have changed parent.
7628 * Check if left page also changed parent.
7630 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7631 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7632 for (i=0; i<ptop; i++) {
7633 mc->mc_pg[i] = mn.mc_pg[i];
7634 mc->mc_ki[i] = mn.mc_ki[i];
7636 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7637 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7641 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7644 mc->mc_flags ^= C_SPLITTING;
7645 if (rc != MDB_SUCCESS) {
7648 if (nflags & MDB_APPEND) {
7649 mc->mc_pg[mc->mc_top] = rp;
7650 mc->mc_ki[mc->mc_top] = 0;
7651 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7654 for (i=0; i<mc->mc_top; i++)
7655 mc->mc_ki[i] = mn.mc_ki[i];
7662 /* Move half of the keys to the right sibling. */
7664 /* grab a page to hold a temporary copy */
7665 copy = mdb_page_malloc(mc->mc_txn, 1);
7669 copy->mp_pgno = mp->mp_pgno;
7670 copy->mp_flags = mp->mp_flags;
7671 copy->mp_lower = PAGEHDRSZ;
7672 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7673 mc->mc_pg[mc->mc_top] = copy;
7674 for (i = j = 0; i <= nkeys; j++) {
7675 if (i == split_indx) {
7676 /* Insert in right sibling. */
7677 /* Reset insert index for right sibling. */
7678 if (i != newindx || (newpos ^ ins_new)) {
7680 mc->mc_pg[mc->mc_top] = rp;
7684 if (i == newindx && !ins_new) {
7685 /* Insert the original entry that caused the split. */
7686 rkey.mv_data = newkey->mv_data;
7687 rkey.mv_size = newkey->mv_size;
7696 /* Update index for the new key. */
7697 mc->mc_ki[mc->mc_top] = j;
7698 } else if (i == nkeys) {
7701 node = NODEPTR(mp, i);
7702 rkey.mv_data = NODEKEY(node);
7703 rkey.mv_size = node->mn_ksize;
7705 xdata.mv_data = NODEDATA(node);
7706 xdata.mv_size = NODEDSZ(node);
7709 pgno = NODEPGNO(node);
7710 flags = node->mn_flags;
7715 if (!IS_LEAF(mp) && j == 0) {
7716 /* First branch index doesn't need key data. */
7720 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7724 nkeys = NUMKEYS(copy);
7725 for (i=0; i<nkeys; i++)
7726 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7727 mp->mp_lower = copy->mp_lower;
7728 mp->mp_upper = copy->mp_upper;
7729 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7730 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7732 /* reset back to original page */
7733 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7734 mc->mc_pg[mc->mc_top] = mp;
7735 if (nflags & MDB_RESERVE) {
7736 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7737 if (!(node->mn_flags & F_BIGDATA))
7738 newdata->mv_data = NODEDATA(node);
7742 /* Make sure mc_ki is still valid.
7744 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7745 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7746 for (i=0; i<ptop; i++) {
7747 mc->mc_pg[i] = mn.mc_pg[i];
7748 mc->mc_ki[i] = mn.mc_ki[i];
7750 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7751 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7755 /* return tmp page to freelist */
7756 mdb_page_free(mc->mc_txn->mt_env, copy);
7759 /* Adjust other cursors pointing to mp */
7760 MDB_cursor *m2, *m3;
7761 MDB_dbi dbi = mc->mc_dbi;
7762 int fixup = NUMKEYS(mp);
7764 if (mc->mc_flags & C_SUB)
7767 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7768 if (mc->mc_flags & C_SUB)
7769 m3 = &m2->mc_xcursor->mx_cursor;
7774 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7776 if (m3->mc_flags & C_SPLITTING)
7781 for (k=m3->mc_top; k>=0; k--) {
7782 m3->mc_ki[k+1] = m3->mc_ki[k];
7783 m3->mc_pg[k+1] = m3->mc_pg[k];
7785 if (m3->mc_ki[0] >= split_indx) {
7790 m3->mc_pg[0] = mc->mc_pg[0];
7794 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7795 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7796 m3->mc_ki[mc->mc_top]++;
7797 if (m3->mc_ki[mc->mc_top] >= fixup) {
7798 m3->mc_pg[mc->mc_top] = rp;
7799 m3->mc_ki[mc->mc_top] -= fixup;
7800 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7802 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7803 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7812 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7813 MDB_val *key, MDB_val *data, unsigned int flags)
7818 assert(key != NULL);
7819 assert(data != NULL);
7821 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7824 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7825 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7827 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7828 return MDB_BAD_VALSIZE;
7831 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7834 mdb_cursor_init(&mc, txn, dbi, &mx);
7835 return mdb_cursor_put(&mc, key, data, flags);
7839 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7841 if ((flag & CHANGEABLE) != flag)
7844 env->me_flags |= flag;
7846 env->me_flags &= ~flag;
7851 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7856 *arg = env->me_flags;
7861 mdb_env_get_path(MDB_env *env, const char **arg)
7866 *arg = env->me_path;
7870 /** Common code for #mdb_stat() and #mdb_env_stat().
7871 * @param[in] env the environment to operate in.
7872 * @param[in] db the #MDB_db record containing the stats to return.
7873 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7874 * @return 0, this function always succeeds.
7877 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7879 arg->ms_psize = env->me_psize;
7880 arg->ms_depth = db->md_depth;
7881 arg->ms_branch_pages = db->md_branch_pages;
7882 arg->ms_leaf_pages = db->md_leaf_pages;
7883 arg->ms_overflow_pages = db->md_overflow_pages;
7884 arg->ms_entries = db->md_entries;
7889 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7893 if (env == NULL || arg == NULL)
7896 toggle = mdb_env_pick_meta(env);
7898 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7902 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7906 if (env == NULL || arg == NULL)
7909 toggle = mdb_env_pick_meta(env);
7910 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7911 arg->me_mapsize = env->me_mapsize;
7912 arg->me_maxreaders = env->me_maxreaders;
7914 /* me_numreaders may be zero if this process never used any readers. Use
7915 * the shared numreader count if it exists.
7917 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7919 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7920 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7924 /** Set the default comparison functions for a database.
7925 * Called immediately after a database is opened to set the defaults.
7926 * The user can then override them with #mdb_set_compare() or
7927 * #mdb_set_dupsort().
7928 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7929 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7932 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7934 uint16_t f = txn->mt_dbs[dbi].md_flags;
7936 txn->mt_dbxs[dbi].md_cmp =
7937 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7938 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7940 txn->mt_dbxs[dbi].md_dcmp =
7941 !(f & MDB_DUPSORT) ? 0 :
7942 ((f & MDB_INTEGERDUP)
7943 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7944 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7947 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7952 int rc, dbflag, exact;
7953 unsigned int unused = 0;
7956 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7957 mdb_default_cmp(txn, FREE_DBI);
7960 if ((flags & VALID_FLAGS) != flags)
7962 if (txn->mt_flags & MDB_TXN_ERROR)
7968 if (flags & PERSISTENT_FLAGS) {
7969 uint16_t f2 = flags & PERSISTENT_FLAGS;
7970 /* make sure flag changes get committed */
7971 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7972 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7973 txn->mt_flags |= MDB_TXN_DIRTY;
7976 mdb_default_cmp(txn, MAIN_DBI);
7980 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7981 mdb_default_cmp(txn, MAIN_DBI);
7984 /* Is the DB already open? */
7986 for (i=2; i<txn->mt_numdbs; i++) {
7987 if (!txn->mt_dbxs[i].md_name.mv_size) {
7988 /* Remember this free slot */
7989 if (!unused) unused = i;
7992 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7993 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7999 /* If no free slot and max hit, fail */
8000 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8001 return MDB_DBS_FULL;
8003 /* Cannot mix named databases with some mainDB flags */
8004 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8005 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8007 /* Find the DB info */
8008 dbflag = DB_NEW|DB_VALID;
8011 key.mv_data = (void *)name;
8012 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8013 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8014 if (rc == MDB_SUCCESS) {
8015 /* make sure this is actually a DB */
8016 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8017 if (!(node->mn_flags & F_SUBDATA))
8018 return MDB_INCOMPATIBLE;
8019 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8020 /* Create if requested */
8022 data.mv_size = sizeof(MDB_db);
8023 data.mv_data = &dummy;
8024 memset(&dummy, 0, sizeof(dummy));
8025 dummy.md_root = P_INVALID;
8026 dummy.md_flags = flags & PERSISTENT_FLAGS;
8027 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8031 /* OK, got info, add to table */
8032 if (rc == MDB_SUCCESS) {
8033 unsigned int slot = unused ? unused : txn->mt_numdbs;
8034 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8035 txn->mt_dbxs[slot].md_name.mv_size = len;
8036 txn->mt_dbxs[slot].md_rel = NULL;
8037 txn->mt_dbflags[slot] = dbflag;
8038 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8040 mdb_default_cmp(txn, slot);
8049 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8051 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8054 if (txn->mt_dbflags[dbi] & DB_STALE) {
8057 /* Stale, must read the DB's root. cursor_init does it for us. */
8058 mdb_cursor_init(&mc, txn, dbi, &mx);
8060 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8063 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8066 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8068 ptr = env->me_dbxs[dbi].md_name.mv_data;
8069 env->me_dbxs[dbi].md_name.mv_data = NULL;
8070 env->me_dbxs[dbi].md_name.mv_size = 0;
8071 env->me_dbflags[dbi] = 0;
8075 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8077 /* We could return the flags for the FREE_DBI too but what's the point? */
8078 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8080 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8084 /** Add all the DB's pages to the free list.
8085 * @param[in] mc Cursor on the DB to free.
8086 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8087 * @return 0 on success, non-zero on failure.
8090 mdb_drop0(MDB_cursor *mc, int subs)
8094 rc = mdb_page_search(mc, NULL, 0);
8095 if (rc == MDB_SUCCESS) {
8096 MDB_txn *txn = mc->mc_txn;
8101 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8102 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8105 mdb_cursor_copy(mc, &mx);
8106 while (mc->mc_snum > 0) {
8107 MDB_page *mp = mc->mc_pg[mc->mc_top];
8108 unsigned n = NUMKEYS(mp);
8110 for (i=0; i<n; i++) {
8111 ni = NODEPTR(mp, i);
8112 if (ni->mn_flags & F_BIGDATA) {
8115 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8116 rc = mdb_page_get(txn, pg, &omp, NULL);
8119 assert(IS_OVERFLOW(omp));
8120 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8124 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8125 mdb_xcursor_init1(mc, ni);
8126 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8132 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8134 for (i=0; i<n; i++) {
8136 ni = NODEPTR(mp, i);
8139 mdb_midl_xappend(txn->mt_free_pgs, pg);
8144 mc->mc_ki[mc->mc_top] = i;
8145 rc = mdb_cursor_sibling(mc, 1);
8147 /* no more siblings, go back to beginning
8148 * of previous level.
8152 for (i=1; i<mc->mc_snum; i++) {
8154 mc->mc_pg[i] = mx.mc_pg[i];
8159 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8160 } else if (rc == MDB_NOTFOUND) {
8166 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8168 MDB_cursor *mc, *m2;
8171 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8174 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8177 rc = mdb_cursor_open(txn, dbi, &mc);
8181 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8182 /* Invalidate the dropped DB's cursors */
8183 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8184 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8188 /* Can't delete the main DB */
8189 if (del && dbi > MAIN_DBI) {
8190 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8192 txn->mt_dbflags[dbi] = DB_STALE;
8193 mdb_dbi_close(txn->mt_env, dbi);
8196 /* reset the DB record, mark it dirty */
8197 txn->mt_dbflags[dbi] |= DB_DIRTY;
8198 txn->mt_dbs[dbi].md_depth = 0;
8199 txn->mt_dbs[dbi].md_branch_pages = 0;
8200 txn->mt_dbs[dbi].md_leaf_pages = 0;
8201 txn->mt_dbs[dbi].md_overflow_pages = 0;
8202 txn->mt_dbs[dbi].md_entries = 0;
8203 txn->mt_dbs[dbi].md_root = P_INVALID;
8205 txn->mt_flags |= MDB_TXN_DIRTY;
8208 mdb_cursor_close(mc);
8212 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8214 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8217 txn->mt_dbxs[dbi].md_cmp = cmp;
8221 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8223 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8226 txn->mt_dbxs[dbi].md_dcmp = cmp;
8230 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8232 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8235 txn->mt_dbxs[dbi].md_rel = rel;
8239 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8241 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8244 txn->mt_dbxs[dbi].md_relctx = ctx;
8248 int mdb_env_get_maxkeysize(MDB_env *env)
8250 return MDB_MAXKEYSIZE;
8253 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8255 unsigned int i, rdrs;
8262 if (!env->me_txns) {
8263 return func("(no reader locks)\n", ctx);
8265 rdrs = env->me_txns->mti_numreaders;
8266 mr = env->me_txns->mti_readers;
8267 for (i=0; i<rdrs; i++) {
8272 if (mr[i].mr_txnid == (txnid_t)-1) {
8273 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8275 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8279 func(" pid thread txnid\n", ctx);
8281 rc = func(buf, ctx);
8287 func("(no active readers)\n", ctx);
8292 /** Insert pid into list if not already present.
8293 * return -1 if already present.
8295 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8297 /* binary search of pid in list */
8299 unsigned cursor = 1;
8301 unsigned n = ids[0];
8304 unsigned pivot = n >> 1;
8305 cursor = base + pivot + 1;
8306 val = pid - ids[cursor];
8311 } else if ( val > 0 ) {
8316 /* found, so it's a duplicate */
8325 for (n = ids[0]; n > cursor; n--)
8331 int mdb_reader_check(MDB_env *env, int *dead)
8333 unsigned int i, j, rdrs;
8344 rdrs = env->me_txns->mti_numreaders;
8345 pids = malloc((rdrs+1) * sizeof(pid_t));
8349 mr = env->me_txns->mti_readers;
8351 for (i=0; i<rdrs; i++) {
8352 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8354 if (mdb_pid_insert(pids, pid) == 0) {
8355 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8357 /* Recheck, a new process may have reused pid */
8358 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8359 for (j=i; j<rdrs; j++)
8360 if (mr[j].mr_pid == pid) {
8365 UNLOCK_MUTEX_R(env);