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))
320 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
322 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
325 /** A default memory page size.
326 * The actual size is platform-dependent, but we use this for
327 * boot-strapping. We probably should not be using this any more.
328 * The #GET_PAGESIZE() macro is used to get the actual size.
330 * Note that we don't currently support Huge pages. On Linux,
331 * regular data files cannot use Huge pages, and in general
332 * Huge pages aren't actually pageable. We rely on the OS
333 * demand-pager to read our data and page it out when memory
334 * pressure from other processes is high. So until OSs have
335 * actual paging support for Huge pages, they're not viable.
337 #define MDB_PAGESIZE 4096
339 /** The minimum number of keys required in a database page.
340 * Setting this to a larger value will place a smaller bound on the
341 * maximum size of a data item. Data items larger than this size will
342 * be pushed into overflow pages instead of being stored directly in
343 * the B-tree node. This value used to default to 4. With a page size
344 * of 4096 bytes that meant that any item larger than 1024 bytes would
345 * go into an overflow page. That also meant that on average 2-3KB of
346 * each overflow page was wasted space. The value cannot be lower than
347 * 2 because then there would no longer be a tree structure. With this
348 * value, items larger than 2KB will go into overflow pages, and on
349 * average only 1KB will be wasted.
351 #define MDB_MINKEYS 2
353 /** A stamp that identifies a file as an MDB file.
354 * There's nothing special about this value other than that it is easily
355 * recognizable, and it will reflect any byte order mismatches.
357 #define MDB_MAGIC 0xBEEFC0DE
359 /** The version number for a database's datafile format. */
360 #define MDB_DATA_VERSION 1
361 /** The version number for a database's lockfile format. */
362 #define MDB_LOCK_VERSION 1
364 /** @brief The maximum size of a key in the database.
366 * The library rejects bigger keys, and cannot deal with records
367 * with bigger keys stored by a library with bigger max keysize.
369 * We require that keys all fit onto a regular page. This limit
370 * could be raised a bit further if needed; to something just
371 * under #MDB_PAGESIZE / #MDB_MINKEYS.
373 * Note that data items in an #MDB_DUPSORT database are actually keys
374 * of a subDB, so they're also limited to this size.
376 #ifndef MDB_MAXKEYSIZE
377 #define MDB_MAXKEYSIZE 511
380 /** @brief The maximum size of a data item.
382 * We only store a 32 bit value for node sizes.
384 #define MAXDATASIZE 0xffffffffUL
389 * This is used for printing a hex dump of a key's contents.
391 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
392 /** Display a key in hex.
394 * Invoke a function to display a key in hex.
396 #define DKEY(x) mdb_dkey(x, kbuf)
402 /** An invalid page number.
403 * Mainly used to denote an empty tree.
405 #define P_INVALID (~(pgno_t)0)
407 /** Test if the flags \b f are set in a flag word \b w. */
408 #define F_ISSET(w, f) (((w) & (f)) == (f))
410 /** Used for offsets within a single page.
411 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
414 typedef uint16_t indx_t;
416 /** Default size of memory map.
417 * This is certainly too small for any actual applications. Apps should always set
418 * the size explicitly using #mdb_env_set_mapsize().
420 #define DEFAULT_MAPSIZE 1048576
422 /** @defgroup readers Reader Lock Table
423 * Readers don't acquire any locks for their data access. Instead, they
424 * simply record their transaction ID in the reader table. The reader
425 * mutex is needed just to find an empty slot in the reader table. The
426 * slot's address is saved in thread-specific data so that subsequent read
427 * transactions started by the same thread need no further locking to proceed.
429 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
431 * No reader table is used if the database is on a read-only filesystem, or
432 * if #MDB_NOLOCK is set.
434 * Since the database uses multi-version concurrency control, readers don't
435 * actually need any locking. This table is used to keep track of which
436 * readers are using data from which old transactions, so that we'll know
437 * when a particular old transaction is no longer in use. Old transactions
438 * that have discarded any data pages can then have those pages reclaimed
439 * for use by a later write transaction.
441 * The lock table is constructed such that reader slots are aligned with the
442 * processor's cache line size. Any slot is only ever used by one thread.
443 * This alignment guarantees that there will be no contention or cache
444 * thrashing as threads update their own slot info, and also eliminates
445 * any need for locking when accessing a slot.
447 * A writer thread will scan every slot in the table to determine the oldest
448 * outstanding reader transaction. Any freed pages older than this will be
449 * reclaimed by the writer. The writer doesn't use any locks when scanning
450 * this table. This means that there's no guarantee that the writer will
451 * see the most up-to-date reader info, but that's not required for correct
452 * operation - all we need is to know the upper bound on the oldest reader,
453 * we don't care at all about the newest reader. So the only consequence of
454 * reading stale information here is that old pages might hang around a
455 * while longer before being reclaimed. That's actually good anyway, because
456 * the longer we delay reclaiming old pages, the more likely it is that a
457 * string of contiguous pages can be found after coalescing old pages from
458 * many old transactions together.
461 /** Number of slots in the reader table.
462 * This value was chosen somewhat arbitrarily. 126 readers plus a
463 * couple mutexes fit exactly into 8KB on my development machine.
464 * Applications should set the table size using #mdb_env_set_maxreaders().
466 #define DEFAULT_READERS 126
468 /** The size of a CPU cache line in bytes. We want our lock structures
469 * aligned to this size to avoid false cache line sharing in the
471 * This value works for most CPUs. For Itanium this should be 128.
477 /** The information we store in a single slot of the reader table.
478 * In addition to a transaction ID, we also record the process and
479 * thread ID that owns a slot, so that we can detect stale information,
480 * e.g. threads or processes that went away without cleaning up.
481 * @note We currently don't check for stale records. We simply re-init
482 * the table when we know that we're the only process opening the
485 typedef struct MDB_rxbody {
486 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
487 * Multiple readers that start at the same time will probably have the
488 * same ID here. Again, it's not important to exclude them from
489 * anything; all we need to know is which version of the DB they
490 * started from so we can avoid overwriting any data used in that
491 * particular version.
494 /** The process ID of the process owning this reader txn. */
496 /** The thread ID of the thread owning this txn. */
500 /** The actual reader record, with cacheline padding. */
501 typedef struct MDB_reader {
504 /** shorthand for mrb_txnid */
505 #define mr_txnid mru.mrx.mrb_txnid
506 #define mr_pid mru.mrx.mrb_pid
507 #define mr_tid mru.mrx.mrb_tid
508 /** cache line alignment */
509 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
513 /** The header for the reader table.
514 * The table resides in a memory-mapped file. (This is a different file
515 * than is used for the main database.)
517 * For POSIX the actual mutexes reside in the shared memory of this
518 * mapped file. On Windows, mutexes are named objects allocated by the
519 * kernel; we store the mutex names in this mapped file so that other
520 * processes can grab them. This same approach is also used on
521 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
522 * process-shared POSIX mutexes. For these cases where a named object
523 * is used, the object name is derived from a 64 bit FNV hash of the
524 * environment pathname. As such, naming collisions are extremely
525 * unlikely. If a collision occurs, the results are unpredictable.
527 typedef struct MDB_txbody {
528 /** Stamp identifying this as an MDB file. It must be set
531 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
533 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
534 char mtb_rmname[MNAME_LEN];
536 /** Mutex protecting access to this table.
537 * This is the reader lock that #LOCK_MUTEX_R acquires.
539 pthread_mutex_t mtb_mutex;
541 /** The ID of the last transaction committed to the database.
542 * This is recorded here only for convenience; the value can always
543 * be determined by reading the main database meta pages.
546 /** The number of slots that have been used in the reader table.
547 * This always records the maximum count, it is not decremented
548 * when readers release their slots.
550 unsigned mtb_numreaders;
553 /** The actual reader table definition. */
554 typedef struct MDB_txninfo {
557 #define mti_magic mt1.mtb.mtb_magic
558 #define mti_format mt1.mtb.mtb_format
559 #define mti_mutex mt1.mtb.mtb_mutex
560 #define mti_rmname mt1.mtb.mtb_rmname
561 #define mti_txnid mt1.mtb.mtb_txnid
562 #define mti_numreaders mt1.mtb.mtb_numreaders
563 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
566 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
567 char mt2_wmname[MNAME_LEN];
568 #define mti_wmname mt2.mt2_wmname
570 pthread_mutex_t mt2_wmutex;
571 #define mti_wmutex mt2.mt2_wmutex
573 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
575 MDB_reader mti_readers[1];
578 /** Lockfile format signature: version, features and field layout */
579 #define MDB_LOCK_FORMAT \
581 ((MDB_LOCK_VERSION) \
582 /* Flags which describe functionality */ \
583 + (((MDB_PIDLOCK) != 0) << 16)))
586 /** Common header for all page types.
587 * Overflow records occupy a number of contiguous pages with no
588 * headers on any page after the first.
590 typedef struct MDB_page {
591 #define mp_pgno mp_p.p_pgno
592 #define mp_next mp_p.p_next
594 pgno_t p_pgno; /**< page number */
595 void * p_next; /**< for in-memory list of freed structs */
598 /** @defgroup mdb_page Page Flags
600 * Flags for the page headers.
603 #define P_BRANCH 0x01 /**< branch page */
604 #define P_LEAF 0x02 /**< leaf page */
605 #define P_OVERFLOW 0x04 /**< overflow page */
606 #define P_META 0x08 /**< meta page */
607 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
608 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
609 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
610 #define P_KEEP 0x8000 /**< leave this page alone during spill */
612 uint16_t mp_flags; /**< @ref mdb_page */
613 #define mp_lower mp_pb.pb.pb_lower
614 #define mp_upper mp_pb.pb.pb_upper
615 #define mp_pages mp_pb.pb_pages
618 indx_t pb_lower; /**< lower bound of free space */
619 indx_t pb_upper; /**< upper bound of free space */
621 uint32_t pb_pages; /**< number of overflow pages */
623 indx_t mp_ptrs[1]; /**< dynamic size */
626 /** Size of the page header, excluding dynamic data at the end */
627 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
629 /** Address of first usable data byte in a page, after the header */
630 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
632 /** Number of nodes on a page */
633 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
635 /** The amount of space remaining in the page */
636 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
638 /** The percentage of space used in the page, in tenths of a percent. */
639 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
640 ((env)->me_psize - PAGEHDRSZ))
641 /** The minimum page fill factor, in tenths of a percent.
642 * Pages emptier than this are candidates for merging.
644 #define FILL_THRESHOLD 250
646 /** Test if a page is a leaf page */
647 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
648 /** Test if a page is a LEAF2 page */
649 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
650 /** Test if a page is a branch page */
651 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
652 /** Test if a page is an overflow page */
653 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
654 /** Test if a page is a sub page */
655 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
657 /** The number of overflow pages needed to store the given size. */
658 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
660 /** Header for a single key/data pair within a page.
661 * We guarantee 2-byte alignment for nodes.
663 typedef struct MDB_node {
664 /** lo and hi are used for data size on leaf nodes and for
665 * child pgno on branch nodes. On 64 bit platforms, flags
666 * is also used for pgno. (Branch nodes have no flags).
667 * They are in host byte order in case that lets some
668 * accesses be optimized into a 32-bit word access.
670 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
671 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
672 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
673 /** @defgroup mdb_node Node Flags
675 * Flags for node headers.
678 #define F_BIGDATA 0x01 /**< data put on overflow page */
679 #define F_SUBDATA 0x02 /**< data is a sub-database */
680 #define F_DUPDATA 0x04 /**< data has duplicates */
682 /** valid flags for #mdb_node_add() */
683 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
686 unsigned short mn_flags; /**< @ref mdb_node */
687 unsigned short mn_ksize; /**< key size */
688 char mn_data[1]; /**< key and data are appended here */
691 /** Size of the node header, excluding dynamic data at the end */
692 #define NODESIZE offsetof(MDB_node, mn_data)
694 /** Bit position of top word in page number, for shifting mn_flags */
695 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
697 /** Size of a node in a branch page with a given key.
698 * This is just the node header plus the key, there is no data.
700 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
702 /** Size of a node in a leaf page with a given key and data.
703 * This is node header plus key plus data size.
705 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
707 /** Address of node \b i in page \b p */
708 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
710 /** Address of the key for the node */
711 #define NODEKEY(node) (void *)((node)->mn_data)
713 /** Address of the data for a node */
714 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
716 /** Get the page number pointed to by a branch node */
717 #define NODEPGNO(node) \
718 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
719 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
720 /** Set the page number in a branch node */
721 #define SETPGNO(node,pgno) do { \
722 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
723 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
725 /** Get the size of the data in a leaf node */
726 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
727 /** Set the size of the data for a leaf node */
728 #define SETDSZ(node,size) do { \
729 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
730 /** The size of a key in a node */
731 #define NODEKSZ(node) ((node)->mn_ksize)
733 /** Copy a page number from src to dst */
735 #define COPY_PGNO(dst,src) dst = src
737 #if SIZE_MAX > 4294967295UL
738 #define COPY_PGNO(dst,src) do { \
739 unsigned short *s, *d; \
740 s = (unsigned short *)&(src); \
741 d = (unsigned short *)&(dst); \
748 #define COPY_PGNO(dst,src) do { \
749 unsigned short *s, *d; \
750 s = (unsigned short *)&(src); \
751 d = (unsigned short *)&(dst); \
757 /** The address of a key in a LEAF2 page.
758 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
759 * There are no node headers, keys are stored contiguously.
761 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
763 /** Set the \b node's key into \b keyptr, if requested. */
764 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
765 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
767 /** Set the \b node's key into \b key. */
768 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
770 /** Information about a single database in the environment. */
771 typedef struct MDB_db {
772 uint32_t md_pad; /**< also ksize for LEAF2 pages */
773 uint16_t md_flags; /**< @ref mdb_dbi_open */
774 uint16_t md_depth; /**< depth of this tree */
775 pgno_t md_branch_pages; /**< number of internal pages */
776 pgno_t md_leaf_pages; /**< number of leaf pages */
777 pgno_t md_overflow_pages; /**< number of overflow pages */
778 size_t md_entries; /**< number of data items */
779 pgno_t md_root; /**< the root page of this tree */
782 /** mdb_dbi_open flags */
783 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
784 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
785 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
786 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
788 /** Handle for the DB used to track free pages. */
790 /** Handle for the default DB. */
793 /** Meta page content.
794 * A meta page is the start point for accessing a database snapshot.
795 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
797 typedef struct MDB_meta {
798 /** Stamp identifying this as an MDB file. It must be set
801 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
803 void *mm_address; /**< address for fixed mapping */
804 size_t mm_mapsize; /**< size of mmap region */
805 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
806 /** The size of pages used in this DB */
807 #define mm_psize mm_dbs[0].md_pad
808 /** Any persistent environment flags. @ref mdb_env */
809 #define mm_flags mm_dbs[0].md_flags
810 pgno_t mm_last_pg; /**< last used page in file */
811 txnid_t mm_txnid; /**< txnid that committed this page */
814 /** Buffer for a stack-allocated dirty page.
815 * The members define size and alignment, and silence type
816 * aliasing warnings. They are not used directly; that could
817 * mean incorrectly using several union members in parallel.
819 typedef union MDB_pagebuf {
820 char mb_raw[MDB_PAGESIZE];
823 char mm_pad[PAGEHDRSZ];
828 /** Auxiliary DB info.
829 * The information here is mostly static/read-only. There is
830 * only a single copy of this record in the environment.
832 typedef struct MDB_dbx {
833 MDB_val md_name; /**< name of the database */
834 MDB_cmp_func *md_cmp; /**< function for comparing keys */
835 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
836 MDB_rel_func *md_rel; /**< user relocate function */
837 void *md_relctx; /**< user-provided context for md_rel */
840 /** A database transaction.
841 * Every operation requires a transaction handle.
844 MDB_txn *mt_parent; /**< parent of a nested txn */
845 MDB_txn *mt_child; /**< nested txn under this txn */
846 pgno_t mt_next_pgno; /**< next unallocated page */
847 /** The ID of this transaction. IDs are integers incrementing from 1.
848 * Only committed write transactions increment the ID. If a transaction
849 * aborts, the ID may be re-used by the next writer.
852 MDB_env *mt_env; /**< the DB environment */
853 /** The list of pages that became unused during this transaction.
856 /** The sorted list of dirty pages we temporarily wrote to disk
857 * because the dirty list was full. page numbers in here are
858 * shifted left by 1, deleted slots have the LSB set.
860 MDB_IDL mt_spill_pgs;
862 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
864 /** For read txns: This thread/txn's reader table slot, or NULL. */
867 /** Array of records for each DB known in the environment. */
869 /** Array of MDB_db records for each known DB */
871 /** @defgroup mt_dbflag Transaction DB Flags
875 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
876 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
877 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
878 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
880 /** In write txns, array of cursors for each DB */
881 MDB_cursor **mt_cursors;
882 /** Array of flags for each DB */
883 unsigned char *mt_dbflags;
884 /** Number of DB records in use. This number only ever increments;
885 * we don't decrement it when individual DB handles are closed.
889 /** @defgroup mdb_txn Transaction Flags
893 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
894 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
895 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
896 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
898 unsigned int mt_flags; /**< @ref mdb_txn */
899 /** dirty_list room: Array size - #dirty pages visible to this txn.
900 * Includes ancestor txns' dirty pages not hidden by other txns'
901 * dirty/spilled pages. Thus commit(nested txn) has room to merge
902 * dirty_list into mt_parent after freeing hidden mt_parent pages.
904 unsigned int mt_dirty_room;
907 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
908 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
909 * raise this on a 64 bit machine.
911 #define CURSOR_STACK 32
915 /** Cursors are used for all DB operations.
916 * A cursor holds a path of (page pointer, key index) from the DB
917 * root to a position in the DB, plus other state. #MDB_DUPSORT
918 * cursors include an xcursor to the current data item. Write txns
919 * track their cursors and keep them up to date when data moves.
920 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
921 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
924 /** Next cursor on this DB in this txn */
926 /** Backup of the original cursor if this cursor is a shadow */
927 MDB_cursor *mc_backup;
928 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
929 struct MDB_xcursor *mc_xcursor;
930 /** The transaction that owns this cursor */
932 /** The database handle this cursor operates on */
934 /** The database record for this cursor */
936 /** The database auxiliary record for this cursor */
938 /** The @ref mt_dbflag for this database */
939 unsigned char *mc_dbflag;
940 unsigned short mc_snum; /**< number of pushed pages */
941 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
942 /** @defgroup mdb_cursor Cursor Flags
944 * Cursor state flags.
947 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
948 #define C_EOF 0x02 /**< No more data */
949 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
950 #define C_DEL 0x08 /**< last op was a cursor_del */
951 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
952 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
954 unsigned int mc_flags; /**< @ref mdb_cursor */
955 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
956 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
959 /** Context for sorted-dup records.
960 * We could have gone to a fully recursive design, with arbitrarily
961 * deep nesting of sub-databases. But for now we only handle these
962 * levels - main DB, optional sub-DB, sorted-duplicate DB.
964 typedef struct MDB_xcursor {
965 /** A sub-cursor for traversing the Dup DB */
966 MDB_cursor mx_cursor;
967 /** The database record for this Dup DB */
969 /** The auxiliary DB record for this Dup DB */
971 /** The @ref mt_dbflag for this Dup DB */
972 unsigned char mx_dbflag;
975 /** State of FreeDB old pages, stored in the MDB_env */
976 typedef struct MDB_pgstate {
977 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
978 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
981 /** The database environment. */
983 HANDLE me_fd; /**< The main data file */
984 HANDLE me_lfd; /**< The lock file */
985 HANDLE me_mfd; /**< just for writing the meta pages */
986 /** Failed to update the meta page. Probably an I/O error. */
987 #define MDB_FATAL_ERROR 0x80000000U
988 /** Some fields are initialized. */
989 #define MDB_ENV_ACTIVE 0x20000000U
990 /** me_txkey is set */
991 #define MDB_ENV_TXKEY 0x10000000U
992 /** Have liveness lock in reader table */
993 #define MDB_LIVE_READER 0x08000000U
994 uint32_t me_flags; /**< @ref mdb_env */
995 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
996 unsigned int me_maxreaders; /**< size of the reader table */
997 unsigned int me_numreaders; /**< max numreaders set by this env */
998 MDB_dbi me_numdbs; /**< number of DBs opened */
999 MDB_dbi me_maxdbs; /**< size of the DB table */
1000 pid_t me_pid; /**< process ID of this env */
1001 char *me_path; /**< path to the DB files */
1002 char *me_map; /**< the memory map of the data file */
1003 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1004 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1005 MDB_txn *me_txn; /**< current write transaction */
1006 size_t me_mapsize; /**< size of the data memory map */
1007 off_t me_size; /**< current file size */
1008 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1009 MDB_dbx *me_dbxs; /**< array of static DB info */
1010 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1011 pthread_key_t me_txkey; /**< thread-key for readers */
1012 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1013 # define me_pglast me_pgstate.mf_pglast
1014 # define me_pghead me_pgstate.mf_pghead
1015 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1016 /** IDL of pages that became unused in a write txn */
1017 MDB_IDL me_free_pgs;
1018 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1019 MDB_ID2L me_dirty_list;
1020 /** Max number of freelist items that can fit in a single overflow page */
1022 /** Max size of a node on a page */
1023 unsigned int me_nodemax;
1025 int me_pidquery; /**< Used in OpenProcess */
1026 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1028 #elif defined(MDB_USE_POSIX_SEM)
1029 sem_t *me_rmutex; /* Shared mutexes are not supported */
1034 /** Nested transaction */
1035 typedef struct MDB_ntxn {
1036 MDB_txn mnt_txn; /**< the transaction */
1037 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1040 /** max number of pages to commit in one writev() call */
1041 #define MDB_COMMIT_PAGES 64
1042 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1043 #undef MDB_COMMIT_PAGES
1044 #define MDB_COMMIT_PAGES IOV_MAX
1047 /* max bytes to write in one call */
1048 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1050 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1051 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1052 static int mdb_page_touch(MDB_cursor *mc);
1054 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1055 static int mdb_page_search_root(MDB_cursor *mc,
1056 MDB_val *key, int modify);
1057 #define MDB_PS_MODIFY 1
1058 #define MDB_PS_ROOTONLY 2
1059 #define MDB_PS_FIRST 4
1060 #define MDB_PS_LAST 8
1061 static int mdb_page_search(MDB_cursor *mc,
1062 MDB_val *key, int flags);
1063 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1065 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1066 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1067 pgno_t newpgno, unsigned int nflags);
1069 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1070 static int mdb_env_pick_meta(const MDB_env *env);
1071 static int mdb_env_write_meta(MDB_txn *txn);
1072 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1073 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1075 static void mdb_env_close0(MDB_env *env, int excl);
1077 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1078 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1079 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1080 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1081 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1082 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1083 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1084 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1085 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1087 static int mdb_rebalance(MDB_cursor *mc);
1088 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1090 static void mdb_cursor_pop(MDB_cursor *mc);
1091 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1093 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1094 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1095 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1096 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1097 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1099 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1100 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1102 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1103 static void mdb_xcursor_init0(MDB_cursor *mc);
1104 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1106 static int mdb_drop0(MDB_cursor *mc, int subs);
1107 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1110 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1114 static SECURITY_DESCRIPTOR mdb_null_sd;
1115 static SECURITY_ATTRIBUTES mdb_all_sa;
1116 static int mdb_sec_inited;
1119 /** Return the library version info. */
1121 mdb_version(int *major, int *minor, int *patch)
1123 if (major) *major = MDB_VERSION_MAJOR;
1124 if (minor) *minor = MDB_VERSION_MINOR;
1125 if (patch) *patch = MDB_VERSION_PATCH;
1126 return MDB_VERSION_STRING;
1129 /** Table of descriptions for MDB @ref errors */
1130 static char *const mdb_errstr[] = {
1131 "MDB_KEYEXIST: Key/data pair already exists",
1132 "MDB_NOTFOUND: No matching key/data pair found",
1133 "MDB_PAGE_NOTFOUND: Requested page not found",
1134 "MDB_CORRUPTED: Located page was wrong type",
1135 "MDB_PANIC: Update of meta page failed",
1136 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1137 "MDB_INVALID: File is not an MDB file",
1138 "MDB_MAP_FULL: Environment mapsize limit reached",
1139 "MDB_DBS_FULL: Environment maxdbs limit reached",
1140 "MDB_READERS_FULL: Environment maxreaders limit reached",
1141 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1142 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1143 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1144 "MDB_PAGE_FULL: Internal error - page has no more space",
1145 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1146 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1147 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1148 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1149 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1153 mdb_strerror(int err)
1157 return ("Successful return: 0");
1159 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1160 i = err - MDB_KEYEXIST;
1161 return mdb_errstr[i];
1164 return strerror(err);
1168 /** Display a key in hexadecimal and return the address of the result.
1169 * @param[in] key the key to display
1170 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1171 * @return The key in hexadecimal form.
1174 mdb_dkey(MDB_val *key, char *buf)
1177 unsigned char *c = key->mv_data;
1183 if (key->mv_size > MDB_MAXKEYSIZE)
1184 return "MDB_MAXKEYSIZE";
1185 /* may want to make this a dynamic check: if the key is mostly
1186 * printable characters, print it as-is instead of converting to hex.
1190 for (i=0; i<key->mv_size; i++)
1191 ptr += sprintf(ptr, "%02x", *c++);
1193 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1198 /** Display all the keys in the page. */
1200 mdb_page_list(MDB_page *mp)
1203 unsigned int i, nkeys, nsize;
1207 nkeys = NUMKEYS(mp);
1208 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1209 for (i=0; i<nkeys; i++) {
1210 node = NODEPTR(mp, i);
1211 key.mv_size = node->mn_ksize;
1212 key.mv_data = node->mn_data;
1213 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1214 if (IS_BRANCH(mp)) {
1215 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1218 if (F_ISSET(node->mn_flags, F_BIGDATA))
1219 nsize += sizeof(pgno_t);
1221 nsize += NODEDSZ(node);
1222 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1228 mdb_cursor_chk(MDB_cursor *mc)
1234 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1235 for (i=0; i<mc->mc_top; i++) {
1237 node = NODEPTR(mp, mc->mc_ki[i]);
1238 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1241 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1247 /** Count all the pages in each DB and in the freelist
1248 * and make sure it matches the actual number of pages
1251 static void mdb_audit(MDB_txn *txn)
1255 MDB_ID freecount, count;
1260 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1261 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1262 freecount += *(MDB_ID *)data.mv_data;
1265 for (i = 0; i<txn->mt_numdbs; i++) {
1267 mdb_cursor_init(&mc, txn, i, &mx);
1268 if (txn->mt_dbs[i].md_root == P_INVALID)
1270 count += txn->mt_dbs[i].md_branch_pages +
1271 txn->mt_dbs[i].md_leaf_pages +
1272 txn->mt_dbs[i].md_overflow_pages;
1273 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1274 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1278 mp = mc.mc_pg[mc.mc_top];
1279 for (j=0; j<NUMKEYS(mp); j++) {
1280 MDB_node *leaf = NODEPTR(mp, j);
1281 if (leaf->mn_flags & F_SUBDATA) {
1283 memcpy(&db, NODEDATA(leaf), sizeof(db));
1284 count += db.md_branch_pages + db.md_leaf_pages +
1285 db.md_overflow_pages;
1289 while (mdb_cursor_sibling(&mc, 1) == 0);
1292 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1293 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1294 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1300 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1302 return txn->mt_dbxs[dbi].md_cmp(a, b);
1306 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1308 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1311 /** Allocate memory for a page.
1312 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1315 mdb_page_malloc(MDB_txn *txn, unsigned num)
1317 MDB_env *env = txn->mt_env;
1318 MDB_page *ret = env->me_dpages;
1319 size_t sz = env->me_psize;
1322 VGMEMP_ALLOC(env, ret, sz);
1323 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1324 env->me_dpages = ret->mp_next;
1330 if ((ret = malloc(sz)) != NULL) {
1331 VGMEMP_ALLOC(env, ret, sz);
1336 /** Free a single page.
1337 * Saves single pages to a list, for future reuse.
1338 * (This is not used for multi-page overflow pages.)
1341 mdb_page_free(MDB_env *env, MDB_page *mp)
1343 mp->mp_next = env->me_dpages;
1344 VGMEMP_FREE(env, mp);
1345 env->me_dpages = mp;
1348 /** Free a dirty page */
1350 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1352 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1353 mdb_page_free(env, dp);
1355 /* large pages just get freed directly */
1356 VGMEMP_FREE(env, dp);
1361 /** Return all dirty pages to dpage list */
1363 mdb_dlist_free(MDB_txn *txn)
1365 MDB_env *env = txn->mt_env;
1366 MDB_ID2L dl = txn->mt_u.dirty_list;
1367 unsigned i, n = dl[0].mid;
1369 for (i = 1; i <= n; i++) {
1370 mdb_dpage_free(env, dl[i].mptr);
1375 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1376 * @param[in] mc A cursor handle for the current operation.
1377 * @param[in] pflags Flags of the pages to update:
1378 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1379 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1380 * @return 0 on success, non-zero on failure.
1383 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1385 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1386 MDB_txn *txn = mc->mc_txn;
1392 int rc = MDB_SUCCESS, level;
1394 /* Mark pages seen by cursors */
1395 if (mc->mc_flags & C_UNTRACK)
1396 mc = NULL; /* will find mc in mt_cursors */
1397 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1398 for (; mc; mc=mc->mc_next) {
1399 if (!(mc->mc_flags & C_INITIALIZED))
1401 for (m3 = mc;; m3 = &mx->mx_cursor) {
1403 for (j=0; j<m3->mc_snum; j++) {
1405 if ((mp->mp_flags & Mask) == pflags)
1406 mp->mp_flags ^= P_KEEP;
1408 mx = m3->mc_xcursor;
1409 /* Proceed to mx if it is at a sub-database */
1410 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1412 if (! (mp && (mp->mp_flags & P_LEAF)))
1414 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1415 if (!(leaf->mn_flags & F_SUBDATA))
1424 /* Mark dirty root pages */
1425 for (i=0; i<txn->mt_numdbs; i++) {
1426 if (txn->mt_dbflags[i] & DB_DIRTY) {
1427 pgno_t pgno = txn->mt_dbs[i].md_root;
1428 if (pgno == P_INVALID)
1430 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1432 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1433 dp->mp_flags ^= P_KEEP;
1441 static int mdb_page_flush(MDB_txn *txn, int keep);
1443 /** Spill pages from the dirty list back to disk.
1444 * This is intended to prevent running into #MDB_TXN_FULL situations,
1445 * but note that they may still occur in a few cases:
1446 * 1) our estimate of the txn size could be too small. Currently this
1447 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1448 * 2) child txns may run out of space if their parents dirtied a
1449 * lot of pages and never spilled them. TODO: we probably should do
1450 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1451 * the parent's dirty_room is below a given threshold.
1453 * Otherwise, if not using nested txns, it is expected that apps will
1454 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1455 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1456 * If the txn never references them again, they can be left alone.
1457 * If the txn only reads them, they can be used without any fuss.
1458 * If the txn writes them again, they can be dirtied immediately without
1459 * going thru all of the work of #mdb_page_touch(). Such references are
1460 * handled by #mdb_page_unspill().
1462 * Also note, we never spill DB root pages, nor pages of active cursors,
1463 * because we'll need these back again soon anyway. And in nested txns,
1464 * we can't spill a page in a child txn if it was already spilled in a
1465 * parent txn. That would alter the parent txns' data even though
1466 * the child hasn't committed yet, and we'd have no way to undo it if
1467 * the child aborted.
1469 * @param[in] m0 cursor A cursor handle identifying the transaction and
1470 * database for which we are checking space.
1471 * @param[in] key For a put operation, the key being stored.
1472 * @param[in] data For a put operation, the data being stored.
1473 * @return 0 on success, non-zero on failure.
1476 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1478 MDB_txn *txn = m0->mc_txn;
1480 MDB_ID2L dl = txn->mt_u.dirty_list;
1481 unsigned int i, j, need;
1484 if (m0->mc_flags & C_SUB)
1487 /* Estimate how much space this op will take */
1488 i = m0->mc_db->md_depth;
1489 /* Named DBs also dirty the main DB */
1490 if (m0->mc_dbi > MAIN_DBI)
1491 i += txn->mt_dbs[MAIN_DBI].md_depth;
1492 /* For puts, roughly factor in the key+data size */
1494 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1495 i += i; /* double it for good measure */
1498 if (txn->mt_dirty_room > i)
1501 if (!txn->mt_spill_pgs) {
1502 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1503 if (!txn->mt_spill_pgs)
1506 /* purge deleted slots */
1507 MDB_IDL sl = txn->mt_spill_pgs;
1508 unsigned int num = sl[0];
1510 for (i=1; i<=num; i++) {
1517 /* Preserve pages which may soon be dirtied again */
1518 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1521 /* Less aggressive spill - we originally spilled the entire dirty list,
1522 * with a few exceptions for cursor pages and DB root pages. But this
1523 * turns out to be a lot of wasted effort because in a large txn many
1524 * of those pages will need to be used again. So now we spill only 1/8th
1525 * of the dirty pages. Testing revealed this to be a good tradeoff,
1526 * better than 1/2, 1/4, or 1/10.
1528 if (need < MDB_IDL_UM_MAX / 8)
1529 need = MDB_IDL_UM_MAX / 8;
1531 /* Save the page IDs of all the pages we're flushing */
1532 /* flush from the tail forward, this saves a lot of shifting later on. */
1533 for (i=dl[0].mid; i && need; i--) {
1534 MDB_ID pn = dl[i].mid << 1;
1536 if (dp->mp_flags & P_KEEP)
1538 /* Can't spill twice, make sure it's not already in a parent's
1541 if (txn->mt_parent) {
1543 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1544 if (tx2->mt_spill_pgs) {
1545 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1546 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1547 dp->mp_flags |= P_KEEP;
1555 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1559 mdb_midl_sort(txn->mt_spill_pgs);
1561 /* Flush the spilled part of dirty list */
1562 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1565 /* Reset any dirty pages we kept that page_flush didn't see */
1566 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1569 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1573 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1575 mdb_find_oldest(MDB_txn *txn)
1578 txnid_t mr, oldest = txn->mt_txnid - 1;
1579 if (txn->mt_env->me_txns) {
1580 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1581 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1592 /** Add a page to the txn's dirty list */
1594 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1597 int (*insert)(MDB_ID2L, MDB_ID2 *);
1599 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1600 insert = mdb_mid2l_append;
1602 insert = mdb_mid2l_insert;
1604 mid.mid = mp->mp_pgno;
1606 insert(txn->mt_u.dirty_list, &mid);
1607 txn->mt_dirty_room--;
1610 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1611 * me_pghead and mt_next_pgno.
1613 * If there are free pages available from older transactions, they
1614 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1615 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1616 * and move me_pglast to say which records were consumed. Only this
1617 * function can create me_pghead and move me_pglast/mt_next_pgno.
1618 * @param[in] mc cursor A cursor handle identifying the transaction and
1619 * database for which we are allocating.
1620 * @param[in] num the number of pages to allocate.
1621 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1622 * will always be satisfied by a single contiguous chunk of memory.
1623 * @return 0 on success, non-zero on failure.
1626 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1628 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1629 /* Get at most <Max_retries> more freeDB records once me_pghead
1630 * has enough pages. If not enough, use new pages from the map.
1631 * If <Paranoid> and mc is updating the freeDB, only get new
1632 * records if me_pghead is empty. Then the freelist cannot play
1633 * catch-up with itself by growing while trying to save it.
1635 enum { Paranoid = 1, Max_retries = 500 };
1637 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1639 int rc, n2 = num-1, retry = Max_retries;
1640 MDB_txn *txn = mc->mc_txn;
1641 MDB_env *env = txn->mt_env;
1642 pgno_t pgno, *mop = env->me_pghead;
1643 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1645 txnid_t oldest = 0, last;
1651 /* If our dirty list is already full, we can't do anything */
1652 if (txn->mt_dirty_room == 0)
1653 return MDB_TXN_FULL;
1655 for (op = MDB_FIRST;; op = MDB_NEXT) {
1658 pgno_t *idl, old_id, new_id;
1660 /* Seek a big enough contiguous page range. Prefer
1661 * pages at the tail, just truncating the list.
1663 if (mop_len >= (unsigned)num) {
1667 if (mop[i-n2] == pgno+n2)
1669 } while (--i >= (unsigned)num);
1670 if (Max_retries < INT_MAX && --retry < 0)
1674 if (op == MDB_FIRST) { /* 1st iteration */
1675 /* Prepare to fetch more and coalesce */
1676 oldest = mdb_find_oldest(txn);
1677 last = env->me_pglast;
1678 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1681 key.mv_data = &last; /* will look up last+1 */
1682 key.mv_size = sizeof(last);
1684 if (Paranoid && mc->mc_dbi == FREE_DBI)
1687 if (Paranoid && retry < 0 && mop_len)
1691 /* Do not fetch more if the record will be too recent */
1694 rc = mdb_cursor_get(&m2, &key, NULL, op);
1696 if (rc == MDB_NOTFOUND)
1700 last = *(txnid_t*)key.mv_data;
1703 np = m2.mc_pg[m2.mc_top];
1704 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1705 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1708 idl = (MDB_ID *) data.mv_data;
1711 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1714 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1716 mop = env->me_pghead;
1718 env->me_pglast = last;
1720 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1721 last, txn->mt_dbs[FREE_DBI].md_root, i));
1723 DPRINTF(("IDL %"Z"u", idl[k]));
1725 /* Merge in descending sorted order */
1728 mop[0] = (pgno_t)-1;
1732 for (; old_id < new_id; old_id = mop[--j])
1739 /* Use new pages from the map when nothing suitable in the freeDB */
1741 pgno = txn->mt_next_pgno;
1742 if (pgno + num >= env->me_maxpg) {
1743 DPUTS("DB size maxed out");
1744 return MDB_MAP_FULL;
1748 if (env->me_flags & MDB_WRITEMAP) {
1749 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1751 if (!(np = mdb_page_malloc(txn, num)))
1755 mop[0] = mop_len -= num;
1756 /* Move any stragglers down */
1757 for (j = i-num; j < mop_len; )
1758 mop[++j] = mop[++i];
1760 txn->mt_next_pgno = pgno + num;
1763 mdb_page_dirty(txn, np);
1769 /** Copy the used portions of a non-overflow page.
1770 * @param[in] dst page to copy into
1771 * @param[in] src page to copy from
1772 * @param[in] psize size of a page
1775 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1777 enum { Align = sizeof(pgno_t) };
1778 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1780 /* If page isn't full, just copy the used portion. Adjust
1781 * alignment so memcpy may copy words instead of bytes.
1783 if ((unused &= -Align) && !IS_LEAF2(src)) {
1785 memcpy(dst, src, (lower + (Align-1)) & -Align);
1786 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1789 memcpy(dst, src, psize - unused);
1793 /** Pull a page off the txn's spill list, if present.
1794 * If a page being referenced was spilled to disk in this txn, bring
1795 * it back and make it dirty/writable again.
1796 * @param[in] txn the transaction handle.
1797 * @param[in] mp the page being referenced.
1798 * @param[out] ret the writable page, if any. ret is unchanged if
1799 * mp wasn't spilled.
1802 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1804 MDB_env *env = txn->mt_env;
1807 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1809 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1810 if (!tx2->mt_spill_pgs)
1812 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1813 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1816 if (txn->mt_dirty_room == 0)
1817 return MDB_TXN_FULL;
1818 if (IS_OVERFLOW(mp))
1822 if (env->me_flags & MDB_WRITEMAP) {
1825 np = mdb_page_malloc(txn, num);
1829 memcpy(np, mp, num * env->me_psize);
1831 mdb_page_copy(np, mp, env->me_psize);
1834 /* If in current txn, this page is no longer spilled.
1835 * If it happens to be the last page, truncate the spill list.
1836 * Otherwise mark it as deleted by setting the LSB.
1838 if (x == txn->mt_spill_pgs[0])
1839 txn->mt_spill_pgs[0]--;
1841 txn->mt_spill_pgs[x] |= 1;
1842 } /* otherwise, if belonging to a parent txn, the
1843 * page remains spilled until child commits
1846 mdb_page_dirty(txn, np);
1847 np->mp_flags |= P_DIRTY;
1855 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1856 * @param[in] mc cursor pointing to the page to be touched
1857 * @return 0 on success, non-zero on failure.
1860 mdb_page_touch(MDB_cursor *mc)
1862 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1863 MDB_txn *txn = mc->mc_txn;
1864 MDB_cursor *m2, *m3;
1868 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1869 if (txn->mt_flags & MDB_TXN_SPILLS) {
1871 rc = mdb_page_unspill(txn, mp, &np);
1877 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1878 (rc = mdb_page_alloc(mc, 1, &np)))
1881 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
1882 mp->mp_pgno, pgno));
1883 assert(mp->mp_pgno != pgno);
1884 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1885 /* Update the parent page, if any, to point to the new page */
1887 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1888 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1889 SETPGNO(node, pgno);
1891 mc->mc_db->md_root = pgno;
1893 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1894 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1896 /* If txn has a parent, make sure the page is in our
1900 unsigned x = mdb_mid2l_search(dl, pgno);
1901 if (x <= dl[0].mid && dl[x].mid == pgno) {
1902 if (mp != dl[x].mptr) { /* bad cursor? */
1903 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1904 return MDB_CORRUPTED;
1909 assert(dl[0].mid < MDB_IDL_UM_MAX);
1911 np = mdb_page_malloc(txn, 1);
1916 mdb_mid2l_insert(dl, &mid);
1921 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1923 np->mp_flags |= P_DIRTY;
1926 /* Adjust cursors pointing to mp */
1927 mc->mc_pg[mc->mc_top] = np;
1928 m2 = txn->mt_cursors[mc->mc_dbi];
1929 if (mc->mc_flags & C_SUB) {
1930 for (; m2; m2=m2->mc_next) {
1931 m3 = &m2->mc_xcursor->mx_cursor;
1932 if (m3->mc_snum < mc->mc_snum) continue;
1933 if (m3->mc_pg[mc->mc_top] == mp)
1934 m3->mc_pg[mc->mc_top] = np;
1937 for (; m2; m2=m2->mc_next) {
1938 if (m2->mc_snum < mc->mc_snum) continue;
1939 if (m2->mc_pg[mc->mc_top] == mp) {
1940 m2->mc_pg[mc->mc_top] = np;
1941 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1942 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1944 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1945 if (!(leaf->mn_flags & F_SUBDATA))
1946 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1955 mdb_env_sync(MDB_env *env, int force)
1958 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1959 if (env->me_flags & MDB_WRITEMAP) {
1960 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1961 ? MS_ASYNC : MS_SYNC;
1962 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1965 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1969 if (MDB_FDATASYNC(env->me_fd))
1976 /** Back up parent txn's cursors, then grab the originals for tracking */
1978 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1980 MDB_cursor *mc, *bk;
1985 for (i = src->mt_numdbs; --i >= 0; ) {
1986 if ((mc = src->mt_cursors[i]) != NULL) {
1987 size = sizeof(MDB_cursor);
1989 size += sizeof(MDB_xcursor);
1990 for (; mc; mc = bk->mc_next) {
1996 mc->mc_db = &dst->mt_dbs[i];
1997 /* Kill pointers into src - and dst to reduce abuse: The
1998 * user may not use mc until dst ends. Otherwise we'd...
2000 mc->mc_txn = NULL; /* ...set this to dst */
2001 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2002 if ((mx = mc->mc_xcursor) != NULL) {
2003 *(MDB_xcursor *)(bk+1) = *mx;
2004 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2006 mc->mc_next = dst->mt_cursors[i];
2007 dst->mt_cursors[i] = mc;
2014 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2015 * @param[in] txn the transaction handle.
2016 * @param[in] merge true to keep changes to parent cursors, false to revert.
2017 * @return 0 on success, non-zero on failure.
2020 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2022 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2026 for (i = txn->mt_numdbs; --i >= 0; ) {
2027 for (mc = cursors[i]; mc; mc = next) {
2029 if ((bk = mc->mc_backup) != NULL) {
2031 /* Commit changes to parent txn */
2032 mc->mc_next = bk->mc_next;
2033 mc->mc_backup = bk->mc_backup;
2034 mc->mc_txn = bk->mc_txn;
2035 mc->mc_db = bk->mc_db;
2036 mc->mc_dbflag = bk->mc_dbflag;
2037 if ((mx = mc->mc_xcursor) != NULL)
2038 mx->mx_cursor.mc_txn = bk->mc_txn;
2040 /* Abort nested txn */
2042 if ((mx = mc->mc_xcursor) != NULL)
2043 *mx = *(MDB_xcursor *)(bk+1);
2047 /* Only malloced cursors are permanently tracked. */
2055 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2058 mdb_txn_reset0(MDB_txn *txn, const char *act);
2060 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2066 Pidset = F_SETLK, Pidcheck = F_GETLK
2070 /** Set or check a pid lock. Set returns 0 on success.
2071 * Check returns 0 if the process is certainly dead, nonzero if it may
2072 * be alive (the lock exists or an error happened so we do not know).
2074 * On Windows Pidset is a no-op, we merely check for the existence
2075 * of the process with the given pid. On POSIX we use a single byte
2076 * lock on the lockfile, set at an offset equal to the pid.
2079 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2081 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2084 if (op == Pidcheck) {
2085 h = OpenProcess(env->me_pidquery, FALSE, pid);
2086 /* No documented "no such process" code, but other program use this: */
2088 return ErrCode() != ERROR_INVALID_PARAMETER;
2089 /* A process exists until all handles to it close. Has it exited? */
2090 ret = WaitForSingleObject(h, 0) != 0;
2097 struct flock lock_info;
2098 memset(&lock_info, 0, sizeof(lock_info));
2099 lock_info.l_type = F_WRLCK;
2100 lock_info.l_whence = SEEK_SET;
2101 lock_info.l_start = pid;
2102 lock_info.l_len = 1;
2103 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2104 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2106 } else if ((rc = ErrCode()) == EINTR) {
2114 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2115 * @param[in] txn the transaction handle to initialize
2116 * @return 0 on success, non-zero on failure.
2119 mdb_txn_renew0(MDB_txn *txn)
2121 MDB_env *env = txn->mt_env;
2122 MDB_txninfo *ti = env->me_txns;
2126 int rc, new_notls = 0;
2129 txn->mt_numdbs = env->me_numdbs;
2130 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2132 if (txn->mt_flags & MDB_TXN_RDONLY) {
2134 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2135 txn->mt_txnid = meta->mm_txnid;
2136 txn->mt_u.reader = NULL;
2138 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2139 pthread_getspecific(env->me_txkey);
2141 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2142 return MDB_BAD_RSLOT;
2144 pid_t pid = env->me_pid;
2145 pthread_t tid = pthread_self();
2147 if (!(env->me_flags & MDB_LIVE_READER)) {
2148 rc = mdb_reader_pid(env, Pidset, pid);
2150 UNLOCK_MUTEX_R(env);
2153 env->me_flags |= MDB_LIVE_READER;
2157 nr = ti->mti_numreaders;
2158 for (i=0; i<nr; i++)
2159 if (ti->mti_readers[i].mr_pid == 0)
2161 if (i == env->me_maxreaders) {
2162 UNLOCK_MUTEX_R(env);
2163 return MDB_READERS_FULL;
2165 ti->mti_readers[i].mr_pid = pid;
2166 ti->mti_readers[i].mr_tid = tid;
2168 ti->mti_numreaders = ++nr;
2169 /* Save numreaders for un-mutexed mdb_env_close() */
2170 env->me_numreaders = nr;
2171 UNLOCK_MUTEX_R(env);
2173 r = &ti->mti_readers[i];
2174 new_notls = (env->me_flags & MDB_NOTLS);
2175 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2180 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2181 txn->mt_u.reader = r;
2182 meta = env->me_metas[txn->mt_txnid & 1];
2188 txn->mt_txnid = ti->mti_txnid;
2189 meta = env->me_metas[txn->mt_txnid & 1];
2191 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2192 txn->mt_txnid = meta->mm_txnid;
2196 if (txn->mt_txnid == mdb_debug_start)
2199 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2200 txn->mt_u.dirty_list = env->me_dirty_list;
2201 txn->mt_u.dirty_list[0].mid = 0;
2202 txn->mt_free_pgs = env->me_free_pgs;
2203 txn->mt_free_pgs[0] = 0;
2204 txn->mt_spill_pgs = NULL;
2208 /* Copy the DB info and flags */
2209 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2211 /* Moved to here to avoid a data race in read TXNs */
2212 txn->mt_next_pgno = meta->mm_last_pg+1;
2214 for (i=2; i<txn->mt_numdbs; i++) {
2215 x = env->me_dbflags[i];
2216 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2217 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2219 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2221 if (env->me_maxpg < txn->mt_next_pgno) {
2222 mdb_txn_reset0(txn, "renew0-mapfail");
2224 txn->mt_u.reader->mr_pid = 0;
2225 txn->mt_u.reader = NULL;
2227 return MDB_MAP_RESIZED;
2234 mdb_txn_renew(MDB_txn *txn)
2238 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2241 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2242 DPUTS("environment had fatal error, must shutdown!");
2246 rc = mdb_txn_renew0(txn);
2247 if (rc == MDB_SUCCESS) {
2248 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2249 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2250 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2256 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2260 int rc, size, tsize = sizeof(MDB_txn);
2262 if (env->me_flags & MDB_FATAL_ERROR) {
2263 DPUTS("environment had fatal error, must shutdown!");
2266 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2269 /* Nested transactions: Max 1 child, write txns only, no writemap */
2270 if (parent->mt_child ||
2271 (flags & MDB_RDONLY) ||
2272 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2273 (env->me_flags & MDB_WRITEMAP))
2275 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2277 tsize = sizeof(MDB_ntxn);
2279 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2280 if (!(flags & MDB_RDONLY))
2281 size += env->me_maxdbs * sizeof(MDB_cursor *);
2283 if ((txn = calloc(1, size)) == NULL) {
2284 DPRINTF(("calloc: %s", strerror(ErrCode())));
2287 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2288 if (flags & MDB_RDONLY) {
2289 txn->mt_flags |= MDB_TXN_RDONLY;
2290 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2292 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2293 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2299 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2300 if (!txn->mt_u.dirty_list ||
2301 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2303 free(txn->mt_u.dirty_list);
2307 txn->mt_txnid = parent->mt_txnid;
2308 txn->mt_dirty_room = parent->mt_dirty_room;
2309 txn->mt_u.dirty_list[0].mid = 0;
2310 txn->mt_spill_pgs = NULL;
2311 txn->mt_next_pgno = parent->mt_next_pgno;
2312 parent->mt_child = txn;
2313 txn->mt_parent = parent;
2314 txn->mt_numdbs = parent->mt_numdbs;
2315 txn->mt_flags = parent->mt_flags;
2316 txn->mt_dbxs = parent->mt_dbxs;
2317 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2318 /* Copy parent's mt_dbflags, but clear DB_NEW */
2319 for (i=0; i<txn->mt_numdbs; i++)
2320 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2322 ntxn = (MDB_ntxn *)txn;
2323 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2324 if (env->me_pghead) {
2325 size = MDB_IDL_SIZEOF(env->me_pghead);
2326 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2328 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2333 rc = mdb_cursor_shadow(parent, txn);
2335 mdb_txn_reset0(txn, "beginchild-fail");
2337 rc = mdb_txn_renew0(txn);
2343 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2344 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2345 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2352 mdb_txn_env(MDB_txn *txn)
2354 if(!txn) return NULL;
2358 /** Export or close DBI handles opened in this txn. */
2360 mdb_dbis_update(MDB_txn *txn, int keep)
2363 MDB_dbi n = txn->mt_numdbs;
2364 MDB_env *env = txn->mt_env;
2365 unsigned char *tdbflags = txn->mt_dbflags;
2367 for (i = n; --i >= 2;) {
2368 if (tdbflags[i] & DB_NEW) {
2370 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2372 char *ptr = env->me_dbxs[i].md_name.mv_data;
2373 env->me_dbxs[i].md_name.mv_data = NULL;
2374 env->me_dbxs[i].md_name.mv_size = 0;
2375 env->me_dbflags[i] = 0;
2380 if (keep && env->me_numdbs < n)
2384 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2385 * May be called twice for readonly txns: First reset it, then abort.
2386 * @param[in] txn the transaction handle to reset
2387 * @param[in] act why the transaction is being reset
2390 mdb_txn_reset0(MDB_txn *txn, const char *act)
2392 MDB_env *env = txn->mt_env;
2394 /* Close any DBI handles opened in this txn */
2395 mdb_dbis_update(txn, 0);
2397 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2398 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2399 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2401 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2402 if (txn->mt_u.reader) {
2403 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2404 if (!(env->me_flags & MDB_NOTLS))
2405 txn->mt_u.reader = NULL; /* txn does not own reader */
2407 txn->mt_numdbs = 0; /* close nothing if called again */
2408 txn->mt_dbxs = NULL; /* mark txn as reset */
2410 mdb_cursors_close(txn, 0);
2412 if (!(env->me_flags & MDB_WRITEMAP)) {
2413 mdb_dlist_free(txn);
2415 mdb_midl_free(env->me_pghead);
2417 if (txn->mt_parent) {
2418 txn->mt_parent->mt_child = NULL;
2419 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2420 mdb_midl_free(txn->mt_free_pgs);
2421 mdb_midl_free(txn->mt_spill_pgs);
2422 free(txn->mt_u.dirty_list);
2426 if (mdb_midl_shrink(&txn->mt_free_pgs))
2427 env->me_free_pgs = txn->mt_free_pgs;
2428 env->me_pghead = NULL;
2432 /* The writer mutex was locked in mdb_txn_begin. */
2434 UNLOCK_MUTEX_W(env);
2439 mdb_txn_reset(MDB_txn *txn)
2444 /* This call is only valid for read-only txns */
2445 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2448 mdb_txn_reset0(txn, "reset");
2452 mdb_txn_abort(MDB_txn *txn)
2458 mdb_txn_abort(txn->mt_child);
2460 mdb_txn_reset0(txn, "abort");
2461 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2462 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2463 txn->mt_u.reader->mr_pid = 0;
2468 /** Save the freelist as of this transaction to the freeDB.
2469 * This changes the freelist. Keep trying until it stabilizes.
2472 mdb_freelist_save(MDB_txn *txn)
2474 /* env->me_pghead[] can grow and shrink during this call.
2475 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2476 * Page numbers cannot disappear from txn->mt_free_pgs[].
2479 MDB_env *env = txn->mt_env;
2480 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2481 txnid_t pglast = 0, head_id = 0;
2482 pgno_t freecnt = 0, *free_pgs, *mop;
2483 ssize_t head_room = 0, total_room = 0, mop_len;
2485 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2487 if (env->me_pghead) {
2488 /* Make sure first page of freeDB is touched and on freelist */
2489 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2490 if (rc && rc != MDB_NOTFOUND)
2495 /* Come back here after each Put() in case freelist changed */
2498 /* If using records from freeDB which we have not yet
2499 * deleted, delete them and any we reserved for me_pghead.
2501 while (pglast < env->me_pglast) {
2502 rc = mdb_cursor_first(&mc, &key, NULL);
2505 pglast = head_id = *(txnid_t *)key.mv_data;
2506 total_room = head_room = 0;
2507 assert(pglast <= env->me_pglast);
2508 rc = mdb_cursor_del(&mc, 0);
2513 /* Save the IDL of pages freed by this txn, to a single record */
2514 if (freecnt < txn->mt_free_pgs[0]) {
2516 /* Make sure last page of freeDB is touched and on freelist */
2517 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2518 if (rc && rc != MDB_NOTFOUND)
2521 free_pgs = txn->mt_free_pgs;
2522 /* Write to last page of freeDB */
2523 key.mv_size = sizeof(txn->mt_txnid);
2524 key.mv_data = &txn->mt_txnid;
2526 freecnt = free_pgs[0];
2527 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2528 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2531 /* Retry if mt_free_pgs[] grew during the Put() */
2532 free_pgs = txn->mt_free_pgs;
2533 } while (freecnt < free_pgs[0]);
2534 mdb_midl_sort(free_pgs);
2535 memcpy(data.mv_data, free_pgs, data.mv_size);
2538 unsigned int i = free_pgs[0];
2539 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2540 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2542 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2548 mop = env->me_pghead;
2549 mop_len = mop ? mop[0] : 0;
2551 /* Reserve records for me_pghead[]. Split it if multi-page,
2552 * to avoid searching freeDB for a page range. Use keys in
2553 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2555 if (total_room >= mop_len) {
2556 if (total_room == mop_len || --more < 0)
2558 } else if (head_room >= maxfree_1pg && head_id > 1) {
2559 /* Keep current record (overflow page), add a new one */
2563 /* (Re)write {key = head_id, IDL length = head_room} */
2564 total_room -= head_room;
2565 head_room = mop_len - total_room;
2566 if (head_room > maxfree_1pg && head_id > 1) {
2567 /* Overflow multi-page for part of me_pghead */
2568 head_room /= head_id; /* amortize page sizes */
2569 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2570 } else if (head_room < 0) {
2571 /* Rare case, not bothering to delete this record */
2574 key.mv_size = sizeof(head_id);
2575 key.mv_data = &head_id;
2576 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2577 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2580 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2581 total_room += head_room;
2584 /* Fill in the reserved me_pghead records */
2590 rc = mdb_cursor_first(&mc, &key, &data);
2591 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2592 unsigned flags = MDB_CURRENT;
2593 txnid_t id = *(txnid_t *)key.mv_data;
2594 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2597 assert(len >= 0 && id <= env->me_pglast);
2599 if (len > mop_len) {
2601 data.mv_size = (len + 1) * sizeof(MDB_ID);
2604 data.mv_data = mop -= len;
2607 rc = mdb_cursor_put(&mc, &key, &data, flags);
2609 if (rc || !(mop_len -= len))
2616 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2617 * @param[in] txn the transaction that's being committed
2618 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2619 * @return 0 on success, non-zero on failure.
2622 mdb_page_flush(MDB_txn *txn, int keep)
2624 MDB_env *env = txn->mt_env;
2625 MDB_ID2L dl = txn->mt_u.dirty_list;
2626 unsigned psize = env->me_psize, j;
2627 int i, pagecount = dl[0].mid, rc;
2628 size_t size = 0, pos = 0;
2630 MDB_page *dp = NULL;
2634 struct iovec iov[MDB_COMMIT_PAGES];
2635 ssize_t wpos = 0, wsize = 0, wres;
2636 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2642 if (env->me_flags & MDB_WRITEMAP) {
2643 /* Clear dirty flags */
2644 while (++i <= pagecount) {
2646 /* Don't flush this page yet */
2647 if (dp->mp_flags & P_KEEP) {
2648 dp->mp_flags ^= P_KEEP;
2652 dp->mp_flags &= ~P_DIRTY;
2657 /* Write the pages */
2659 if (++i <= pagecount) {
2661 /* Don't flush this page yet */
2662 if (dp->mp_flags & P_KEEP) {
2663 dp->mp_flags ^= P_KEEP;
2668 /* clear dirty flag */
2669 dp->mp_flags &= ~P_DIRTY;
2672 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2677 /* Windows actually supports scatter/gather I/O, but only on
2678 * unbuffered file handles. Since we're relying on the OS page
2679 * cache for all our data, that's self-defeating. So we just
2680 * write pages one at a time. We use the ov structure to set
2681 * the write offset, to at least save the overhead of a Seek
2684 DPRINTF(("committing page %"Z"u", pgno));
2685 memset(&ov, 0, sizeof(ov));
2686 ov.Offset = pos & 0xffffffff;
2687 ov.OffsetHigh = pos >> 16 >> 16;
2688 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2690 DPRINTF(("WriteFile: %d", rc));
2694 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2695 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2697 /* Write previous page(s) */
2698 #ifdef MDB_USE_PWRITEV
2699 wres = pwritev(env->me_fd, iov, n, wpos);
2702 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2704 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2706 DPRINTF(("lseek: %s", strerror(rc)));
2709 wres = writev(env->me_fd, iov, n);
2712 if (wres != wsize) {
2715 DPRINTF(("Write error: %s", strerror(rc)));
2717 rc = EIO; /* TODO: Use which error code? */
2718 DPUTS("short write, filesystem full?");
2729 DPRINTF(("committing page %"Z"u", pgno));
2730 next_pos = pos + size;
2731 iov[n].iov_len = size;
2732 iov[n].iov_base = (char *)dp;
2738 for (i = keep; ++i <= pagecount; ) {
2740 /* This is a page we skipped above */
2743 dl[j].mid = dp->mp_pgno;
2746 mdb_dpage_free(env, dp);
2751 txn->mt_dirty_room += i - j;
2757 mdb_txn_commit(MDB_txn *txn)
2763 assert(txn != NULL);
2764 assert(txn->mt_env != NULL);
2766 if (txn->mt_child) {
2767 rc = mdb_txn_commit(txn->mt_child);
2768 txn->mt_child = NULL;
2775 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2776 mdb_dbis_update(txn, 1);
2777 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2782 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2783 DPUTS("error flag is set, can't commit");
2785 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2790 if (txn->mt_parent) {
2791 MDB_txn *parent = txn->mt_parent;
2794 unsigned x, y, len, ps_len;
2796 /* Append our free list to parent's */
2797 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2800 mdb_midl_free(txn->mt_free_pgs);
2801 /* Failures after this must either undo the changes
2802 * to the parent or set MDB_TXN_ERROR in the parent.
2805 parent->mt_next_pgno = txn->mt_next_pgno;
2806 parent->mt_flags = txn->mt_flags;
2808 /* Merge our cursors into parent's and close them */
2809 mdb_cursors_close(txn, 1);
2811 /* Update parent's DB table. */
2812 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2813 parent->mt_numdbs = txn->mt_numdbs;
2814 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2815 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2816 for (i=2; i<txn->mt_numdbs; i++) {
2817 /* preserve parent's DB_NEW status */
2818 x = parent->mt_dbflags[i] & DB_NEW;
2819 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2822 dst = parent->mt_u.dirty_list;
2823 src = txn->mt_u.dirty_list;
2824 /* Remove anything in our dirty list from parent's spill list */
2825 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2827 pspill[0] = (pgno_t)-1;
2828 /* Mark our dirty pages as deleted in parent spill list */
2829 for (i=0, len=src[0].mid; ++i <= len; ) {
2830 MDB_ID pn = src[i].mid << 1;
2831 while (pn > pspill[x])
2833 if (pn == pspill[x]) {
2838 /* Squash deleted pagenums if we deleted any */
2839 for (x=y; ++x <= ps_len; )
2840 if (!(pspill[x] & 1))
2841 pspill[++y] = pspill[x];
2845 /* Find len = length of merging our dirty list with parent's */
2847 dst[0].mid = 0; /* simplify loops */
2848 if (parent->mt_parent) {
2849 len = x + src[0].mid;
2850 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2851 for (i = x; y && i; y--) {
2852 pgno_t yp = src[y].mid;
2853 while (yp < dst[i].mid)
2855 if (yp == dst[i].mid) {
2860 } else { /* Simplify the above for single-ancestor case */
2861 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2863 /* Merge our dirty list with parent's */
2865 for (i = len; y; dst[i--] = src[y--]) {
2866 pgno_t yp = src[y].mid;
2867 while (yp < dst[x].mid)
2868 dst[i--] = dst[x--];
2869 if (yp == dst[x].mid)
2870 free(dst[x--].mptr);
2874 free(txn->mt_u.dirty_list);
2875 parent->mt_dirty_room = txn->mt_dirty_room;
2876 if (txn->mt_spill_pgs) {
2877 if (parent->mt_spill_pgs) {
2878 /* TODO: Prevent failure here, so parent does not fail */
2879 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2881 parent->mt_flags |= MDB_TXN_ERROR;
2882 mdb_midl_free(txn->mt_spill_pgs);
2883 mdb_midl_sort(parent->mt_spill_pgs);
2885 parent->mt_spill_pgs = txn->mt_spill_pgs;
2889 parent->mt_child = NULL;
2890 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2895 if (txn != env->me_txn) {
2896 DPUTS("attempt to commit unknown transaction");
2901 mdb_cursors_close(txn, 0);
2903 if (!txn->mt_u.dirty_list[0].mid &&
2904 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2907 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2908 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2910 /* Update DB root pointers */
2911 if (txn->mt_numdbs > 2) {
2915 data.mv_size = sizeof(MDB_db);
2917 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2918 for (i = 2; i < txn->mt_numdbs; i++) {
2919 if (txn->mt_dbflags[i] & DB_DIRTY) {
2920 data.mv_data = &txn->mt_dbs[i];
2921 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2928 rc = mdb_freelist_save(txn);
2932 mdb_midl_free(env->me_pghead);
2933 env->me_pghead = NULL;
2934 if (mdb_midl_shrink(&txn->mt_free_pgs))
2935 env->me_free_pgs = txn->mt_free_pgs;
2941 if ((rc = mdb_page_flush(txn, 0)) ||
2942 (rc = mdb_env_sync(env, 0)) ||
2943 (rc = mdb_env_write_meta(txn)))
2949 mdb_dbis_update(txn, 1);
2952 UNLOCK_MUTEX_W(env);
2962 /** Read the environment parameters of a DB environment before
2963 * mapping it into memory.
2964 * @param[in] env the environment handle
2965 * @param[out] meta address of where to store the meta information
2966 * @return 0 on success, non-zero on failure.
2969 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2976 /* We don't know the page size yet, so use a minimum value.
2977 * Read both meta pages so we can use the latest one.
2980 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2984 memset(&ov, 0, sizeof(ov));
2986 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2987 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2990 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2992 if (rc != MDB_PAGESIZE) {
2993 if (rc == 0 && off == 0)
2995 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2996 DPRINTF(("read: %s", mdb_strerror(rc)));
3000 p = (MDB_page *)&pbuf;
3002 if (!F_ISSET(p->mp_flags, P_META)) {
3003 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3008 if (m->mm_magic != MDB_MAGIC) {
3009 DPUTS("meta has invalid magic");
3013 if (m->mm_version != MDB_DATA_VERSION) {
3014 DPRINTF(("database is version %u, expected version %u",
3015 m->mm_version, MDB_DATA_VERSION));
3016 return MDB_VERSION_MISMATCH;
3019 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3025 /** Write the environment parameters of a freshly created DB environment.
3026 * @param[in] env the environment handle
3027 * @param[out] meta address of where to store the meta information
3028 * @return 0 on success, non-zero on failure.
3031 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3039 memset(&ov, 0, sizeof(ov));
3040 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3042 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3045 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3046 len = pwrite(fd, ptr, size, pos); \
3047 rc = (len >= 0); } while(0)
3050 DPUTS("writing new meta page");
3052 psize = env->me_psize;
3054 meta->mm_magic = MDB_MAGIC;
3055 meta->mm_version = MDB_DATA_VERSION;
3056 meta->mm_mapsize = env->me_mapsize;
3057 meta->mm_psize = psize;
3058 meta->mm_last_pg = 1;
3059 meta->mm_flags = env->me_flags & 0xffff;
3060 meta->mm_flags |= MDB_INTEGERKEY;
3061 meta->mm_dbs[0].md_root = P_INVALID;
3062 meta->mm_dbs[1].md_root = P_INVALID;
3064 p = calloc(2, psize);
3066 p->mp_flags = P_META;
3067 *(MDB_meta *)METADATA(p) = *meta;
3069 q = (MDB_page *)((char *)p + psize);
3071 q->mp_flags = P_META;
3072 *(MDB_meta *)METADATA(q) = *meta;
3074 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3077 else if ((unsigned) len == psize * 2)
3085 /** Update the environment info to commit a transaction.
3086 * @param[in] txn the transaction that's being committed
3087 * @return 0 on success, non-zero on failure.
3090 mdb_env_write_meta(MDB_txn *txn)
3093 MDB_meta meta, metab, *mp;
3095 int rc, len, toggle;
3104 assert(txn != NULL);
3105 assert(txn->mt_env != NULL);
3107 toggle = txn->mt_txnid & 1;
3108 DPRINTF(("writing meta page %d for root page %"Z"u",
3109 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3112 mp = env->me_metas[toggle];
3114 if (env->me_flags & MDB_WRITEMAP) {
3115 /* Persist any increases of mapsize config */
3116 if (env->me_mapsize > mp->mm_mapsize)
3117 mp->mm_mapsize = env->me_mapsize;
3118 mp->mm_dbs[0] = txn->mt_dbs[0];
3119 mp->mm_dbs[1] = txn->mt_dbs[1];
3120 mp->mm_last_pg = txn->mt_next_pgno - 1;
3121 mp->mm_txnid = txn->mt_txnid;
3122 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3123 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3126 ptr += env->me_psize;
3127 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3134 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3135 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3137 ptr = (char *)&meta;
3138 if (env->me_mapsize > mp->mm_mapsize) {
3139 /* Persist any increases of mapsize config */
3140 meta.mm_mapsize = env->me_mapsize;
3141 off = offsetof(MDB_meta, mm_mapsize);
3143 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3145 len = sizeof(MDB_meta) - off;
3148 meta.mm_dbs[0] = txn->mt_dbs[0];
3149 meta.mm_dbs[1] = txn->mt_dbs[1];
3150 meta.mm_last_pg = txn->mt_next_pgno - 1;
3151 meta.mm_txnid = txn->mt_txnid;
3154 off += env->me_psize;
3157 /* Write to the SYNC fd */
3158 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3159 env->me_fd : env->me_mfd;
3162 memset(&ov, 0, sizeof(ov));
3164 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3168 rc = pwrite(mfd, ptr, len, off);
3171 rc = rc < 0 ? ErrCode() : EIO;
3172 DPUTS("write failed, disk error?");
3173 /* On a failure, the pagecache still contains the new data.
3174 * Write some old data back, to prevent it from being used.
3175 * Use the non-SYNC fd; we know it will fail anyway.
3177 meta.mm_last_pg = metab.mm_last_pg;
3178 meta.mm_txnid = metab.mm_txnid;
3180 memset(&ov, 0, sizeof(ov));
3182 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3184 r2 = pwrite(env->me_fd, ptr, len, off);
3185 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3188 env->me_flags |= MDB_FATAL_ERROR;
3192 /* Memory ordering issues are irrelevant; since the entire writer
3193 * is wrapped by wmutex, all of these changes will become visible
3194 * after the wmutex is unlocked. Since the DB is multi-version,
3195 * readers will get consistent data regardless of how fresh or
3196 * how stale their view of these values is.
3199 env->me_txns->mti_txnid = txn->mt_txnid;
3204 /** Check both meta pages to see which one is newer.
3205 * @param[in] env the environment handle
3206 * @return meta toggle (0 or 1).
3209 mdb_env_pick_meta(const MDB_env *env)
3211 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3215 mdb_env_create(MDB_env **env)
3219 e = calloc(1, sizeof(MDB_env));
3223 e->me_maxreaders = DEFAULT_READERS;
3224 e->me_maxdbs = e->me_numdbs = 2;
3225 e->me_fd = INVALID_HANDLE_VALUE;
3226 e->me_lfd = INVALID_HANDLE_VALUE;
3227 e->me_mfd = INVALID_HANDLE_VALUE;
3228 #ifdef MDB_USE_POSIX_SEM
3229 e->me_rmutex = SEM_FAILED;
3230 e->me_wmutex = SEM_FAILED;
3232 e->me_pid = getpid();
3233 VGMEMP_CREATE(e,0,0);
3239 mdb_env_map(MDB_env *env, void *addr, int newsize)
3242 unsigned int flags = env->me_flags;
3246 LONG sizelo, sizehi;
3247 sizelo = env->me_mapsize & 0xffffffff;
3248 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3250 /* Windows won't create mappings for zero length files.
3251 * Just allocate the maxsize right now.
3254 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3255 || !SetEndOfFile(env->me_fd)
3256 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3259 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3260 PAGE_READWRITE : PAGE_READONLY,
3261 sizehi, sizelo, NULL);
3264 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3265 FILE_MAP_WRITE : FILE_MAP_READ,
3266 0, 0, env->me_mapsize, addr);
3267 rc = env->me_map ? 0 : ErrCode();
3272 int prot = PROT_READ;
3273 if (flags & MDB_WRITEMAP) {
3275 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3278 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3280 if (env->me_map == MAP_FAILED) {
3284 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3286 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3288 #ifdef POSIX_MADV_RANDOM
3289 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3290 #endif /* POSIX_MADV_RANDOM */
3291 #endif /* MADV_RANDOM */
3294 /* Can happen because the address argument to mmap() is just a
3295 * hint. mmap() can pick another, e.g. if the range is in use.
3296 * The MAP_FIXED flag would prevent that, but then mmap could
3297 * instead unmap existing pages to make room for the new map.
3299 if (addr && env->me_map != addr)
3300 return EBUSY; /* TODO: Make a new MDB_* error code? */
3302 p = (MDB_page *)env->me_map;
3303 env->me_metas[0] = METADATA(p);
3304 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3310 mdb_env_set_mapsize(MDB_env *env, size_t size)
3312 /* If env is already open, caller is responsible for making
3313 * sure there are no active txns.
3321 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3322 else if (size < env->me_mapsize) {
3323 /* If the configured size is smaller, make sure it's
3324 * still big enough. Silently round up to minimum if not.
3326 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3330 munmap(env->me_map, env->me_mapsize);
3331 env->me_mapsize = size;
3332 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3333 rc = mdb_env_map(env, old, 1);
3337 env->me_mapsize = size;
3339 env->me_maxpg = env->me_mapsize / env->me_psize;
3344 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3348 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3353 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3355 if (env->me_map || readers < 1)
3357 env->me_maxreaders = readers;
3362 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3364 if (!env || !readers)
3366 *readers = env->me_maxreaders;
3370 /** Further setup required for opening an MDB environment
3373 mdb_env_open2(MDB_env *env)
3375 unsigned int flags = env->me_flags;
3376 int i, newenv = 0, rc;
3380 /* See if we should use QueryLimited */
3382 if ((rc & 0xff) > 5)
3383 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3385 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3388 memset(&meta, 0, sizeof(meta));
3390 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3393 DPUTS("new mdbenv");
3395 GET_PAGESIZE(env->me_psize);
3397 env->me_psize = meta.mm_psize;
3400 /* Was a mapsize configured? */
3401 if (!env->me_mapsize) {
3402 /* If this is a new environment, take the default,
3403 * else use the size recorded in the existing env.
3405 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3406 } else if (env->me_mapsize < meta.mm_mapsize) {
3407 /* If the configured size is smaller, make sure it's
3408 * still big enough. Silently round up to minimum if not.
3410 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3411 if (env->me_mapsize < minsize)
3412 env->me_mapsize = minsize;
3415 rc = mdb_env_map(env, meta.mm_address, newenv);
3420 if (flags & MDB_FIXEDMAP)
3421 meta.mm_address = env->me_map;
3422 i = mdb_env_init_meta(env, &meta);
3423 if (i != MDB_SUCCESS) {
3427 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3428 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3430 env->me_maxpg = env->me_mapsize / env->me_psize;
3433 int toggle = mdb_env_pick_meta(env);
3434 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3436 DPRINTF(("opened database version %u, pagesize %u",
3437 env->me_metas[0]->mm_version, env->me_psize));
3438 DPRINTF(("using meta page %d", toggle));
3439 DPRINTF(("depth: %u", db->md_depth));
3440 DPRINTF(("entries: %"Z"u", db->md_entries));
3441 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3442 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3443 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3444 DPRINTF(("root: %"Z"u", db->md_root));
3452 /** Release a reader thread's slot in the reader lock table.
3453 * This function is called automatically when a thread exits.
3454 * @param[in] ptr This points to the slot in the reader lock table.
3457 mdb_env_reader_dest(void *ptr)
3459 MDB_reader *reader = ptr;
3465 /** Junk for arranging thread-specific callbacks on Windows. This is
3466 * necessarily platform and compiler-specific. Windows supports up
3467 * to 1088 keys. Let's assume nobody opens more than 64 environments
3468 * in a single process, for now. They can override this if needed.
3470 #ifndef MAX_TLS_KEYS
3471 #define MAX_TLS_KEYS 64
3473 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3474 static int mdb_tls_nkeys;
3476 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3480 case DLL_PROCESS_ATTACH: break;
3481 case DLL_THREAD_ATTACH: break;
3482 case DLL_THREAD_DETACH:
3483 for (i=0; i<mdb_tls_nkeys; i++) {
3484 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3485 mdb_env_reader_dest(r);
3488 case DLL_PROCESS_DETACH: break;
3493 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3495 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3499 /* Force some symbol references.
3500 * _tls_used forces the linker to create the TLS directory if not already done
3501 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3503 #pragma comment(linker, "/INCLUDE:_tls_used")
3504 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3505 #pragma const_seg(".CRT$XLB")
3506 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3507 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3510 #pragma comment(linker, "/INCLUDE:__tls_used")
3511 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3512 #pragma data_seg(".CRT$XLB")
3513 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3515 #endif /* WIN 32/64 */
3516 #endif /* !__GNUC__ */
3519 /** Downgrade the exclusive lock on the region back to shared */
3521 mdb_env_share_locks(MDB_env *env, int *excl)
3523 int rc = 0, toggle = mdb_env_pick_meta(env);
3525 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3530 /* First acquire a shared lock. The Unlock will
3531 * then release the existing exclusive lock.
3533 memset(&ov, 0, sizeof(ov));
3534 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3537 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3543 struct flock lock_info;
3544 /* The shared lock replaces the existing lock */
3545 memset((void *)&lock_info, 0, sizeof(lock_info));
3546 lock_info.l_type = F_RDLCK;
3547 lock_info.l_whence = SEEK_SET;
3548 lock_info.l_start = 0;
3549 lock_info.l_len = 1;
3550 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3551 (rc = ErrCode()) == EINTR) ;
3552 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3559 /** Try to get exlusive lock, otherwise shared.
3560 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3563 mdb_env_excl_lock(MDB_env *env, int *excl)
3567 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3571 memset(&ov, 0, sizeof(ov));
3572 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3579 struct flock lock_info;
3580 memset((void *)&lock_info, 0, sizeof(lock_info));
3581 lock_info.l_type = F_WRLCK;
3582 lock_info.l_whence = SEEK_SET;
3583 lock_info.l_start = 0;
3584 lock_info.l_len = 1;
3585 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3586 (rc = ErrCode()) == EINTR) ;
3590 # ifdef MDB_USE_POSIX_SEM
3591 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3594 lock_info.l_type = F_RDLCK;
3595 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3596 (rc = ErrCode()) == EINTR) ;
3604 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3606 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3608 * @(#) $Revision: 5.1 $
3609 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3610 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3612 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3616 * Please do not copyright this code. This code is in the public domain.
3618 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3619 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3620 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3621 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3622 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3623 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3624 * PERFORMANCE OF THIS SOFTWARE.
3627 * chongo <Landon Curt Noll> /\oo/\
3628 * http://www.isthe.com/chongo/
3630 * Share and Enjoy! :-)
3633 typedef unsigned long long mdb_hash_t;
3634 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3636 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3637 * @param[in] val value to hash
3638 * @param[in] hval initial value for hash
3639 * @return 64 bit hash
3641 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3642 * hval arg on the first call.
3645 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3647 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3648 unsigned char *end = s + val->mv_size;
3650 * FNV-1a hash each octet of the string
3653 /* xor the bottom with the current octet */
3654 hval ^= (mdb_hash_t)*s++;
3656 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3657 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3658 (hval << 7) + (hval << 8) + (hval << 40);
3660 /* return our new hash value */
3664 /** Hash the string and output the encoded hash.
3665 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3666 * very short name limits. We don't care about the encoding being reversible,
3667 * we just want to preserve as many bits of the input as possible in a
3668 * small printable string.
3669 * @param[in] str string to hash
3670 * @param[out] encbuf an array of 11 chars to hold the hash
3672 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3675 mdb_pack85(unsigned long l, char *out)
3679 for (i=0; i<5; i++) {
3680 *out++ = mdb_a85[l % 85];
3686 mdb_hash_enc(MDB_val *val, char *encbuf)
3688 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3690 mdb_pack85(h, encbuf);
3691 mdb_pack85(h>>32, encbuf+5);
3696 /** Open and/or initialize the lock region for the environment.
3697 * @param[in] env The MDB environment.
3698 * @param[in] lpath The pathname of the file used for the lock region.
3699 * @param[in] mode The Unix permissions for the file, if we create it.
3700 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3701 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3702 * @return 0 on success, non-zero on failure.
3705 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3708 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3710 # define MDB_ERRCODE_ROFS EROFS
3711 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3712 # define MDB_CLOEXEC O_CLOEXEC
3715 # define MDB_CLOEXEC 0
3722 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3723 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3724 FILE_ATTRIBUTE_NORMAL, NULL);
3726 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3728 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3730 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3735 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3736 /* Lose record locks when exec*() */
3737 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3738 fcntl(env->me_lfd, F_SETFD, fdflags);
3741 if (!(env->me_flags & MDB_NOTLS)) {
3742 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3745 env->me_flags |= MDB_ENV_TXKEY;
3747 /* Windows TLS callbacks need help finding their TLS info. */
3748 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3752 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3756 /* Try to get exclusive lock. If we succeed, then
3757 * nobody is using the lock region and we should initialize it.
3759 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3762 size = GetFileSize(env->me_lfd, NULL);
3764 size = lseek(env->me_lfd, 0, SEEK_END);
3765 if (size == -1) goto fail_errno;
3767 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3768 if (size < rsize && *excl > 0) {
3770 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3771 || !SetEndOfFile(env->me_lfd))
3774 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3778 size = rsize - sizeof(MDB_txninfo);
3779 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3784 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3786 if (!mh) goto fail_errno;
3787 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3789 if (!env->me_txns) goto fail_errno;
3791 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3793 if (m == MAP_FAILED) goto fail_errno;
3799 BY_HANDLE_FILE_INFORMATION stbuf;
3808 if (!mdb_sec_inited) {
3809 InitializeSecurityDescriptor(&mdb_null_sd,
3810 SECURITY_DESCRIPTOR_REVISION);
3811 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3812 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3813 mdb_all_sa.bInheritHandle = FALSE;
3814 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3817 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3818 idbuf.volume = stbuf.dwVolumeSerialNumber;
3819 idbuf.nhigh = stbuf.nFileIndexHigh;
3820 idbuf.nlow = stbuf.nFileIndexLow;
3821 val.mv_data = &idbuf;
3822 val.mv_size = sizeof(idbuf);
3823 mdb_hash_enc(&val, encbuf);
3824 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3825 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3826 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3827 if (!env->me_rmutex) goto fail_errno;
3828 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3829 if (!env->me_wmutex) goto fail_errno;
3830 #elif defined(MDB_USE_POSIX_SEM)
3839 #if defined(__NetBSD__)
3840 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3842 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3843 idbuf.dev = stbuf.st_dev;
3844 idbuf.ino = stbuf.st_ino;
3845 val.mv_data = &idbuf;
3846 val.mv_size = sizeof(idbuf);
3847 mdb_hash_enc(&val, encbuf);
3848 #ifdef MDB_SHORT_SEMNAMES
3849 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3851 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3852 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3853 /* Clean up after a previous run, if needed: Try to
3854 * remove both semaphores before doing anything else.
3856 sem_unlink(env->me_txns->mti_rmname);
3857 sem_unlink(env->me_txns->mti_wmname);
3858 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3859 O_CREAT|O_EXCL, mode, 1);
3860 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3861 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3862 O_CREAT|O_EXCL, mode, 1);
3863 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3864 #else /* MDB_USE_POSIX_SEM */
3865 pthread_mutexattr_t mattr;
3867 if ((rc = pthread_mutexattr_init(&mattr))
3868 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3869 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3870 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3872 pthread_mutexattr_destroy(&mattr);
3873 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3875 env->me_txns->mti_magic = MDB_MAGIC;
3876 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3877 env->me_txns->mti_txnid = 0;
3878 env->me_txns->mti_numreaders = 0;
3881 if (env->me_txns->mti_magic != MDB_MAGIC) {
3882 DPUTS("lock region has invalid magic");
3886 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3887 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3888 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3889 rc = MDB_VERSION_MISMATCH;
3893 if (rc && rc != EACCES && rc != EAGAIN) {
3897 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3898 if (!env->me_rmutex) goto fail_errno;
3899 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3900 if (!env->me_wmutex) goto fail_errno;
3901 #elif defined(MDB_USE_POSIX_SEM)
3902 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3903 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3904 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3905 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3916 /** The name of the lock file in the DB environment */
3917 #define LOCKNAME "/lock.mdb"
3918 /** The name of the data file in the DB environment */
3919 #define DATANAME "/data.mdb"
3920 /** The suffix of the lock file when no subdir is used */
3921 #define LOCKSUFF "-lock"
3922 /** Only a subset of the @ref mdb_env flags can be changed
3923 * at runtime. Changing other flags requires closing the
3924 * environment and re-opening it with the new flags.
3926 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3927 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK)
3930 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3932 int oflags, rc, len, excl = -1;
3933 char *lpath, *dpath;
3935 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3939 if (flags & MDB_NOSUBDIR) {
3940 rc = len + sizeof(LOCKSUFF) + len + 1;
3942 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3947 if (flags & MDB_NOSUBDIR) {
3948 dpath = lpath + len + sizeof(LOCKSUFF);
3949 sprintf(lpath, "%s" LOCKSUFF, path);
3950 strcpy(dpath, path);
3952 dpath = lpath + len + sizeof(LOCKNAME);
3953 sprintf(lpath, "%s" LOCKNAME, path);
3954 sprintf(dpath, "%s" DATANAME, path);
3958 flags |= env->me_flags;
3959 if (flags & MDB_RDONLY) {
3960 /* silently ignore WRITEMAP when we're only getting read access */
3961 flags &= ~MDB_WRITEMAP;
3963 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3964 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3967 env->me_flags = flags |= MDB_ENV_ACTIVE;
3971 env->me_path = strdup(path);
3972 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3973 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3974 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3979 /* For RDONLY, get lockfile after we know datafile exists */
3980 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
3981 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3987 if (F_ISSET(flags, MDB_RDONLY)) {
3988 oflags = GENERIC_READ;
3989 len = OPEN_EXISTING;
3991 oflags = GENERIC_READ|GENERIC_WRITE;
3994 mode = FILE_ATTRIBUTE_NORMAL;
3995 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3996 NULL, len, mode, NULL);
3998 if (F_ISSET(flags, MDB_RDONLY))
4001 oflags = O_RDWR | O_CREAT;
4003 env->me_fd = open(dpath, oflags, mode);
4005 if (env->me_fd == INVALID_HANDLE_VALUE) {
4010 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4011 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4016 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4017 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4018 env->me_mfd = env->me_fd;
4020 /* Synchronous fd for meta writes. Needed even with
4021 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4024 len = OPEN_EXISTING;
4025 env->me_mfd = CreateFile(dpath, oflags,
4026 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4027 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4030 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4032 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4037 DPRINTF(("opened dbenv %p", (void *) env));
4039 rc = mdb_env_share_locks(env, &excl);
4045 mdb_env_close0(env, excl);
4051 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4053 mdb_env_close0(MDB_env *env, int excl)
4057 if (!(env->me_flags & MDB_ENV_ACTIVE))
4060 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4061 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4062 free(env->me_dbxs[i].md_name.mv_data);
4064 free(env->me_dbflags);
4067 free(env->me_dirty_list);
4068 mdb_midl_free(env->me_free_pgs);
4070 if (env->me_flags & MDB_ENV_TXKEY) {
4071 pthread_key_delete(env->me_txkey);
4073 /* Delete our key from the global list */
4074 for (i=0; i<mdb_tls_nkeys; i++)
4075 if (mdb_tls_keys[i] == env->me_txkey) {
4076 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4084 munmap(env->me_map, env->me_mapsize);
4086 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4087 (void) close(env->me_mfd);
4088 if (env->me_fd != INVALID_HANDLE_VALUE)
4089 (void) close(env->me_fd);
4091 pid_t pid = env->me_pid;
4092 /* Clearing readers is done in this function because
4093 * me_txkey with its destructor must be disabled first.
4095 for (i = env->me_numreaders; --i >= 0; )
4096 if (env->me_txns->mti_readers[i].mr_pid == pid)
4097 env->me_txns->mti_readers[i].mr_pid = 0;
4099 if (env->me_rmutex) {
4100 CloseHandle(env->me_rmutex);
4101 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4103 /* Windows automatically destroys the mutexes when
4104 * the last handle closes.
4106 #elif defined(MDB_USE_POSIX_SEM)
4107 if (env->me_rmutex != SEM_FAILED) {
4108 sem_close(env->me_rmutex);
4109 if (env->me_wmutex != SEM_FAILED)
4110 sem_close(env->me_wmutex);
4111 /* If we have the filelock: If we are the
4112 * only remaining user, clean up semaphores.
4115 mdb_env_excl_lock(env, &excl);
4117 sem_unlink(env->me_txns->mti_rmname);
4118 sem_unlink(env->me_txns->mti_wmname);
4122 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4124 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4127 /* Unlock the lockfile. Windows would have unlocked it
4128 * after closing anyway, but not necessarily at once.
4130 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4133 (void) close(env->me_lfd);
4136 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4140 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4142 MDB_txn *txn = NULL;
4148 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4152 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4155 /* Do the lock/unlock of the reader mutex before starting the
4156 * write txn. Otherwise other read txns could block writers.
4158 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4163 /* We must start the actual read txn after blocking writers */
4164 mdb_txn_reset0(txn, "reset-stage1");
4166 /* Temporarily block writers until we snapshot the meta pages */
4169 rc = mdb_txn_renew0(txn);
4171 UNLOCK_MUTEX_W(env);
4176 wsize = env->me_psize * 2;
4180 DO_WRITE(rc, fd, ptr, w2, len);
4184 } else if (len > 0) {
4190 /* Non-blocking or async handles are not supported */
4196 UNLOCK_MUTEX_W(env);
4201 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4203 if (wsize > MAX_WRITE)
4207 DO_WRITE(rc, fd, ptr, w2, len);
4211 } else if (len > 0) {
4228 mdb_env_copy(MDB_env *env, const char *path)
4232 HANDLE newfd = INVALID_HANDLE_VALUE;
4234 if (env->me_flags & MDB_NOSUBDIR) {
4235 lpath = (char *)path;
4238 len += sizeof(DATANAME);
4239 lpath = malloc(len);
4242 sprintf(lpath, "%s" DATANAME, path);
4245 /* The destination path must exist, but the destination file must not.
4246 * We don't want the OS to cache the writes, since the source data is
4247 * already in the OS cache.
4250 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4251 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4253 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4255 if (newfd == INVALID_HANDLE_VALUE) {
4261 /* Set O_DIRECT if the file system supports it */
4262 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4263 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4265 #ifdef F_NOCACHE /* __APPLE__ */
4266 rc = fcntl(newfd, F_NOCACHE, 1);
4273 rc = mdb_env_copyfd(env, newfd);
4276 if (!(env->me_flags & MDB_NOSUBDIR))
4278 if (newfd != INVALID_HANDLE_VALUE)
4279 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4286 mdb_env_close(MDB_env *env)
4293 VGMEMP_DESTROY(env);
4294 while ((dp = env->me_dpages) != NULL) {
4295 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4296 env->me_dpages = dp->mp_next;
4300 mdb_env_close0(env, 0);
4304 /** Compare two items pointing at aligned size_t's */
4306 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4308 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4309 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4312 /** Compare two items pointing at aligned unsigned int's */
4314 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4316 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4317 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4320 /** Compare two items pointing at unsigned ints of unknown alignment.
4321 * Nodes and keys are guaranteed to be 2-byte aligned.
4324 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4326 #if BYTE_ORDER == LITTLE_ENDIAN
4327 unsigned short *u, *c;
4330 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4331 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4334 } while(!x && u > (unsigned short *)a->mv_data);
4337 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4341 /** Compare two items lexically */
4343 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4350 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4356 diff = memcmp(a->mv_data, b->mv_data, len);
4357 return diff ? diff : len_diff<0 ? -1 : len_diff;
4360 /** Compare two items in reverse byte order */
4362 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4364 const unsigned char *p1, *p2, *p1_lim;
4368 p1_lim = (const unsigned char *)a->mv_data;
4369 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4370 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4372 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4378 while (p1 > p1_lim) {
4379 diff = *--p1 - *--p2;
4383 return len_diff<0 ? -1 : len_diff;
4386 /** Search for key within a page, using binary search.
4387 * Returns the smallest entry larger or equal to the key.
4388 * If exactp is non-null, stores whether the found entry was an exact match
4389 * in *exactp (1 or 0).
4390 * Updates the cursor index with the index of the found entry.
4391 * If no entry larger or equal to the key is found, returns NULL.
4394 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4396 unsigned int i = 0, nkeys;
4399 MDB_page *mp = mc->mc_pg[mc->mc_top];
4400 MDB_node *node = NULL;
4405 nkeys = NUMKEYS(mp);
4410 COPY_PGNO(pgno, mp->mp_pgno);
4411 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4412 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4419 low = IS_LEAF(mp) ? 0 : 1;
4421 cmp = mc->mc_dbx->md_cmp;
4423 /* Branch pages have no data, so if using integer keys,
4424 * alignment is guaranteed. Use faster mdb_cmp_int.
4426 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4427 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4434 nodekey.mv_size = mc->mc_db->md_pad;
4435 node = NODEPTR(mp, 0); /* fake */
4436 while (low <= high) {
4437 i = (low + high) >> 1;
4438 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4439 rc = cmp(key, &nodekey);
4440 DPRINTF(("found leaf index %u [%s], rc = %i",
4441 i, DKEY(&nodekey), rc));
4450 while (low <= high) {
4451 i = (low + high) >> 1;
4453 node = NODEPTR(mp, i);
4454 nodekey.mv_size = NODEKSZ(node);
4455 nodekey.mv_data = NODEKEY(node);
4457 rc = cmp(key, &nodekey);
4460 DPRINTF(("found leaf index %u [%s], rc = %i",
4461 i, DKEY(&nodekey), rc));
4463 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4464 i, DKEY(&nodekey), NODEPGNO(node), rc));
4475 if (rc > 0) { /* Found entry is less than the key. */
4476 i++; /* Skip to get the smallest entry larger than key. */
4478 node = NODEPTR(mp, i);
4481 *exactp = (rc == 0);
4482 /* store the key index */
4483 mc->mc_ki[mc->mc_top] = i;
4485 /* There is no entry larger or equal to the key. */
4488 /* nodeptr is fake for LEAF2 */
4494 mdb_cursor_adjust(MDB_cursor *mc, func)
4498 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4499 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4506 /** Pop a page off the top of the cursor's stack. */
4508 mdb_cursor_pop(MDB_cursor *mc)
4512 MDB_page *top = mc->mc_pg[mc->mc_top];
4518 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4519 DDBI(mc), (void *) mc));
4523 /** Push a page onto the top of the cursor's stack. */
4525 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4527 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4528 DDBI(mc), (void *) mc));
4530 if (mc->mc_snum >= CURSOR_STACK) {
4531 assert(mc->mc_snum < CURSOR_STACK);
4532 return MDB_CURSOR_FULL;
4535 mc->mc_top = mc->mc_snum++;
4536 mc->mc_pg[mc->mc_top] = mp;
4537 mc->mc_ki[mc->mc_top] = 0;
4542 /** Find the address of the page corresponding to a given page number.
4543 * @param[in] txn the transaction for this access.
4544 * @param[in] pgno the page number for the page to retrieve.
4545 * @param[out] ret address of a pointer where the page's address will be stored.
4546 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4547 * @return 0 on success, non-zero on failure.
4550 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4552 MDB_env *env = txn->mt_env;
4556 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4560 MDB_ID2L dl = tx2->mt_u.dirty_list;
4562 /* Spilled pages were dirtied in this txn and flushed
4563 * because the dirty list got full. Bring this page
4564 * back in from the map (but don't unspill it here,
4565 * leave that unless page_touch happens again).
4567 if (tx2->mt_spill_pgs) {
4568 MDB_ID pn = pgno << 1;
4569 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4570 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4571 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4576 unsigned x = mdb_mid2l_search(dl, pgno);
4577 if (x <= dl[0].mid && dl[x].mid == pgno) {
4583 } while ((tx2 = tx2->mt_parent) != NULL);
4586 if (pgno < txn->mt_next_pgno) {
4588 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4590 DPRINTF(("page %"Z"u not found", pgno));
4592 return MDB_PAGE_NOTFOUND;
4602 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4603 * The cursor is at the root page, set up the rest of it.
4606 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4608 MDB_page *mp = mc->mc_pg[mc->mc_top];
4612 while (IS_BRANCH(mp)) {
4616 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4617 assert(NUMKEYS(mp) > 1);
4618 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4620 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4622 if (flags & MDB_PS_LAST)
4623 i = NUMKEYS(mp) - 1;
4626 node = mdb_node_search(mc, key, &exact);
4628 i = NUMKEYS(mp) - 1;
4630 i = mc->mc_ki[mc->mc_top];
4636 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4639 assert(i < NUMKEYS(mp));
4640 node = NODEPTR(mp, i);
4642 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4645 mc->mc_ki[mc->mc_top] = i;
4646 if ((rc = mdb_cursor_push(mc, mp)))
4649 if (flags & MDB_PS_MODIFY) {
4650 if ((rc = mdb_page_touch(mc)) != 0)
4652 mp = mc->mc_pg[mc->mc_top];
4657 DPRINTF(("internal error, index points to a %02X page!?",
4659 return MDB_CORRUPTED;
4662 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4663 key ? DKEY(key) : "null"));
4664 mc->mc_flags |= C_INITIALIZED;
4665 mc->mc_flags &= ~C_EOF;
4670 /** Search for the lowest key under the current branch page.
4671 * This just bypasses a NUMKEYS check in the current page
4672 * before calling mdb_page_search_root(), because the callers
4673 * are all in situations where the current page is known to
4677 mdb_page_search_lowest(MDB_cursor *mc)
4679 MDB_page *mp = mc->mc_pg[mc->mc_top];
4680 MDB_node *node = NODEPTR(mp, 0);
4683 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4686 mc->mc_ki[mc->mc_top] = 0;
4687 if ((rc = mdb_cursor_push(mc, mp)))
4689 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4692 /** Search for the page a given key should be in.
4693 * Push it and its parent pages on the cursor stack.
4694 * @param[in,out] mc the cursor for this operation.
4695 * @param[in] key the key to search for, or NULL for first/last page.
4696 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4697 * are touched (updated with new page numbers).
4698 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4699 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4700 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4701 * @return 0 on success, non-zero on failure.
4704 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4709 /* Make sure the txn is still viable, then find the root from
4710 * the txn's db table and set it as the root of the cursor's stack.
4712 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4713 DPUTS("transaction has failed, must abort");
4716 /* Make sure we're using an up-to-date root */
4717 if (*mc->mc_dbflag & DB_STALE) {
4719 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4720 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4727 MDB_node *leaf = mdb_node_search(&mc2,
4728 &mc->mc_dbx->md_name, &exact);
4730 return MDB_NOTFOUND;
4731 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4734 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4736 /* The txn may not know this DBI, or another process may
4737 * have dropped and recreated the DB with other flags.
4739 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4740 return MDB_INCOMPATIBLE;
4741 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4743 *mc->mc_dbflag &= ~DB_STALE;
4745 root = mc->mc_db->md_root;
4747 if (root == P_INVALID) { /* Tree is empty. */
4748 DPUTS("tree is empty");
4749 return MDB_NOTFOUND;
4754 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4755 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4761 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4762 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4764 if (flags & MDB_PS_MODIFY) {
4765 if ((rc = mdb_page_touch(mc)))
4769 if (flags & MDB_PS_ROOTONLY)
4772 return mdb_page_search_root(mc, key, flags);
4776 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4778 MDB_txn *txn = mc->mc_txn;
4779 pgno_t pg = mp->mp_pgno;
4780 unsigned x = 0, ovpages = mp->mp_pages;
4781 MDB_env *env = txn->mt_env;
4782 MDB_IDL sl = txn->mt_spill_pgs;
4783 MDB_ID pn = pg << 1;
4786 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4787 /* If the page is dirty or on the spill list we just acquired it,
4788 * so we should give it back to our current free list, if any.
4789 * Otherwise put it onto the list of pages we freed in this txn.
4791 * Won't create me_pghead: me_pglast must be inited along with it.
4792 * Unsupported in nested txns: They would need to hide the page
4793 * range in ancestor txns' dirty and spilled lists.
4795 if (env->me_pghead &&
4797 ((mp->mp_flags & P_DIRTY) ||
4798 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4802 MDB_ID2 *dl, ix, iy;
4803 rc = mdb_midl_need(&env->me_pghead, ovpages);
4806 if (!(mp->mp_flags & P_DIRTY)) {
4807 /* This page is no longer spilled */
4814 /* Remove from dirty list */
4815 dl = txn->mt_u.dirty_list;
4817 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4825 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4826 txn->mt_flags |= MDB_TXN_ERROR;
4827 return MDB_CORRUPTED;
4830 if (!(env->me_flags & MDB_WRITEMAP))
4831 mdb_dpage_free(env, mp);
4833 /* Insert in me_pghead */
4834 mop = env->me_pghead;
4835 j = mop[0] + ovpages;
4836 for (i = mop[0]; i && mop[i] < pg; i--)
4842 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4846 mc->mc_db->md_overflow_pages -= ovpages;
4850 /** Return the data associated with a given node.
4851 * @param[in] txn The transaction for this operation.
4852 * @param[in] leaf The node being read.
4853 * @param[out] data Updated to point to the node's data.
4854 * @return 0 on success, non-zero on failure.
4857 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4859 MDB_page *omp; /* overflow page */
4863 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4864 data->mv_size = NODEDSZ(leaf);
4865 data->mv_data = NODEDATA(leaf);
4869 /* Read overflow data.
4871 data->mv_size = NODEDSZ(leaf);
4872 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4873 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4874 DPRINTF(("read overflow page %"Z"u failed", pgno));
4877 data->mv_data = METADATA(omp);
4883 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4884 MDB_val *key, MDB_val *data)
4893 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4895 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4898 if (txn->mt_flags & MDB_TXN_ERROR)
4901 if (key->mv_size > MDB_MAXKEYSIZE) {
4902 return MDB_BAD_VALSIZE;
4905 mdb_cursor_init(&mc, txn, dbi, &mx);
4906 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4909 /** Find a sibling for a page.
4910 * Replaces the page at the top of the cursor's stack with the
4911 * specified sibling, if one exists.
4912 * @param[in] mc The cursor for this operation.
4913 * @param[in] move_right Non-zero if the right sibling is requested,
4914 * otherwise the left sibling.
4915 * @return 0 on success, non-zero on failure.
4918 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4924 if (mc->mc_snum < 2) {
4925 return MDB_NOTFOUND; /* root has no siblings */
4929 DPRINTF(("parent page is page %"Z"u, index %u",
4930 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4932 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4933 : (mc->mc_ki[mc->mc_top] == 0)) {
4934 DPRINTF(("no more keys left, moving to %s sibling",
4935 move_right ? "right" : "left"));
4936 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4937 /* undo cursor_pop before returning */
4944 mc->mc_ki[mc->mc_top]++;
4946 mc->mc_ki[mc->mc_top]--;
4947 DPRINTF(("just moving to %s index key %u",
4948 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4950 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4952 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4953 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
4954 /* mc will be inconsistent if caller does mc_snum++ as above */
4955 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
4959 mdb_cursor_push(mc, mp);
4961 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4966 /** Move the cursor to the next data item. */
4968 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4974 if (mc->mc_flags & C_EOF) {
4975 return MDB_NOTFOUND;
4978 assert(mc->mc_flags & C_INITIALIZED);
4980 mp = mc->mc_pg[mc->mc_top];
4982 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4983 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4984 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4985 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4986 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4987 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
4988 if (rc == MDB_SUCCESS)
4989 MDB_GET_KEY(leaf, key);
4994 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4995 if (op == MDB_NEXT_DUP)
4996 return MDB_NOTFOUND;
5000 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5001 if (mc->mc_flags & C_DEL)
5004 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5005 DPUTS("=====> move to next sibling page");
5006 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5007 mc->mc_flags |= C_EOF;
5010 mp = mc->mc_pg[mc->mc_top];
5011 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5013 mc->mc_ki[mc->mc_top]++;
5016 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5017 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5020 key->mv_size = mc->mc_db->md_pad;
5021 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5025 assert(IS_LEAF(mp));
5026 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5028 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5029 mdb_xcursor_init1(mc, leaf);
5032 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5035 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5036 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5037 if (rc != MDB_SUCCESS)
5042 MDB_GET_KEY(leaf, key);
5046 /** Move the cursor to the previous data item. */
5048 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5054 assert(mc->mc_flags & C_INITIALIZED);
5056 mp = mc->mc_pg[mc->mc_top];
5058 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5059 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5060 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5061 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5062 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5063 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5064 if (rc == MDB_SUCCESS)
5065 MDB_GET_KEY(leaf, key);
5069 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5070 if (op == MDB_PREV_DUP)
5071 return MDB_NOTFOUND;
5076 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5078 if (mc->mc_ki[mc->mc_top] == 0) {
5079 DPUTS("=====> move to prev sibling page");
5080 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5083 mp = mc->mc_pg[mc->mc_top];
5084 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5085 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5087 mc->mc_ki[mc->mc_top]--;
5089 mc->mc_flags &= ~C_EOF;
5091 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5092 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5095 key->mv_size = mc->mc_db->md_pad;
5096 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5100 assert(IS_LEAF(mp));
5101 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5103 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5104 mdb_xcursor_init1(mc, leaf);
5107 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5110 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5111 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5112 if (rc != MDB_SUCCESS)
5117 MDB_GET_KEY(leaf, key);
5121 /** Set the cursor on a specific data item. */
5123 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5124 MDB_cursor_op op, int *exactp)
5128 MDB_node *leaf = NULL;
5133 if (key->mv_size == 0)
5134 return MDB_BAD_VALSIZE;
5137 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5139 /* See if we're already on the right page */
5140 if (mc->mc_flags & C_INITIALIZED) {
5143 mp = mc->mc_pg[mc->mc_top];
5145 mc->mc_ki[mc->mc_top] = 0;
5146 return MDB_NOTFOUND;
5148 if (mp->mp_flags & P_LEAF2) {
5149 nodekey.mv_size = mc->mc_db->md_pad;
5150 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5152 leaf = NODEPTR(mp, 0);
5153 MDB_GET_KEY2(leaf, nodekey);
5155 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5157 /* Probably happens rarely, but first node on the page
5158 * was the one we wanted.
5160 mc->mc_ki[mc->mc_top] = 0;
5167 unsigned int nkeys = NUMKEYS(mp);
5169 if (mp->mp_flags & P_LEAF2) {
5170 nodekey.mv_data = LEAF2KEY(mp,
5171 nkeys-1, nodekey.mv_size);
5173 leaf = NODEPTR(mp, nkeys-1);
5174 MDB_GET_KEY2(leaf, nodekey);
5176 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5178 /* last node was the one we wanted */
5179 mc->mc_ki[mc->mc_top] = nkeys-1;
5185 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5186 /* This is definitely the right page, skip search_page */
5187 if (mp->mp_flags & P_LEAF2) {
5188 nodekey.mv_data = LEAF2KEY(mp,
5189 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5191 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5192 MDB_GET_KEY2(leaf, nodekey);
5194 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5196 /* current node was the one we wanted */
5206 /* If any parents have right-sibs, search.
5207 * Otherwise, there's nothing further.
5209 for (i=0; i<mc->mc_top; i++)
5211 NUMKEYS(mc->mc_pg[i])-1)
5213 if (i == mc->mc_top) {
5214 /* There are no other pages */
5215 mc->mc_ki[mc->mc_top] = nkeys;
5216 return MDB_NOTFOUND;
5220 /* There are no other pages */
5221 mc->mc_ki[mc->mc_top] = 0;
5222 if (op == MDB_SET_RANGE) {
5226 return MDB_NOTFOUND;
5230 rc = mdb_page_search(mc, key, 0);
5231 if (rc != MDB_SUCCESS)
5234 mp = mc->mc_pg[mc->mc_top];
5235 assert(IS_LEAF(mp));
5238 leaf = mdb_node_search(mc, key, exactp);
5239 if (exactp != NULL && !*exactp) {
5240 /* MDB_SET specified and not an exact match. */
5241 return MDB_NOTFOUND;
5245 DPUTS("===> inexact leaf not found, goto sibling");
5246 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5247 return rc; /* no entries matched */
5248 mp = mc->mc_pg[mc->mc_top];
5249 assert(IS_LEAF(mp));
5250 leaf = NODEPTR(mp, 0);
5254 mc->mc_flags |= C_INITIALIZED;
5255 mc->mc_flags &= ~C_EOF;
5258 key->mv_size = mc->mc_db->md_pad;
5259 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5263 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5264 mdb_xcursor_init1(mc, leaf);
5267 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5268 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5269 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5272 if (op == MDB_GET_BOTH) {
5278 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5279 if (rc != MDB_SUCCESS)
5282 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5284 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5286 rc = mc->mc_dbx->md_dcmp(data, &d2);
5288 if (op == MDB_GET_BOTH || rc > 0)
5289 return MDB_NOTFOUND;
5295 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5296 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5301 /* The key already matches in all other cases */
5302 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5303 MDB_GET_KEY(leaf, key);
5304 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5309 /** Move the cursor to the first item in the database. */
5311 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5317 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5319 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5320 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5321 if (rc != MDB_SUCCESS)
5324 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5326 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5327 mc->mc_flags |= C_INITIALIZED;
5328 mc->mc_flags &= ~C_EOF;
5330 mc->mc_ki[mc->mc_top] = 0;
5332 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5333 key->mv_size = mc->mc_db->md_pad;
5334 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5339 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5340 mdb_xcursor_init1(mc, leaf);
5341 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5345 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5349 MDB_GET_KEY(leaf, key);
5353 /** Move the cursor to the last item in the database. */
5355 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5361 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5363 if (!(mc->mc_flags & C_EOF)) {
5365 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5366 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5367 if (rc != MDB_SUCCESS)
5370 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5373 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5374 mc->mc_flags |= C_INITIALIZED|C_EOF;
5375 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5377 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5378 key->mv_size = mc->mc_db->md_pad;
5379 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5384 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5385 mdb_xcursor_init1(mc, leaf);
5386 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5390 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5395 MDB_GET_KEY(leaf, key);
5400 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5405 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5409 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5413 case MDB_GET_CURRENT:
5414 if (!(mc->mc_flags & C_INITIALIZED)) {
5417 MDB_page *mp = mc->mc_pg[mc->mc_top];
5419 mc->mc_ki[mc->mc_top] = 0;
5425 key->mv_size = mc->mc_db->md_pad;
5426 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5428 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5429 MDB_GET_KEY(leaf, key);
5431 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5432 if (mc->mc_flags & C_DEL)
5433 mdb_xcursor_init1(mc, leaf);
5434 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5436 rc = mdb_node_read(mc->mc_txn, leaf, data);
5443 case MDB_GET_BOTH_RANGE:
5448 if (mc->mc_xcursor == NULL) {
5449 rc = MDB_INCOMPATIBLE;
5458 } else if (key->mv_size > MDB_MAXKEYSIZE) {
5459 rc = MDB_BAD_VALSIZE;
5460 } else if (op == MDB_SET_RANGE)
5461 rc = mdb_cursor_set(mc, key, data, op, NULL);
5463 rc = mdb_cursor_set(mc, key, data, op, &exact);
5465 case MDB_GET_MULTIPLE:
5466 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5470 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5471 rc = MDB_INCOMPATIBLE;
5475 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5476 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5479 case MDB_NEXT_MULTIPLE:
5484 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5485 rc = MDB_INCOMPATIBLE;
5488 if (!(mc->mc_flags & C_INITIALIZED))
5489 rc = mdb_cursor_first(mc, key, data);
5491 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5492 if (rc == MDB_SUCCESS) {
5493 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5496 mx = &mc->mc_xcursor->mx_cursor;
5497 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5499 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5500 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5508 case MDB_NEXT_NODUP:
5509 if (!(mc->mc_flags & C_INITIALIZED))
5510 rc = mdb_cursor_first(mc, key, data);
5512 rc = mdb_cursor_next(mc, key, data, op);
5516 case MDB_PREV_NODUP:
5517 if (!(mc->mc_flags & C_INITIALIZED)) {
5518 rc = mdb_cursor_last(mc, key, data);
5521 mc->mc_flags |= C_INITIALIZED;
5522 mc->mc_ki[mc->mc_top]++;
5524 rc = mdb_cursor_prev(mc, key, data, op);
5527 rc = mdb_cursor_first(mc, key, data);
5530 mfunc = mdb_cursor_first;
5532 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5536 if (mc->mc_xcursor == NULL) {
5537 rc = MDB_INCOMPATIBLE;
5540 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5544 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5547 rc = mdb_cursor_last(mc, key, data);
5550 mfunc = mdb_cursor_last;
5553 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5558 if (mc->mc_flags & C_DEL)
5559 mc->mc_flags ^= C_DEL;
5564 /** Touch all the pages in the cursor stack. Set mc_top.
5565 * Makes sure all the pages are writable, before attempting a write operation.
5566 * @param[in] mc The cursor to operate on.
5569 mdb_cursor_touch(MDB_cursor *mc)
5571 int rc = MDB_SUCCESS;
5573 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5576 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5577 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5580 *mc->mc_dbflag |= DB_DIRTY;
5585 rc = mdb_page_touch(mc);
5586 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5587 mc->mc_top = mc->mc_snum-1;
5592 /** Do not spill pages to disk if txn is getting full, may fail instead */
5593 #define MDB_NOSPILL 0x8000
5596 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5599 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5600 MDB_node *leaf = NULL;
5601 MDB_val xdata, *rdata, dkey;
5604 int do_sub = 0, insert = 0;
5605 unsigned int mcount = 0, dcount = 0, nospill;
5609 char dbuf[MDB_MAXKEYSIZE+1];
5610 unsigned int nflags;
5613 /* Check this first so counter will always be zero on any
5616 if (flags & MDB_MULTIPLE) {
5617 dcount = data[1].mv_size;
5618 data[1].mv_size = 0;
5619 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5620 return MDB_INCOMPATIBLE;
5623 nospill = flags & MDB_NOSPILL;
5624 flags &= ~MDB_NOSPILL;
5626 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5627 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5629 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5630 return MDB_BAD_VALSIZE;
5632 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5633 return MDB_BAD_VALSIZE;
5635 #if SIZE_MAX > MAXDATASIZE
5636 if (data->mv_size > MAXDATASIZE)
5637 return MDB_BAD_VALSIZE;
5640 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5641 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5645 if (flags == MDB_CURRENT) {
5646 if (!(mc->mc_flags & C_INITIALIZED))
5649 } else if (mc->mc_db->md_root == P_INVALID) {
5650 /* new database, cursor has nothing to point to */
5653 mc->mc_flags &= ~C_INITIALIZED;
5658 if (flags & MDB_APPEND) {
5660 rc = mdb_cursor_last(mc, &k2, &d2);
5662 rc = mc->mc_dbx->md_cmp(key, &k2);
5665 mc->mc_ki[mc->mc_top]++;
5667 /* new key is <= last key */
5672 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5674 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5675 DPRINTF(("duplicate key [%s]", DKEY(key)));
5677 return MDB_KEYEXIST;
5679 if (rc && rc != MDB_NOTFOUND)
5683 if (mc->mc_flags & C_DEL)
5684 mc->mc_flags ^= C_DEL;
5686 /* Cursor is positioned, check for room in the dirty list */
5688 if (flags & MDB_MULTIPLE) {
5690 xdata.mv_size = data->mv_size * dcount;
5694 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5698 if (rc == MDB_NO_ROOT) {
5700 /* new database, write a root leaf page */
5701 DPUTS("allocating new root leaf page");
5702 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5705 mdb_cursor_push(mc, np);
5706 mc->mc_db->md_root = np->mp_pgno;
5707 mc->mc_db->md_depth++;
5708 *mc->mc_dbflag |= DB_DIRTY;
5709 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5711 np->mp_flags |= P_LEAF2;
5712 mc->mc_flags |= C_INITIALIZED;
5714 /* make sure all cursor pages are writable */
5715 rc2 = mdb_cursor_touch(mc);
5720 /* The key already exists */
5721 if (rc == MDB_SUCCESS) {
5722 /* there's only a key anyway, so this is a no-op */
5723 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5724 unsigned int ksize = mc->mc_db->md_pad;
5725 if (key->mv_size != ksize)
5726 return MDB_BAD_VALSIZE;
5727 if (flags == MDB_CURRENT) {
5728 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5729 memcpy(ptr, key->mv_data, ksize);
5734 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5737 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5738 /* Was a single item before, must convert now */
5740 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5741 /* Just overwrite the current item */
5742 if (flags == MDB_CURRENT)
5745 dkey.mv_size = NODEDSZ(leaf);
5746 dkey.mv_data = NODEDATA(leaf);
5747 #if UINT_MAX < SIZE_MAX
5748 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5749 #ifdef MISALIGNED_OK
5750 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5752 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5755 /* if data matches, skip it */
5756 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5757 if (flags & MDB_NODUPDATA)
5759 else if (flags & MDB_MULTIPLE)
5766 /* create a fake page for the dup items */
5767 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5768 dkey.mv_data = dbuf;
5769 fp = (MDB_page *)&pbuf;
5770 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5771 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5772 fp->mp_lower = PAGEHDRSZ;
5773 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5774 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5775 fp->mp_flags |= P_LEAF2;
5776 fp->mp_pad = data->mv_size;
5777 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5779 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5780 (dkey.mv_size & 1) + (data->mv_size & 1);
5782 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5785 xdata.mv_size = fp->mp_upper;
5790 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5791 /* See if we need to convert from fake page to subDB */
5793 unsigned int offset;
5797 fp = NODEDATA(leaf);
5798 if (flags == MDB_CURRENT) {
5800 fp->mp_flags |= P_DIRTY;
5801 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5802 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5806 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5807 offset = fp->mp_pad;
5808 if (SIZELEFT(fp) >= offset)
5810 offset *= 4; /* space for 4 more */
5812 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5814 offset += offset & 1;
5815 fp_flags = fp->mp_flags;
5816 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5817 offset >= mc->mc_txn->mt_env->me_nodemax) {
5818 /* yes, convert it */
5820 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5821 dummy.md_pad = fp->mp_pad;
5822 dummy.md_flags = MDB_DUPFIXED;
5823 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5824 dummy.md_flags |= MDB_INTEGERKEY;
5827 dummy.md_branch_pages = 0;
5828 dummy.md_leaf_pages = 1;
5829 dummy.md_overflow_pages = 0;
5830 dummy.md_entries = NUMKEYS(fp);
5832 xdata.mv_size = sizeof(MDB_db);
5833 xdata.mv_data = &dummy;
5834 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5836 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5837 flags |= F_DUPDATA|F_SUBDATA;
5838 dummy.md_root = mp->mp_pgno;
5839 fp_flags &= ~P_SUBP;
5841 /* no, just grow it */
5843 xdata.mv_size = NODEDSZ(leaf) + offset;
5844 xdata.mv_data = &pbuf;
5845 mp = (MDB_page *)&pbuf;
5846 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5849 mp->mp_flags = fp_flags | P_DIRTY;
5850 mp->mp_pad = fp->mp_pad;
5851 mp->mp_lower = fp->mp_lower;
5852 mp->mp_upper = fp->mp_upper + offset;
5854 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5856 nsize = NODEDSZ(leaf) - fp->mp_upper;
5857 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5858 for (i=0; i<NUMKEYS(fp); i++)
5859 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5861 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5865 /* data is on sub-DB, just store it */
5866 flags |= F_DUPDATA|F_SUBDATA;
5870 /* overflow page overwrites need special handling */
5871 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5874 unsigned psize = mc->mc_txn->mt_env->me_psize;
5875 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5877 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5878 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5880 ovpages = omp->mp_pages;
5882 /* Is the ov page large enough? */
5883 if (ovpages >= dpages) {
5884 if (!(omp->mp_flags & P_DIRTY) &&
5885 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5887 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5890 level = 0; /* dirty in this txn or clean */
5893 if (omp->mp_flags & P_DIRTY) {
5894 /* yes, overwrite it. Note in this case we don't
5895 * bother to try shrinking the page if the new data
5896 * is smaller than the overflow threshold.
5899 /* It is writable only in a parent txn */
5900 size_t sz = (size_t) psize * ovpages, off;
5901 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5907 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5908 if (!(flags & MDB_RESERVE)) {
5909 /* Copy end of page, adjusting alignment so
5910 * compiler may copy words instead of bytes.
5912 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5913 memcpy((size_t *)((char *)np + off),
5914 (size_t *)((char *)omp + off), sz - off);
5917 memcpy(np, omp, sz); /* Copy beginning of page */
5920 SETDSZ(leaf, data->mv_size);
5921 if (F_ISSET(flags, MDB_RESERVE))
5922 data->mv_data = METADATA(omp);
5924 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5928 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5930 } else if (NODEDSZ(leaf) == data->mv_size) {
5931 /* same size, just replace it. Note that we could
5932 * also reuse this node if the new data is smaller,
5933 * but instead we opt to shrink the node in that case.
5935 if (F_ISSET(flags, MDB_RESERVE))
5936 data->mv_data = NODEDATA(leaf);
5937 else if (data->mv_size)
5938 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5940 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5943 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5944 mc->mc_db->md_entries--;
5946 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5953 nflags = flags & NODE_ADD_FLAGS;
5954 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5955 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5956 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5957 nflags &= ~MDB_APPEND;
5959 nflags |= MDB_SPLIT_REPLACE;
5960 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5962 /* There is room already in this leaf page. */
5963 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5964 if (rc == 0 && !do_sub && insert) {
5965 /* Adjust other cursors pointing to mp */
5966 MDB_cursor *m2, *m3;
5967 MDB_dbi dbi = mc->mc_dbi;
5968 unsigned i = mc->mc_top;
5969 MDB_page *mp = mc->mc_pg[i];
5971 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5972 if (mc->mc_flags & C_SUB)
5973 m3 = &m2->mc_xcursor->mx_cursor;
5976 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5977 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5984 if (rc != MDB_SUCCESS)
5985 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5987 /* Now store the actual data in the child DB. Note that we're
5988 * storing the user data in the keys field, so there are strict
5989 * size limits on dupdata. The actual data fields of the child
5990 * DB are all zero size.
5997 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5998 if (flags & MDB_CURRENT) {
5999 xflags = MDB_CURRENT|MDB_NOSPILL;
6001 mdb_xcursor_init1(mc, leaf);
6002 xflags = (flags & MDB_NODUPDATA) ?
6003 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6005 /* converted, write the original data first */
6007 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6011 /* Adjust other cursors pointing to mp */
6013 unsigned i = mc->mc_top;
6014 MDB_page *mp = mc->mc_pg[i];
6016 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6017 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6018 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6019 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6020 mdb_xcursor_init1(m2, leaf);
6024 /* we've done our job */
6027 if (flags & MDB_APPENDDUP)
6028 xflags |= MDB_APPEND;
6029 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6030 if (flags & F_SUBDATA) {
6031 void *db = NODEDATA(leaf);
6032 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6035 /* sub-writes might have failed so check rc again.
6036 * Don't increment count if we just replaced an existing item.
6038 if (!rc && !(flags & MDB_CURRENT))
6039 mc->mc_db->md_entries++;
6040 if (flags & MDB_MULTIPLE) {
6044 /* let caller know how many succeeded, if any */
6045 data[1].mv_size = mcount;
6046 if (mcount < dcount) {
6047 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6048 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6055 /* If we succeeded and the key didn't exist before, make sure
6056 * the cursor is marked valid.
6059 mc->mc_flags |= C_INITIALIZED;
6064 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6069 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6070 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6072 if (!(mc->mc_flags & C_INITIALIZED))
6075 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6078 rc = mdb_cursor_touch(mc);
6082 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6084 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6085 if (!(flags & MDB_NODUPDATA)) {
6086 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6087 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6089 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6090 /* If sub-DB still has entries, we're done */
6091 if (mc->mc_xcursor->mx_db.md_entries) {
6092 if (leaf->mn_flags & F_SUBDATA) {
6093 /* update subDB info */
6094 void *db = NODEDATA(leaf);
6095 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6098 /* shrink fake page */
6099 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6100 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6101 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6102 /* fix other sub-DB cursors pointed at this fake page */
6103 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6104 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6105 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6106 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6107 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6110 mc->mc_db->md_entries--;
6111 mc->mc_flags |= C_DEL;
6114 /* otherwise fall thru and delete the sub-DB */
6117 if (leaf->mn_flags & F_SUBDATA) {
6118 /* add all the child DB's pages to the free list */
6119 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6120 if (rc == MDB_SUCCESS) {
6121 mc->mc_db->md_entries -=
6122 mc->mc_xcursor->mx_db.md_entries;
6127 return mdb_cursor_del0(mc, leaf);
6130 /** Allocate and initialize new pages for a database.
6131 * @param[in] mc a cursor on the database being added to.
6132 * @param[in] flags flags defining what type of page is being allocated.
6133 * @param[in] num the number of pages to allocate. This is usually 1,
6134 * unless allocating overflow pages for a large record.
6135 * @param[out] mp Address of a page, or NULL on failure.
6136 * @return 0 on success, non-zero on failure.
6139 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6144 if ((rc = mdb_page_alloc(mc, num, &np)))
6146 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6147 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6148 np->mp_flags = flags | P_DIRTY;
6149 np->mp_lower = PAGEHDRSZ;
6150 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6153 mc->mc_db->md_branch_pages++;
6154 else if (IS_LEAF(np))
6155 mc->mc_db->md_leaf_pages++;
6156 else if (IS_OVERFLOW(np)) {
6157 mc->mc_db->md_overflow_pages += num;
6165 /** Calculate the size of a leaf node.
6166 * The size depends on the environment's page size; if a data item
6167 * is too large it will be put onto an overflow page and the node
6168 * size will only include the key and not the data. Sizes are always
6169 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6170 * of the #MDB_node headers.
6171 * @param[in] env The environment handle.
6172 * @param[in] key The key for the node.
6173 * @param[in] data The data for the node.
6174 * @return The number of bytes needed to store the node.
6177 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6181 sz = LEAFSIZE(key, data);
6182 if (sz >= env->me_nodemax) {
6183 /* put on overflow page */
6184 sz -= data->mv_size - sizeof(pgno_t);
6188 return sz + sizeof(indx_t);
6191 /** Calculate the size of a branch node.
6192 * The size should depend on the environment's page size but since
6193 * we currently don't support spilling large keys onto overflow
6194 * pages, it's simply the size of the #MDB_node header plus the
6195 * size of the key. Sizes are always rounded up to an even number
6196 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6197 * @param[in] env The environment handle.
6198 * @param[in] key The key for the node.
6199 * @return The number of bytes needed to store the node.
6202 mdb_branch_size(MDB_env *env, MDB_val *key)
6207 if (sz >= env->me_nodemax) {
6208 /* put on overflow page */
6209 /* not implemented */
6210 /* sz -= key->size - sizeof(pgno_t); */
6213 return sz + sizeof(indx_t);
6216 /** Add a node to the page pointed to by the cursor.
6217 * @param[in] mc The cursor for this operation.
6218 * @param[in] indx The index on the page where the new node should be added.
6219 * @param[in] key The key for the new node.
6220 * @param[in] data The data for the new node, if any.
6221 * @param[in] pgno The page number, if adding a branch node.
6222 * @param[in] flags Flags for the node.
6223 * @return 0 on success, non-zero on failure. Possible errors are:
6225 * <li>ENOMEM - failed to allocate overflow pages for the node.
6226 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6227 * should never happen since all callers already calculate the
6228 * page's free space before calling this function.
6232 mdb_node_add(MDB_cursor *mc, indx_t indx,
6233 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6236 size_t node_size = NODESIZE;
6240 MDB_page *mp = mc->mc_pg[mc->mc_top];
6241 MDB_page *ofp = NULL; /* overflow page */
6244 assert(mp->mp_upper >= mp->mp_lower);
6246 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6247 IS_LEAF(mp) ? "leaf" : "branch",
6248 IS_SUBP(mp) ? "sub-" : "",
6249 mp->mp_pgno, indx, data ? data->mv_size : 0,
6250 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6253 /* Move higher keys up one slot. */
6254 int ksize = mc->mc_db->md_pad, dif;
6255 char *ptr = LEAF2KEY(mp, indx, ksize);
6256 dif = NUMKEYS(mp) - indx;
6258 memmove(ptr+ksize, ptr, dif*ksize);
6259 /* insert new key */
6260 memcpy(ptr, key->mv_data, ksize);
6262 /* Just using these for counting */
6263 mp->mp_lower += sizeof(indx_t);
6264 mp->mp_upper -= ksize - sizeof(indx_t);
6268 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6270 node_size += key->mv_size;
6273 if (F_ISSET(flags, F_BIGDATA)) {
6274 /* Data already on overflow page. */
6275 node_size += sizeof(pgno_t);
6276 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6277 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6279 /* Put data on overflow page. */
6280 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6281 data->mv_size, node_size+data->mv_size));
6282 node_size += sizeof(pgno_t) + (node_size & 1);
6283 if ((ssize_t)node_size > room)
6285 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6287 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6291 node_size += data->mv_size;
6294 node_size += node_size & 1;
6295 if ((ssize_t)node_size > room)
6299 /* Move higher pointers up one slot. */
6300 for (i = NUMKEYS(mp); i > indx; i--)
6301 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6303 /* Adjust free space offsets. */
6304 ofs = mp->mp_upper - node_size;
6305 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6306 mp->mp_ptrs[indx] = ofs;
6308 mp->mp_lower += sizeof(indx_t);
6310 /* Write the node data. */
6311 node = NODEPTR(mp, indx);
6312 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6313 node->mn_flags = flags;
6315 SETDSZ(node,data->mv_size);
6320 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6325 if (F_ISSET(flags, F_BIGDATA))
6326 memcpy(node->mn_data + key->mv_size, data->mv_data,
6328 else if (F_ISSET(flags, MDB_RESERVE))
6329 data->mv_data = node->mn_data + key->mv_size;
6331 memcpy(node->mn_data + key->mv_size, data->mv_data,
6334 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6336 if (F_ISSET(flags, MDB_RESERVE))
6337 data->mv_data = METADATA(ofp);
6339 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6346 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6347 mp->mp_pgno, NUMKEYS(mp)));
6348 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6349 DPRINTF(("node size = %"Z"u", node_size));
6350 return MDB_PAGE_FULL;
6353 /** Delete the specified node from a page.
6354 * @param[in] mp The page to operate on.
6355 * @param[in] indx The index of the node to delete.
6356 * @param[in] ksize The size of a node. Only used if the page is
6357 * part of a #MDB_DUPFIXED database.
6360 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6363 indx_t i, j, numkeys, ptr;
6370 COPY_PGNO(pgno, mp->mp_pgno);
6371 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6372 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6375 assert(indx < NUMKEYS(mp));
6378 int x = NUMKEYS(mp) - 1 - indx;
6379 base = LEAF2KEY(mp, indx, ksize);
6381 memmove(base, base + ksize, x * ksize);
6382 mp->mp_lower -= sizeof(indx_t);
6383 mp->mp_upper += ksize - sizeof(indx_t);
6387 node = NODEPTR(mp, indx);
6388 sz = NODESIZE + node->mn_ksize;
6390 if (F_ISSET(node->mn_flags, F_BIGDATA))
6391 sz += sizeof(pgno_t);
6393 sz += NODEDSZ(node);
6397 ptr = mp->mp_ptrs[indx];
6398 numkeys = NUMKEYS(mp);
6399 for (i = j = 0; i < numkeys; i++) {
6401 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6402 if (mp->mp_ptrs[i] < ptr)
6403 mp->mp_ptrs[j] += sz;
6408 base = (char *)mp + mp->mp_upper;
6409 memmove(base + sz, base, ptr - mp->mp_upper);
6411 mp->mp_lower -= sizeof(indx_t);
6415 /** Compact the main page after deleting a node on a subpage.
6416 * @param[in] mp The main page to operate on.
6417 * @param[in] indx The index of the subpage on the main page.
6420 mdb_node_shrink(MDB_page *mp, indx_t indx)
6427 indx_t i, numkeys, ptr;
6429 node = NODEPTR(mp, indx);
6430 sp = (MDB_page *)NODEDATA(node);
6431 osize = NODEDSZ(node);
6433 delta = sp->mp_upper - sp->mp_lower;
6434 SETDSZ(node, osize - delta);
6435 xp = (MDB_page *)((char *)sp + delta);
6437 /* shift subpage upward */
6439 nsize = NUMKEYS(sp) * sp->mp_pad;
6440 memmove(METADATA(xp), METADATA(sp), nsize);
6443 nsize = osize - sp->mp_upper;
6444 numkeys = NUMKEYS(sp);
6445 for (i=numkeys-1; i>=0; i--)
6446 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6448 xp->mp_upper = sp->mp_lower;
6449 xp->mp_lower = sp->mp_lower;
6450 xp->mp_flags = sp->mp_flags;
6451 xp->mp_pad = sp->mp_pad;
6452 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6454 /* shift lower nodes upward */
6455 ptr = mp->mp_ptrs[indx];
6456 numkeys = NUMKEYS(mp);
6457 for (i = 0; i < numkeys; i++) {
6458 if (mp->mp_ptrs[i] <= ptr)
6459 mp->mp_ptrs[i] += delta;
6462 base = (char *)mp + mp->mp_upper;
6463 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6464 mp->mp_upper += delta;
6467 /** Initial setup of a sorted-dups cursor.
6468 * Sorted duplicates are implemented as a sub-database for the given key.
6469 * The duplicate data items are actually keys of the sub-database.
6470 * Operations on the duplicate data items are performed using a sub-cursor
6471 * initialized when the sub-database is first accessed. This function does
6472 * the preliminary setup of the sub-cursor, filling in the fields that
6473 * depend only on the parent DB.
6474 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6477 mdb_xcursor_init0(MDB_cursor *mc)
6479 MDB_xcursor *mx = mc->mc_xcursor;
6481 mx->mx_cursor.mc_xcursor = NULL;
6482 mx->mx_cursor.mc_txn = mc->mc_txn;
6483 mx->mx_cursor.mc_db = &mx->mx_db;
6484 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6485 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6486 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6487 mx->mx_cursor.mc_snum = 0;
6488 mx->mx_cursor.mc_top = 0;
6489 mx->mx_cursor.mc_flags = C_SUB;
6490 mx->mx_dbx.md_name.mv_size = 0;
6491 mx->mx_dbx.md_name.mv_data = NULL;
6492 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6493 mx->mx_dbx.md_dcmp = NULL;
6494 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6497 /** Final setup of a sorted-dups cursor.
6498 * Sets up the fields that depend on the data from the main cursor.
6499 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6500 * @param[in] node The data containing the #MDB_db record for the
6501 * sorted-dup database.
6504 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6506 MDB_xcursor *mx = mc->mc_xcursor;
6508 if (node->mn_flags & F_SUBDATA) {
6509 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6510 mx->mx_cursor.mc_pg[0] = 0;
6511 mx->mx_cursor.mc_snum = 0;
6512 mx->mx_cursor.mc_top = 0;
6513 mx->mx_cursor.mc_flags = C_SUB;
6515 MDB_page *fp = NODEDATA(node);
6516 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6517 mx->mx_db.md_flags = 0;
6518 mx->mx_db.md_depth = 1;
6519 mx->mx_db.md_branch_pages = 0;
6520 mx->mx_db.md_leaf_pages = 1;
6521 mx->mx_db.md_overflow_pages = 0;
6522 mx->mx_db.md_entries = NUMKEYS(fp);
6523 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6524 mx->mx_cursor.mc_snum = 1;
6525 mx->mx_cursor.mc_top = 0;
6526 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6527 mx->mx_cursor.mc_pg[0] = fp;
6528 mx->mx_cursor.mc_ki[0] = 0;
6529 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6530 mx->mx_db.md_flags = MDB_DUPFIXED;
6531 mx->mx_db.md_pad = fp->mp_pad;
6532 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6533 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6536 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6537 mx->mx_db.md_root));
6538 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6539 #if UINT_MAX < SIZE_MAX
6540 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6541 #ifdef MISALIGNED_OK
6542 mx->mx_dbx.md_cmp = mdb_cmp_long;
6544 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6549 /** Initialize a cursor for a given transaction and database. */
6551 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6554 mc->mc_backup = NULL;
6557 mc->mc_db = &txn->mt_dbs[dbi];
6558 mc->mc_dbx = &txn->mt_dbxs[dbi];
6559 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6564 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6566 mc->mc_xcursor = mx;
6567 mdb_xcursor_init0(mc);
6569 mc->mc_xcursor = NULL;
6571 if (*mc->mc_dbflag & DB_STALE) {
6572 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6577 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6580 size_t size = sizeof(MDB_cursor);
6582 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6585 if (txn->mt_flags & MDB_TXN_ERROR)
6588 /* Allow read access to the freelist */
6589 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6592 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6593 size += sizeof(MDB_xcursor);
6595 if ((mc = malloc(size)) != NULL) {
6596 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6597 if (txn->mt_cursors) {
6598 mc->mc_next = txn->mt_cursors[dbi];
6599 txn->mt_cursors[dbi] = mc;
6600 mc->mc_flags |= C_UNTRACK;
6612 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6614 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6617 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6620 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6624 /* Return the count of duplicate data items for the current key */
6626 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6630 if (mc == NULL || countp == NULL)
6633 if (mc->mc_xcursor == NULL)
6634 return MDB_INCOMPATIBLE;
6636 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6637 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6640 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6643 *countp = mc->mc_xcursor->mx_db.md_entries;
6649 mdb_cursor_close(MDB_cursor *mc)
6651 if (mc && !mc->mc_backup) {
6652 /* remove from txn, if tracked */
6653 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6654 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6655 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6657 *prev = mc->mc_next;
6664 mdb_cursor_txn(MDB_cursor *mc)
6666 if (!mc) return NULL;
6671 mdb_cursor_dbi(MDB_cursor *mc)
6677 /** Replace the key for a node with a new key.
6678 * @param[in] mc Cursor pointing to the node to operate on.
6679 * @param[in] key The new key to use.
6680 * @return 0 on success, non-zero on failure.
6683 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6690 indx_t ptr, i, numkeys, indx;
6693 indx = mc->mc_ki[mc->mc_top];
6694 mp = mc->mc_pg[mc->mc_top];
6695 node = NODEPTR(mp, indx);
6696 ptr = mp->mp_ptrs[indx];
6700 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6701 k2.mv_data = NODEKEY(node);
6702 k2.mv_size = node->mn_ksize;
6703 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6705 mdb_dkey(&k2, kbuf2),
6711 delta0 = delta = key->mv_size - node->mn_ksize;
6713 /* Must be 2-byte aligned. If new key is
6714 * shorter by 1, the shift will be skipped.
6716 delta += (delta & 1);
6718 if (delta > 0 && SIZELEFT(mp) < delta) {
6720 /* not enough space left, do a delete and split */
6721 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6722 pgno = NODEPGNO(node);
6723 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6724 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6727 numkeys = NUMKEYS(mp);
6728 for (i = 0; i < numkeys; i++) {
6729 if (mp->mp_ptrs[i] <= ptr)
6730 mp->mp_ptrs[i] -= delta;
6733 base = (char *)mp + mp->mp_upper;
6734 len = ptr - mp->mp_upper + NODESIZE;
6735 memmove(base - delta, base, len);
6736 mp->mp_upper -= delta;
6738 node = NODEPTR(mp, indx);
6741 /* But even if no shift was needed, update ksize */
6743 node->mn_ksize = key->mv_size;
6746 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6752 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6754 /** Move a node from csrc to cdst.
6757 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6764 unsigned short flags;
6768 /* Mark src and dst as dirty. */
6769 if ((rc = mdb_page_touch(csrc)) ||
6770 (rc = mdb_page_touch(cdst)))
6773 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6774 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6775 key.mv_size = csrc->mc_db->md_pad;
6776 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6778 data.mv_data = NULL;
6782 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6783 assert(!((long)srcnode&1));
6784 srcpg = NODEPGNO(srcnode);
6785 flags = srcnode->mn_flags;
6786 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6787 unsigned int snum = csrc->mc_snum;
6789 /* must find the lowest key below src */
6790 mdb_page_search_lowest(csrc);
6791 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6792 key.mv_size = csrc->mc_db->md_pad;
6793 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6795 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6796 key.mv_size = NODEKSZ(s2);
6797 key.mv_data = NODEKEY(s2);
6799 csrc->mc_snum = snum--;
6800 csrc->mc_top = snum;
6802 key.mv_size = NODEKSZ(srcnode);
6803 key.mv_data = NODEKEY(srcnode);
6805 data.mv_size = NODEDSZ(srcnode);
6806 data.mv_data = NODEDATA(srcnode);
6808 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6809 unsigned int snum = cdst->mc_snum;
6812 /* must find the lowest key below dst */
6813 mdb_page_search_lowest(cdst);
6814 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6815 bkey.mv_size = cdst->mc_db->md_pad;
6816 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6818 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6819 bkey.mv_size = NODEKSZ(s2);
6820 bkey.mv_data = NODEKEY(s2);
6822 cdst->mc_snum = snum--;
6823 cdst->mc_top = snum;
6824 mdb_cursor_copy(cdst, &mn);
6826 rc = mdb_update_key(&mn, &bkey);
6831 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6832 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6833 csrc->mc_ki[csrc->mc_top],
6835 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6836 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6838 /* Add the node to the destination page.
6840 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6841 if (rc != MDB_SUCCESS)
6844 /* Delete the node from the source page.
6846 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6849 /* Adjust other cursors pointing to mp */
6850 MDB_cursor *m2, *m3;
6851 MDB_dbi dbi = csrc->mc_dbi;
6852 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6854 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6855 if (csrc->mc_flags & C_SUB)
6856 m3 = &m2->mc_xcursor->mx_cursor;
6859 if (m3 == csrc) continue;
6860 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6861 csrc->mc_ki[csrc->mc_top]) {
6862 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6863 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6868 /* Update the parent separators.
6870 if (csrc->mc_ki[csrc->mc_top] == 0) {
6871 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6872 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6873 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6875 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6876 key.mv_size = NODEKSZ(srcnode);
6877 key.mv_data = NODEKEY(srcnode);
6879 DPRINTF(("update separator for source page %"Z"u to [%s]",
6880 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6881 mdb_cursor_copy(csrc, &mn);
6884 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6887 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6889 indx_t ix = csrc->mc_ki[csrc->mc_top];
6890 nullkey.mv_size = 0;
6891 csrc->mc_ki[csrc->mc_top] = 0;
6892 rc = mdb_update_key(csrc, &nullkey);
6893 csrc->mc_ki[csrc->mc_top] = ix;
6894 assert(rc == MDB_SUCCESS);
6898 if (cdst->mc_ki[cdst->mc_top] == 0) {
6899 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6900 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6901 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6903 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6904 key.mv_size = NODEKSZ(srcnode);
6905 key.mv_data = NODEKEY(srcnode);
6907 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6908 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6909 mdb_cursor_copy(cdst, &mn);
6912 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6915 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6917 indx_t ix = cdst->mc_ki[cdst->mc_top];
6918 nullkey.mv_size = 0;
6919 cdst->mc_ki[cdst->mc_top] = 0;
6920 rc = mdb_update_key(cdst, &nullkey);
6921 cdst->mc_ki[cdst->mc_top] = ix;
6922 assert(rc == MDB_SUCCESS);
6929 /** Merge one page into another.
6930 * The nodes from the page pointed to by \b csrc will
6931 * be copied to the page pointed to by \b cdst and then
6932 * the \b csrc page will be freed.
6933 * @param[in] csrc Cursor pointing to the source page.
6934 * @param[in] cdst Cursor pointing to the destination page.
6937 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6945 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6946 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6948 assert(csrc->mc_snum > 1); /* can't merge root page */
6949 assert(cdst->mc_snum > 1);
6951 /* Mark dst as dirty. */
6952 if ((rc = mdb_page_touch(cdst)))
6955 /* Move all nodes from src to dst.
6957 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6958 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6959 key.mv_size = csrc->mc_db->md_pad;
6960 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6961 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6962 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6963 if (rc != MDB_SUCCESS)
6965 key.mv_data = (char *)key.mv_data + key.mv_size;
6968 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6969 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6970 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6971 unsigned int snum = csrc->mc_snum;
6973 /* must find the lowest key below src */
6974 mdb_page_search_lowest(csrc);
6975 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6976 key.mv_size = csrc->mc_db->md_pad;
6977 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6979 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6980 key.mv_size = NODEKSZ(s2);
6981 key.mv_data = NODEKEY(s2);
6983 csrc->mc_snum = snum--;
6984 csrc->mc_top = snum;
6986 key.mv_size = srcnode->mn_ksize;
6987 key.mv_data = NODEKEY(srcnode);
6990 data.mv_size = NODEDSZ(srcnode);
6991 data.mv_data = NODEDATA(srcnode);
6992 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6993 if (rc != MDB_SUCCESS)
6998 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6999 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7000 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7002 /* Unlink the src page from parent and add to free list.
7004 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7005 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7008 rc = mdb_update_key(csrc, &key);
7014 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7015 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7018 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7019 csrc->mc_db->md_leaf_pages--;
7021 csrc->mc_db->md_branch_pages--;
7023 /* Adjust other cursors pointing to mp */
7024 MDB_cursor *m2, *m3;
7025 MDB_dbi dbi = csrc->mc_dbi;
7026 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7028 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7029 if (csrc->mc_flags & C_SUB)
7030 m3 = &m2->mc_xcursor->mx_cursor;
7033 if (m3 == csrc) continue;
7034 if (m3->mc_snum < csrc->mc_snum) continue;
7035 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7036 m3->mc_pg[csrc->mc_top] = mp;
7037 m3->mc_ki[csrc->mc_top] += nkeys;
7041 mdb_cursor_pop(csrc);
7043 return mdb_rebalance(csrc);
7046 /** Copy the contents of a cursor.
7047 * @param[in] csrc The cursor to copy from.
7048 * @param[out] cdst The cursor to copy to.
7051 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7055 cdst->mc_txn = csrc->mc_txn;
7056 cdst->mc_dbi = csrc->mc_dbi;
7057 cdst->mc_db = csrc->mc_db;
7058 cdst->mc_dbx = csrc->mc_dbx;
7059 cdst->mc_snum = csrc->mc_snum;
7060 cdst->mc_top = csrc->mc_top;
7061 cdst->mc_flags = csrc->mc_flags;
7063 for (i=0; i<csrc->mc_snum; i++) {
7064 cdst->mc_pg[i] = csrc->mc_pg[i];
7065 cdst->mc_ki[i] = csrc->mc_ki[i];
7069 /** Rebalance the tree after a delete operation.
7070 * @param[in] mc Cursor pointing to the page where rebalancing
7072 * @return 0 on success, non-zero on failure.
7075 mdb_rebalance(MDB_cursor *mc)
7079 unsigned int ptop, minkeys;
7082 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7086 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7087 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7088 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7089 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7090 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7094 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7095 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7098 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7099 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7105 if (mc->mc_snum < 2) {
7106 MDB_page *mp = mc->mc_pg[0];
7108 DPUTS("Can't rebalance a subpage, ignoring");
7111 if (NUMKEYS(mp) == 0) {
7112 DPUTS("tree is completely empty");
7113 mc->mc_db->md_root = P_INVALID;
7114 mc->mc_db->md_depth = 0;
7115 mc->mc_db->md_leaf_pages = 0;
7116 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7119 /* Adjust cursors pointing to mp */
7122 mc->mc_flags &= ~C_INITIALIZED;
7124 MDB_cursor *m2, *m3;
7125 MDB_dbi dbi = mc->mc_dbi;
7127 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7128 if (mc->mc_flags & C_SUB)
7129 m3 = &m2->mc_xcursor->mx_cursor;
7132 if (m3->mc_snum < mc->mc_snum) continue;
7133 if (m3->mc_pg[0] == mp) {
7136 m3->mc_flags &= ~C_INITIALIZED;
7140 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7141 DPUTS("collapsing root page!");
7142 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7145 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7146 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7149 mc->mc_db->md_depth--;
7150 mc->mc_db->md_branch_pages--;
7151 mc->mc_ki[0] = mc->mc_ki[1];
7153 /* Adjust other cursors pointing to mp */
7154 MDB_cursor *m2, *m3;
7155 MDB_dbi dbi = mc->mc_dbi;
7157 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7158 if (mc->mc_flags & C_SUB)
7159 m3 = &m2->mc_xcursor->mx_cursor;
7162 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7163 if (m3->mc_pg[0] == mp) {
7164 m3->mc_pg[0] = mc->mc_pg[0];
7167 m3->mc_ki[0] = m3->mc_ki[1];
7172 DPUTS("root page doesn't need rebalancing");
7176 /* The parent (branch page) must have at least 2 pointers,
7177 * otherwise the tree is invalid.
7179 ptop = mc->mc_top-1;
7180 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7182 /* Leaf page fill factor is below the threshold.
7183 * Try to move keys from left or right neighbor, or
7184 * merge with a neighbor page.
7189 mdb_cursor_copy(mc, &mn);
7190 mn.mc_xcursor = NULL;
7192 if (mc->mc_ki[ptop] == 0) {
7193 /* We're the leftmost leaf in our parent.
7195 DPUTS("reading right neighbor");
7197 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7198 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7201 mn.mc_ki[mn.mc_top] = 0;
7202 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7204 /* There is at least one neighbor to the left.
7206 DPUTS("reading left neighbor");
7208 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7209 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7212 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7213 mc->mc_ki[mc->mc_top] = 0;
7216 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7217 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7218 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7220 /* If the neighbor page is above threshold and has enough keys,
7221 * move one key from it. Otherwise we should try to merge them.
7222 * (A branch page must never have less than 2 keys.)
7224 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7225 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7226 return mdb_node_move(&mn, mc);
7228 if (mc->mc_ki[ptop] == 0)
7229 rc = mdb_page_merge(&mn, mc);
7231 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7232 rc = mdb_page_merge(mc, &mn);
7233 mdb_cursor_copy(&mn, mc);
7235 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7240 /** Complete a delete operation started by #mdb_cursor_del(). */
7242 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7249 mp = mc->mc_pg[mc->mc_top];
7250 ki = mc->mc_ki[mc->mc_top];
7252 /* add overflow pages to free list */
7253 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7257 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7258 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7259 (rc = mdb_ovpage_free(mc, omp)))
7262 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7263 mc->mc_db->md_entries--;
7264 rc = mdb_rebalance(mc);
7265 if (rc != MDB_SUCCESS)
7266 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7269 MDB_dbi dbi = mc->mc_dbi;
7271 mp = mc->mc_pg[mc->mc_top];
7272 nkeys = NUMKEYS(mp);
7274 /* if mc points past last node in page, find next sibling */
7275 if (mc->mc_ki[mc->mc_top] >= nkeys)
7276 mdb_cursor_sibling(mc, 1);
7278 /* Adjust other cursors pointing to mp */
7279 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7282 if (!(m2->mc_flags & C_INITIALIZED))
7284 if (m2->mc_pg[mc->mc_top] == mp) {
7285 if (m2->mc_ki[mc->mc_top] >= ki) {
7286 m2->mc_flags |= C_DEL;
7287 if (m2->mc_ki[mc->mc_top] > ki)
7288 m2->mc_ki[mc->mc_top]--;
7290 if (m2->mc_ki[mc->mc_top] >= nkeys)
7291 mdb_cursor_sibling(m2, 1);
7294 mc->mc_flags |= C_DEL;
7301 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7302 MDB_val *key, MDB_val *data)
7307 MDB_val rdata, *xdata;
7311 assert(key != NULL);
7313 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7315 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7318 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7319 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7321 if (key->mv_size > MDB_MAXKEYSIZE) {
7322 return MDB_BAD_VALSIZE;
7325 mdb_cursor_init(&mc, txn, dbi, &mx);
7328 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7329 /* must ignore any data */
7340 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7342 /* let mdb_page_split know about this cursor if needed:
7343 * delete will trigger a rebalance; if it needs to move
7344 * a node from one page to another, it will have to
7345 * update the parent's separator key(s). If the new sepkey
7346 * is larger than the current one, the parent page may
7347 * run out of space, triggering a split. We need this
7348 * cursor to be consistent until the end of the rebalance.
7350 mc.mc_flags |= C_UNTRACK;
7351 mc.mc_next = txn->mt_cursors[dbi];
7352 txn->mt_cursors[dbi] = &mc;
7353 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7354 txn->mt_cursors[dbi] = mc.mc_next;
7359 /** Split a page and insert a new node.
7360 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7361 * The cursor will be updated to point to the actual page and index where
7362 * the node got inserted after the split.
7363 * @param[in] newkey The key for the newly inserted node.
7364 * @param[in] newdata The data for the newly inserted node.
7365 * @param[in] newpgno The page number, if the new node is a branch node.
7366 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7367 * @return 0 on success, non-zero on failure.
7370 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7371 unsigned int nflags)
7374 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7377 unsigned int i, j, split_indx, nkeys, pmax;
7379 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7381 MDB_page *mp, *rp, *pp;
7386 mp = mc->mc_pg[mc->mc_top];
7387 newindx = mc->mc_ki[mc->mc_top];
7389 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7390 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7391 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7393 /* Create a right sibling. */
7394 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7396 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7398 if (mc->mc_snum < 2) {
7399 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7401 /* shift current top to make room for new parent */
7402 mc->mc_pg[1] = mc->mc_pg[0];
7403 mc->mc_ki[1] = mc->mc_ki[0];
7406 mc->mc_db->md_root = pp->mp_pgno;
7407 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7408 mc->mc_db->md_depth++;
7411 /* Add left (implicit) pointer. */
7412 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7413 /* undo the pre-push */
7414 mc->mc_pg[0] = mc->mc_pg[1];
7415 mc->mc_ki[0] = mc->mc_ki[1];
7416 mc->mc_db->md_root = mp->mp_pgno;
7417 mc->mc_db->md_depth--;
7424 ptop = mc->mc_top-1;
7425 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7428 mc->mc_flags |= C_SPLITTING;
7429 mdb_cursor_copy(mc, &mn);
7430 mn.mc_pg[mn.mc_top] = rp;
7431 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7433 if (nflags & MDB_APPEND) {
7434 mn.mc_ki[mn.mc_top] = 0;
7436 split_indx = newindx;
7441 nkeys = NUMKEYS(mp);
7442 split_indx = nkeys / 2;
7443 if (newindx < split_indx)
7449 unsigned int lsize, rsize, ksize;
7450 /* Move half of the keys to the right sibling */
7452 x = mc->mc_ki[mc->mc_top] - split_indx;
7453 ksize = mc->mc_db->md_pad;
7454 split = LEAF2KEY(mp, split_indx, ksize);
7455 rsize = (nkeys - split_indx) * ksize;
7456 lsize = (nkeys - split_indx) * sizeof(indx_t);
7457 mp->mp_lower -= lsize;
7458 rp->mp_lower += lsize;
7459 mp->mp_upper += rsize - lsize;
7460 rp->mp_upper -= rsize - lsize;
7461 sepkey.mv_size = ksize;
7462 if (newindx == split_indx) {
7463 sepkey.mv_data = newkey->mv_data;
7465 sepkey.mv_data = split;
7468 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7469 memcpy(rp->mp_ptrs, split, rsize);
7470 sepkey.mv_data = rp->mp_ptrs;
7471 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7472 memcpy(ins, newkey->mv_data, ksize);
7473 mp->mp_lower += sizeof(indx_t);
7474 mp->mp_upper -= ksize - sizeof(indx_t);
7477 memcpy(rp->mp_ptrs, split, x * ksize);
7478 ins = LEAF2KEY(rp, x, ksize);
7479 memcpy(ins, newkey->mv_data, ksize);
7480 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7481 rp->mp_lower += sizeof(indx_t);
7482 rp->mp_upper -= ksize - sizeof(indx_t);
7483 mc->mc_ki[mc->mc_top] = x;
7484 mc->mc_pg[mc->mc_top] = rp;
7489 /* For leaf pages, check the split point based on what
7490 * fits where, since otherwise mdb_node_add can fail.
7492 * This check is only needed when the data items are
7493 * relatively large, such that being off by one will
7494 * make the difference between success or failure.
7496 * It's also relevant if a page happens to be laid out
7497 * such that one half of its nodes are all "small" and
7498 * the other half of its nodes are "large." If the new
7499 * item is also "large" and falls on the half with
7500 * "large" nodes, it also may not fit.
7503 unsigned int psize, nsize;
7504 /* Maximum free space in an empty page */
7505 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7506 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7507 if ((nkeys < 20) || (nsize > pmax/16)) {
7508 if (newindx <= split_indx) {
7511 for (i=0; i<split_indx; i++) {
7512 node = NODEPTR(mp, i);
7513 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7514 if (F_ISSET(node->mn_flags, F_BIGDATA))
7515 psize += sizeof(pgno_t);
7517 psize += NODEDSZ(node);
7521 split_indx = newindx;
7532 for (i=nkeys-1; i>=split_indx; i--) {
7533 node = NODEPTR(mp, i);
7534 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7535 if (F_ISSET(node->mn_flags, F_BIGDATA))
7536 psize += sizeof(pgno_t);
7538 psize += NODEDSZ(node);
7542 split_indx = newindx;
7553 /* First find the separating key between the split pages.
7554 * The case where newindx == split_indx is ambiguous; the
7555 * new item could go to the new page or stay on the original
7556 * page. If newpos == 1 it goes to the new page.
7558 if (newindx == split_indx && newpos) {
7559 sepkey.mv_size = newkey->mv_size;
7560 sepkey.mv_data = newkey->mv_data;
7562 node = NODEPTR(mp, split_indx);
7563 sepkey.mv_size = node->mn_ksize;
7564 sepkey.mv_data = NODEKEY(node);
7568 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7570 /* Copy separator key to the parent.
7572 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7576 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7579 if (mn.mc_snum == mc->mc_snum) {
7580 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7581 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7582 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7583 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7588 /* Right page might now have changed parent.
7589 * Check if left page also changed parent.
7591 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7592 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7593 for (i=0; i<ptop; i++) {
7594 mc->mc_pg[i] = mn.mc_pg[i];
7595 mc->mc_ki[i] = mn.mc_ki[i];
7597 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7598 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7602 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7605 mc->mc_flags ^= C_SPLITTING;
7606 if (rc != MDB_SUCCESS) {
7609 if (nflags & MDB_APPEND) {
7610 mc->mc_pg[mc->mc_top] = rp;
7611 mc->mc_ki[mc->mc_top] = 0;
7612 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7615 for (i=0; i<mc->mc_top; i++)
7616 mc->mc_ki[i] = mn.mc_ki[i];
7623 /* Move half of the keys to the right sibling. */
7625 /* grab a page to hold a temporary copy */
7626 copy = mdb_page_malloc(mc->mc_txn, 1);
7630 copy->mp_pgno = mp->mp_pgno;
7631 copy->mp_flags = mp->mp_flags;
7632 copy->mp_lower = PAGEHDRSZ;
7633 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7634 mc->mc_pg[mc->mc_top] = copy;
7635 for (i = j = 0; i <= nkeys; j++) {
7636 if (i == split_indx) {
7637 /* Insert in right sibling. */
7638 /* Reset insert index for right sibling. */
7639 if (i != newindx || (newpos ^ ins_new)) {
7641 mc->mc_pg[mc->mc_top] = rp;
7645 if (i == newindx && !ins_new) {
7646 /* Insert the original entry that caused the split. */
7647 rkey.mv_data = newkey->mv_data;
7648 rkey.mv_size = newkey->mv_size;
7657 /* Update index for the new key. */
7658 mc->mc_ki[mc->mc_top] = j;
7659 } else if (i == nkeys) {
7662 node = NODEPTR(mp, i);
7663 rkey.mv_data = NODEKEY(node);
7664 rkey.mv_size = node->mn_ksize;
7666 xdata.mv_data = NODEDATA(node);
7667 xdata.mv_size = NODEDSZ(node);
7670 pgno = NODEPGNO(node);
7671 flags = node->mn_flags;
7676 if (!IS_LEAF(mp) && j == 0) {
7677 /* First branch index doesn't need key data. */
7681 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7685 nkeys = NUMKEYS(copy);
7686 for (i=0; i<nkeys; i++)
7687 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7688 mp->mp_lower = copy->mp_lower;
7689 mp->mp_upper = copy->mp_upper;
7690 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7691 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7693 /* reset back to original page */
7694 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7695 mc->mc_pg[mc->mc_top] = mp;
7696 if (nflags & MDB_RESERVE) {
7697 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7698 if (!(node->mn_flags & F_BIGDATA))
7699 newdata->mv_data = NODEDATA(node);
7703 /* Make sure mc_ki is still valid.
7705 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7706 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7707 for (i=0; i<ptop; i++) {
7708 mc->mc_pg[i] = mn.mc_pg[i];
7709 mc->mc_ki[i] = mn.mc_ki[i];
7711 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7712 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7716 /* return tmp page to freelist */
7717 mdb_page_free(mc->mc_txn->mt_env, copy);
7720 /* Adjust other cursors pointing to mp */
7721 MDB_cursor *m2, *m3;
7722 MDB_dbi dbi = mc->mc_dbi;
7723 int fixup = NUMKEYS(mp);
7725 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7726 if (mc->mc_flags & C_SUB)
7727 m3 = &m2->mc_xcursor->mx_cursor;
7732 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7734 if (m3->mc_flags & C_SPLITTING)
7739 for (k=m3->mc_top; k>=0; k--) {
7740 m3->mc_ki[k+1] = m3->mc_ki[k];
7741 m3->mc_pg[k+1] = m3->mc_pg[k];
7743 if (m3->mc_ki[0] >= split_indx) {
7748 m3->mc_pg[0] = mc->mc_pg[0];
7752 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7753 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7754 m3->mc_ki[mc->mc_top]++;
7755 if (m3->mc_ki[mc->mc_top] >= fixup) {
7756 m3->mc_pg[mc->mc_top] = rp;
7757 m3->mc_ki[mc->mc_top] -= fixup;
7758 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7760 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7761 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7770 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7771 MDB_val *key, MDB_val *data, unsigned int flags)
7776 assert(key != NULL);
7777 assert(data != NULL);
7779 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7782 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7785 mdb_cursor_init(&mc, txn, dbi, &mx);
7786 return mdb_cursor_put(&mc, key, data, flags);
7790 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7792 if ((flag & CHANGEABLE) != flag)
7795 env->me_flags |= flag;
7797 env->me_flags &= ~flag;
7802 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7807 *arg = env->me_flags;
7812 mdb_env_get_path(MDB_env *env, const char **arg)
7817 *arg = env->me_path;
7821 /** Common code for #mdb_stat() and #mdb_env_stat().
7822 * @param[in] env the environment to operate in.
7823 * @param[in] db the #MDB_db record containing the stats to return.
7824 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7825 * @return 0, this function always succeeds.
7828 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7830 arg->ms_psize = env->me_psize;
7831 arg->ms_depth = db->md_depth;
7832 arg->ms_branch_pages = db->md_branch_pages;
7833 arg->ms_leaf_pages = db->md_leaf_pages;
7834 arg->ms_overflow_pages = db->md_overflow_pages;
7835 arg->ms_entries = db->md_entries;
7840 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7844 if (env == NULL || arg == NULL)
7847 toggle = mdb_env_pick_meta(env);
7849 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7853 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7857 if (env == NULL || arg == NULL)
7860 toggle = mdb_env_pick_meta(env);
7861 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7862 arg->me_mapsize = env->me_mapsize;
7863 arg->me_maxreaders = env->me_maxreaders;
7865 /* me_numreaders may be zero if this process never used any readers. Use
7866 * the shared numreader count if it exists.
7868 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7870 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7871 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7875 /** Set the default comparison functions for a database.
7876 * Called immediately after a database is opened to set the defaults.
7877 * The user can then override them with #mdb_set_compare() or
7878 * #mdb_set_dupsort().
7879 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7880 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7883 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7885 uint16_t f = txn->mt_dbs[dbi].md_flags;
7887 txn->mt_dbxs[dbi].md_cmp =
7888 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7889 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7891 txn->mt_dbxs[dbi].md_dcmp =
7892 !(f & MDB_DUPSORT) ? 0 :
7893 ((f & MDB_INTEGERDUP)
7894 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7895 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7898 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7903 int rc, dbflag, exact;
7904 unsigned int unused = 0;
7907 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7908 mdb_default_cmp(txn, FREE_DBI);
7911 if ((flags & VALID_FLAGS) != flags)
7913 if (txn->mt_flags & MDB_TXN_ERROR)
7919 if (flags & PERSISTENT_FLAGS) {
7920 uint16_t f2 = flags & PERSISTENT_FLAGS;
7921 /* make sure flag changes get committed */
7922 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7923 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7924 txn->mt_flags |= MDB_TXN_DIRTY;
7927 mdb_default_cmp(txn, MAIN_DBI);
7931 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7932 mdb_default_cmp(txn, MAIN_DBI);
7935 /* Is the DB already open? */
7937 for (i=2; i<txn->mt_numdbs; i++) {
7938 if (!txn->mt_dbxs[i].md_name.mv_size) {
7939 /* Remember this free slot */
7940 if (!unused) unused = i;
7943 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7944 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7950 /* If no free slot and max hit, fail */
7951 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7952 return MDB_DBS_FULL;
7954 /* Cannot mix named databases with some mainDB flags */
7955 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7956 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7958 /* Find the DB info */
7959 dbflag = DB_NEW|DB_VALID;
7962 key.mv_data = (void *)name;
7963 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7964 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7965 if (rc == MDB_SUCCESS) {
7966 /* make sure this is actually a DB */
7967 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7968 if (!(node->mn_flags & F_SUBDATA))
7969 return MDB_INCOMPATIBLE;
7970 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7971 /* Create if requested */
7973 data.mv_size = sizeof(MDB_db);
7974 data.mv_data = &dummy;
7975 memset(&dummy, 0, sizeof(dummy));
7976 dummy.md_root = P_INVALID;
7977 dummy.md_flags = flags & PERSISTENT_FLAGS;
7978 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7982 /* OK, got info, add to table */
7983 if (rc == MDB_SUCCESS) {
7984 unsigned int slot = unused ? unused : txn->mt_numdbs;
7985 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7986 txn->mt_dbxs[slot].md_name.mv_size = len;
7987 txn->mt_dbxs[slot].md_rel = NULL;
7988 txn->mt_dbflags[slot] = dbflag;
7989 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7991 mdb_default_cmp(txn, slot);
8000 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8002 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8005 if (txn->mt_dbflags[dbi] & DB_STALE) {
8008 /* Stale, must read the DB's root. cursor_init does it for us. */
8009 mdb_cursor_init(&mc, txn, dbi, &mx);
8011 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8014 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8017 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8019 ptr = env->me_dbxs[dbi].md_name.mv_data;
8020 env->me_dbxs[dbi].md_name.mv_data = NULL;
8021 env->me_dbxs[dbi].md_name.mv_size = 0;
8022 env->me_dbflags[dbi] = 0;
8026 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8028 /* We could return the flags for the FREE_DBI too but what's the point? */
8029 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8031 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8035 /** Add all the DB's pages to the free list.
8036 * @param[in] mc Cursor on the DB to free.
8037 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8038 * @return 0 on success, non-zero on failure.
8041 mdb_drop0(MDB_cursor *mc, int subs)
8045 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8046 if (rc == MDB_SUCCESS) {
8047 MDB_txn *txn = mc->mc_txn;
8052 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8053 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8056 mdb_cursor_copy(mc, &mx);
8057 while (mc->mc_snum > 0) {
8058 MDB_page *mp = mc->mc_pg[mc->mc_top];
8059 unsigned n = NUMKEYS(mp);
8061 for (i=0; i<n; i++) {
8062 ni = NODEPTR(mp, i);
8063 if (ni->mn_flags & F_BIGDATA) {
8066 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8067 rc = mdb_page_get(txn, pg, &omp, NULL);
8070 assert(IS_OVERFLOW(omp));
8071 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8075 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8076 mdb_xcursor_init1(mc, ni);
8077 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8083 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8085 for (i=0; i<n; i++) {
8087 ni = NODEPTR(mp, i);
8090 mdb_midl_xappend(txn->mt_free_pgs, pg);
8095 mc->mc_ki[mc->mc_top] = i;
8096 rc = mdb_cursor_sibling(mc, 1);
8098 /* no more siblings, go back to beginning
8099 * of previous level.
8103 for (i=1; i<mc->mc_snum; i++) {
8105 mc->mc_pg[i] = mx.mc_pg[i];
8110 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8111 } else if (rc == MDB_NOTFOUND) {
8117 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8119 MDB_cursor *mc, *m2;
8122 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8125 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8128 rc = mdb_cursor_open(txn, dbi, &mc);
8132 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8133 /* Invalidate the dropped DB's cursors */
8134 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8135 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8139 /* Can't delete the main DB */
8140 if (del && dbi > MAIN_DBI) {
8141 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8143 txn->mt_dbflags[dbi] = DB_STALE;
8144 mdb_dbi_close(txn->mt_env, dbi);
8147 /* reset the DB record, mark it dirty */
8148 txn->mt_dbflags[dbi] |= DB_DIRTY;
8149 txn->mt_dbs[dbi].md_depth = 0;
8150 txn->mt_dbs[dbi].md_branch_pages = 0;
8151 txn->mt_dbs[dbi].md_leaf_pages = 0;
8152 txn->mt_dbs[dbi].md_overflow_pages = 0;
8153 txn->mt_dbs[dbi].md_entries = 0;
8154 txn->mt_dbs[dbi].md_root = P_INVALID;
8156 txn->mt_flags |= MDB_TXN_DIRTY;
8159 mdb_cursor_close(mc);
8163 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8165 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8168 txn->mt_dbxs[dbi].md_cmp = cmp;
8172 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8174 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8177 txn->mt_dbxs[dbi].md_dcmp = cmp;
8181 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8183 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8186 txn->mt_dbxs[dbi].md_rel = rel;
8190 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8192 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8195 txn->mt_dbxs[dbi].md_relctx = ctx;
8199 int mdb_env_get_maxkeysize(MDB_env *env)
8201 return MDB_MAXKEYSIZE;
8204 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8206 unsigned int i, rdrs;
8213 if (!env->me_txns) {
8214 return func("(no reader locks)\n", ctx);
8216 rdrs = env->me_txns->mti_numreaders;
8217 mr = env->me_txns->mti_readers;
8218 for (i=0; i<rdrs; i++) {
8223 if (mr[i].mr_txnid == (txnid_t)-1) {
8224 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8226 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8230 func(" pid thread txnid\n", ctx);
8232 rc = func(buf, ctx);
8238 func("(no active readers)\n", ctx);
8243 /** Insert pid into list if not already present.
8244 * return -1 if already present.
8246 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8248 /* binary search of pid in list */
8250 unsigned cursor = 1;
8252 unsigned n = ids[0];
8255 unsigned pivot = n >> 1;
8256 cursor = base + pivot + 1;
8257 val = pid - ids[cursor];
8262 } else if ( val > 0 ) {
8267 /* found, so it's a duplicate */
8276 for (n = ids[0]; n > cursor; n--)
8282 int mdb_reader_check(MDB_env *env, int *dead)
8284 unsigned int i, j, rdrs;
8295 rdrs = env->me_txns->mti_numreaders;
8296 pids = malloc((rdrs+1) * sizeof(pid_t));
8300 mr = env->me_txns->mti_readers;
8302 for (i=0; i<rdrs; i++) {
8303 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8305 if (mdb_pid_insert(pids, pid) == 0) {
8306 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8308 /* Recheck, a new process may have reused pid */
8309 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8310 for (j=i; j<rdrs; j++)
8311 if (mr[j].mr_pid == pid) {
8312 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8313 (unsigned) pid, mr[j].mr_txnid));
8318 UNLOCK_MUTEX_R(env);