2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define MDB_PIDLOCK 0
144 #define pthread_t DWORD
145 #define pthread_mutex_t HANDLE
146 #define pthread_key_t DWORD
147 #define pthread_self() GetCurrentThreadId()
148 #define pthread_key_create(x,y) \
149 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
150 #define pthread_key_delete(x) TlsFree(x)
151 #define pthread_getspecific(x) TlsGetValue(x)
152 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
153 #define pthread_mutex_unlock(x) ReleaseMutex(x)
154 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
155 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
156 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
157 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
158 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
159 #define getpid() GetCurrentProcessId()
160 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
161 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
162 #define ErrCode() GetLastError()
163 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
164 #define close(fd) (CloseHandle(fd) ? 0 : -1)
165 #define munmap(ptr,len) UnmapViewOfFile(ptr)
166 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
167 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
169 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
174 #define Z "z" /**< printf format modifier for size_t */
176 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
177 #define MDB_PIDLOCK 1
179 #ifdef MDB_USE_POSIX_SEM
181 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
182 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
183 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
184 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
187 mdb_sem_wait(sem_t *sem)
190 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
195 /** Lock the reader mutex.
197 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
198 /** Unlock the reader mutex.
200 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
202 /** Lock the writer mutex.
203 * Only a single write transaction is allowed at a time. Other writers
204 * will block waiting for this mutex.
206 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
207 /** Unlock the writer mutex.
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
210 #endif /* MDB_USE_POSIX_SEM */
212 /** Get the error code for the last failed system function.
214 #define ErrCode() errno
216 /** An abstraction for a file handle.
217 * On POSIX systems file handles are small integers. On Windows
218 * they're opaque pointers.
222 /** A value for an invalid file handle.
223 * Mainly used to initialize file variables and signify that they are
226 #define INVALID_HANDLE_VALUE (-1)
228 /** Get the size of a memory page for the system.
229 * This is the basic size that the platform's memory manager uses, and is
230 * fundamental to the use of memory-mapped files.
232 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
235 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
238 #define MNAME_LEN (sizeof(pthread_mutex_t))
244 /** A flag for opening a file and requesting synchronous data writes.
245 * This is only used when writing a meta page. It's not strictly needed;
246 * we could just do a normal write and then immediately perform a flush.
247 * But if this flag is available it saves us an extra system call.
249 * @note If O_DSYNC is undefined but exists in /usr/include,
250 * preferably set some compiler flag to get the definition.
251 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
254 # define MDB_DSYNC O_DSYNC
258 /** Function for flushing the data of a file. Define this to fsync
259 * if fdatasync() is not supported.
261 #ifndef MDB_FDATASYNC
262 # define MDB_FDATASYNC fdatasync
266 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
277 /** A page number in the database.
278 * Note that 64 bit page numbers are overkill, since pages themselves
279 * already represent 12-13 bits of addressable memory, and the OS will
280 * always limit applications to a maximum of 63 bits of address space.
282 * @note In the #MDB_node structure, we only store 48 bits of this value,
283 * which thus limits us to only 60 bits of addressable data.
285 typedef MDB_ID pgno_t;
287 /** A transaction ID.
288 * See struct MDB_txn.mt_txnid for details.
290 typedef MDB_ID txnid_t;
292 /** @defgroup debug Debug Macros
296 /** Enable debug output. Needs variable argument macros (a C99 feature).
297 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
298 * read from and written to the database (used for free space management).
304 static int mdb_debug;
305 static txnid_t mdb_debug_start;
307 /** Print a debug message with printf formatting.
308 * Requires double parenthesis around 2 or more args.
310 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
311 # define DPRINTF0(fmt, ...) \
312 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
314 # define DPRINTF(args) ((void) 0)
316 /** Print a debug string.
317 * The string is printed literally, with no format processing.
319 #define DPUTS(arg) DPRINTF(("%s", arg))
322 /** A default memory page size.
323 * The actual size is platform-dependent, but we use this for
324 * boot-strapping. We probably should not be using this any more.
325 * The #GET_PAGESIZE() macro is used to get the actual size.
327 * Note that we don't currently support Huge pages. On Linux,
328 * regular data files cannot use Huge pages, and in general
329 * Huge pages aren't actually pageable. We rely on the OS
330 * demand-pager to read our data and page it out when memory
331 * pressure from other processes is high. So until OSs have
332 * actual paging support for Huge pages, they're not viable.
334 #define MDB_PAGESIZE 4096
336 /** The minimum number of keys required in a database page.
337 * Setting this to a larger value will place a smaller bound on the
338 * maximum size of a data item. Data items larger than this size will
339 * be pushed into overflow pages instead of being stored directly in
340 * the B-tree node. This value used to default to 4. With a page size
341 * of 4096 bytes that meant that any item larger than 1024 bytes would
342 * go into an overflow page. That also meant that on average 2-3KB of
343 * each overflow page was wasted space. The value cannot be lower than
344 * 2 because then there would no longer be a tree structure. With this
345 * value, items larger than 2KB will go into overflow pages, and on
346 * average only 1KB will be wasted.
348 #define MDB_MINKEYS 2
350 /** A stamp that identifies a file as an MDB file.
351 * There's nothing special about this value other than that it is easily
352 * recognizable, and it will reflect any byte order mismatches.
354 #define MDB_MAGIC 0xBEEFC0DE
356 /** The version number for a database's datafile format. */
357 #define MDB_DATA_VERSION 1
358 /** The version number for a database's lockfile format. */
359 #define MDB_LOCK_VERSION 1
361 /** @brief The maximum size of a key in the database.
363 * The library rejects bigger keys, and cannot deal with records
364 * with bigger keys stored by a library with bigger max keysize.
366 * We require that keys all fit onto a regular page. This limit
367 * could be raised a bit further if needed; to something just
368 * under #MDB_PAGESIZE / #MDB_MINKEYS.
370 * Note that data items in an #MDB_DUPSORT database are actually keys
371 * of a subDB, so they're also limited to this size.
373 #ifndef MDB_MAXKEYSIZE
374 #define MDB_MAXKEYSIZE 511
377 /** @brief The maximum size of a data item.
379 * We only store a 32 bit value for node sizes.
381 #define MAXDATASIZE 0xffffffffUL
386 * This is used for printing a hex dump of a key's contents.
388 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
389 /** Display a key in hex.
391 * Invoke a function to display a key in hex.
393 #define DKEY(x) mdb_dkey(x, kbuf)
399 /** An invalid page number.
400 * Mainly used to denote an empty tree.
402 #define P_INVALID (~(pgno_t)0)
404 /** Test if the flags \b f are set in a flag word \b w. */
405 #define F_ISSET(w, f) (((w) & (f)) == (f))
407 /** Used for offsets within a single page.
408 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
411 typedef uint16_t indx_t;
413 /** Default size of memory map.
414 * This is certainly too small for any actual applications. Apps should always set
415 * the size explicitly using #mdb_env_set_mapsize().
417 #define DEFAULT_MAPSIZE 1048576
419 /** @defgroup readers Reader Lock Table
420 * Readers don't acquire any locks for their data access. Instead, they
421 * simply record their transaction ID in the reader table. The reader
422 * mutex is needed just to find an empty slot in the reader table. The
423 * slot's address is saved in thread-specific data so that subsequent read
424 * transactions started by the same thread need no further locking to proceed.
426 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
428 * No reader table is used if the database is on a read-only filesystem.
430 * Since the database uses multi-version concurrency control, readers don't
431 * actually need any locking. This table is used to keep track of which
432 * readers are using data from which old transactions, so that we'll know
433 * when a particular old transaction is no longer in use. Old transactions
434 * that have discarded any data pages can then have those pages reclaimed
435 * for use by a later write transaction.
437 * The lock table is constructed such that reader slots are aligned with the
438 * processor's cache line size. Any slot is only ever used by one thread.
439 * This alignment guarantees that there will be no contention or cache
440 * thrashing as threads update their own slot info, and also eliminates
441 * any need for locking when accessing a slot.
443 * A writer thread will scan every slot in the table to determine the oldest
444 * outstanding reader transaction. Any freed pages older than this will be
445 * reclaimed by the writer. The writer doesn't use any locks when scanning
446 * this table. This means that there's no guarantee that the writer will
447 * see the most up-to-date reader info, but that's not required for correct
448 * operation - all we need is to know the upper bound on the oldest reader,
449 * we don't care at all about the newest reader. So the only consequence of
450 * reading stale information here is that old pages might hang around a
451 * while longer before being reclaimed. That's actually good anyway, because
452 * the longer we delay reclaiming old pages, the more likely it is that a
453 * string of contiguous pages can be found after coalescing old pages from
454 * many old transactions together.
457 /** Number of slots in the reader table.
458 * This value was chosen somewhat arbitrarily. 126 readers plus a
459 * couple mutexes fit exactly into 8KB on my development machine.
460 * Applications should set the table size using #mdb_env_set_maxreaders().
462 #define DEFAULT_READERS 126
464 /** The size of a CPU cache line in bytes. We want our lock structures
465 * aligned to this size to avoid false cache line sharing in the
467 * This value works for most CPUs. For Itanium this should be 128.
473 /** The information we store in a single slot of the reader table.
474 * In addition to a transaction ID, we also record the process and
475 * thread ID that owns a slot, so that we can detect stale information,
476 * e.g. threads or processes that went away without cleaning up.
477 * @note We currently don't check for stale records. We simply re-init
478 * the table when we know that we're the only process opening the
481 typedef struct MDB_rxbody {
482 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
483 * Multiple readers that start at the same time will probably have the
484 * same ID here. Again, it's not important to exclude them from
485 * anything; all we need to know is which version of the DB they
486 * started from so we can avoid overwriting any data used in that
487 * particular version.
490 /** The process ID of the process owning this reader txn. */
492 /** The thread ID of the thread owning this txn. */
496 /** The actual reader record, with cacheline padding. */
497 typedef struct MDB_reader {
500 /** shorthand for mrb_txnid */
501 #define mr_txnid mru.mrx.mrb_txnid
502 #define mr_pid mru.mrx.mrb_pid
503 #define mr_tid mru.mrx.mrb_tid
504 /** cache line alignment */
505 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
509 /** The header for the reader table.
510 * The table resides in a memory-mapped file. (This is a different file
511 * than is used for the main database.)
513 * For POSIX the actual mutexes reside in the shared memory of this
514 * mapped file. On Windows, mutexes are named objects allocated by the
515 * kernel; we store the mutex names in this mapped file so that other
516 * processes can grab them. This same approach is also used on
517 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
518 * process-shared POSIX mutexes. For these cases where a named object
519 * is used, the object name is derived from a 64 bit FNV hash of the
520 * environment pathname. As such, naming collisions are extremely
521 * unlikely. If a collision occurs, the results are unpredictable.
523 typedef struct MDB_txbody {
524 /** Stamp identifying this as an MDB file. It must be set
527 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
529 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
530 char mtb_rmname[MNAME_LEN];
532 /** Mutex protecting access to this table.
533 * This is the reader lock that #LOCK_MUTEX_R acquires.
535 pthread_mutex_t mtb_mutex;
537 /** The ID of the last transaction committed to the database.
538 * This is recorded here only for convenience; the value can always
539 * be determined by reading the main database meta pages.
542 /** The number of slots that have been used in the reader table.
543 * This always records the maximum count, it is not decremented
544 * when readers release their slots.
546 unsigned mtb_numreaders;
549 /** The actual reader table definition. */
550 typedef struct MDB_txninfo {
553 #define mti_magic mt1.mtb.mtb_magic
554 #define mti_format mt1.mtb.mtb_format
555 #define mti_mutex mt1.mtb.mtb_mutex
556 #define mti_rmname mt1.mtb.mtb_rmname
557 #define mti_txnid mt1.mtb.mtb_txnid
558 #define mti_numreaders mt1.mtb.mtb_numreaders
559 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
562 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
563 char mt2_wmname[MNAME_LEN];
564 #define mti_wmname mt2.mt2_wmname
566 pthread_mutex_t mt2_wmutex;
567 #define mti_wmutex mt2.mt2_wmutex
569 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
571 MDB_reader mti_readers[1];
574 /** Lockfile format signature: version, features and field layout */
575 #define MDB_LOCK_FORMAT \
577 ((MDB_LOCK_VERSION) \
578 /* Flags which describe functionality */ \
579 + (((MDB_PIDLOCK) != 0) << 16)))
582 /** Common header for all page types.
583 * Overflow records occupy a number of contiguous pages with no
584 * headers on any page after the first.
586 typedef struct MDB_page {
587 #define mp_pgno mp_p.p_pgno
588 #define mp_next mp_p.p_next
590 pgno_t p_pgno; /**< page number */
591 void * p_next; /**< for in-memory list of freed structs */
594 /** @defgroup mdb_page Page Flags
596 * Flags for the page headers.
599 #define P_BRANCH 0x01 /**< branch page */
600 #define P_LEAF 0x02 /**< leaf page */
601 #define P_OVERFLOW 0x04 /**< overflow page */
602 #define P_META 0x08 /**< meta page */
603 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
604 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
605 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
606 #define P_KEEP 0x8000 /**< leave this page alone during spill */
608 uint16_t mp_flags; /**< @ref mdb_page */
609 #define mp_lower mp_pb.pb.pb_lower
610 #define mp_upper mp_pb.pb.pb_upper
611 #define mp_pages mp_pb.pb_pages
614 indx_t pb_lower; /**< lower bound of free space */
615 indx_t pb_upper; /**< upper bound of free space */
617 uint32_t pb_pages; /**< number of overflow pages */
619 indx_t mp_ptrs[1]; /**< dynamic size */
622 /** Size of the page header, excluding dynamic data at the end */
623 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
625 /** Address of first usable data byte in a page, after the header */
626 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
628 /** Number of nodes on a page */
629 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
631 /** The amount of space remaining in the page */
632 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
634 /** The percentage of space used in the page, in tenths of a percent. */
635 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
636 ((env)->me_psize - PAGEHDRSZ))
637 /** The minimum page fill factor, in tenths of a percent.
638 * Pages emptier than this are candidates for merging.
640 #define FILL_THRESHOLD 250
642 /** Test if a page is a leaf page */
643 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
644 /** Test if a page is a LEAF2 page */
645 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
646 /** Test if a page is a branch page */
647 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
648 /** Test if a page is an overflow page */
649 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
650 /** Test if a page is a sub page */
651 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
653 /** The number of overflow pages needed to store the given size. */
654 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
656 /** Header for a single key/data pair within a page.
657 * We guarantee 2-byte alignment for nodes.
659 typedef struct MDB_node {
660 /** lo and hi are used for data size on leaf nodes and for
661 * child pgno on branch nodes. On 64 bit platforms, flags
662 * is also used for pgno. (Branch nodes have no flags).
663 * They are in host byte order in case that lets some
664 * accesses be optimized into a 32-bit word access.
666 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
667 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
668 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
669 /** @defgroup mdb_node Node Flags
671 * Flags for node headers.
674 #define F_BIGDATA 0x01 /**< data put on overflow page */
675 #define F_SUBDATA 0x02 /**< data is a sub-database */
676 #define F_DUPDATA 0x04 /**< data has duplicates */
678 /** valid flags for #mdb_node_add() */
679 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
682 unsigned short mn_flags; /**< @ref mdb_node */
683 unsigned short mn_ksize; /**< key size */
684 char mn_data[1]; /**< key and data are appended here */
687 /** Size of the node header, excluding dynamic data at the end */
688 #define NODESIZE offsetof(MDB_node, mn_data)
690 /** Bit position of top word in page number, for shifting mn_flags */
691 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
693 /** Size of a node in a branch page with a given key.
694 * This is just the node header plus the key, there is no data.
696 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
698 /** Size of a node in a leaf page with a given key and data.
699 * This is node header plus key plus data size.
701 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
703 /** Address of node \b i in page \b p */
704 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
706 /** Address of the key for the node */
707 #define NODEKEY(node) (void *)((node)->mn_data)
709 /** Address of the data for a node */
710 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
712 /** Get the page number pointed to by a branch node */
713 #define NODEPGNO(node) \
714 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
715 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
716 /** Set the page number in a branch node */
717 #define SETPGNO(node,pgno) do { \
718 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
719 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
721 /** Get the size of the data in a leaf node */
722 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
723 /** Set the size of the data for a leaf node */
724 #define SETDSZ(node,size) do { \
725 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
726 /** The size of a key in a node */
727 #define NODEKSZ(node) ((node)->mn_ksize)
729 /** Copy a page number from src to dst */
731 #define COPY_PGNO(dst,src) dst = src
733 #if SIZE_MAX > 4294967295UL
734 #define COPY_PGNO(dst,src) do { \
735 unsigned short *s, *d; \
736 s = (unsigned short *)&(src); \
737 d = (unsigned short *)&(dst); \
744 #define COPY_PGNO(dst,src) do { \
745 unsigned short *s, *d; \
746 s = (unsigned short *)&(src); \
747 d = (unsigned short *)&(dst); \
753 /** The address of a key in a LEAF2 page.
754 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
755 * There are no node headers, keys are stored contiguously.
757 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
759 /** Set the \b node's key into \b keyptr, if requested. */
760 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
761 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
763 /** Set the \b node's key into \b key. */
764 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
766 /** Information about a single database in the environment. */
767 typedef struct MDB_db {
768 uint32_t md_pad; /**< also ksize for LEAF2 pages */
769 uint16_t md_flags; /**< @ref mdb_dbi_open */
770 uint16_t md_depth; /**< depth of this tree */
771 pgno_t md_branch_pages; /**< number of internal pages */
772 pgno_t md_leaf_pages; /**< number of leaf pages */
773 pgno_t md_overflow_pages; /**< number of overflow pages */
774 size_t md_entries; /**< number of data items */
775 pgno_t md_root; /**< the root page of this tree */
778 /** mdb_dbi_open flags */
779 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
780 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
781 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
782 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
784 /** Handle for the DB used to track free pages. */
786 /** Handle for the default DB. */
789 /** Meta page content.
790 * A meta page is the start point for accessing a database snapshot.
791 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
793 typedef struct MDB_meta {
794 /** Stamp identifying this as an MDB file. It must be set
797 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
799 void *mm_address; /**< address for fixed mapping */
800 size_t mm_mapsize; /**< size of mmap region */
801 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
802 /** The size of pages used in this DB */
803 #define mm_psize mm_dbs[0].md_pad
804 /** Any persistent environment flags. @ref mdb_env */
805 #define mm_flags mm_dbs[0].md_flags
806 pgno_t mm_last_pg; /**< last used page in file */
807 txnid_t mm_txnid; /**< txnid that committed this page */
810 /** Buffer for a stack-allocated dirty page.
811 * The members define size and alignment, and silence type
812 * aliasing warnings. They are not used directly; that could
813 * mean incorrectly using several union members in parallel.
815 typedef union MDB_pagebuf {
816 char mb_raw[MDB_PAGESIZE];
819 char mm_pad[PAGEHDRSZ];
824 /** Auxiliary DB info.
825 * The information here is mostly static/read-only. There is
826 * only a single copy of this record in the environment.
828 typedef struct MDB_dbx {
829 MDB_val md_name; /**< name of the database */
830 MDB_cmp_func *md_cmp; /**< function for comparing keys */
831 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
832 MDB_rel_func *md_rel; /**< user relocate function */
833 void *md_relctx; /**< user-provided context for md_rel */
836 /** A database transaction.
837 * Every operation requires a transaction handle.
840 MDB_txn *mt_parent; /**< parent of a nested txn */
841 MDB_txn *mt_child; /**< nested txn under this txn */
842 pgno_t mt_next_pgno; /**< next unallocated page */
843 /** The ID of this transaction. IDs are integers incrementing from 1.
844 * Only committed write transactions increment the ID. If a transaction
845 * aborts, the ID may be re-used by the next writer.
848 MDB_env *mt_env; /**< the DB environment */
849 /** The list of pages that became unused during this transaction.
852 /** The sorted list of dirty pages we temporarily wrote to disk
853 * because the dirty list was full. page numbers in here are
854 * shifted left by 1, deleted slots have the LSB set.
856 MDB_IDL mt_spill_pgs;
858 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
860 /** For read txns: This thread/txn's reader table slot, or NULL. */
863 /** Array of records for each DB known in the environment. */
865 /** Array of MDB_db records for each known DB */
867 /** @defgroup mt_dbflag Transaction DB Flags
871 #define DB_DIRTY 0x01 /**< DB was written in this txn */
872 #define DB_STALE 0x02 /**< DB record is older than txnID */
873 #define DB_NEW 0x04 /**< DB handle opened in this txn */
874 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
876 /** In write txns, array of cursors for each DB */
877 MDB_cursor **mt_cursors;
878 /** Array of flags for each DB */
879 unsigned char *mt_dbflags;
880 /** Number of DB records in use. This number only ever increments;
881 * we don't decrement it when individual DB handles are closed.
885 /** @defgroup mdb_txn Transaction Flags
889 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
890 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
891 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
892 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
894 unsigned int mt_flags; /**< @ref mdb_txn */
895 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
896 unsigned int mt_dirty_room;
897 /** Tracks which of the two meta pages was used at the start
898 * of this transaction.
900 unsigned int mt_toggle;
903 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
904 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
905 * raise this on a 64 bit machine.
907 #define CURSOR_STACK 32
911 /** Cursors are used for all DB operations.
912 * A cursor holds a path of (page pointer, key index) from the DB
913 * root to a position in the DB, plus other state. #MDB_DUPSORT
914 * cursors include an xcursor to the current data item. Write txns
915 * track their cursors and keep them up to date when data moves.
916 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
917 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
920 /** Next cursor on this DB in this txn */
922 /** Backup of the original cursor if this cursor is a shadow */
923 MDB_cursor *mc_backup;
924 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
925 struct MDB_xcursor *mc_xcursor;
926 /** The transaction that owns this cursor */
928 /** The database handle this cursor operates on */
930 /** The database record for this cursor */
932 /** The database auxiliary record for this cursor */
934 /** The @ref mt_dbflag for this database */
935 unsigned char *mc_dbflag;
936 unsigned short mc_snum; /**< number of pushed pages */
937 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
938 /** @defgroup mdb_cursor Cursor Flags
940 * Cursor state flags.
943 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
944 #define C_EOF 0x02 /**< No more data */
945 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
946 #define C_DEL 0x08 /**< last op was a cursor_del */
947 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
948 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
950 unsigned int mc_flags; /**< @ref mdb_cursor */
951 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
952 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
955 /** Context for sorted-dup records.
956 * We could have gone to a fully recursive design, with arbitrarily
957 * deep nesting of sub-databases. But for now we only handle these
958 * levels - main DB, optional sub-DB, sorted-duplicate DB.
960 typedef struct MDB_xcursor {
961 /** A sub-cursor for traversing the Dup DB */
962 MDB_cursor mx_cursor;
963 /** The database record for this Dup DB */
965 /** The auxiliary DB record for this Dup DB */
967 /** The @ref mt_dbflag for this Dup DB */
968 unsigned char mx_dbflag;
971 /** State of FreeDB old pages, stored in the MDB_env */
972 typedef struct MDB_pgstate {
973 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
974 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
977 /** The database environment. */
979 HANDLE me_fd; /**< The main data file */
980 HANDLE me_lfd; /**< The lock file */
981 HANDLE me_mfd; /**< just for writing the meta pages */
982 /** Failed to update the meta page. Probably an I/O error. */
983 #define MDB_FATAL_ERROR 0x80000000U
984 /** Some fields are initialized. */
985 #define MDB_ENV_ACTIVE 0x20000000U
986 /** me_txkey is set */
987 #define MDB_ENV_TXKEY 0x10000000U
988 /** Have liveness lock in reader table */
989 #define MDB_LIVE_READER 0x08000000U
990 uint32_t me_flags; /**< @ref mdb_env */
991 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
992 unsigned int me_maxreaders; /**< size of the reader table */
993 unsigned int me_numreaders; /**< max numreaders set by this env */
994 MDB_dbi me_numdbs; /**< number of DBs opened */
995 MDB_dbi me_maxdbs; /**< size of the DB table */
996 pid_t me_pid; /**< process ID of this env */
997 char *me_path; /**< path to the DB files */
998 char *me_map; /**< the memory map of the data file */
999 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1000 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1001 MDB_txn *me_txn; /**< current write transaction */
1002 size_t me_mapsize; /**< size of the data memory map */
1003 off_t me_size; /**< current file size */
1004 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1005 MDB_dbx *me_dbxs; /**< array of static DB info */
1006 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1007 pthread_key_t me_txkey; /**< thread-key for readers */
1008 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1009 # define me_pglast me_pgstate.mf_pglast
1010 # define me_pghead me_pgstate.mf_pghead
1011 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1012 /** IDL of pages that became unused in a write txn */
1013 MDB_IDL me_free_pgs;
1014 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1015 MDB_ID2L me_dirty_list;
1016 /** Max number of freelist items that can fit in a single overflow page */
1018 /** Max size of a node on a page */
1019 unsigned int me_nodemax;
1021 int me_pidquery; /**< Used in OpenProcess */
1022 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1024 #elif defined(MDB_USE_POSIX_SEM)
1025 sem_t *me_rmutex; /* Shared mutexes are not supported */
1030 /** Nested transaction */
1031 typedef struct MDB_ntxn {
1032 MDB_txn mnt_txn; /**< the transaction */
1033 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1036 /** max number of pages to commit in one writev() call */
1037 #define MDB_COMMIT_PAGES 64
1038 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1039 #undef MDB_COMMIT_PAGES
1040 #define MDB_COMMIT_PAGES IOV_MAX
1043 /* max bytes to write in one call */
1044 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1046 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1047 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1048 static int mdb_page_touch(MDB_cursor *mc);
1050 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1051 static int mdb_page_search_root(MDB_cursor *mc,
1052 MDB_val *key, int modify);
1053 #define MDB_PS_MODIFY 1
1054 #define MDB_PS_ROOTONLY 2
1055 static int mdb_page_search(MDB_cursor *mc,
1056 MDB_val *key, int flags);
1057 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1059 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1060 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1061 pgno_t newpgno, unsigned int nflags);
1063 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1064 static int mdb_env_pick_meta(const MDB_env *env);
1065 static int mdb_env_write_meta(MDB_txn *txn);
1066 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1067 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1069 static void mdb_env_close0(MDB_env *env, int excl);
1071 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1072 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1073 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1074 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1075 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1076 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1077 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1078 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1079 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1081 static int mdb_rebalance(MDB_cursor *mc);
1082 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1084 static void mdb_cursor_pop(MDB_cursor *mc);
1085 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1087 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1088 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1089 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1090 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1091 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1093 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1094 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1096 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1097 static void mdb_xcursor_init0(MDB_cursor *mc);
1098 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1100 static int mdb_drop0(MDB_cursor *mc, int subs);
1101 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1104 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1108 static SECURITY_DESCRIPTOR mdb_null_sd;
1109 static SECURITY_ATTRIBUTES mdb_all_sa;
1110 static int mdb_sec_inited;
1113 /** Return the library version info. */
1115 mdb_version(int *major, int *minor, int *patch)
1117 if (major) *major = MDB_VERSION_MAJOR;
1118 if (minor) *minor = MDB_VERSION_MINOR;
1119 if (patch) *patch = MDB_VERSION_PATCH;
1120 return MDB_VERSION_STRING;
1123 /** Table of descriptions for MDB @ref errors */
1124 static char *const mdb_errstr[] = {
1125 "MDB_KEYEXIST: Key/data pair already exists",
1126 "MDB_NOTFOUND: No matching key/data pair found",
1127 "MDB_PAGE_NOTFOUND: Requested page not found",
1128 "MDB_CORRUPTED: Located page was wrong type",
1129 "MDB_PANIC: Update of meta page failed",
1130 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1131 "MDB_INVALID: File is not an MDB file",
1132 "MDB_MAP_FULL: Environment mapsize limit reached",
1133 "MDB_DBS_FULL: Environment maxdbs limit reached",
1134 "MDB_READERS_FULL: Environment maxreaders limit reached",
1135 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1136 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1137 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1138 "MDB_PAGE_FULL: Internal error - page has no more space",
1139 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1140 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1141 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1142 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1143 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1147 mdb_strerror(int err)
1151 return ("Successful return: 0");
1153 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1154 i = err - MDB_KEYEXIST;
1155 return mdb_errstr[i];
1158 return strerror(err);
1162 /** Display a key in hexadecimal and return the address of the result.
1163 * @param[in] key the key to display
1164 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1165 * @return The key in hexadecimal form.
1168 mdb_dkey(MDB_val *key, char *buf)
1171 unsigned char *c = key->mv_data;
1177 if (key->mv_size > MDB_MAXKEYSIZE)
1178 return "MDB_MAXKEYSIZE";
1179 /* may want to make this a dynamic check: if the key is mostly
1180 * printable characters, print it as-is instead of converting to hex.
1184 for (i=0; i<key->mv_size; i++)
1185 ptr += sprintf(ptr, "%02x", *c++);
1187 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1192 /** Display all the keys in the page. */
1194 mdb_page_list(MDB_page *mp)
1197 unsigned int i, nkeys, nsize;
1201 nkeys = NUMKEYS(mp);
1202 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1203 for (i=0; i<nkeys; i++) {
1204 node = NODEPTR(mp, i);
1205 key.mv_size = node->mn_ksize;
1206 key.mv_data = node->mn_data;
1207 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1208 if (IS_BRANCH(mp)) {
1209 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1212 if (F_ISSET(node->mn_flags, F_BIGDATA))
1213 nsize += sizeof(pgno_t);
1215 nsize += NODEDSZ(node);
1216 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1222 mdb_cursor_chk(MDB_cursor *mc)
1228 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1229 for (i=0; i<mc->mc_top; i++) {
1231 node = NODEPTR(mp, mc->mc_ki[i]);
1232 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1235 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1241 /** Count all the pages in each DB and in the freelist
1242 * and make sure it matches the actual number of pages
1245 static void mdb_audit(MDB_txn *txn)
1249 MDB_ID freecount, count;
1254 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1255 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1256 freecount += *(MDB_ID *)data.mv_data;
1259 for (i = 0; i<txn->mt_numdbs; i++) {
1261 mdb_cursor_init(&mc, txn, i, &mx);
1262 if (txn->mt_dbs[i].md_root == P_INVALID)
1264 count += txn->mt_dbs[i].md_branch_pages +
1265 txn->mt_dbs[i].md_leaf_pages +
1266 txn->mt_dbs[i].md_overflow_pages;
1267 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1268 mdb_page_search(&mc, NULL, 0);
1272 mp = mc.mc_pg[mc.mc_top];
1273 for (j=0; j<NUMKEYS(mp); j++) {
1274 MDB_node *leaf = NODEPTR(mp, j);
1275 if (leaf->mn_flags & F_SUBDATA) {
1277 memcpy(&db, NODEDATA(leaf), sizeof(db));
1278 count += db.md_branch_pages + db.md_leaf_pages +
1279 db.md_overflow_pages;
1283 while (mdb_cursor_sibling(&mc, 1) == 0);
1286 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1287 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1288 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1294 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1296 return txn->mt_dbxs[dbi].md_cmp(a, b);
1300 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1302 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1305 /** Allocate memory for a page.
1306 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1309 mdb_page_malloc(MDB_txn *txn, unsigned num)
1311 MDB_env *env = txn->mt_env;
1312 MDB_page *ret = env->me_dpages;
1313 size_t sz = env->me_psize;
1316 VGMEMP_ALLOC(env, ret, sz);
1317 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1318 env->me_dpages = ret->mp_next;
1324 if ((ret = malloc(sz)) != NULL) {
1325 VGMEMP_ALLOC(env, ret, sz);
1330 /** Free a single page.
1331 * Saves single pages to a list, for future reuse.
1332 * (This is not used for multi-page overflow pages.)
1335 mdb_page_free(MDB_env *env, MDB_page *mp)
1337 mp->mp_next = env->me_dpages;
1338 VGMEMP_FREE(env, mp);
1339 env->me_dpages = mp;
1342 /** Free a dirty page */
1344 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1346 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1347 mdb_page_free(env, dp);
1349 /* large pages just get freed directly */
1350 VGMEMP_FREE(env, dp);
1355 /** Return all dirty pages to dpage list */
1357 mdb_dlist_free(MDB_txn *txn)
1359 MDB_env *env = txn->mt_env;
1360 MDB_ID2L dl = txn->mt_u.dirty_list;
1361 unsigned i, n = dl[0].mid;
1363 for (i = 1; i <= n; i++) {
1364 mdb_dpage_free(env, dl[i].mptr);
1369 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1370 * @param[in] mc A cursor handle for the current operation.
1371 * @param[in] pflags Flags of the pages to update:
1372 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1373 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1374 * @return 0 on success, non-zero on failure.
1377 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1379 MDB_txn *txn = mc->mc_txn;
1384 int rc = MDB_SUCCESS, level;
1386 /* Mark pages seen by cursors */
1387 if (mc->mc_flags & C_UNTRACK)
1388 mc = NULL; /* will find mc in mt_cursors */
1389 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1390 for (; mc; mc=mc->mc_next) {
1391 for (m3 = mc; m3->mc_flags & C_INITIALIZED; m3 = &mx->mx_cursor) {
1392 for (j=0; j<m3->mc_snum; j++)
1393 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1395 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1396 mx = m3->mc_xcursor;
1406 /* Mark dirty root pages */
1407 for (i=0; i<txn->mt_numdbs; i++) {
1408 if (txn->mt_dbflags[i] & DB_DIRTY) {
1409 pgno_t pgno = txn->mt_dbs[i].md_root;
1410 if (pgno == P_INVALID)
1412 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1414 if ((dp->mp_flags & (P_DIRTY|P_KEEP)) == pflags && level <= 1)
1415 dp->mp_flags ^= P_KEEP;
1423 static int mdb_page_flush(MDB_txn *txn, int keep);
1425 /** Spill pages from the dirty list back to disk.
1426 * This is intended to prevent running into #MDB_TXN_FULL situations,
1427 * but note that they may still occur in a few cases:
1428 * 1) our estimate of the txn size could be too small. Currently this
1429 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1430 * 2) child txns may run out of space if their parents dirtied a
1431 * lot of pages and never spilled them. TODO: we probably should do
1432 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1433 * the parent's dirty_room is below a given threshold.
1435 * Otherwise, if not using nested txns, it is expected that apps will
1436 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1437 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1438 * If the txn never references them again, they can be left alone.
1439 * If the txn only reads them, they can be used without any fuss.
1440 * If the txn writes them again, they can be dirtied immediately without
1441 * going thru all of the work of #mdb_page_touch(). Such references are
1442 * handled by #mdb_page_unspill().
1444 * Also note, we never spill DB root pages, nor pages of active cursors,
1445 * because we'll need these back again soon anyway. And in nested txns,
1446 * we can't spill a page in a child txn if it was already spilled in a
1447 * parent txn. That would alter the parent txns' data even though
1448 * the child hasn't committed yet, and we'd have no way to undo it if
1449 * the child aborted.
1451 * @param[in] m0 cursor A cursor handle identifying the transaction and
1452 * database for which we are checking space.
1453 * @param[in] key For a put operation, the key being stored.
1454 * @param[in] data For a put operation, the data being stored.
1455 * @return 0 on success, non-zero on failure.
1458 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1460 MDB_txn *txn = m0->mc_txn;
1462 MDB_ID2L dl = txn->mt_u.dirty_list;
1463 unsigned int i, j, need;
1466 if (m0->mc_flags & C_SUB)
1469 /* Estimate how much space this op will take */
1470 i = m0->mc_db->md_depth;
1471 /* Named DBs also dirty the main DB */
1472 if (m0->mc_dbi > MAIN_DBI)
1473 i += txn->mt_dbs[MAIN_DBI].md_depth;
1474 /* For puts, roughly factor in the key+data size */
1476 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1477 i += i; /* double it for good measure */
1480 if (txn->mt_dirty_room > i)
1483 if (!txn->mt_spill_pgs) {
1484 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1485 if (!txn->mt_spill_pgs)
1488 /* purge deleted slots */
1489 MDB_IDL sl = txn->mt_spill_pgs;
1490 unsigned int num = sl[0];
1492 for (i=1; i<=num; i++) {
1499 /* Preserve pages which may soon be dirtied again */
1500 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1503 /* Less aggressive spill - we originally spilled the entire dirty list,
1504 * with a few exceptions for cursor pages and DB root pages. But this
1505 * turns out to be a lot of wasted effort because in a large txn many
1506 * of those pages will need to be used again. So now we spill only 1/8th
1507 * of the dirty pages. Testing revealed this to be a good tradeoff,
1508 * better than 1/2, 1/4, or 1/10.
1510 if (need < MDB_IDL_UM_MAX / 8)
1511 need = MDB_IDL_UM_MAX / 8;
1513 /* Save the page IDs of all the pages we're flushing */
1514 /* flush from the tail forward, this saves a lot of shifting later on. */
1515 for (i=dl[0].mid; i && need; i--) {
1516 MDB_ID pn = dl[i].mid << 1;
1518 if (dp->mp_flags & P_KEEP)
1520 /* Can't spill twice, make sure it's not already in a parent's
1523 if (txn->mt_parent) {
1525 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1526 if (tx2->mt_spill_pgs) {
1527 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1528 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1529 dp->mp_flags |= P_KEEP;
1537 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1541 mdb_midl_sort(txn->mt_spill_pgs);
1543 /* Flush the spilled part of dirty list */
1544 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1547 /* Reset any dirty pages we kept that page_flush didn't see */
1548 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1552 if (txn->mt_parent) {
1553 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1554 /* dirty pages that are dirty in an ancestor don't
1555 * count against this txn's dirty_room.
1557 for (i=1; i<=dl[0].mid; i++) {
1558 pgno_t pgno = dl[i].mid;
1560 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1561 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1562 if (j <= tx2->mt_u.dirty_list[0].mid &&
1563 tx2->mt_u.dirty_list[j].mid == pgno) {
1564 txn->mt_dirty_room++;
1570 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1572 txn->mt_flags |= MDB_TXN_SPILLS;
1574 txn->mt_flags |= MDB_TXN_ERROR;
1579 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1581 mdb_find_oldest(MDB_txn *txn)
1584 txnid_t mr, oldest = txn->mt_txnid - 1;
1585 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1586 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1596 /** Add a page to the txn's dirty list */
1598 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1601 int (*insert)(MDB_ID2L, MDB_ID2 *);
1603 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1604 insert = mdb_mid2l_append;
1606 insert = mdb_mid2l_insert;
1608 mid.mid = mp->mp_pgno;
1610 insert(txn->mt_u.dirty_list, &mid);
1611 txn->mt_dirty_room--;
1614 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1615 * me_pghead and mt_next_pgno.
1617 * If there are free pages available from older transactions, they
1618 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1619 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1620 * and move me_pglast to say which records were consumed. Only this
1621 * function can create me_pghead and move me_pglast/mt_next_pgno.
1622 * @param[in] mc cursor A cursor handle identifying the transaction and
1623 * database for which we are allocating.
1624 * @param[in] num the number of pages to allocate.
1625 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1626 * will always be satisfied by a single contiguous chunk of memory.
1627 * @return 0 on success, non-zero on failure.
1630 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1632 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1633 /* Get at most <Max_retries> more freeDB records once me_pghead
1634 * has enough pages. If not enough, use new pages from the map.
1635 * If <Paranoid> and mc is updating the freeDB, only get new
1636 * records if me_pghead is empty. Then the freelist cannot play
1637 * catch-up with itself by growing while trying to save it.
1639 enum { Paranoid = 1, Max_retries = 500 };
1641 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1643 int rc, n2 = num-1, retry = Max_retries;
1644 MDB_txn *txn = mc->mc_txn;
1645 MDB_env *env = txn->mt_env;
1646 pgno_t pgno, *mop = env->me_pghead;
1647 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1649 txnid_t oldest = 0, last;
1655 /* If our dirty list is already full, we can't do anything */
1656 if (txn->mt_dirty_room == 0)
1657 return MDB_TXN_FULL;
1659 for (op = MDB_FIRST;; op = MDB_NEXT) {
1662 pgno_t *idl, old_id, new_id;
1664 /* Seek a big enough contiguous page range. Prefer
1665 * pages at the tail, just truncating the list.
1667 if (mop_len >= (unsigned)num) {
1671 if (mop[i-n2] == pgno+n2)
1673 } while (--i >= (unsigned)num);
1674 if (Max_retries < INT_MAX && --retry < 0)
1678 if (op == MDB_FIRST) { /* 1st iteration */
1679 /* Prepare to fetch more and coalesce */
1680 oldest = mdb_find_oldest(txn);
1681 last = env->me_pglast;
1682 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1685 key.mv_data = &last; /* will look up last+1 */
1686 key.mv_size = sizeof(last);
1688 if (Paranoid && mc->mc_dbi == FREE_DBI)
1691 if (Paranoid && retry < 0 && mop_len)
1695 /* Do not fetch more if the record will be too recent */
1698 rc = mdb_cursor_get(&m2, &key, NULL, op);
1700 if (rc == MDB_NOTFOUND)
1704 last = *(txnid_t*)key.mv_data;
1707 np = m2.mc_pg[m2.mc_top];
1708 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1709 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1712 idl = (MDB_ID *) data.mv_data;
1715 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1718 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1720 mop = env->me_pghead;
1722 env->me_pglast = last;
1724 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1725 last, txn->mt_dbs[FREE_DBI].md_root, i));
1727 DPRINTF(("IDL %"Z"u", idl[k]));
1729 /* Merge in descending sorted order */
1732 mop[0] = (pgno_t)-1;
1736 for (; old_id < new_id; old_id = mop[--j])
1743 /* Use new pages from the map when nothing suitable in the freeDB */
1745 pgno = txn->mt_next_pgno;
1746 if (pgno + num >= env->me_maxpg) {
1747 DPUTS("DB size maxed out");
1748 return MDB_MAP_FULL;
1752 if (env->me_flags & MDB_WRITEMAP) {
1753 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1755 if (!(np = mdb_page_malloc(txn, num)))
1759 mop[0] = mop_len -= num;
1760 /* Move any stragglers down */
1761 for (j = i-num; j < mop_len; )
1762 mop[++j] = mop[++i];
1764 txn->mt_next_pgno = pgno + num;
1767 mdb_page_dirty(txn, np);
1773 /** Copy the used portions of a non-overflow page.
1774 * @param[in] dst page to copy into
1775 * @param[in] src page to copy from
1776 * @param[in] psize size of a page
1779 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1781 enum { Align = sizeof(pgno_t) };
1782 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1784 /* If page isn't full, just copy the used portion. Adjust
1785 * alignment so memcpy may copy words instead of bytes.
1787 if ((unused &= -Align) && !IS_LEAF2(src)) {
1789 memcpy(dst, src, (lower + (Align-1)) & -Align);
1790 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1793 memcpy(dst, src, psize - unused);
1797 /** Pull a page off the txn's spill list, if present.
1798 * If a page being referenced was spilled to disk in this txn, bring
1799 * it back and make it dirty/writable again.
1800 * @param[in] tx0 the transaction handle.
1801 * @param[in] mp the page being referenced.
1802 * @param[out] ret the writable page, if any. ret is unchanged if
1803 * mp wasn't spilled.
1806 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1808 MDB_env *env = tx0->mt_env;
1811 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1813 for (txn = tx0; txn; txn=txn->mt_parent) {
1814 if (!txn->mt_spill_pgs)
1816 x = mdb_midl_search(txn->mt_spill_pgs, pn);
1817 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pn) {
1820 if (IS_OVERFLOW(mp))
1824 if (env->me_flags & MDB_WRITEMAP) {
1827 np = mdb_page_malloc(txn, num);
1831 memcpy(np, mp, num * env->me_psize);
1833 mdb_page_copy(np, mp, env->me_psize);
1836 /* If in current txn, this page is no longer spilled.
1837 * If it happens to be the last page, truncate the spill list.
1838 * Otherwise mark it as deleted by setting the LSB.
1840 if (x == txn->mt_spill_pgs[0])
1841 txn->mt_spill_pgs[0]--;
1843 txn->mt_spill_pgs[x] |= 1;
1844 } /* otherwise, if belonging to a parent txn, the
1845 * page remains spilled until child commits
1848 if (txn->mt_parent) {
1850 /* If this page is also in a parent's dirty list, then
1851 * it's already accounted in dirty_room, and we need to
1852 * cancel out the decrement that mdb_page_dirty does.
1854 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1855 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1856 if (x <= tx2->mt_u.dirty_list[0].mid &&
1857 tx2->mt_u.dirty_list[x].mid == pgno) {
1858 tx0->mt_dirty_room++;
1863 mdb_page_dirty(tx0, np);
1864 np->mp_flags |= P_DIRTY;
1872 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1873 * @param[in] mc cursor pointing to the page to be touched
1874 * @return 0 on success, non-zero on failure.
1877 mdb_page_touch(MDB_cursor *mc)
1879 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1880 MDB_txn *txn = mc->mc_txn;
1881 MDB_cursor *m2, *m3;
1886 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1887 if (txn->mt_flags & MDB_TXN_SPILLS) {
1889 rc = mdb_page_unspill(txn, mp, &np);
1895 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1896 (rc = mdb_page_alloc(mc, 1, &np)))
1899 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1900 assert(mp->mp_pgno != pgno);
1901 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1902 /* Update the parent page, if any, to point to the new page */
1904 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1905 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1906 SETPGNO(node, pgno);
1908 mc->mc_db->md_root = pgno;
1910 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1911 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1913 /* If txn has a parent, make sure the page is in our
1917 unsigned x = mdb_mid2l_search(dl, pgno);
1918 if (x <= dl[0].mid && dl[x].mid == pgno) {
1919 if (mp != dl[x].mptr) { /* bad cursor? */
1920 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1921 return MDB_CORRUPTED;
1926 assert(dl[0].mid < MDB_IDL_UM_MAX);
1928 np = mdb_page_malloc(txn, 1);
1933 mdb_mid2l_insert(dl, &mid);
1938 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1940 np->mp_flags |= P_DIRTY;
1943 /* Adjust cursors pointing to mp */
1944 mc->mc_pg[mc->mc_top] = np;
1946 if (mc->mc_flags & C_SUB) {
1948 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1949 m3 = &m2->mc_xcursor->mx_cursor;
1950 if (m3->mc_snum < mc->mc_snum) continue;
1951 if (m3->mc_pg[mc->mc_top] == mp)
1952 m3->mc_pg[mc->mc_top] = np;
1955 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1956 if (m2->mc_snum < mc->mc_snum) continue;
1957 if (m2->mc_pg[mc->mc_top] == mp) {
1958 m2->mc_pg[mc->mc_top] = np;
1959 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1960 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1962 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1963 if (!(leaf->mn_flags & F_SUBDATA))
1964 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1973 mdb_env_sync(MDB_env *env, int force)
1976 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1977 if (env->me_flags & MDB_WRITEMAP) {
1978 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1979 ? MS_ASYNC : MS_SYNC;
1980 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1983 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1987 if (MDB_FDATASYNC(env->me_fd))
1994 /** Back up parent txn's cursors, then grab the originals for tracking */
1996 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1998 MDB_cursor *mc, *bk;
2003 for (i = src->mt_numdbs; --i >= 0; ) {
2004 if ((mc = src->mt_cursors[i]) != NULL) {
2005 size = sizeof(MDB_cursor);
2007 size += sizeof(MDB_xcursor);
2008 for (; mc; mc = bk->mc_next) {
2014 mc->mc_db = &dst->mt_dbs[i];
2015 /* Kill pointers into src - and dst to reduce abuse: The
2016 * user may not use mc until dst ends. Otherwise we'd...
2018 mc->mc_txn = NULL; /* ...set this to dst */
2019 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2020 if ((mx = mc->mc_xcursor) != NULL) {
2021 *(MDB_xcursor *)(bk+1) = *mx;
2022 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2024 mc->mc_next = dst->mt_cursors[i];
2025 dst->mt_cursors[i] = mc;
2032 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2033 * @param[in] txn the transaction handle.
2034 * @param[in] merge true to keep changes to parent cursors, false to revert.
2035 * @return 0 on success, non-zero on failure.
2038 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2040 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2044 for (i = txn->mt_numdbs; --i >= 0; ) {
2045 for (mc = cursors[i]; mc; mc = next) {
2047 if ((bk = mc->mc_backup) != NULL) {
2049 /* Commit changes to parent txn */
2050 mc->mc_next = bk->mc_next;
2051 mc->mc_backup = bk->mc_backup;
2052 mc->mc_txn = bk->mc_txn;
2053 mc->mc_db = bk->mc_db;
2054 mc->mc_dbflag = bk->mc_dbflag;
2055 if ((mx = mc->mc_xcursor) != NULL)
2056 mx->mx_cursor.mc_txn = bk->mc_txn;
2058 /* Abort nested txn */
2060 if ((mx = mc->mc_xcursor) != NULL)
2061 *mx = *(MDB_xcursor *)(bk+1);
2065 /* Only malloced cursors are permanently tracked. */
2073 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2076 mdb_txn_reset0(MDB_txn *txn, const char *act);
2078 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2084 Pidset = F_SETLK, Pidcheck = F_GETLK
2088 /** Set or check a pid lock. Set returns 0 on success.
2089 * Check returns 0 if the process is certainly dead, nonzero if it may
2090 * be alive (the lock exists or an error happened so we do not know).
2092 * On Windows Pidset is a no-op, we merely check for the existence
2093 * of the process with the given pid. On POSIX we use a single byte
2094 * lock on the lockfile, set at an offset equal to the pid.
2097 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2099 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2102 if (op == Pidcheck) {
2103 h = OpenProcess(env->me_pidquery, FALSE, pid);
2104 /* No documented "no such process" code, but other program use this: */
2106 return ErrCode() != ERROR_INVALID_PARAMETER;
2107 /* A process exists until all handles to it close. Has it exited? */
2108 ret = WaitForSingleObject(h, 0) != 0;
2115 struct flock lock_info;
2116 memset(&lock_info, 0, sizeof(lock_info));
2117 lock_info.l_type = F_WRLCK;
2118 lock_info.l_whence = SEEK_SET;
2119 lock_info.l_start = pid;
2120 lock_info.l_len = 1;
2121 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2122 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2124 } else if ((rc = ErrCode()) == EINTR) {
2132 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2133 * @param[in] txn the transaction handle to initialize
2134 * @return 0 on success, non-zero on failure.
2137 mdb_txn_renew0(MDB_txn *txn)
2139 MDB_env *env = txn->mt_env;
2142 int rc, new_notls = 0;
2145 txn->mt_numdbs = env->me_numdbs;
2146 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2148 if (txn->mt_flags & MDB_TXN_RDONLY) {
2149 if (!env->me_txns) {
2150 i = mdb_env_pick_meta(env);
2151 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2152 txn->mt_u.reader = NULL;
2154 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2155 pthread_getspecific(env->me_txkey);
2157 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2158 return MDB_BAD_RSLOT;
2160 pid_t pid = env->me_pid;
2161 pthread_t tid = pthread_self();
2163 if (!(env->me_flags & MDB_LIVE_READER)) {
2164 rc = mdb_reader_pid(env, Pidset, pid);
2166 UNLOCK_MUTEX_R(env);
2169 env->me_flags |= MDB_LIVE_READER;
2173 for (i=0; i<env->me_txns->mti_numreaders; i++)
2174 if (env->me_txns->mti_readers[i].mr_pid == 0)
2176 if (i == env->me_maxreaders) {
2177 UNLOCK_MUTEX_R(env);
2178 return MDB_READERS_FULL;
2180 env->me_txns->mti_readers[i].mr_pid = pid;
2181 env->me_txns->mti_readers[i].mr_tid = tid;
2182 if (i >= env->me_txns->mti_numreaders)
2183 env->me_txns->mti_numreaders = i+1;
2184 /* Save numreaders for un-mutexed mdb_env_close() */
2185 env->me_numreaders = env->me_txns->mti_numreaders;
2186 UNLOCK_MUTEX_R(env);
2187 r = &env->me_txns->mti_readers[i];
2188 new_notls = (env->me_flags & MDB_NOTLS);
2189 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2194 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2195 txn->mt_u.reader = r;
2197 txn->mt_toggle = txn->mt_txnid & 1;
2201 txn->mt_txnid = env->me_txns->mti_txnid;
2202 txn->mt_toggle = txn->mt_txnid & 1;
2205 if (txn->mt_txnid == mdb_debug_start)
2208 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2209 txn->mt_u.dirty_list = env->me_dirty_list;
2210 txn->mt_u.dirty_list[0].mid = 0;
2211 txn->mt_free_pgs = env->me_free_pgs;
2212 txn->mt_free_pgs[0] = 0;
2213 txn->mt_spill_pgs = NULL;
2217 /* Copy the DB info and flags */
2218 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2220 /* Moved to here to avoid a data race in read TXNs */
2221 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2223 for (i=2; i<txn->mt_numdbs; i++) {
2224 x = env->me_dbflags[i];
2225 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2226 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2228 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2230 if (env->me_maxpg < txn->mt_next_pgno) {
2231 mdb_txn_reset0(txn, "renew0-mapfail");
2233 txn->mt_u.reader->mr_pid = 0;
2234 txn->mt_u.reader = NULL;
2236 return MDB_MAP_RESIZED;
2243 mdb_txn_renew(MDB_txn *txn)
2247 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2250 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2251 DPUTS("environment had fatal error, must shutdown!");
2255 rc = mdb_txn_renew0(txn);
2256 if (rc == MDB_SUCCESS) {
2257 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2258 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2259 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2265 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2269 int rc, size, tsize = sizeof(MDB_txn);
2271 if (env->me_flags & MDB_FATAL_ERROR) {
2272 DPUTS("environment had fatal error, must shutdown!");
2275 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2278 /* Nested transactions: Max 1 child, write txns only, no writemap */
2279 if (parent->mt_child ||
2280 (flags & MDB_RDONLY) ||
2281 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2282 (env->me_flags & MDB_WRITEMAP))
2284 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2286 tsize = sizeof(MDB_ntxn);
2288 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2289 if (!(flags & MDB_RDONLY))
2290 size += env->me_maxdbs * sizeof(MDB_cursor *);
2292 if ((txn = calloc(1, size)) == NULL) {
2293 DPRINTF(("calloc: %s", strerror(ErrCode())));
2296 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2297 if (flags & MDB_RDONLY) {
2298 txn->mt_flags |= MDB_TXN_RDONLY;
2299 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2301 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2302 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2308 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2309 if (!txn->mt_u.dirty_list ||
2310 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2312 free(txn->mt_u.dirty_list);
2316 txn->mt_txnid = parent->mt_txnid;
2317 txn->mt_toggle = parent->mt_toggle;
2318 txn->mt_dirty_room = parent->mt_dirty_room;
2319 txn->mt_u.dirty_list[0].mid = 0;
2320 txn->mt_spill_pgs = NULL;
2321 txn->mt_next_pgno = parent->mt_next_pgno;
2322 parent->mt_child = txn;
2323 txn->mt_parent = parent;
2324 txn->mt_numdbs = parent->mt_numdbs;
2325 txn->mt_flags = parent->mt_flags;
2326 txn->mt_dbxs = parent->mt_dbxs;
2327 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2328 /* Copy parent's mt_dbflags, but clear DB_NEW */
2329 for (i=0; i<txn->mt_numdbs; i++)
2330 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2332 ntxn = (MDB_ntxn *)txn;
2333 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2334 if (env->me_pghead) {
2335 size = MDB_IDL_SIZEOF(env->me_pghead);
2336 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2338 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2343 rc = mdb_cursor_shadow(parent, txn);
2345 mdb_txn_reset0(txn, "beginchild-fail");
2347 rc = mdb_txn_renew0(txn);
2353 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2354 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2355 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2362 mdb_txn_env(MDB_txn *txn)
2364 if(!txn) return NULL;
2368 /** Export or close DBI handles opened in this txn. */
2370 mdb_dbis_update(MDB_txn *txn, int keep)
2373 MDB_dbi n = txn->mt_numdbs;
2374 MDB_env *env = txn->mt_env;
2375 unsigned char *tdbflags = txn->mt_dbflags;
2377 for (i = n; --i >= 2;) {
2378 if (tdbflags[i] & DB_NEW) {
2380 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2382 char *ptr = env->me_dbxs[i].md_name.mv_data;
2383 env->me_dbxs[i].md_name.mv_data = NULL;
2384 env->me_dbxs[i].md_name.mv_size = 0;
2385 env->me_dbflags[i] = 0;
2390 if (keep && env->me_numdbs < n)
2394 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2395 * May be called twice for readonly txns: First reset it, then abort.
2396 * @param[in] txn the transaction handle to reset
2397 * @param[in] act why the transaction is being reset
2400 mdb_txn_reset0(MDB_txn *txn, const char *act)
2402 MDB_env *env = txn->mt_env;
2404 /* Close any DBI handles opened in this txn */
2405 mdb_dbis_update(txn, 0);
2407 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2408 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2409 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2411 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2412 if (txn->mt_u.reader) {
2413 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2414 if (!(env->me_flags & MDB_NOTLS))
2415 txn->mt_u.reader = NULL; /* txn does not own reader */
2417 txn->mt_numdbs = 0; /* close nothing if called again */
2418 txn->mt_dbxs = NULL; /* mark txn as reset */
2420 mdb_cursors_close(txn, 0);
2422 if (!(env->me_flags & MDB_WRITEMAP)) {
2423 mdb_dlist_free(txn);
2425 mdb_midl_free(env->me_pghead);
2427 if (txn->mt_parent) {
2428 txn->mt_parent->mt_child = NULL;
2429 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2430 mdb_midl_free(txn->mt_free_pgs);
2431 mdb_midl_free(txn->mt_spill_pgs);
2432 free(txn->mt_u.dirty_list);
2436 if (mdb_midl_shrink(&txn->mt_free_pgs))
2437 env->me_free_pgs = txn->mt_free_pgs;
2438 env->me_pghead = NULL;
2442 /* The writer mutex was locked in mdb_txn_begin. */
2443 UNLOCK_MUTEX_W(env);
2448 mdb_txn_reset(MDB_txn *txn)
2453 /* This call is only valid for read-only txns */
2454 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2457 mdb_txn_reset0(txn, "reset");
2461 mdb_txn_abort(MDB_txn *txn)
2467 mdb_txn_abort(txn->mt_child);
2469 mdb_txn_reset0(txn, "abort");
2470 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2471 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2472 txn->mt_u.reader->mr_pid = 0;
2477 /** Save the freelist as of this transaction to the freeDB.
2478 * This changes the freelist. Keep trying until it stabilizes.
2481 mdb_freelist_save(MDB_txn *txn)
2483 /* env->me_pghead[] can grow and shrink during this call.
2484 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2485 * Page numbers cannot disappear from txn->mt_free_pgs[].
2488 MDB_env *env = txn->mt_env;
2489 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2490 txnid_t pglast = 0, head_id = 0;
2491 pgno_t freecnt = 0, *free_pgs, *mop;
2492 ssize_t head_room = 0, total_room = 0, mop_len;
2494 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2496 if (env->me_pghead) {
2497 /* Make sure first page of freeDB is touched and on freelist */
2498 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2499 if (rc && rc != MDB_NOTFOUND)
2504 /* Come back here after each Put() in case freelist changed */
2507 /* If using records from freeDB which we have not yet
2508 * deleted, delete them and any we reserved for me_pghead.
2510 while (pglast < env->me_pglast) {
2511 rc = mdb_cursor_first(&mc, &key, NULL);
2514 pglast = head_id = *(txnid_t *)key.mv_data;
2515 total_room = head_room = 0;
2516 assert(pglast <= env->me_pglast);
2517 rc = mdb_cursor_del(&mc, 0);
2522 /* Save the IDL of pages freed by this txn, to a single record */
2523 if (freecnt < txn->mt_free_pgs[0]) {
2525 /* Make sure last page of freeDB is touched and on freelist */
2526 key.mv_size = MDB_MAXKEYSIZE+1;
2528 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2529 if (rc && rc != MDB_NOTFOUND)
2532 free_pgs = txn->mt_free_pgs;
2533 /* Write to last page of freeDB */
2534 key.mv_size = sizeof(txn->mt_txnid);
2535 key.mv_data = &txn->mt_txnid;
2537 freecnt = free_pgs[0];
2538 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2539 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2542 /* Retry if mt_free_pgs[] grew during the Put() */
2543 free_pgs = txn->mt_free_pgs;
2544 } while (freecnt < free_pgs[0]);
2545 mdb_midl_sort(free_pgs);
2546 memcpy(data.mv_data, free_pgs, data.mv_size);
2549 unsigned int i = free_pgs[0];
2550 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2551 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2553 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2559 mop = env->me_pghead;
2560 mop_len = mop ? mop[0] : 0;
2562 /* Reserve records for me_pghead[]. Split it if multi-page,
2563 * to avoid searching freeDB for a page range. Use keys in
2564 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2566 if (total_room >= mop_len) {
2567 if (total_room == mop_len || --more < 0)
2569 } else if (head_room >= maxfree_1pg && head_id > 1) {
2570 /* Keep current record (overflow page), add a new one */
2574 /* (Re)write {key = head_id, IDL length = head_room} */
2575 total_room -= head_room;
2576 head_room = mop_len - total_room;
2577 if (head_room > maxfree_1pg && head_id > 1) {
2578 /* Overflow multi-page for part of me_pghead */
2579 head_room /= head_id; /* amortize page sizes */
2580 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2581 } else if (head_room < 0) {
2582 /* Rare case, not bothering to delete this record */
2585 key.mv_size = sizeof(head_id);
2586 key.mv_data = &head_id;
2587 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2588 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2591 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2592 total_room += head_room;
2595 /* Fill in the reserved me_pghead records */
2601 rc = mdb_cursor_first(&mc, &key, &data);
2602 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2603 unsigned flags = MDB_CURRENT;
2604 txnid_t id = *(txnid_t *)key.mv_data;
2605 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2608 assert(len >= 0 && id <= env->me_pglast);
2610 if (len > mop_len) {
2612 data.mv_size = (len + 1) * sizeof(MDB_ID);
2615 data.mv_data = mop -= len;
2618 rc = mdb_cursor_put(&mc, &key, &data, flags);
2620 if (rc || !(mop_len -= len))
2627 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2628 * @param[in] txn the transaction that's being committed
2629 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2630 * @return 0 on success, non-zero on failure.
2633 mdb_page_flush(MDB_txn *txn, int keep)
2635 MDB_env *env = txn->mt_env;
2636 MDB_ID2L dl = txn->mt_u.dirty_list;
2637 unsigned psize = env->me_psize, j;
2638 int i, pagecount = dl[0].mid, rc;
2639 size_t size = 0, pos = 0;
2641 MDB_page *dp = NULL;
2645 struct iovec iov[MDB_COMMIT_PAGES];
2646 ssize_t wpos = 0, wsize = 0, wres;
2647 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2653 if (env->me_flags & MDB_WRITEMAP) {
2654 /* Clear dirty flags */
2655 while (++i <= pagecount) {
2657 /* Don't flush this page yet */
2658 if (dp->mp_flags & P_KEEP) {
2659 dp->mp_flags ^= P_KEEP;
2663 dp->mp_flags &= ~P_DIRTY;
2669 /* Write the pages */
2671 if (++i <= pagecount) {
2673 /* Don't flush this page yet */
2674 if (dp->mp_flags & P_KEEP) {
2675 dp->mp_flags ^= P_KEEP;
2680 /* clear dirty flag */
2681 dp->mp_flags &= ~P_DIRTY;
2684 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2689 /* Windows actually supports scatter/gather I/O, but only on
2690 * unbuffered file handles. Since we're relying on the OS page
2691 * cache for all our data, that's self-defeating. So we just
2692 * write pages one at a time. We use the ov structure to set
2693 * the write offset, to at least save the overhead of a Seek
2696 DPRINTF(("committing page %"Z"u", pgno));
2697 memset(&ov, 0, sizeof(ov));
2698 ov.Offset = pos & 0xffffffff;
2699 ov.OffsetHigh = pos >> 16 >> 16;
2700 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2702 DPRINTF(("WriteFile: %d", rc));
2706 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2707 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2709 /* Write previous page(s) */
2710 #ifdef MDB_USE_PWRITEV
2711 wres = pwritev(env->me_fd, iov, n, wpos);
2714 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2716 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2718 DPRINTF(("lseek: %s", strerror(rc)));
2721 wres = writev(env->me_fd, iov, n);
2724 if (wres != wsize) {
2727 DPRINTF(("Write error: %s", strerror(rc)));
2729 rc = EIO; /* TODO: Use which error code? */
2730 DPUTS("short write, filesystem full?");
2741 DPRINTF(("committing page %"Z"u", pgno));
2742 next_pos = pos + size;
2743 iov[n].iov_len = size;
2744 iov[n].iov_base = (char *)dp;
2750 for (i = keep; ++i <= pagecount; ) {
2752 /* This is a page we skipped above */
2755 dl[j].mid = dp->mp_pgno;
2758 mdb_dpage_free(env, dp);
2766 mdb_txn_commit(MDB_txn *txn)
2772 assert(txn != NULL);
2773 assert(txn->mt_env != NULL);
2775 if (txn->mt_child) {
2776 rc = mdb_txn_commit(txn->mt_child);
2777 txn->mt_child = NULL;
2784 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2785 mdb_dbis_update(txn, 1);
2786 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2791 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2792 DPUTS("error flag is set, can't commit");
2794 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2799 if (txn->mt_parent) {
2800 MDB_txn *parent = txn->mt_parent;
2804 /* Append our free list to parent's */
2805 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2808 mdb_midl_free(txn->mt_free_pgs);
2810 parent->mt_next_pgno = txn->mt_next_pgno;
2811 parent->mt_flags = txn->mt_flags;
2813 /* Merge our cursors into parent's and close them */
2814 mdb_cursors_close(txn, 1);
2816 /* Update parent's DB table. */
2817 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2818 parent->mt_numdbs = txn->mt_numdbs;
2819 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2820 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2821 for (i=2; i<txn->mt_numdbs; i++) {
2822 /* preserve parent's DB_NEW status */
2823 x = parent->mt_dbflags[i] & DB_NEW;
2824 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2827 dst = parent->mt_u.dirty_list;
2828 src = txn->mt_u.dirty_list;
2829 /* Remove anything in our dirty list from parent's spill list */
2830 if (parent->mt_spill_pgs) {
2831 x = parent->mt_spill_pgs[0];
2833 /* zero out our dirty pages in parent spill list */
2834 for (i=1; i<=src[0].mid; i++) {
2835 MDB_ID pn = src[i].mid << 1;
2836 if (pn < parent->mt_spill_pgs[x])
2838 if (pn > parent->mt_spill_pgs[x]) {
2844 parent->mt_spill_pgs[x] = 0;
2847 /* OK, we had a few hits, squash zeros from the spill list */
2848 if (len < parent->mt_spill_pgs[0]) {
2850 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2851 if (parent->mt_spill_pgs[y]) {
2853 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2858 parent->mt_spill_pgs[0] = len;
2861 /* Find len = length of merging our dirty list with parent's */
2863 dst[0].mid = 0; /* simplify loops */
2864 if (parent->mt_parent) {
2865 len = x + src[0].mid;
2866 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2867 for (i = x; y && i; y--) {
2868 pgno_t yp = src[y].mid;
2869 while (yp < dst[i].mid)
2871 if (yp == dst[i].mid) {
2876 } else { /* Simplify the above for single-ancestor case */
2877 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2879 /* Merge our dirty list with parent's */
2881 for (i = len; y; dst[i--] = src[y--]) {
2882 pgno_t yp = src[y].mid;
2883 while (yp < dst[x].mid)
2884 dst[i--] = dst[x--];
2885 if (yp == dst[x].mid)
2886 free(dst[x--].mptr);
2890 free(txn->mt_u.dirty_list);
2891 parent->mt_dirty_room = txn->mt_dirty_room;
2892 if (txn->mt_spill_pgs) {
2893 if (parent->mt_spill_pgs) {
2894 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2895 mdb_midl_free(txn->mt_spill_pgs);
2896 mdb_midl_sort(parent->mt_spill_pgs);
2898 parent->mt_spill_pgs = txn->mt_spill_pgs;
2902 parent->mt_child = NULL;
2903 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2908 if (txn != env->me_txn) {
2909 DPUTS("attempt to commit unknown transaction");
2914 mdb_cursors_close(txn, 0);
2916 if (!txn->mt_u.dirty_list[0].mid &&
2917 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2920 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2921 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2923 /* Update DB root pointers */
2924 if (txn->mt_numdbs > 2) {
2928 data.mv_size = sizeof(MDB_db);
2930 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2931 for (i = 2; i < txn->mt_numdbs; i++) {
2932 if (txn->mt_dbflags[i] & DB_DIRTY) {
2933 data.mv_data = &txn->mt_dbs[i];
2934 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2941 rc = mdb_freelist_save(txn);
2945 mdb_midl_free(env->me_pghead);
2946 env->me_pghead = NULL;
2947 if (mdb_midl_shrink(&txn->mt_free_pgs))
2948 env->me_free_pgs = txn->mt_free_pgs;
2954 if ((rc = mdb_page_flush(txn, 0)) ||
2955 (rc = mdb_env_sync(env, 0)) ||
2956 (rc = mdb_env_write_meta(txn)))
2962 mdb_dbis_update(txn, 1);
2964 UNLOCK_MUTEX_W(env);
2974 /** Read the environment parameters of a DB environment before
2975 * mapping it into memory.
2976 * @param[in] env the environment handle
2977 * @param[out] meta address of where to store the meta information
2978 * @return 0 on success, non-zero on failure.
2981 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2988 /* We don't know the page size yet, so use a minimum value.
2989 * Read both meta pages so we can use the latest one.
2992 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2996 memset(&ov, 0, sizeof(ov));
2998 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2999 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3002 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
3004 if (rc != MDB_PAGESIZE) {
3005 if (rc == 0 && off == 0)
3007 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3008 DPRINTF(("read: %s", mdb_strerror(rc)));
3012 p = (MDB_page *)&pbuf;
3014 if (!F_ISSET(p->mp_flags, P_META)) {
3015 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3020 if (m->mm_magic != MDB_MAGIC) {
3021 DPUTS("meta has invalid magic");
3025 if (m->mm_version != MDB_DATA_VERSION) {
3026 DPRINTF(("database is version %u, expected version %u",
3027 m->mm_version, MDB_DATA_VERSION));
3028 return MDB_VERSION_MISMATCH;
3031 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3037 /** Write the environment parameters of a freshly created DB environment.
3038 * @param[in] env the environment handle
3039 * @param[out] meta address of where to store the meta information
3040 * @return 0 on success, non-zero on failure.
3043 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3051 memset(&ov, 0, sizeof(ov));
3052 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3054 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3057 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3058 len = pwrite(fd, ptr, size, pos); \
3059 rc = (len >= 0); } while(0)
3062 DPUTS("writing new meta page");
3064 psize = env->me_psize;
3066 meta->mm_magic = MDB_MAGIC;
3067 meta->mm_version = MDB_DATA_VERSION;
3068 meta->mm_mapsize = env->me_mapsize;
3069 meta->mm_psize = psize;
3070 meta->mm_last_pg = 1;
3071 meta->mm_flags = env->me_flags & 0xffff;
3072 meta->mm_flags |= MDB_INTEGERKEY;
3073 meta->mm_dbs[0].md_root = P_INVALID;
3074 meta->mm_dbs[1].md_root = P_INVALID;
3076 p = calloc(2, psize);
3078 p->mp_flags = P_META;
3079 *(MDB_meta *)METADATA(p) = *meta;
3081 q = (MDB_page *)((char *)p + psize);
3083 q->mp_flags = P_META;
3084 *(MDB_meta *)METADATA(q) = *meta;
3086 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3089 else if ((unsigned) len == psize * 2)
3097 /** Update the environment info to commit a transaction.
3098 * @param[in] txn the transaction that's being committed
3099 * @return 0 on success, non-zero on failure.
3102 mdb_env_write_meta(MDB_txn *txn)
3105 MDB_meta meta, metab, *mp;
3107 int rc, len, toggle;
3116 assert(txn != NULL);
3117 assert(txn->mt_env != NULL);
3119 toggle = !txn->mt_toggle;
3120 DPRINTF(("writing meta page %d for root page %"Z"u",
3121 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3124 mp = env->me_metas[toggle];
3126 if (env->me_flags & MDB_WRITEMAP) {
3127 /* Persist any increases of mapsize config */
3128 if (env->me_mapsize > mp->mm_mapsize)
3129 mp->mm_mapsize = env->me_mapsize;
3130 mp->mm_dbs[0] = txn->mt_dbs[0];
3131 mp->mm_dbs[1] = txn->mt_dbs[1];
3132 mp->mm_last_pg = txn->mt_next_pgno - 1;
3133 mp->mm_txnid = txn->mt_txnid;
3134 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3135 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3138 ptr += env->me_psize;
3139 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3146 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3147 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3149 ptr = (char *)&meta;
3150 if (env->me_mapsize > mp->mm_mapsize) {
3151 /* Persist any increases of mapsize config */
3152 meta.mm_mapsize = env->me_mapsize;
3153 off = offsetof(MDB_meta, mm_mapsize);
3155 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3157 len = sizeof(MDB_meta) - off;
3160 meta.mm_dbs[0] = txn->mt_dbs[0];
3161 meta.mm_dbs[1] = txn->mt_dbs[1];
3162 meta.mm_last_pg = txn->mt_next_pgno - 1;
3163 meta.mm_txnid = txn->mt_txnid;
3166 off += env->me_psize;
3169 /* Write to the SYNC fd */
3170 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3171 env->me_fd : env->me_mfd;
3174 memset(&ov, 0, sizeof(ov));
3176 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3180 rc = pwrite(mfd, ptr, len, off);
3183 rc = rc < 0 ? ErrCode() : EIO;
3184 DPUTS("write failed, disk error?");
3185 /* On a failure, the pagecache still contains the new data.
3186 * Write some old data back, to prevent it from being used.
3187 * Use the non-SYNC fd; we know it will fail anyway.
3189 meta.mm_last_pg = metab.mm_last_pg;
3190 meta.mm_txnid = metab.mm_txnid;
3192 memset(&ov, 0, sizeof(ov));
3194 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3196 r2 = pwrite(env->me_fd, ptr, len, off);
3197 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3200 env->me_flags |= MDB_FATAL_ERROR;
3204 /* Memory ordering issues are irrelevant; since the entire writer
3205 * is wrapped by wmutex, all of these changes will become visible
3206 * after the wmutex is unlocked. Since the DB is multi-version,
3207 * readers will get consistent data regardless of how fresh or
3208 * how stale their view of these values is.
3210 env->me_txns->mti_txnid = txn->mt_txnid;
3215 /** Check both meta pages to see which one is newer.
3216 * @param[in] env the environment handle
3217 * @return meta toggle (0 or 1).
3220 mdb_env_pick_meta(const MDB_env *env)
3222 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3226 mdb_env_create(MDB_env **env)
3230 e = calloc(1, sizeof(MDB_env));
3234 e->me_maxreaders = DEFAULT_READERS;
3235 e->me_maxdbs = e->me_numdbs = 2;
3236 e->me_fd = INVALID_HANDLE_VALUE;
3237 e->me_lfd = INVALID_HANDLE_VALUE;
3238 e->me_mfd = INVALID_HANDLE_VALUE;
3239 #ifdef MDB_USE_POSIX_SEM
3240 e->me_rmutex = SEM_FAILED;
3241 e->me_wmutex = SEM_FAILED;
3243 e->me_pid = getpid();
3244 VGMEMP_CREATE(e,0,0);
3250 mdb_env_map(MDB_env *env, void *addr, int newsize)
3253 unsigned int flags = env->me_flags;
3257 LONG sizelo, sizehi;
3258 sizelo = env->me_mapsize & 0xffffffff;
3259 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3261 /* Windows won't create mappings for zero length files.
3262 * Just allocate the maxsize right now.
3265 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3266 || !SetEndOfFile(env->me_fd)
3267 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3270 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3271 PAGE_READWRITE : PAGE_READONLY,
3272 sizehi, sizelo, NULL);
3275 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3276 FILE_MAP_WRITE : FILE_MAP_READ,
3277 0, 0, env->me_mapsize, addr);
3278 rc = env->me_map ? 0 : ErrCode();
3283 int prot = PROT_READ;
3284 if (flags & MDB_WRITEMAP) {
3286 if (newsize && ftruncate(env->me_fd, env->me_mapsize) < 0)
3289 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3291 if (env->me_map == MAP_FAILED) {
3295 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3297 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3299 #ifdef POSIX_MADV_RANDOM
3300 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3301 #endif /* POSIX_MADV_RANDOM */
3302 #endif /* MADV_RANDOM */
3305 /* Can happen because the address argument to mmap() is just a
3306 * hint. mmap() can pick another, e.g. if the range is in use.
3307 * The MAP_FIXED flag would prevent that, but then mmap could
3308 * instead unmap existing pages to make room for the new map.
3310 if (addr && env->me_map != addr)
3311 return EBUSY; /* TODO: Make a new MDB_* error code? */
3313 p = (MDB_page *)env->me_map;
3314 env->me_metas[0] = METADATA(p);
3315 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3321 mdb_env_set_mapsize(MDB_env *env, size_t size)
3323 /* If env is already open, caller is responsible for making
3324 * sure there are no active txns.
3332 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3333 munmap(env->me_map, env->me_mapsize);
3334 env->me_mapsize = size;
3335 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3336 rc = mdb_env_map(env, old, 1);
3340 env->me_mapsize = size;
3342 env->me_maxpg = env->me_mapsize / env->me_psize;
3347 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3351 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3356 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3358 if (env->me_map || readers < 1)
3360 env->me_maxreaders = readers;
3365 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3367 if (!env || !readers)
3369 *readers = env->me_maxreaders;
3373 /** Further setup required for opening an MDB environment
3376 mdb_env_open2(MDB_env *env)
3378 unsigned int flags = env->me_flags;
3379 int i, newenv = 0, rc;
3383 /* See if we should use QueryLimited */
3385 if ((rc & 0xff) > 5)
3386 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3388 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3391 memset(&meta, 0, sizeof(meta));
3393 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3396 DPUTS("new mdbenv");
3398 GET_PAGESIZE(env->me_psize);
3400 env->me_psize = meta.mm_psize;
3403 /* Was a mapsize configured? */
3404 if (!env->me_mapsize) {
3405 /* If this is a new environment, take the default,
3406 * else use the size recorded in the existing env.
3408 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3409 } else if (env->me_mapsize < meta.mm_mapsize) {
3410 /* If the configured size is smaller, make sure it's
3411 * still big enough. Silently round up to minimum if not.
3413 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3414 if (env->me_mapsize < minsize)
3415 env->me_mapsize = minsize;
3418 rc = mdb_env_map(env, meta.mm_address, newenv);
3423 if (flags & MDB_FIXEDMAP)
3424 meta.mm_address = env->me_map;
3425 i = mdb_env_init_meta(env, &meta);
3426 if (i != MDB_SUCCESS) {
3430 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3431 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3433 env->me_maxpg = env->me_mapsize / env->me_psize;
3436 int toggle = mdb_env_pick_meta(env);
3437 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3439 DPRINTF(("opened database version %u, pagesize %u",
3440 env->me_metas[0]->mm_version, env->me_psize));
3441 DPRINTF(("using meta page %d", toggle));
3442 DPRINTF(("depth: %u", db->md_depth));
3443 DPRINTF(("entries: %"Z"u", db->md_entries));
3444 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3445 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3446 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3447 DPRINTF(("root: %"Z"u", db->md_root));
3455 /** Release a reader thread's slot in the reader lock table.
3456 * This function is called automatically when a thread exits.
3457 * @param[in] ptr This points to the slot in the reader lock table.
3460 mdb_env_reader_dest(void *ptr)
3462 MDB_reader *reader = ptr;
3468 /** Junk for arranging thread-specific callbacks on Windows. This is
3469 * necessarily platform and compiler-specific. Windows supports up
3470 * to 1088 keys. Let's assume nobody opens more than 64 environments
3471 * in a single process, for now. They can override this if needed.
3473 #ifndef MAX_TLS_KEYS
3474 #define MAX_TLS_KEYS 64
3476 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3477 static int mdb_tls_nkeys;
3479 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3483 case DLL_PROCESS_ATTACH: break;
3484 case DLL_THREAD_ATTACH: break;
3485 case DLL_THREAD_DETACH:
3486 for (i=0; i<mdb_tls_nkeys; i++) {
3487 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3488 mdb_env_reader_dest(r);
3491 case DLL_PROCESS_DETACH: break;
3496 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3498 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3502 /* Force some symbol references.
3503 * _tls_used forces the linker to create the TLS directory if not already done
3504 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3506 #pragma comment(linker, "/INCLUDE:_tls_used")
3507 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3508 #pragma const_seg(".CRT$XLB")
3509 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3510 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3513 #pragma comment(linker, "/INCLUDE:__tls_used")
3514 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3515 #pragma data_seg(".CRT$XLB")
3516 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3518 #endif /* WIN 32/64 */
3519 #endif /* !__GNUC__ */
3522 /** Downgrade the exclusive lock on the region back to shared */
3524 mdb_env_share_locks(MDB_env *env, int *excl)
3526 int rc = 0, toggle = mdb_env_pick_meta(env);
3528 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3533 /* First acquire a shared lock. The Unlock will
3534 * then release the existing exclusive lock.
3536 memset(&ov, 0, sizeof(ov));
3537 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3540 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3546 struct flock lock_info;
3547 /* The shared lock replaces the existing lock */
3548 memset((void *)&lock_info, 0, sizeof(lock_info));
3549 lock_info.l_type = F_RDLCK;
3550 lock_info.l_whence = SEEK_SET;
3551 lock_info.l_start = 0;
3552 lock_info.l_len = 1;
3553 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3554 (rc = ErrCode()) == EINTR) ;
3555 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3562 /** Try to get exlusive lock, otherwise shared.
3563 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3566 mdb_env_excl_lock(MDB_env *env, int *excl)
3570 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3574 memset(&ov, 0, sizeof(ov));
3575 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3582 struct flock lock_info;
3583 memset((void *)&lock_info, 0, sizeof(lock_info));
3584 lock_info.l_type = F_WRLCK;
3585 lock_info.l_whence = SEEK_SET;
3586 lock_info.l_start = 0;
3587 lock_info.l_len = 1;
3588 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3589 (rc = ErrCode()) == EINTR) ;
3593 # ifdef MDB_USE_POSIX_SEM
3594 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3597 lock_info.l_type = F_RDLCK;
3598 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3599 (rc = ErrCode()) == EINTR) ;
3607 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3609 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3611 * @(#) $Revision: 5.1 $
3612 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3613 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3615 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3619 * Please do not copyright this code. This code is in the public domain.
3621 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3622 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3623 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3624 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3625 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3626 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3627 * PERFORMANCE OF THIS SOFTWARE.
3630 * chongo <Landon Curt Noll> /\oo/\
3631 * http://www.isthe.com/chongo/
3633 * Share and Enjoy! :-)
3636 typedef unsigned long long mdb_hash_t;
3637 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3639 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3640 * @param[in] val value to hash
3641 * @param[in] hval initial value for hash
3642 * @return 64 bit hash
3644 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3645 * hval arg on the first call.
3648 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3650 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3651 unsigned char *end = s + val->mv_size;
3653 * FNV-1a hash each octet of the string
3656 /* xor the bottom with the current octet */
3657 hval ^= (mdb_hash_t)*s++;
3659 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3660 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3661 (hval << 7) + (hval << 8) + (hval << 40);
3663 /* return our new hash value */
3667 /** Hash the string and output the encoded hash.
3668 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3669 * very short name limits. We don't care about the encoding being reversible,
3670 * we just want to preserve as many bits of the input as possible in a
3671 * small printable string.
3672 * @param[in] str string to hash
3673 * @param[out] encbuf an array of 11 chars to hold the hash
3675 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3678 mdb_pack85(unsigned long l, char *out)
3682 for (i=0; i<5; i++) {
3683 *out++ = mdb_a85[l % 85];
3689 mdb_hash_enc(MDB_val *val, char *encbuf)
3691 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3693 mdb_pack85(h, encbuf);
3694 mdb_pack85(h>>32, encbuf+5);
3699 /** Open and/or initialize the lock region for the environment.
3700 * @param[in] env The MDB environment.
3701 * @param[in] lpath The pathname of the file used for the lock region.
3702 * @param[in] mode The Unix permissions for the file, if we create it.
3703 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3704 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3705 * @return 0 on success, non-zero on failure.
3708 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3711 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3713 # define MDB_ERRCODE_ROFS EROFS
3714 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3715 # define MDB_CLOEXEC O_CLOEXEC
3718 # define MDB_CLOEXEC 0
3725 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3726 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3727 FILE_ATTRIBUTE_NORMAL, NULL);
3729 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3731 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3733 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3738 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3739 /* Lose record locks when exec*() */
3740 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3741 fcntl(env->me_lfd, F_SETFD, fdflags);
3744 if (!(env->me_flags & MDB_NOTLS)) {
3745 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3748 env->me_flags |= MDB_ENV_TXKEY;
3750 /* Windows TLS callbacks need help finding their TLS info. */
3751 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3755 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3759 /* Try to get exclusive lock. If we succeed, then
3760 * nobody is using the lock region and we should initialize it.
3762 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3765 size = GetFileSize(env->me_lfd, NULL);
3767 size = lseek(env->me_lfd, 0, SEEK_END);
3768 if (size == -1) goto fail_errno;
3770 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3771 if (size < rsize && *excl > 0) {
3773 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3774 || !SetEndOfFile(env->me_lfd))
3777 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3781 size = rsize - sizeof(MDB_txninfo);
3782 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3787 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3789 if (!mh) goto fail_errno;
3790 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3792 if (!env->me_txns) goto fail_errno;
3794 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3796 if (m == MAP_FAILED) goto fail_errno;
3802 BY_HANDLE_FILE_INFORMATION stbuf;
3811 if (!mdb_sec_inited) {
3812 InitializeSecurityDescriptor(&mdb_null_sd,
3813 SECURITY_DESCRIPTOR_REVISION);
3814 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3815 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3816 mdb_all_sa.bInheritHandle = FALSE;
3817 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3820 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3821 idbuf.volume = stbuf.dwVolumeSerialNumber;
3822 idbuf.nhigh = stbuf.nFileIndexHigh;
3823 idbuf.nlow = stbuf.nFileIndexLow;
3824 val.mv_data = &idbuf;
3825 val.mv_size = sizeof(idbuf);
3826 mdb_hash_enc(&val, encbuf);
3827 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3828 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3829 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3830 if (!env->me_rmutex) goto fail_errno;
3831 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3832 if (!env->me_wmutex) goto fail_errno;
3833 #elif defined(MDB_USE_POSIX_SEM)
3842 #if defined(__NetBSD__)
3843 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3845 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3846 idbuf.dev = stbuf.st_dev;
3847 idbuf.ino = stbuf.st_ino;
3848 val.mv_data = &idbuf;
3849 val.mv_size = sizeof(idbuf);
3850 mdb_hash_enc(&val, encbuf);
3851 #ifdef MDB_SHORT_SEMNAMES
3852 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3854 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3855 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3856 /* Clean up after a previous run, if needed: Try to
3857 * remove both semaphores before doing anything else.
3859 sem_unlink(env->me_txns->mti_rmname);
3860 sem_unlink(env->me_txns->mti_wmname);
3861 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3862 O_CREAT|O_EXCL, mode, 1);
3863 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3864 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3865 O_CREAT|O_EXCL, mode, 1);
3866 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3867 #else /* MDB_USE_POSIX_SEM */
3868 pthread_mutexattr_t mattr;
3870 if ((rc = pthread_mutexattr_init(&mattr))
3871 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3872 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3873 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3875 pthread_mutexattr_destroy(&mattr);
3876 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3878 env->me_txns->mti_magic = MDB_MAGIC;
3879 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3880 env->me_txns->mti_txnid = 0;
3881 env->me_txns->mti_numreaders = 0;
3884 if (env->me_txns->mti_magic != MDB_MAGIC) {
3885 DPUTS("lock region has invalid magic");
3889 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3890 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3891 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3892 rc = MDB_VERSION_MISMATCH;
3896 if (rc && rc != EACCES && rc != EAGAIN) {
3900 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3901 if (!env->me_rmutex) goto fail_errno;
3902 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3903 if (!env->me_wmutex) goto fail_errno;
3904 #elif defined(MDB_USE_POSIX_SEM)
3905 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3906 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3907 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3908 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3919 /** The name of the lock file in the DB environment */
3920 #define LOCKNAME "/lock.mdb"
3921 /** The name of the data file in the DB environment */
3922 #define DATANAME "/data.mdb"
3923 /** The suffix of the lock file when no subdir is used */
3924 #define LOCKSUFF "-lock"
3925 /** Only a subset of the @ref mdb_env flags can be changed
3926 * at runtime. Changing other flags requires closing the
3927 * environment and re-opening it with the new flags.
3929 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3930 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3933 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3935 int oflags, rc, len, excl = -1;
3936 char *lpath, *dpath;
3938 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3942 if (flags & MDB_NOSUBDIR) {
3943 rc = len + sizeof(LOCKSUFF) + len + 1;
3945 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3950 if (flags & MDB_NOSUBDIR) {
3951 dpath = lpath + len + sizeof(LOCKSUFF);
3952 sprintf(lpath, "%s" LOCKSUFF, path);
3953 strcpy(dpath, path);
3955 dpath = lpath + len + sizeof(LOCKNAME);
3956 sprintf(lpath, "%s" LOCKNAME, path);
3957 sprintf(dpath, "%s" DATANAME, path);
3961 flags |= env->me_flags;
3962 if (flags & MDB_RDONLY) {
3963 /* silently ignore WRITEMAP when we're only getting read access */
3964 flags &= ~MDB_WRITEMAP;
3966 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3967 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3970 env->me_flags = flags |= MDB_ENV_ACTIVE;
3974 env->me_path = strdup(path);
3975 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3976 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3977 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3982 /* For RDONLY, get lockfile after we know datafile exists */
3983 if (!F_ISSET(flags, MDB_RDONLY)) {
3984 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3990 if (F_ISSET(flags, MDB_RDONLY)) {
3991 oflags = GENERIC_READ;
3992 len = OPEN_EXISTING;
3994 oflags = GENERIC_READ|GENERIC_WRITE;
3997 mode = FILE_ATTRIBUTE_NORMAL;
3998 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3999 NULL, len, mode, NULL);
4001 if (F_ISSET(flags, MDB_RDONLY))
4004 oflags = O_RDWR | O_CREAT;
4006 env->me_fd = open(dpath, oflags, mode);
4008 if (env->me_fd == INVALID_HANDLE_VALUE) {
4013 if (F_ISSET(flags, MDB_RDONLY)) {
4014 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4019 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4020 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4021 env->me_mfd = env->me_fd;
4023 /* Synchronous fd for meta writes. Needed even with
4024 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4027 len = OPEN_EXISTING;
4028 env->me_mfd = CreateFile(dpath, oflags,
4029 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4030 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4033 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4035 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4040 DPRINTF(("opened dbenv %p", (void *) env));
4042 rc = mdb_env_share_locks(env, &excl);
4048 mdb_env_close0(env, excl);
4054 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4056 mdb_env_close0(MDB_env *env, int excl)
4060 if (!(env->me_flags & MDB_ENV_ACTIVE))
4063 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4064 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4065 free(env->me_dbxs[i].md_name.mv_data);
4067 free(env->me_dbflags);
4070 free(env->me_dirty_list);
4071 mdb_midl_free(env->me_free_pgs);
4073 if (env->me_flags & MDB_ENV_TXKEY) {
4074 pthread_key_delete(env->me_txkey);
4076 /* Delete our key from the global list */
4077 for (i=0; i<mdb_tls_nkeys; i++)
4078 if (mdb_tls_keys[i] == env->me_txkey) {
4079 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4087 munmap(env->me_map, env->me_mapsize);
4089 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4090 (void) close(env->me_mfd);
4091 if (env->me_fd != INVALID_HANDLE_VALUE)
4092 (void) close(env->me_fd);
4094 pid_t pid = env->me_pid;
4095 /* Clearing readers is done in this function because
4096 * me_txkey with its destructor must be disabled first.
4098 for (i = env->me_numreaders; --i >= 0; )
4099 if (env->me_txns->mti_readers[i].mr_pid == pid)
4100 env->me_txns->mti_readers[i].mr_pid = 0;
4102 if (env->me_rmutex) {
4103 CloseHandle(env->me_rmutex);
4104 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4106 /* Windows automatically destroys the mutexes when
4107 * the last handle closes.
4109 #elif defined(MDB_USE_POSIX_SEM)
4110 if (env->me_rmutex != SEM_FAILED) {
4111 sem_close(env->me_rmutex);
4112 if (env->me_wmutex != SEM_FAILED)
4113 sem_close(env->me_wmutex);
4114 /* If we have the filelock: If we are the
4115 * only remaining user, clean up semaphores.
4118 mdb_env_excl_lock(env, &excl);
4120 sem_unlink(env->me_txns->mti_rmname);
4121 sem_unlink(env->me_txns->mti_wmname);
4125 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4127 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4130 /* Unlock the lockfile. Windows would have unlocked it
4131 * after closing anyway, but not necessarily at once.
4133 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4136 (void) close(env->me_lfd);
4139 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4143 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4145 MDB_txn *txn = NULL;
4151 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4155 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4158 /* Do the lock/unlock of the reader mutex before starting the
4159 * write txn. Otherwise other read txns could block writers.
4161 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4166 /* We must start the actual read txn after blocking writers */
4167 mdb_txn_reset0(txn, "reset-stage1");
4169 /* Temporarily block writers until we snapshot the meta pages */
4172 rc = mdb_txn_renew0(txn);
4174 UNLOCK_MUTEX_W(env);
4179 wsize = env->me_psize * 2;
4183 DO_WRITE(rc, fd, ptr, w2, len);
4187 } else if (len > 0) {
4193 /* Non-blocking or async handles are not supported */
4199 UNLOCK_MUTEX_W(env);
4204 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4206 if (wsize > MAX_WRITE)
4210 DO_WRITE(rc, fd, ptr, w2, len);
4214 } else if (len > 0) {
4231 mdb_env_copy(MDB_env *env, const char *path)
4235 HANDLE newfd = INVALID_HANDLE_VALUE;
4237 if (env->me_flags & MDB_NOSUBDIR) {
4238 lpath = (char *)path;
4241 len += sizeof(DATANAME);
4242 lpath = malloc(len);
4245 sprintf(lpath, "%s" DATANAME, path);
4248 /* The destination path must exist, but the destination file must not.
4249 * We don't want the OS to cache the writes, since the source data is
4250 * already in the OS cache.
4253 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4254 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4256 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4258 if (newfd == INVALID_HANDLE_VALUE) {
4264 /* Set O_DIRECT if the file system supports it */
4265 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4266 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4268 #ifdef F_NOCACHE /* __APPLE__ */
4269 rc = fcntl(newfd, F_NOCACHE, 1);
4276 rc = mdb_env_copyfd(env, newfd);
4279 if (!(env->me_flags & MDB_NOSUBDIR))
4281 if (newfd != INVALID_HANDLE_VALUE)
4282 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4289 mdb_env_close(MDB_env *env)
4296 VGMEMP_DESTROY(env);
4297 while ((dp = env->me_dpages) != NULL) {
4298 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4299 env->me_dpages = dp->mp_next;
4303 mdb_env_close0(env, 0);
4307 /** Compare two items pointing at aligned size_t's */
4309 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4311 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4312 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4315 /** Compare two items pointing at aligned unsigned int's */
4317 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4319 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4320 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4323 /** Compare two items pointing at unsigned ints of unknown alignment.
4324 * Nodes and keys are guaranteed to be 2-byte aligned.
4327 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4329 #if BYTE_ORDER == LITTLE_ENDIAN
4330 unsigned short *u, *c;
4333 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4334 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4337 } while(!x && u > (unsigned short *)a->mv_data);
4340 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4344 /** Compare two items lexically */
4346 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4353 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4359 diff = memcmp(a->mv_data, b->mv_data, len);
4360 return diff ? diff : len_diff<0 ? -1 : len_diff;
4363 /** Compare two items in reverse byte order */
4365 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4367 const unsigned char *p1, *p2, *p1_lim;
4371 p1_lim = (const unsigned char *)a->mv_data;
4372 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4373 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4375 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4381 while (p1 > p1_lim) {
4382 diff = *--p1 - *--p2;
4386 return len_diff<0 ? -1 : len_diff;
4389 /** Search for key within a page, using binary search.
4390 * Returns the smallest entry larger or equal to the key.
4391 * If exactp is non-null, stores whether the found entry was an exact match
4392 * in *exactp (1 or 0).
4393 * Updates the cursor index with the index of the found entry.
4394 * If no entry larger or equal to the key is found, returns NULL.
4397 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4399 unsigned int i = 0, nkeys;
4402 MDB_page *mp = mc->mc_pg[mc->mc_top];
4403 MDB_node *node = NULL;
4408 nkeys = NUMKEYS(mp);
4413 COPY_PGNO(pgno, mp->mp_pgno);
4414 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4415 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4422 low = IS_LEAF(mp) ? 0 : 1;
4424 cmp = mc->mc_dbx->md_cmp;
4426 /* Branch pages have no data, so if using integer keys,
4427 * alignment is guaranteed. Use faster mdb_cmp_int.
4429 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4430 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4437 nodekey.mv_size = mc->mc_db->md_pad;
4438 node = NODEPTR(mp, 0); /* fake */
4439 while (low <= high) {
4440 i = (low + high) >> 1;
4441 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4442 rc = cmp(key, &nodekey);
4443 DPRINTF(("found leaf index %u [%s], rc = %i",
4444 i, DKEY(&nodekey), rc));
4453 while (low <= high) {
4454 i = (low + high) >> 1;
4456 node = NODEPTR(mp, i);
4457 nodekey.mv_size = NODEKSZ(node);
4458 nodekey.mv_data = NODEKEY(node);
4460 rc = cmp(key, &nodekey);
4463 DPRINTF(("found leaf index %u [%s], rc = %i",
4464 i, DKEY(&nodekey), rc));
4466 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4467 i, DKEY(&nodekey), NODEPGNO(node), rc));
4478 if (rc > 0) { /* Found entry is less than the key. */
4479 i++; /* Skip to get the smallest entry larger than key. */
4481 node = NODEPTR(mp, i);
4484 *exactp = (rc == 0);
4485 /* store the key index */
4486 mc->mc_ki[mc->mc_top] = i;
4488 /* There is no entry larger or equal to the key. */
4491 /* nodeptr is fake for LEAF2 */
4497 mdb_cursor_adjust(MDB_cursor *mc, func)
4501 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4502 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4509 /** Pop a page off the top of the cursor's stack. */
4511 mdb_cursor_pop(MDB_cursor *mc)
4515 MDB_page *top = mc->mc_pg[mc->mc_top];
4521 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4522 mc->mc_dbi, (void *) mc));
4526 /** Push a page onto the top of the cursor's stack. */
4528 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4530 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4531 mc->mc_dbi, (void *) mc));
4533 if (mc->mc_snum >= CURSOR_STACK) {
4534 assert(mc->mc_snum < CURSOR_STACK);
4535 return MDB_CURSOR_FULL;
4538 mc->mc_top = mc->mc_snum++;
4539 mc->mc_pg[mc->mc_top] = mp;
4540 mc->mc_ki[mc->mc_top] = 0;
4545 /** Find the address of the page corresponding to a given page number.
4546 * @param[in] txn the transaction for this access.
4547 * @param[in] pgno the page number for the page to retrieve.
4548 * @param[out] ret address of a pointer where the page's address will be stored.
4549 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4550 * @return 0 on success, non-zero on failure.
4553 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4555 MDB_env *env = txn->mt_env;
4559 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4563 MDB_ID2L dl = tx2->mt_u.dirty_list;
4565 /* Spilled pages were dirtied in this txn and flushed
4566 * because the dirty list got full. Bring this page
4567 * back in from the map (but don't unspill it here,
4568 * leave that unless page_touch happens again).
4570 if (tx2->mt_spill_pgs) {
4571 MDB_ID pn = pgno << 1;
4572 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4573 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4574 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4579 unsigned x = mdb_mid2l_search(dl, pgno);
4580 if (x <= dl[0].mid && dl[x].mid == pgno) {
4586 } while ((tx2 = tx2->mt_parent) != NULL);
4589 if (pgno < txn->mt_next_pgno) {
4591 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4593 DPRINTF(("page %"Z"u not found", pgno));
4595 return MDB_PAGE_NOTFOUND;
4605 /** Search for the page a given key should be in.
4606 * Pushes parent pages on the cursor stack. This function continues a
4607 * search on a cursor that has already been initialized. (Usually by
4608 * #mdb_page_search() but also by #mdb_node_move().)
4609 * @param[in,out] mc the cursor for this operation.
4610 * @param[in] key the key to search for. If NULL, search for the lowest
4611 * page. (This is used by #mdb_cursor_first().)
4612 * @param[in] modify If true, visited pages are updated with new page numbers.
4613 * @return 0 on success, non-zero on failure.
4616 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4618 MDB_page *mp = mc->mc_pg[mc->mc_top];
4622 while (IS_BRANCH(mp)) {
4626 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4627 assert(NUMKEYS(mp) > 1);
4628 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4630 if (key == NULL) /* Initialize cursor to first page. */
4632 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4633 /* cursor to last page */
4637 node = mdb_node_search(mc, key, &exact);
4639 i = NUMKEYS(mp) - 1;
4641 i = mc->mc_ki[mc->mc_top];
4650 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4651 assert(i < NUMKEYS(mp));
4652 node = NODEPTR(mp, i);
4654 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4657 mc->mc_ki[mc->mc_top] = i;
4658 if ((rc = mdb_cursor_push(mc, mp)))
4662 if ((rc = mdb_page_touch(mc)) != 0)
4664 mp = mc->mc_pg[mc->mc_top];
4669 DPRINTF(("internal error, index points to a %02X page!?",
4671 return MDB_CORRUPTED;
4674 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4675 key ? DKEY(key) : NULL));
4676 mc->mc_flags |= C_INITIALIZED;
4677 mc->mc_flags &= ~C_EOF;
4682 /** Search for the lowest key under the current branch page.
4683 * This just bypasses a NUMKEYS check in the current page
4684 * before calling mdb_page_search_root(), because the callers
4685 * are all in situations where the current page is known to
4689 mdb_page_search_lowest(MDB_cursor *mc)
4691 MDB_page *mp = mc->mc_pg[mc->mc_top];
4692 MDB_node *node = NODEPTR(mp, 0);
4695 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4698 mc->mc_ki[mc->mc_top] = 0;
4699 if ((rc = mdb_cursor_push(mc, mp)))
4701 return mdb_page_search_root(mc, NULL, 0);
4704 /** Search for the page a given key should be in.
4705 * Pushes parent pages on the cursor stack. This function just sets up
4706 * the search; it finds the root page for \b mc's database and sets this
4707 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4708 * called to complete the search.
4709 * @param[in,out] mc the cursor for this operation.
4710 * @param[in] key the key to search for. If NULL, search for the lowest
4711 * page. (This is used by #mdb_cursor_first().)
4712 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4713 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4714 * @return 0 on success, non-zero on failure.
4717 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4722 /* Make sure the txn is still viable, then find the root from
4723 * the txn's db table.
4725 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4726 DPUTS("transaction has failed, must abort");
4729 /* Make sure we're using an up-to-date root */
4730 if (mc->mc_dbi > MAIN_DBI) {
4731 if ((*mc->mc_dbflag & DB_STALE) ||
4732 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4734 unsigned char dbflag = 0;
4735 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4736 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4739 if (*mc->mc_dbflag & DB_STALE) {
4743 MDB_node *leaf = mdb_node_search(&mc2,
4744 &mc->mc_dbx->md_name, &exact);
4746 return MDB_NOTFOUND;
4747 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4750 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4752 /* The txn may not know this DBI, or another process may
4753 * have dropped and recreated the DB with other flags.
4755 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4756 return MDB_INCOMPATIBLE;
4757 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4759 if (flags & MDB_PS_MODIFY)
4761 *mc->mc_dbflag &= ~DB_STALE;
4762 *mc->mc_dbflag |= dbflag;
4765 root = mc->mc_db->md_root;
4767 if (root == P_INVALID) { /* Tree is empty. */
4768 DPUTS("tree is empty");
4769 return MDB_NOTFOUND;
4774 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4775 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4781 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4782 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4784 if (flags & MDB_PS_MODIFY) {
4785 if ((rc = mdb_page_touch(mc)))
4789 if (flags & MDB_PS_ROOTONLY)
4792 return mdb_page_search_root(mc, key, flags);
4796 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4798 MDB_txn *txn = mc->mc_txn;
4799 pgno_t pg = mp->mp_pgno;
4800 unsigned x = 0, ovpages = mp->mp_pages;
4801 MDB_env *env = txn->mt_env;
4802 MDB_IDL sl = txn->mt_spill_pgs;
4803 MDB_ID pn = pg << 1;
4806 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4807 /* If the page is dirty or on the spill list we just acquired it,
4808 * so we should give it back to our current free list, if any.
4809 * Otherwise put it onto the list of pages we freed in this txn.
4811 * Won't create me_pghead: me_pglast must be inited along with it.
4812 * Unsupported in nested txns: They would need to hide the page
4813 * range in ancestor txns' dirty and spilled lists.
4815 if (env->me_pghead &&
4817 ((mp->mp_flags & P_DIRTY) ||
4818 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4822 MDB_ID2 *dl, ix, iy;
4823 rc = mdb_midl_need(&env->me_pghead, ovpages);
4826 if (!(mp->mp_flags & P_DIRTY)) {
4827 /* This page is no longer spilled */
4834 /* Remove from dirty list */
4835 dl = txn->mt_u.dirty_list;
4837 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4845 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4846 txn->mt_flags |= MDB_TXN_ERROR;
4847 return MDB_CORRUPTED;
4850 if (!(env->me_flags & MDB_WRITEMAP))
4851 mdb_dpage_free(env, mp);
4853 /* Insert in me_pghead */
4854 mop = env->me_pghead;
4855 j = mop[0] + ovpages;
4856 for (i = mop[0]; i && mop[i] < pg; i--)
4862 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4866 mc->mc_db->md_overflow_pages -= ovpages;
4870 /** Return the data associated with a given node.
4871 * @param[in] txn The transaction for this operation.
4872 * @param[in] leaf The node being read.
4873 * @param[out] data Updated to point to the node's data.
4874 * @return 0 on success, non-zero on failure.
4877 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4879 MDB_page *omp; /* overflow page */
4883 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4884 data->mv_size = NODEDSZ(leaf);
4885 data->mv_data = NODEDATA(leaf);
4889 /* Read overflow data.
4891 data->mv_size = NODEDSZ(leaf);
4892 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4893 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4894 DPRINTF(("read overflow page %"Z"u failed", pgno));
4897 data->mv_data = METADATA(omp);
4903 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4904 MDB_val *key, MDB_val *data)
4913 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4915 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4918 if (txn->mt_flags & MDB_TXN_ERROR)
4921 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4922 return MDB_BAD_VALSIZE;
4925 mdb_cursor_init(&mc, txn, dbi, &mx);
4926 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4929 /** Find a sibling for a page.
4930 * Replaces the page at the top of the cursor's stack with the
4931 * specified sibling, if one exists.
4932 * @param[in] mc The cursor for this operation.
4933 * @param[in] move_right Non-zero if the right sibling is requested,
4934 * otherwise the left sibling.
4935 * @return 0 on success, non-zero on failure.
4938 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4944 if (mc->mc_snum < 2) {
4945 return MDB_NOTFOUND; /* root has no siblings */
4949 DPRINTF(("parent page is page %"Z"u, index %u",
4950 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4952 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4953 : (mc->mc_ki[mc->mc_top] == 0)) {
4954 DPRINTF(("no more keys left, moving to %s sibling",
4955 move_right ? "right" : "left"));
4956 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4957 /* undo cursor_pop before returning */
4964 mc->mc_ki[mc->mc_top]++;
4966 mc->mc_ki[mc->mc_top]--;
4967 DPRINTF(("just moving to %s index key %u",
4968 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4970 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4972 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4973 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4976 mdb_cursor_push(mc, mp);
4978 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4983 /** Move the cursor to the next data item. */
4985 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4991 if (mc->mc_flags & C_EOF) {
4992 return MDB_NOTFOUND;
4995 assert(mc->mc_flags & C_INITIALIZED);
4997 mp = mc->mc_pg[mc->mc_top];
4999 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5000 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5001 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5002 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5003 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5004 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5005 if (rc == MDB_SUCCESS)
5006 MDB_GET_KEY(leaf, key);
5011 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5012 if (op == MDB_NEXT_DUP)
5013 return MDB_NOTFOUND;
5017 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5018 if (mc->mc_flags & C_DEL)
5021 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5022 DPUTS("=====> move to next sibling page");
5023 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5024 mc->mc_flags |= C_EOF;
5027 mp = mc->mc_pg[mc->mc_top];
5028 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5030 mc->mc_ki[mc->mc_top]++;
5033 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5034 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5037 key->mv_size = mc->mc_db->md_pad;
5038 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5042 assert(IS_LEAF(mp));
5043 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5045 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5046 mdb_xcursor_init1(mc, leaf);
5049 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5052 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5053 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5054 if (rc != MDB_SUCCESS)
5059 MDB_GET_KEY(leaf, key);
5063 /** Move the cursor to the previous data item. */
5065 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5071 assert(mc->mc_flags & C_INITIALIZED);
5073 mp = mc->mc_pg[mc->mc_top];
5075 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5076 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5077 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5078 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5079 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5080 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5081 if (rc == MDB_SUCCESS)
5082 MDB_GET_KEY(leaf, key);
5086 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5087 if (op == MDB_PREV_DUP)
5088 return MDB_NOTFOUND;
5093 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5095 if (mc->mc_ki[mc->mc_top] == 0) {
5096 DPUTS("=====> move to prev sibling page");
5097 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5100 mp = mc->mc_pg[mc->mc_top];
5101 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5102 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5104 mc->mc_ki[mc->mc_top]--;
5106 mc->mc_flags &= ~C_EOF;
5108 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5109 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5112 key->mv_size = mc->mc_db->md_pad;
5113 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5117 assert(IS_LEAF(mp));
5118 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5120 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5121 mdb_xcursor_init1(mc, leaf);
5124 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5127 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5128 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5129 if (rc != MDB_SUCCESS)
5134 MDB_GET_KEY(leaf, key);
5138 /** Set the cursor on a specific data item. */
5140 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5141 MDB_cursor_op op, int *exactp)
5145 MDB_node *leaf = NULL;
5150 assert(key->mv_size > 0);
5153 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5155 /* See if we're already on the right page */
5156 if (mc->mc_flags & C_INITIALIZED) {
5159 mp = mc->mc_pg[mc->mc_top];
5161 mc->mc_ki[mc->mc_top] = 0;
5162 return MDB_NOTFOUND;
5164 if (mp->mp_flags & P_LEAF2) {
5165 nodekey.mv_size = mc->mc_db->md_pad;
5166 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5168 leaf = NODEPTR(mp, 0);
5169 MDB_GET_KEY2(leaf, nodekey);
5171 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5173 /* Probably happens rarely, but first node on the page
5174 * was the one we wanted.
5176 mc->mc_ki[mc->mc_top] = 0;
5183 unsigned int nkeys = NUMKEYS(mp);
5185 if (mp->mp_flags & P_LEAF2) {
5186 nodekey.mv_data = LEAF2KEY(mp,
5187 nkeys-1, nodekey.mv_size);
5189 leaf = NODEPTR(mp, nkeys-1);
5190 MDB_GET_KEY2(leaf, nodekey);
5192 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5194 /* last node was the one we wanted */
5195 mc->mc_ki[mc->mc_top] = nkeys-1;
5201 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5202 /* This is definitely the right page, skip search_page */
5203 if (mp->mp_flags & P_LEAF2) {
5204 nodekey.mv_data = LEAF2KEY(mp,
5205 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5207 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5208 MDB_GET_KEY2(leaf, nodekey);
5210 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5212 /* current node was the one we wanted */
5222 /* If any parents have right-sibs, search.
5223 * Otherwise, there's nothing further.
5225 for (i=0; i<mc->mc_top; i++)
5227 NUMKEYS(mc->mc_pg[i])-1)
5229 if (i == mc->mc_top) {
5230 /* There are no other pages */
5231 mc->mc_ki[mc->mc_top] = nkeys;
5232 return MDB_NOTFOUND;
5236 /* There are no other pages */
5237 mc->mc_ki[mc->mc_top] = 0;
5238 if (op == MDB_SET_RANGE) {
5242 return MDB_NOTFOUND;
5246 rc = mdb_page_search(mc, key, 0);
5247 if (rc != MDB_SUCCESS)
5250 mp = mc->mc_pg[mc->mc_top];
5251 assert(IS_LEAF(mp));
5254 leaf = mdb_node_search(mc, key, exactp);
5255 if (exactp != NULL && !*exactp) {
5256 /* MDB_SET specified and not an exact match. */
5257 return MDB_NOTFOUND;
5261 DPUTS("===> inexact leaf not found, goto sibling");
5262 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5263 return rc; /* no entries matched */
5264 mp = mc->mc_pg[mc->mc_top];
5265 assert(IS_LEAF(mp));
5266 leaf = NODEPTR(mp, 0);
5270 mc->mc_flags |= C_INITIALIZED;
5271 mc->mc_flags &= ~C_EOF;
5274 key->mv_size = mc->mc_db->md_pad;
5275 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5279 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5280 mdb_xcursor_init1(mc, leaf);
5283 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5284 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5285 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5288 if (op == MDB_GET_BOTH) {
5294 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5295 if (rc != MDB_SUCCESS)
5298 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5300 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5302 rc = mc->mc_dbx->md_dcmp(data, &d2);
5304 if (op == MDB_GET_BOTH || rc > 0)
5305 return MDB_NOTFOUND;
5311 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5312 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5317 /* The key already matches in all other cases */
5318 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5319 MDB_GET_KEY(leaf, key);
5320 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5325 /** Move the cursor to the first item in the database. */
5327 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5333 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5335 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5336 rc = mdb_page_search(mc, NULL, 0);
5337 if (rc != MDB_SUCCESS)
5340 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5342 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5343 mc->mc_flags |= C_INITIALIZED;
5344 mc->mc_flags &= ~C_EOF;
5346 mc->mc_ki[mc->mc_top] = 0;
5348 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5349 key->mv_size = mc->mc_db->md_pad;
5350 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5355 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5356 mdb_xcursor_init1(mc, leaf);
5357 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5361 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5365 MDB_GET_KEY(leaf, key);
5369 /** Move the cursor to the last item in the database. */
5371 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5377 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5379 if (!(mc->mc_flags & C_EOF)) {
5381 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5384 lkey.mv_size = MDB_MAXKEYSIZE+1;
5385 lkey.mv_data = NULL;
5386 rc = mdb_page_search(mc, &lkey, 0);
5387 if (rc != MDB_SUCCESS)
5390 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5393 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5394 mc->mc_flags |= C_INITIALIZED|C_EOF;
5395 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5397 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5398 key->mv_size = mc->mc_db->md_pad;
5399 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5404 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5405 mdb_xcursor_init1(mc, leaf);
5406 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5410 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5415 MDB_GET_KEY(leaf, key);
5420 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5425 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5429 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5433 case MDB_GET_CURRENT:
5434 if (!(mc->mc_flags & C_INITIALIZED)) {
5437 MDB_page *mp = mc->mc_pg[mc->mc_top];
5439 mc->mc_ki[mc->mc_top] = 0;
5445 key->mv_size = mc->mc_db->md_pad;
5446 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5448 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5449 MDB_GET_KEY(leaf, key);
5451 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5452 if (mc->mc_flags & C_DEL)
5453 mdb_xcursor_init1(mc, leaf);
5454 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5456 rc = mdb_node_read(mc->mc_txn, leaf, data);
5463 case MDB_GET_BOTH_RANGE:
5468 if (mc->mc_xcursor == NULL) {
5469 rc = MDB_INCOMPATIBLE;
5478 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5479 rc = MDB_BAD_VALSIZE;
5480 } else if (op == MDB_SET_RANGE)
5481 rc = mdb_cursor_set(mc, key, data, op, NULL);
5483 rc = mdb_cursor_set(mc, key, data, op, &exact);
5485 case MDB_GET_MULTIPLE:
5486 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5490 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5491 rc = MDB_INCOMPATIBLE;
5495 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5496 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5499 case MDB_NEXT_MULTIPLE:
5504 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5505 rc = MDB_INCOMPATIBLE;
5508 if (!(mc->mc_flags & C_INITIALIZED))
5509 rc = mdb_cursor_first(mc, key, data);
5511 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5512 if (rc == MDB_SUCCESS) {
5513 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5516 mx = &mc->mc_xcursor->mx_cursor;
5517 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5519 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5520 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5528 case MDB_NEXT_NODUP:
5529 if (!(mc->mc_flags & C_INITIALIZED))
5530 rc = mdb_cursor_first(mc, key, data);
5532 rc = mdb_cursor_next(mc, key, data, op);
5536 case MDB_PREV_NODUP:
5537 if (!(mc->mc_flags & C_INITIALIZED)) {
5538 rc = mdb_cursor_last(mc, key, data);
5541 mc->mc_flags |= C_INITIALIZED;
5542 mc->mc_ki[mc->mc_top]++;
5544 rc = mdb_cursor_prev(mc, key, data, op);
5547 rc = mdb_cursor_first(mc, key, data);
5550 mfunc = mdb_cursor_first;
5552 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5556 if (mc->mc_xcursor == NULL) {
5557 rc = MDB_INCOMPATIBLE;
5560 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5564 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5567 rc = mdb_cursor_last(mc, key, data);
5570 mfunc = mdb_cursor_last;
5573 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5578 if (mc->mc_flags & C_DEL)
5579 mc->mc_flags ^= C_DEL;
5584 /** Touch all the pages in the cursor stack.
5585 * Makes sure all the pages are writable, before attempting a write operation.
5586 * @param[in] mc The cursor to operate on.
5589 mdb_cursor_touch(MDB_cursor *mc)
5593 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5596 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5597 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5600 *mc->mc_dbflag |= DB_DIRTY;
5602 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5603 rc = mdb_page_touch(mc);
5607 mc->mc_top = mc->mc_snum-1;
5611 /** Do not spill pages to disk if txn is getting full, may fail instead */
5612 #define MDB_NOSPILL 0x8000
5615 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5618 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5619 MDB_node *leaf = NULL;
5620 MDB_val xdata, *rdata, dkey;
5623 int do_sub = 0, insert = 0;
5624 unsigned int mcount = 0, dcount = 0, nospill;
5628 char dbuf[MDB_MAXKEYSIZE+1];
5629 unsigned int nflags;
5632 /* Check this first so counter will always be zero on any
5635 if (flags & MDB_MULTIPLE) {
5636 dcount = data[1].mv_size;
5637 data[1].mv_size = 0;
5638 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5639 return MDB_INCOMPATIBLE;
5642 nospill = flags & MDB_NOSPILL;
5643 flags &= ~MDB_NOSPILL;
5645 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5646 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5648 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5649 return MDB_BAD_VALSIZE;
5651 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5652 return MDB_BAD_VALSIZE;
5654 #if SIZE_MAX > MAXDATASIZE
5655 if (data->mv_size > MAXDATASIZE)
5656 return MDB_BAD_VALSIZE;
5659 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5660 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5664 if (flags == MDB_CURRENT) {
5665 if (!(mc->mc_flags & C_INITIALIZED))
5668 } else if (mc->mc_db->md_root == P_INVALID) {
5669 /* new database, cursor has nothing to point to */
5671 mc->mc_flags &= ~C_INITIALIZED;
5676 if (flags & MDB_APPEND) {
5678 rc = mdb_cursor_last(mc, &k2, &d2);
5680 rc = mc->mc_dbx->md_cmp(key, &k2);
5683 mc->mc_ki[mc->mc_top]++;
5685 /* new key is <= last key */
5690 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5692 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5693 DPRINTF(("duplicate key [%s]", DKEY(key)));
5695 return MDB_KEYEXIST;
5697 if (rc && rc != MDB_NOTFOUND)
5701 if (mc->mc_flags & C_DEL)
5702 mc->mc_flags ^= C_DEL;
5704 /* Cursor is positioned, check for room in the dirty list */
5706 if (flags & MDB_MULTIPLE) {
5708 xdata.mv_size = data->mv_size * dcount;
5712 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5716 if (rc == MDB_NO_ROOT) {
5718 /* new database, write a root leaf page */
5719 DPUTS("allocating new root leaf page");
5720 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5723 mdb_cursor_push(mc, np);
5724 mc->mc_db->md_root = np->mp_pgno;
5725 mc->mc_db->md_depth++;
5726 *mc->mc_dbflag |= DB_DIRTY;
5727 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5729 np->mp_flags |= P_LEAF2;
5730 mc->mc_flags |= C_INITIALIZED;
5732 /* make sure all cursor pages are writable */
5733 rc2 = mdb_cursor_touch(mc);
5738 /* The key already exists */
5739 if (rc == MDB_SUCCESS) {
5740 /* there's only a key anyway, so this is a no-op */
5741 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5742 unsigned int ksize = mc->mc_db->md_pad;
5743 if (key->mv_size != ksize)
5744 return MDB_BAD_VALSIZE;
5745 if (flags == MDB_CURRENT) {
5746 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5747 memcpy(ptr, key->mv_data, ksize);
5752 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5755 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5756 /* Was a single item before, must convert now */
5758 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5759 /* Just overwrite the current item */
5760 if (flags == MDB_CURRENT)
5763 dkey.mv_size = NODEDSZ(leaf);
5764 dkey.mv_data = NODEDATA(leaf);
5765 #if UINT_MAX < SIZE_MAX
5766 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5767 #ifdef MISALIGNED_OK
5768 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5770 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5773 /* if data matches, skip it */
5774 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5775 if (flags & MDB_NODUPDATA)
5777 else if (flags & MDB_MULTIPLE)
5784 /* create a fake page for the dup items */
5785 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5786 dkey.mv_data = dbuf;
5787 fp = (MDB_page *)&pbuf;
5788 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5789 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5790 fp->mp_lower = PAGEHDRSZ;
5791 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5792 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5793 fp->mp_flags |= P_LEAF2;
5794 fp->mp_pad = data->mv_size;
5795 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5797 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5798 (dkey.mv_size & 1) + (data->mv_size & 1);
5800 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5803 xdata.mv_size = fp->mp_upper;
5808 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5809 /* See if we need to convert from fake page to subDB */
5811 unsigned int offset;
5815 fp = NODEDATA(leaf);
5816 if (flags == MDB_CURRENT) {
5818 fp->mp_flags |= P_DIRTY;
5819 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5820 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5824 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5825 offset = fp->mp_pad;
5826 if (SIZELEFT(fp) >= offset)
5828 offset *= 4; /* space for 4 more */
5830 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5832 offset += offset & 1;
5833 fp_flags = fp->mp_flags;
5834 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5835 offset >= mc->mc_txn->mt_env->me_nodemax) {
5836 /* yes, convert it */
5838 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5839 dummy.md_pad = fp->mp_pad;
5840 dummy.md_flags = MDB_DUPFIXED;
5841 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5842 dummy.md_flags |= MDB_INTEGERKEY;
5845 dummy.md_branch_pages = 0;
5846 dummy.md_leaf_pages = 1;
5847 dummy.md_overflow_pages = 0;
5848 dummy.md_entries = NUMKEYS(fp);
5850 xdata.mv_size = sizeof(MDB_db);
5851 xdata.mv_data = &dummy;
5852 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5854 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5855 flags |= F_DUPDATA|F_SUBDATA;
5856 dummy.md_root = mp->mp_pgno;
5857 fp_flags &= ~P_SUBP;
5859 /* no, just grow it */
5861 xdata.mv_size = NODEDSZ(leaf) + offset;
5862 xdata.mv_data = &pbuf;
5863 mp = (MDB_page *)&pbuf;
5864 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5867 mp->mp_flags = fp_flags | P_DIRTY;
5868 mp->mp_pad = fp->mp_pad;
5869 mp->mp_lower = fp->mp_lower;
5870 mp->mp_upper = fp->mp_upper + offset;
5872 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5874 nsize = NODEDSZ(leaf) - fp->mp_upper;
5875 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5876 for (i=0; i<NUMKEYS(fp); i++)
5877 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5879 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5883 /* data is on sub-DB, just store it */
5884 flags |= F_DUPDATA|F_SUBDATA;
5888 /* overflow page overwrites need special handling */
5889 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5892 unsigned psize = mc->mc_txn->mt_env->me_psize;
5893 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5895 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5896 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5898 ovpages = omp->mp_pages;
5900 /* Is the ov page large enough? */
5901 if (ovpages >= dpages) {
5902 if (!(omp->mp_flags & P_DIRTY) &&
5903 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5905 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5908 level = 0; /* dirty in this txn or clean */
5911 if (omp->mp_flags & P_DIRTY) {
5912 /* yes, overwrite it. Note in this case we don't
5913 * bother to try shrinking the page if the new data
5914 * is smaller than the overflow threshold.
5917 /* It is writable only in a parent txn */
5918 size_t sz = (size_t) psize * ovpages, off;
5919 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5925 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5926 if (!(flags & MDB_RESERVE)) {
5927 /* Copy end of page, adjusting alignment so
5928 * compiler may copy words instead of bytes.
5930 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5931 memcpy((size_t *)((char *)np + off),
5932 (size_t *)((char *)omp + off), sz - off);
5935 memcpy(np, omp, sz); /* Copy beginning of page */
5938 SETDSZ(leaf, data->mv_size);
5939 if (F_ISSET(flags, MDB_RESERVE))
5940 data->mv_data = METADATA(omp);
5942 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5946 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5948 } else if (NODEDSZ(leaf) == data->mv_size) {
5949 /* same size, just replace it. Note that we could
5950 * also reuse this node if the new data is smaller,
5951 * but instead we opt to shrink the node in that case.
5953 if (F_ISSET(flags, MDB_RESERVE))
5954 data->mv_data = NODEDATA(leaf);
5955 else if (data->mv_size)
5956 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5958 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5961 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5962 mc->mc_db->md_entries--;
5964 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5971 nflags = flags & NODE_ADD_FLAGS;
5972 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5973 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5974 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5975 nflags &= ~MDB_APPEND;
5977 nflags |= MDB_SPLIT_REPLACE;
5978 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5980 /* There is room already in this leaf page. */
5981 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5982 if (rc == 0 && !do_sub && insert) {
5983 /* Adjust other cursors pointing to mp */
5984 MDB_cursor *m2, *m3;
5985 MDB_dbi dbi = mc->mc_dbi;
5986 unsigned i = mc->mc_top;
5987 MDB_page *mp = mc->mc_pg[i];
5989 if (mc->mc_flags & C_SUB)
5992 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5993 if (mc->mc_flags & C_SUB)
5994 m3 = &m2->mc_xcursor->mx_cursor;
5997 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5998 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6005 if (rc != MDB_SUCCESS)
6006 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6008 /* Now store the actual data in the child DB. Note that we're
6009 * storing the user data in the keys field, so there are strict
6010 * size limits on dupdata. The actual data fields of the child
6011 * DB are all zero size.
6018 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6019 if (flags & MDB_CURRENT) {
6020 xflags = MDB_CURRENT|MDB_NOSPILL;
6022 mdb_xcursor_init1(mc, leaf);
6023 xflags = (flags & MDB_NODUPDATA) ?
6024 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6026 /* converted, write the original data first */
6028 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6032 /* Adjust other cursors pointing to mp */
6034 unsigned i = mc->mc_top;
6035 MDB_page *mp = mc->mc_pg[i];
6037 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6038 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6039 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6040 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6041 mdb_xcursor_init1(m2, leaf);
6045 /* we've done our job */
6048 if (flags & MDB_APPENDDUP)
6049 xflags |= MDB_APPEND;
6050 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6051 if (flags & F_SUBDATA) {
6052 void *db = NODEDATA(leaf);
6053 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6056 /* sub-writes might have failed so check rc again.
6057 * Don't increment count if we just replaced an existing item.
6059 if (!rc && !(flags & MDB_CURRENT))
6060 mc->mc_db->md_entries++;
6061 if (flags & MDB_MULTIPLE) {
6065 /* let caller know how many succeeded, if any */
6066 data[1].mv_size = mcount;
6067 if (mcount < dcount) {
6068 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6069 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6076 /* If we succeeded and the key didn't exist before, make sure
6077 * the cursor is marked valid.
6080 mc->mc_flags |= C_INITIALIZED;
6085 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6090 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6091 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6093 if (!(mc->mc_flags & C_INITIALIZED))
6096 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6098 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
6100 rc = mdb_cursor_touch(mc);
6104 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6106 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6107 if (!(flags & MDB_NODUPDATA)) {
6108 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6109 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6111 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6112 /* If sub-DB still has entries, we're done */
6113 if (mc->mc_xcursor->mx_db.md_entries) {
6114 if (leaf->mn_flags & F_SUBDATA) {
6115 /* update subDB info */
6116 void *db = NODEDATA(leaf);
6117 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6120 /* shrink fake page */
6121 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6122 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6123 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6124 /* fix other sub-DB cursors pointed at this fake page */
6125 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6126 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6127 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6128 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6129 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6132 mc->mc_db->md_entries--;
6133 mc->mc_flags |= C_DEL;
6136 /* otherwise fall thru and delete the sub-DB */
6139 if (leaf->mn_flags & F_SUBDATA) {
6140 /* add all the child DB's pages to the free list */
6141 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6142 if (rc == MDB_SUCCESS) {
6143 mc->mc_db->md_entries -=
6144 mc->mc_xcursor->mx_db.md_entries;
6149 return mdb_cursor_del0(mc, leaf);
6152 /** Allocate and initialize new pages for a database.
6153 * @param[in] mc a cursor on the database being added to.
6154 * @param[in] flags flags defining what type of page is being allocated.
6155 * @param[in] num the number of pages to allocate. This is usually 1,
6156 * unless allocating overflow pages for a large record.
6157 * @param[out] mp Address of a page, or NULL on failure.
6158 * @return 0 on success, non-zero on failure.
6161 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6166 if ((rc = mdb_page_alloc(mc, num, &np)))
6168 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6169 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6170 np->mp_flags = flags | P_DIRTY;
6171 np->mp_lower = PAGEHDRSZ;
6172 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6175 mc->mc_db->md_branch_pages++;
6176 else if (IS_LEAF(np))
6177 mc->mc_db->md_leaf_pages++;
6178 else if (IS_OVERFLOW(np)) {
6179 mc->mc_db->md_overflow_pages += num;
6187 /** Calculate the size of a leaf node.
6188 * The size depends on the environment's page size; if a data item
6189 * is too large it will be put onto an overflow page and the node
6190 * size will only include the key and not the data. Sizes are always
6191 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6192 * of the #MDB_node headers.
6193 * @param[in] env The environment handle.
6194 * @param[in] key The key for the node.
6195 * @param[in] data The data for the node.
6196 * @return The number of bytes needed to store the node.
6199 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6203 sz = LEAFSIZE(key, data);
6204 if (sz >= env->me_nodemax) {
6205 /* put on overflow page */
6206 sz -= data->mv_size - sizeof(pgno_t);
6210 return sz + sizeof(indx_t);
6213 /** Calculate the size of a branch node.
6214 * The size should depend on the environment's page size but since
6215 * we currently don't support spilling large keys onto overflow
6216 * pages, it's simply the size of the #MDB_node header plus the
6217 * size of the key. Sizes are always rounded up to an even number
6218 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6219 * @param[in] env The environment handle.
6220 * @param[in] key The key for the node.
6221 * @return The number of bytes needed to store the node.
6224 mdb_branch_size(MDB_env *env, MDB_val *key)
6229 if (sz >= env->me_nodemax) {
6230 /* put on overflow page */
6231 /* not implemented */
6232 /* sz -= key->size - sizeof(pgno_t); */
6235 return sz + sizeof(indx_t);
6238 /** Add a node to the page pointed to by the cursor.
6239 * @param[in] mc The cursor for this operation.
6240 * @param[in] indx The index on the page where the new node should be added.
6241 * @param[in] key The key for the new node.
6242 * @param[in] data The data for the new node, if any.
6243 * @param[in] pgno The page number, if adding a branch node.
6244 * @param[in] flags Flags for the node.
6245 * @return 0 on success, non-zero on failure. Possible errors are:
6247 * <li>ENOMEM - failed to allocate overflow pages for the node.
6248 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6249 * should never happen since all callers already calculate the
6250 * page's free space before calling this function.
6254 mdb_node_add(MDB_cursor *mc, indx_t indx,
6255 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6258 size_t node_size = NODESIZE;
6261 MDB_page *mp = mc->mc_pg[mc->mc_top];
6262 MDB_page *ofp = NULL; /* overflow page */
6265 assert(mp->mp_upper >= mp->mp_lower);
6267 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6268 IS_LEAF(mp) ? "leaf" : "branch",
6269 IS_SUBP(mp) ? "sub-" : "",
6270 mp->mp_pgno, indx, data ? data->mv_size : 0,
6271 key ? key->mv_size : 0, key ? DKEY(key) : NULL));
6274 /* Move higher keys up one slot. */
6275 int ksize = mc->mc_db->md_pad, dif;
6276 char *ptr = LEAF2KEY(mp, indx, ksize);
6277 dif = NUMKEYS(mp) - indx;
6279 memmove(ptr+ksize, ptr, dif*ksize);
6280 /* insert new key */
6281 memcpy(ptr, key->mv_data, ksize);
6283 /* Just using these for counting */
6284 mp->mp_lower += sizeof(indx_t);
6285 mp->mp_upper -= ksize - sizeof(indx_t);
6290 node_size += key->mv_size;
6294 if (F_ISSET(flags, F_BIGDATA)) {
6295 /* Data already on overflow page. */
6296 node_size += sizeof(pgno_t);
6297 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6298 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6300 /* Put data on overflow page. */
6301 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6302 data->mv_size, node_size+data->mv_size));
6303 node_size += sizeof(pgno_t);
6304 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6306 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6309 node_size += data->mv_size;
6312 node_size += node_size & 1;
6314 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6315 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6316 mp->mp_pgno, NUMKEYS(mp)));
6317 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6318 mp->mp_upper - mp->mp_lower));
6319 DPRINTF(("node size = %"Z"u", node_size));
6320 return MDB_PAGE_FULL;
6323 /* Move higher pointers up one slot. */
6324 for (i = NUMKEYS(mp); i > indx; i--)
6325 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6327 /* Adjust free space offsets. */
6328 ofs = mp->mp_upper - node_size;
6329 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6330 mp->mp_ptrs[indx] = ofs;
6332 mp->mp_lower += sizeof(indx_t);
6334 /* Write the node data. */
6335 node = NODEPTR(mp, indx);
6336 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6337 node->mn_flags = flags;
6339 SETDSZ(node,data->mv_size);
6344 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6349 if (F_ISSET(flags, F_BIGDATA))
6350 memcpy(node->mn_data + key->mv_size, data->mv_data,
6352 else if (F_ISSET(flags, MDB_RESERVE))
6353 data->mv_data = node->mn_data + key->mv_size;
6355 memcpy(node->mn_data + key->mv_size, data->mv_data,
6358 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6360 if (F_ISSET(flags, MDB_RESERVE))
6361 data->mv_data = METADATA(ofp);
6363 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6370 /** Delete the specified node from a page.
6371 * @param[in] mp The page to operate on.
6372 * @param[in] indx The index of the node to delete.
6373 * @param[in] ksize The size of a node. Only used if the page is
6374 * part of a #MDB_DUPFIXED database.
6377 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6380 indx_t i, j, numkeys, ptr;
6387 COPY_PGNO(pgno, mp->mp_pgno);
6388 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6389 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6392 assert(indx < NUMKEYS(mp));
6395 int x = NUMKEYS(mp) - 1 - indx;
6396 base = LEAF2KEY(mp, indx, ksize);
6398 memmove(base, base + ksize, x * ksize);
6399 mp->mp_lower -= sizeof(indx_t);
6400 mp->mp_upper += ksize - sizeof(indx_t);
6404 node = NODEPTR(mp, indx);
6405 sz = NODESIZE + node->mn_ksize;
6407 if (F_ISSET(node->mn_flags, F_BIGDATA))
6408 sz += sizeof(pgno_t);
6410 sz += NODEDSZ(node);
6414 ptr = mp->mp_ptrs[indx];
6415 numkeys = NUMKEYS(mp);
6416 for (i = j = 0; i < numkeys; i++) {
6418 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6419 if (mp->mp_ptrs[i] < ptr)
6420 mp->mp_ptrs[j] += sz;
6425 base = (char *)mp + mp->mp_upper;
6426 memmove(base + sz, base, ptr - mp->mp_upper);
6428 mp->mp_lower -= sizeof(indx_t);
6432 /** Compact the main page after deleting a node on a subpage.
6433 * @param[in] mp The main page to operate on.
6434 * @param[in] indx The index of the subpage on the main page.
6437 mdb_node_shrink(MDB_page *mp, indx_t indx)
6444 indx_t i, numkeys, ptr;
6446 node = NODEPTR(mp, indx);
6447 sp = (MDB_page *)NODEDATA(node);
6448 osize = NODEDSZ(node);
6450 delta = sp->mp_upper - sp->mp_lower;
6451 SETDSZ(node, osize - delta);
6452 xp = (MDB_page *)((char *)sp + delta);
6454 /* shift subpage upward */
6456 nsize = NUMKEYS(sp) * sp->mp_pad;
6457 memmove(METADATA(xp), METADATA(sp), nsize);
6460 nsize = osize - sp->mp_upper;
6461 numkeys = NUMKEYS(sp);
6462 for (i=numkeys-1; i>=0; i--)
6463 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6465 xp->mp_upper = sp->mp_lower;
6466 xp->mp_lower = sp->mp_lower;
6467 xp->mp_flags = sp->mp_flags;
6468 xp->mp_pad = sp->mp_pad;
6469 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6471 /* shift lower nodes upward */
6472 ptr = mp->mp_ptrs[indx];
6473 numkeys = NUMKEYS(mp);
6474 for (i = 0; i < numkeys; i++) {
6475 if (mp->mp_ptrs[i] <= ptr)
6476 mp->mp_ptrs[i] += delta;
6479 base = (char *)mp + mp->mp_upper;
6480 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6481 mp->mp_upper += delta;
6484 /** Initial setup of a sorted-dups cursor.
6485 * Sorted duplicates are implemented as a sub-database for the given key.
6486 * The duplicate data items are actually keys of the sub-database.
6487 * Operations on the duplicate data items are performed using a sub-cursor
6488 * initialized when the sub-database is first accessed. This function does
6489 * the preliminary setup of the sub-cursor, filling in the fields that
6490 * depend only on the parent DB.
6491 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6494 mdb_xcursor_init0(MDB_cursor *mc)
6496 MDB_xcursor *mx = mc->mc_xcursor;
6498 mx->mx_cursor.mc_xcursor = NULL;
6499 mx->mx_cursor.mc_txn = mc->mc_txn;
6500 mx->mx_cursor.mc_db = &mx->mx_db;
6501 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6502 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6503 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6504 mx->mx_cursor.mc_snum = 0;
6505 mx->mx_cursor.mc_top = 0;
6506 mx->mx_cursor.mc_flags = C_SUB;
6507 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6508 mx->mx_dbx.md_dcmp = NULL;
6509 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6512 /** Final setup of a sorted-dups cursor.
6513 * Sets up the fields that depend on the data from the main cursor.
6514 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6515 * @param[in] node The data containing the #MDB_db record for the
6516 * sorted-dup database.
6519 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6521 MDB_xcursor *mx = mc->mc_xcursor;
6523 if (node->mn_flags & F_SUBDATA) {
6524 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6525 mx->mx_cursor.mc_pg[0] = 0;
6526 mx->mx_cursor.mc_snum = 0;
6527 mx->mx_cursor.mc_flags = C_SUB;
6529 MDB_page *fp = NODEDATA(node);
6530 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6531 mx->mx_db.md_flags = 0;
6532 mx->mx_db.md_depth = 1;
6533 mx->mx_db.md_branch_pages = 0;
6534 mx->mx_db.md_leaf_pages = 1;
6535 mx->mx_db.md_overflow_pages = 0;
6536 mx->mx_db.md_entries = NUMKEYS(fp);
6537 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6538 mx->mx_cursor.mc_snum = 1;
6539 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6540 mx->mx_cursor.mc_top = 0;
6541 mx->mx_cursor.mc_pg[0] = fp;
6542 mx->mx_cursor.mc_ki[0] = 0;
6543 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6544 mx->mx_db.md_flags = MDB_DUPFIXED;
6545 mx->mx_db.md_pad = fp->mp_pad;
6546 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6547 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6550 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6551 mx->mx_db.md_root));
6552 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6554 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6555 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6556 #if UINT_MAX < SIZE_MAX
6557 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6558 #ifdef MISALIGNED_OK
6559 mx->mx_dbx.md_cmp = mdb_cmp_long;
6561 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6566 /** Initialize a cursor for a given transaction and database. */
6568 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6571 mc->mc_backup = NULL;
6574 mc->mc_db = &txn->mt_dbs[dbi];
6575 mc->mc_dbx = &txn->mt_dbxs[dbi];
6576 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6581 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6583 mc->mc_xcursor = mx;
6584 mdb_xcursor_init0(mc);
6586 mc->mc_xcursor = NULL;
6588 if (*mc->mc_dbflag & DB_STALE) {
6589 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6594 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6597 size_t size = sizeof(MDB_cursor);
6599 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6602 if (txn->mt_flags & MDB_TXN_ERROR)
6605 /* Allow read access to the freelist */
6606 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6609 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6610 size += sizeof(MDB_xcursor);
6612 if ((mc = malloc(size)) != NULL) {
6613 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6614 if (txn->mt_cursors) {
6615 mc->mc_next = txn->mt_cursors[dbi];
6616 txn->mt_cursors[dbi] = mc;
6617 mc->mc_flags |= C_UNTRACK;
6629 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6631 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6634 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6637 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6641 /* Return the count of duplicate data items for the current key */
6643 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6647 if (mc == NULL || countp == NULL)
6650 if (mc->mc_xcursor == NULL)
6651 return MDB_INCOMPATIBLE;
6653 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6654 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6657 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6660 *countp = mc->mc_xcursor->mx_db.md_entries;
6666 mdb_cursor_close(MDB_cursor *mc)
6668 if (mc && !mc->mc_backup) {
6669 /* remove from txn, if tracked */
6670 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6671 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6672 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6674 *prev = mc->mc_next;
6681 mdb_cursor_txn(MDB_cursor *mc)
6683 if (!mc) return NULL;
6688 mdb_cursor_dbi(MDB_cursor *mc)
6694 /** Replace the key for a node with a new key.
6695 * @param[in] mc Cursor pointing to the node to operate on.
6696 * @param[in] key The new key to use.
6697 * @return 0 on success, non-zero on failure.
6700 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6707 indx_t ptr, i, numkeys, indx;
6710 indx = mc->mc_ki[mc->mc_top];
6711 mp = mc->mc_pg[mc->mc_top];
6712 node = NODEPTR(mp, indx);
6713 ptr = mp->mp_ptrs[indx];
6717 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6718 k2.mv_data = NODEKEY(node);
6719 k2.mv_size = node->mn_ksize;
6720 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6722 mdb_dkey(&k2, kbuf2),
6728 delta0 = delta = key->mv_size - node->mn_ksize;
6730 /* Must be 2-byte aligned. If new key is
6731 * shorter by 1, the shift will be skipped.
6733 delta += (delta & 1);
6735 if (delta > 0 && SIZELEFT(mp) < delta) {
6737 /* not enough space left, do a delete and split */
6738 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6739 pgno = NODEPGNO(node);
6740 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6741 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6744 numkeys = NUMKEYS(mp);
6745 for (i = 0; i < numkeys; i++) {
6746 if (mp->mp_ptrs[i] <= ptr)
6747 mp->mp_ptrs[i] -= delta;
6750 base = (char *)mp + mp->mp_upper;
6751 len = ptr - mp->mp_upper + NODESIZE;
6752 memmove(base - delta, base, len);
6753 mp->mp_upper -= delta;
6755 node = NODEPTR(mp, indx);
6758 /* But even if no shift was needed, update ksize */
6760 node->mn_ksize = key->mv_size;
6763 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6769 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6771 /** Move a node from csrc to cdst.
6774 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6781 unsigned short flags;
6785 /* Mark src and dst as dirty. */
6786 if ((rc = mdb_page_touch(csrc)) ||
6787 (rc = mdb_page_touch(cdst)))
6790 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6791 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6792 key.mv_size = csrc->mc_db->md_pad;
6793 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6795 data.mv_data = NULL;
6799 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6800 assert(!((long)srcnode&1));
6801 srcpg = NODEPGNO(srcnode);
6802 flags = srcnode->mn_flags;
6803 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6804 unsigned int snum = csrc->mc_snum;
6806 /* must find the lowest key below src */
6807 mdb_page_search_lowest(csrc);
6808 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6809 key.mv_size = csrc->mc_db->md_pad;
6810 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6812 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6813 key.mv_size = NODEKSZ(s2);
6814 key.mv_data = NODEKEY(s2);
6816 csrc->mc_snum = snum--;
6817 csrc->mc_top = snum;
6819 key.mv_size = NODEKSZ(srcnode);
6820 key.mv_data = NODEKEY(srcnode);
6822 data.mv_size = NODEDSZ(srcnode);
6823 data.mv_data = NODEDATA(srcnode);
6825 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6826 unsigned int snum = cdst->mc_snum;
6829 /* must find the lowest key below dst */
6830 mdb_page_search_lowest(cdst);
6831 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6832 bkey.mv_size = cdst->mc_db->md_pad;
6833 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6835 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6836 bkey.mv_size = NODEKSZ(s2);
6837 bkey.mv_data = NODEKEY(s2);
6839 cdst->mc_snum = snum--;
6840 cdst->mc_top = snum;
6841 mdb_cursor_copy(cdst, &mn);
6843 rc = mdb_update_key(&mn, &bkey);
6848 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6849 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6850 csrc->mc_ki[csrc->mc_top],
6852 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6853 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6855 /* Add the node to the destination page.
6857 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6858 if (rc != MDB_SUCCESS)
6861 /* Delete the node from the source page.
6863 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6866 /* Adjust other cursors pointing to mp */
6867 MDB_cursor *m2, *m3;
6868 MDB_dbi dbi = csrc->mc_dbi;
6869 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6871 if (csrc->mc_flags & C_SUB)
6874 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6875 if (csrc->mc_flags & C_SUB)
6876 m3 = &m2->mc_xcursor->mx_cursor;
6879 if (m3 == csrc) continue;
6880 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6881 csrc->mc_ki[csrc->mc_top]) {
6882 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6883 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6888 /* Update the parent separators.
6890 if (csrc->mc_ki[csrc->mc_top] == 0) {
6891 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6892 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6893 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6895 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6896 key.mv_size = NODEKSZ(srcnode);
6897 key.mv_data = NODEKEY(srcnode);
6899 DPRINTF(("update separator for source page %"Z"u to [%s]",
6900 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6901 mdb_cursor_copy(csrc, &mn);
6904 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6907 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6909 indx_t ix = csrc->mc_ki[csrc->mc_top];
6910 nullkey.mv_size = 0;
6911 csrc->mc_ki[csrc->mc_top] = 0;
6912 rc = mdb_update_key(csrc, &nullkey);
6913 csrc->mc_ki[csrc->mc_top] = ix;
6914 assert(rc == MDB_SUCCESS);
6918 if (cdst->mc_ki[cdst->mc_top] == 0) {
6919 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6920 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6921 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6923 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6924 key.mv_size = NODEKSZ(srcnode);
6925 key.mv_data = NODEKEY(srcnode);
6927 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6928 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6929 mdb_cursor_copy(cdst, &mn);
6932 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6935 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6937 indx_t ix = cdst->mc_ki[cdst->mc_top];
6938 nullkey.mv_size = 0;
6939 cdst->mc_ki[cdst->mc_top] = 0;
6940 rc = mdb_update_key(cdst, &nullkey);
6941 cdst->mc_ki[cdst->mc_top] = ix;
6942 assert(rc == MDB_SUCCESS);
6949 /** Merge one page into another.
6950 * The nodes from the page pointed to by \b csrc will
6951 * be copied to the page pointed to by \b cdst and then
6952 * the \b csrc page will be freed.
6953 * @param[in] csrc Cursor pointing to the source page.
6954 * @param[in] cdst Cursor pointing to the destination page.
6957 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6965 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6966 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6968 assert(csrc->mc_snum > 1); /* can't merge root page */
6969 assert(cdst->mc_snum > 1);
6971 /* Mark dst as dirty. */
6972 if ((rc = mdb_page_touch(cdst)))
6975 /* Move all nodes from src to dst.
6977 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6978 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6979 key.mv_size = csrc->mc_db->md_pad;
6980 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6981 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6982 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6983 if (rc != MDB_SUCCESS)
6985 key.mv_data = (char *)key.mv_data + key.mv_size;
6988 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6989 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6990 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6991 unsigned int snum = csrc->mc_snum;
6993 /* must find the lowest key below src */
6994 mdb_page_search_lowest(csrc);
6995 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6996 key.mv_size = csrc->mc_db->md_pad;
6997 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6999 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7000 key.mv_size = NODEKSZ(s2);
7001 key.mv_data = NODEKEY(s2);
7003 csrc->mc_snum = snum--;
7004 csrc->mc_top = snum;
7006 key.mv_size = srcnode->mn_ksize;
7007 key.mv_data = NODEKEY(srcnode);
7010 data.mv_size = NODEDSZ(srcnode);
7011 data.mv_data = NODEDATA(srcnode);
7012 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7013 if (rc != MDB_SUCCESS)
7018 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7019 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7020 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7022 /* Unlink the src page from parent and add to free list.
7024 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7025 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7028 rc = mdb_update_key(csrc, &key);
7034 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7035 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7038 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7039 csrc->mc_db->md_leaf_pages--;
7041 csrc->mc_db->md_branch_pages--;
7043 /* Adjust other cursors pointing to mp */
7044 MDB_cursor *m2, *m3;
7045 MDB_dbi dbi = csrc->mc_dbi;
7046 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7048 if (csrc->mc_flags & C_SUB)
7051 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7052 if (csrc->mc_flags & C_SUB)
7053 m3 = &m2->mc_xcursor->mx_cursor;
7056 if (m3 == csrc) continue;
7057 if (m3->mc_snum < csrc->mc_snum) continue;
7058 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7059 m3->mc_pg[csrc->mc_top] = mp;
7060 m3->mc_ki[csrc->mc_top] += nkeys;
7064 mdb_cursor_pop(csrc);
7066 return mdb_rebalance(csrc);
7069 /** Copy the contents of a cursor.
7070 * @param[in] csrc The cursor to copy from.
7071 * @param[out] cdst The cursor to copy to.
7074 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7078 cdst->mc_txn = csrc->mc_txn;
7079 cdst->mc_dbi = csrc->mc_dbi;
7080 cdst->mc_db = csrc->mc_db;
7081 cdst->mc_dbx = csrc->mc_dbx;
7082 cdst->mc_snum = csrc->mc_snum;
7083 cdst->mc_top = csrc->mc_top;
7084 cdst->mc_flags = csrc->mc_flags;
7086 for (i=0; i<csrc->mc_snum; i++) {
7087 cdst->mc_pg[i] = csrc->mc_pg[i];
7088 cdst->mc_ki[i] = csrc->mc_ki[i];
7092 /** Rebalance the tree after a delete operation.
7093 * @param[in] mc Cursor pointing to the page where rebalancing
7095 * @return 0 on success, non-zero on failure.
7098 mdb_rebalance(MDB_cursor *mc)
7102 unsigned int ptop, minkeys;
7105 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7109 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7110 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7111 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7112 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7113 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7117 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7118 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7121 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7122 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7128 if (mc->mc_snum < 2) {
7129 MDB_page *mp = mc->mc_pg[0];
7131 DPUTS("Can't rebalance a subpage, ignoring");
7134 if (NUMKEYS(mp) == 0) {
7135 DPUTS("tree is completely empty");
7136 mc->mc_db->md_root = P_INVALID;
7137 mc->mc_db->md_depth = 0;
7138 mc->mc_db->md_leaf_pages = 0;
7139 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7142 /* Adjust cursors pointing to mp */
7146 MDB_cursor *m2, *m3;
7147 MDB_dbi dbi = mc->mc_dbi;
7149 if (mc->mc_flags & C_SUB)
7152 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7153 if (mc->mc_flags & C_SUB)
7154 m3 = &m2->mc_xcursor->mx_cursor;
7157 if (m3->mc_snum < mc->mc_snum) continue;
7158 if (m3->mc_pg[0] == mp) {
7164 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7165 DPUTS("collapsing root page!");
7166 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7169 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7170 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7173 mc->mc_db->md_depth--;
7174 mc->mc_db->md_branch_pages--;
7175 mc->mc_ki[0] = mc->mc_ki[1];
7177 /* Adjust other cursors pointing to mp */
7178 MDB_cursor *m2, *m3;
7179 MDB_dbi dbi = mc->mc_dbi;
7181 if (mc->mc_flags & C_SUB)
7184 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7185 if (mc->mc_flags & C_SUB)
7186 m3 = &m2->mc_xcursor->mx_cursor;
7189 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7190 if (m3->mc_pg[0] == mp) {
7191 m3->mc_pg[0] = mc->mc_pg[0];
7194 m3->mc_ki[0] = m3->mc_ki[1];
7199 DPUTS("root page doesn't need rebalancing");
7203 /* The parent (branch page) must have at least 2 pointers,
7204 * otherwise the tree is invalid.
7206 ptop = mc->mc_top-1;
7207 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7209 /* Leaf page fill factor is below the threshold.
7210 * Try to move keys from left or right neighbor, or
7211 * merge with a neighbor page.
7216 mdb_cursor_copy(mc, &mn);
7217 mn.mc_xcursor = NULL;
7219 if (mc->mc_ki[ptop] == 0) {
7220 /* We're the leftmost leaf in our parent.
7222 DPUTS("reading right neighbor");
7224 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7225 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7228 mn.mc_ki[mn.mc_top] = 0;
7229 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7231 /* There is at least one neighbor to the left.
7233 DPUTS("reading left neighbor");
7235 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7236 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7239 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7240 mc->mc_ki[mc->mc_top] = 0;
7243 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7244 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7245 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7247 /* If the neighbor page is above threshold and has enough keys,
7248 * move one key from it. Otherwise we should try to merge them.
7249 * (A branch page must never have less than 2 keys.)
7251 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7252 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7253 return mdb_node_move(&mn, mc);
7255 if (mc->mc_ki[ptop] == 0)
7256 rc = mdb_page_merge(&mn, mc);
7258 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7259 rc = mdb_page_merge(mc, &mn);
7260 mdb_cursor_copy(&mn, mc);
7262 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7267 /** Complete a delete operation started by #mdb_cursor_del(). */
7269 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7276 mp = mc->mc_pg[mc->mc_top];
7277 ki = mc->mc_ki[mc->mc_top];
7279 /* add overflow pages to free list */
7280 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7284 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7285 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7286 (rc = mdb_ovpage_free(mc, omp)))
7289 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7290 mc->mc_db->md_entries--;
7291 rc = mdb_rebalance(mc);
7292 if (rc != MDB_SUCCESS)
7293 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7296 MDB_dbi dbi = mc->mc_dbi;
7298 mp = mc->mc_pg[mc->mc_top];
7299 nkeys = NUMKEYS(mp);
7301 /* if mc points past last node in page, find next sibling */
7302 if (mc->mc_ki[mc->mc_top] >= nkeys)
7303 mdb_cursor_sibling(mc, 1);
7305 /* Adjust other cursors pointing to mp */
7306 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7309 if (!(m2->mc_flags & C_INITIALIZED))
7311 if (m2->mc_pg[mc->mc_top] == mp) {
7312 if (m2->mc_ki[mc->mc_top] >= ki) {
7313 m2->mc_flags |= C_DEL;
7314 if (m2->mc_ki[mc->mc_top] > ki)
7315 m2->mc_ki[mc->mc_top]--;
7317 if (m2->mc_ki[mc->mc_top] >= nkeys)
7318 mdb_cursor_sibling(m2, 1);
7321 mc->mc_flags |= C_DEL;
7328 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7329 MDB_val *key, MDB_val *data)
7334 MDB_val rdata, *xdata;
7338 assert(key != NULL);
7340 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7342 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7345 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7346 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7348 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7349 return MDB_BAD_VALSIZE;
7352 mdb_cursor_init(&mc, txn, dbi, &mx);
7355 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7356 /* must ignore any data */
7367 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7369 /* let mdb_page_split know about this cursor if needed:
7370 * delete will trigger a rebalance; if it needs to move
7371 * a node from one page to another, it will have to
7372 * update the parent's separator key(s). If the new sepkey
7373 * is larger than the current one, the parent page may
7374 * run out of space, triggering a split. We need this
7375 * cursor to be consistent until the end of the rebalance.
7377 mc.mc_flags |= C_UNTRACK;
7378 mc.mc_next = txn->mt_cursors[dbi];
7379 txn->mt_cursors[dbi] = &mc;
7380 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7381 txn->mt_cursors[dbi] = mc.mc_next;
7386 /** Split a page and insert a new node.
7387 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7388 * The cursor will be updated to point to the actual page and index where
7389 * the node got inserted after the split.
7390 * @param[in] newkey The key for the newly inserted node.
7391 * @param[in] newdata The data for the newly inserted node.
7392 * @param[in] newpgno The page number, if the new node is a branch node.
7393 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7394 * @return 0 on success, non-zero on failure.
7397 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7398 unsigned int nflags)
7401 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7404 unsigned int i, j, split_indx, nkeys, pmax;
7406 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7408 MDB_page *mp, *rp, *pp;
7413 mp = mc->mc_pg[mc->mc_top];
7414 newindx = mc->mc_ki[mc->mc_top];
7416 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7417 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7418 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7420 /* Create a right sibling. */
7421 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7423 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7425 if (mc->mc_snum < 2) {
7426 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7428 /* shift current top to make room for new parent */
7429 mc->mc_pg[1] = mc->mc_pg[0];
7430 mc->mc_ki[1] = mc->mc_ki[0];
7433 mc->mc_db->md_root = pp->mp_pgno;
7434 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7435 mc->mc_db->md_depth++;
7438 /* Add left (implicit) pointer. */
7439 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7440 /* undo the pre-push */
7441 mc->mc_pg[0] = mc->mc_pg[1];
7442 mc->mc_ki[0] = mc->mc_ki[1];
7443 mc->mc_db->md_root = mp->mp_pgno;
7444 mc->mc_db->md_depth--;
7451 ptop = mc->mc_top-1;
7452 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7455 mc->mc_flags |= C_SPLITTING;
7456 mdb_cursor_copy(mc, &mn);
7457 mn.mc_pg[mn.mc_top] = rp;
7458 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7460 if (nflags & MDB_APPEND) {
7461 mn.mc_ki[mn.mc_top] = 0;
7463 split_indx = newindx;
7468 nkeys = NUMKEYS(mp);
7469 split_indx = nkeys / 2;
7470 if (newindx < split_indx)
7476 unsigned int lsize, rsize, ksize;
7477 /* Move half of the keys to the right sibling */
7479 x = mc->mc_ki[mc->mc_top] - split_indx;
7480 ksize = mc->mc_db->md_pad;
7481 split = LEAF2KEY(mp, split_indx, ksize);
7482 rsize = (nkeys - split_indx) * ksize;
7483 lsize = (nkeys - split_indx) * sizeof(indx_t);
7484 mp->mp_lower -= lsize;
7485 rp->mp_lower += lsize;
7486 mp->mp_upper += rsize - lsize;
7487 rp->mp_upper -= rsize - lsize;
7488 sepkey.mv_size = ksize;
7489 if (newindx == split_indx) {
7490 sepkey.mv_data = newkey->mv_data;
7492 sepkey.mv_data = split;
7495 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7496 memcpy(rp->mp_ptrs, split, rsize);
7497 sepkey.mv_data = rp->mp_ptrs;
7498 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7499 memcpy(ins, newkey->mv_data, ksize);
7500 mp->mp_lower += sizeof(indx_t);
7501 mp->mp_upper -= ksize - sizeof(indx_t);
7504 memcpy(rp->mp_ptrs, split, x * ksize);
7505 ins = LEAF2KEY(rp, x, ksize);
7506 memcpy(ins, newkey->mv_data, ksize);
7507 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7508 rp->mp_lower += sizeof(indx_t);
7509 rp->mp_upper -= ksize - sizeof(indx_t);
7510 mc->mc_ki[mc->mc_top] = x;
7511 mc->mc_pg[mc->mc_top] = rp;
7516 /* For leaf pages, check the split point based on what
7517 * fits where, since otherwise mdb_node_add can fail.
7519 * This check is only needed when the data items are
7520 * relatively large, such that being off by one will
7521 * make the difference between success or failure.
7523 * It's also relevant if a page happens to be laid out
7524 * such that one half of its nodes are all "small" and
7525 * the other half of its nodes are "large." If the new
7526 * item is also "large" and falls on the half with
7527 * "large" nodes, it also may not fit.
7530 unsigned int psize, nsize;
7531 /* Maximum free space in an empty page */
7532 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7533 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7534 if ((nkeys < 20) || (nsize > pmax/16)) {
7535 if (newindx <= split_indx) {
7538 for (i=0; i<split_indx; i++) {
7539 node = NODEPTR(mp, i);
7540 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7541 if (F_ISSET(node->mn_flags, F_BIGDATA))
7542 psize += sizeof(pgno_t);
7544 psize += NODEDSZ(node);
7548 split_indx = newindx;
7559 for (i=nkeys-1; i>=split_indx; i--) {
7560 node = NODEPTR(mp, i);
7561 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7562 if (F_ISSET(node->mn_flags, F_BIGDATA))
7563 psize += sizeof(pgno_t);
7565 psize += NODEDSZ(node);
7569 split_indx = newindx;
7580 /* First find the separating key between the split pages.
7581 * The case where newindx == split_indx is ambiguous; the
7582 * new item could go to the new page or stay on the original
7583 * page. If newpos == 1 it goes to the new page.
7585 if (newindx == split_indx && newpos) {
7586 sepkey.mv_size = newkey->mv_size;
7587 sepkey.mv_data = newkey->mv_data;
7589 node = NODEPTR(mp, split_indx);
7590 sepkey.mv_size = node->mn_ksize;
7591 sepkey.mv_data = NODEKEY(node);
7595 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7597 /* Copy separator key to the parent.
7599 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7603 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7606 if (mn.mc_snum == mc->mc_snum) {
7607 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7608 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7609 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7610 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7615 /* Right page might now have changed parent.
7616 * Check if left page also changed parent.
7618 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7619 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7620 for (i=0; i<ptop; i++) {
7621 mc->mc_pg[i] = mn.mc_pg[i];
7622 mc->mc_ki[i] = mn.mc_ki[i];
7624 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7625 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7629 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7632 mc->mc_flags ^= C_SPLITTING;
7633 if (rc != MDB_SUCCESS) {
7636 if (nflags & MDB_APPEND) {
7637 mc->mc_pg[mc->mc_top] = rp;
7638 mc->mc_ki[mc->mc_top] = 0;
7639 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7642 for (i=0; i<mc->mc_top; i++)
7643 mc->mc_ki[i] = mn.mc_ki[i];
7650 /* Move half of the keys to the right sibling. */
7652 /* grab a page to hold a temporary copy */
7653 copy = mdb_page_malloc(mc->mc_txn, 1);
7657 copy->mp_pgno = mp->mp_pgno;
7658 copy->mp_flags = mp->mp_flags;
7659 copy->mp_lower = PAGEHDRSZ;
7660 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7661 mc->mc_pg[mc->mc_top] = copy;
7662 for (i = j = 0; i <= nkeys; j++) {
7663 if (i == split_indx) {
7664 /* Insert in right sibling. */
7665 /* Reset insert index for right sibling. */
7666 if (i != newindx || (newpos ^ ins_new)) {
7668 mc->mc_pg[mc->mc_top] = rp;
7672 if (i == newindx && !ins_new) {
7673 /* Insert the original entry that caused the split. */
7674 rkey.mv_data = newkey->mv_data;
7675 rkey.mv_size = newkey->mv_size;
7684 /* Update index for the new key. */
7685 mc->mc_ki[mc->mc_top] = j;
7686 } else if (i == nkeys) {
7689 node = NODEPTR(mp, i);
7690 rkey.mv_data = NODEKEY(node);
7691 rkey.mv_size = node->mn_ksize;
7693 xdata.mv_data = NODEDATA(node);
7694 xdata.mv_size = NODEDSZ(node);
7697 pgno = NODEPGNO(node);
7698 flags = node->mn_flags;
7703 if (!IS_LEAF(mp) && j == 0) {
7704 /* First branch index doesn't need key data. */
7708 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7712 nkeys = NUMKEYS(copy);
7713 for (i=0; i<nkeys; i++)
7714 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7715 mp->mp_lower = copy->mp_lower;
7716 mp->mp_upper = copy->mp_upper;
7717 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7718 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7720 /* reset back to original page */
7721 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7722 mc->mc_pg[mc->mc_top] = mp;
7723 if (nflags & MDB_RESERVE) {
7724 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7725 if (!(node->mn_flags & F_BIGDATA))
7726 newdata->mv_data = NODEDATA(node);
7730 /* Make sure mc_ki is still valid.
7732 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7733 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7734 for (i=0; i<ptop; i++) {
7735 mc->mc_pg[i] = mn.mc_pg[i];
7736 mc->mc_ki[i] = mn.mc_ki[i];
7738 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7739 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7743 /* return tmp page to freelist */
7744 mdb_page_free(mc->mc_txn->mt_env, copy);
7747 /* Adjust other cursors pointing to mp */
7748 MDB_cursor *m2, *m3;
7749 MDB_dbi dbi = mc->mc_dbi;
7750 int fixup = NUMKEYS(mp);
7752 if (mc->mc_flags & C_SUB)
7755 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7756 if (mc->mc_flags & C_SUB)
7757 m3 = &m2->mc_xcursor->mx_cursor;
7762 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7764 if (m3->mc_flags & C_SPLITTING)
7769 for (k=m3->mc_top; k>=0; k--) {
7770 m3->mc_ki[k+1] = m3->mc_ki[k];
7771 m3->mc_pg[k+1] = m3->mc_pg[k];
7773 if (m3->mc_ki[0] >= split_indx) {
7778 m3->mc_pg[0] = mc->mc_pg[0];
7782 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7783 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7784 m3->mc_ki[mc->mc_top]++;
7785 if (m3->mc_ki[mc->mc_top] >= fixup) {
7786 m3->mc_pg[mc->mc_top] = rp;
7787 m3->mc_ki[mc->mc_top] -= fixup;
7788 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7790 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7791 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7800 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7801 MDB_val *key, MDB_val *data, unsigned int flags)
7806 assert(key != NULL);
7807 assert(data != NULL);
7809 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7812 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7813 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7815 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7816 return MDB_BAD_VALSIZE;
7819 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7822 mdb_cursor_init(&mc, txn, dbi, &mx);
7823 return mdb_cursor_put(&mc, key, data, flags);
7827 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7829 if ((flag & CHANGEABLE) != flag)
7832 env->me_flags |= flag;
7834 env->me_flags &= ~flag;
7839 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7844 *arg = env->me_flags;
7849 mdb_env_get_path(MDB_env *env, const char **arg)
7854 *arg = env->me_path;
7858 /** Common code for #mdb_stat() and #mdb_env_stat().
7859 * @param[in] env the environment to operate in.
7860 * @param[in] db the #MDB_db record containing the stats to return.
7861 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7862 * @return 0, this function always succeeds.
7865 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7867 arg->ms_psize = env->me_psize;
7868 arg->ms_depth = db->md_depth;
7869 arg->ms_branch_pages = db->md_branch_pages;
7870 arg->ms_leaf_pages = db->md_leaf_pages;
7871 arg->ms_overflow_pages = db->md_overflow_pages;
7872 arg->ms_entries = db->md_entries;
7877 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7881 if (env == NULL || arg == NULL)
7884 toggle = mdb_env_pick_meta(env);
7886 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7890 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7894 if (env == NULL || arg == NULL)
7897 toggle = mdb_env_pick_meta(env);
7898 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7899 arg->me_mapsize = env->me_mapsize;
7900 arg->me_maxreaders = env->me_maxreaders;
7902 /* me_numreaders may be zero if this process never used any readers. Use
7903 * the shared numreader count if it exists.
7905 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7907 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7908 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7912 /** Set the default comparison functions for a database.
7913 * Called immediately after a database is opened to set the defaults.
7914 * The user can then override them with #mdb_set_compare() or
7915 * #mdb_set_dupsort().
7916 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7917 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7920 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7922 uint16_t f = txn->mt_dbs[dbi].md_flags;
7924 txn->mt_dbxs[dbi].md_cmp =
7925 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7926 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7928 txn->mt_dbxs[dbi].md_dcmp =
7929 !(f & MDB_DUPSORT) ? 0 :
7930 ((f & MDB_INTEGERDUP)
7931 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7932 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7935 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7940 int rc, dbflag, exact;
7941 unsigned int unused = 0;
7944 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7945 mdb_default_cmp(txn, FREE_DBI);
7948 if ((flags & VALID_FLAGS) != flags)
7950 if (txn->mt_flags & MDB_TXN_ERROR)
7956 if (flags & PERSISTENT_FLAGS) {
7957 uint16_t f2 = flags & PERSISTENT_FLAGS;
7958 /* make sure flag changes get committed */
7959 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7960 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7961 txn->mt_flags |= MDB_TXN_DIRTY;
7964 mdb_default_cmp(txn, MAIN_DBI);
7968 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7969 mdb_default_cmp(txn, MAIN_DBI);
7972 /* Is the DB already open? */
7974 for (i=2; i<txn->mt_numdbs; i++) {
7975 if (!txn->mt_dbxs[i].md_name.mv_size) {
7976 /* Remember this free slot */
7977 if (!unused) unused = i;
7980 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7981 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7987 /* If no free slot and max hit, fail */
7988 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7989 return MDB_DBS_FULL;
7991 /* Cannot mix named databases with some mainDB flags */
7992 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7993 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7995 /* Find the DB info */
7996 dbflag = DB_NEW|DB_VALID;
7999 key.mv_data = (void *)name;
8000 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8001 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8002 if (rc == MDB_SUCCESS) {
8003 /* make sure this is actually a DB */
8004 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8005 if (!(node->mn_flags & F_SUBDATA))
8006 return MDB_INCOMPATIBLE;
8007 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8008 /* Create if requested */
8010 data.mv_size = sizeof(MDB_db);
8011 data.mv_data = &dummy;
8012 memset(&dummy, 0, sizeof(dummy));
8013 dummy.md_root = P_INVALID;
8014 dummy.md_flags = flags & PERSISTENT_FLAGS;
8015 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8019 /* OK, got info, add to table */
8020 if (rc == MDB_SUCCESS) {
8021 unsigned int slot = unused ? unused : txn->mt_numdbs;
8022 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8023 txn->mt_dbxs[slot].md_name.mv_size = len;
8024 txn->mt_dbxs[slot].md_rel = NULL;
8025 txn->mt_dbflags[slot] = dbflag;
8026 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8028 mdb_default_cmp(txn, slot);
8037 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8039 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8042 if (txn->mt_dbflags[dbi] & DB_STALE) {
8045 /* Stale, must read the DB's root. cursor_init does it for us. */
8046 mdb_cursor_init(&mc, txn, dbi, &mx);
8048 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8051 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8054 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8056 ptr = env->me_dbxs[dbi].md_name.mv_data;
8057 env->me_dbxs[dbi].md_name.mv_data = NULL;
8058 env->me_dbxs[dbi].md_name.mv_size = 0;
8059 env->me_dbflags[dbi] = 0;
8063 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8065 /* We could return the flags for the FREE_DBI too but what's the point? */
8066 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8068 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8072 /** Add all the DB's pages to the free list.
8073 * @param[in] mc Cursor on the DB to free.
8074 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8075 * @return 0 on success, non-zero on failure.
8078 mdb_drop0(MDB_cursor *mc, int subs)
8082 rc = mdb_page_search(mc, NULL, 0);
8083 if (rc == MDB_SUCCESS) {
8084 MDB_txn *txn = mc->mc_txn;
8089 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8090 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8093 mdb_cursor_copy(mc, &mx);
8094 while (mc->mc_snum > 0) {
8095 MDB_page *mp = mc->mc_pg[mc->mc_top];
8096 unsigned n = NUMKEYS(mp);
8098 for (i=0; i<n; i++) {
8099 ni = NODEPTR(mp, i);
8100 if (ni->mn_flags & F_BIGDATA) {
8103 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8104 rc = mdb_page_get(txn, pg, &omp, NULL);
8107 assert(IS_OVERFLOW(omp));
8108 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8112 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8113 mdb_xcursor_init1(mc, ni);
8114 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8120 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8122 for (i=0; i<n; i++) {
8124 ni = NODEPTR(mp, i);
8127 mdb_midl_xappend(txn->mt_free_pgs, pg);
8132 mc->mc_ki[mc->mc_top] = i;
8133 rc = mdb_cursor_sibling(mc, 1);
8135 /* no more siblings, go back to beginning
8136 * of previous level.
8140 for (i=1; i<mc->mc_snum; i++) {
8142 mc->mc_pg[i] = mx.mc_pg[i];
8147 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8148 } else if (rc == MDB_NOTFOUND) {
8154 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8156 MDB_cursor *mc, *m2;
8159 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8162 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8165 rc = mdb_cursor_open(txn, dbi, &mc);
8169 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8170 /* Invalidate the dropped DB's cursors */
8171 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8172 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8176 /* Can't delete the main DB */
8177 if (del && dbi > MAIN_DBI) {
8178 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8180 txn->mt_dbflags[dbi] = DB_STALE;
8181 mdb_dbi_close(txn->mt_env, dbi);
8184 /* reset the DB record, mark it dirty */
8185 txn->mt_dbflags[dbi] |= DB_DIRTY;
8186 txn->mt_dbs[dbi].md_depth = 0;
8187 txn->mt_dbs[dbi].md_branch_pages = 0;
8188 txn->mt_dbs[dbi].md_leaf_pages = 0;
8189 txn->mt_dbs[dbi].md_overflow_pages = 0;
8190 txn->mt_dbs[dbi].md_entries = 0;
8191 txn->mt_dbs[dbi].md_root = P_INVALID;
8193 txn->mt_flags |= MDB_TXN_DIRTY;
8196 mdb_cursor_close(mc);
8200 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8202 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8205 txn->mt_dbxs[dbi].md_cmp = cmp;
8209 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8211 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8214 txn->mt_dbxs[dbi].md_dcmp = cmp;
8218 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8220 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8223 txn->mt_dbxs[dbi].md_rel = rel;
8227 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8229 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8232 txn->mt_dbxs[dbi].md_relctx = ctx;
8236 int mdb_env_get_maxkeysize(MDB_env *env)
8238 return MDB_MAXKEYSIZE;
8241 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8243 unsigned int i, rdrs;
8250 if (!env->me_txns) {
8251 return func("(no reader locks)\n", ctx);
8253 rdrs = env->me_txns->mti_numreaders;
8254 mr = env->me_txns->mti_readers;
8255 for (i=0; i<rdrs; i++) {
8260 if (mr[i].mr_txnid == (txnid_t)-1) {
8261 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8263 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8267 func(" pid thread txnid\n", ctx);
8269 rc = func(buf, ctx);
8275 func("(no active readers)\n", ctx);
8280 /** Insert pid into list if not already present.
8281 * return -1 if already present.
8283 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8285 /* binary search of pid in list */
8287 unsigned cursor = 1;
8289 unsigned n = ids[0];
8292 unsigned pivot = n >> 1;
8293 cursor = base + pivot + 1;
8294 val = pid - ids[cursor];
8299 } else if ( val > 0 ) {
8304 /* found, so it's a duplicate */
8313 for (n = ids[0]; n > cursor; n--)
8319 int mdb_reader_check(MDB_env *env, int *dead)
8321 unsigned int i, j, rdrs;
8332 rdrs = env->me_txns->mti_numreaders;
8333 pids = malloc((rdrs+1) * sizeof(pid_t));
8337 mr = env->me_txns->mti_readers;
8339 for (i=0; i<rdrs; i++) {
8340 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8342 if (mdb_pid_insert(pids, pid) == 0) {
8343 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8345 /* Recheck, a new process may have reused pid */
8346 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8347 for (j=i; j<rdrs; j++)
8348 if (mr[j].mr_pid == pid) {
8353 UNLOCK_MUTEX_R(env);