2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define MDB_PIDLOCK 0
144 #define pthread_t DWORD
145 #define pthread_mutex_t HANDLE
146 #define pthread_key_t DWORD
147 #define pthread_self() GetCurrentThreadId()
148 #define pthread_key_create(x,y) \
149 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
150 #define pthread_key_delete(x) TlsFree(x)
151 #define pthread_getspecific(x) TlsGetValue(x)
152 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
153 #define pthread_mutex_unlock(x) ReleaseMutex(x)
154 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
155 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
156 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
157 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
158 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
159 #define getpid() GetCurrentProcessId()
160 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
161 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
162 #define ErrCode() GetLastError()
163 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
164 #define close(fd) (CloseHandle(fd) ? 0 : -1)
165 #define munmap(ptr,len) UnmapViewOfFile(ptr)
166 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
167 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
169 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
176 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
177 #define MDB_PIDLOCK 1
179 #ifdef MDB_USE_POSIX_SEM
181 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
182 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
183 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
184 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
187 mdb_sem_wait(sem_t *sem)
190 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
195 /** Lock the reader mutex.
197 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
198 /** Unlock the reader mutex.
200 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
202 /** Lock the writer mutex.
203 * Only a single write transaction is allowed at a time. Other writers
204 * will block waiting for this mutex.
206 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
207 /** Unlock the writer mutex.
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
210 #endif /* MDB_USE_POSIX_SEM */
212 /** Get the error code for the last failed system function.
214 #define ErrCode() errno
216 /** An abstraction for a file handle.
217 * On POSIX systems file handles are small integers. On Windows
218 * they're opaque pointers.
222 /** A value for an invalid file handle.
223 * Mainly used to initialize file variables and signify that they are
226 #define INVALID_HANDLE_VALUE (-1)
228 /** Get the size of a memory page for the system.
229 * This is the basic size that the platform's memory manager uses, and is
230 * fundamental to the use of memory-mapped files.
232 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
235 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
238 #define MNAME_LEN (sizeof(pthread_mutex_t))
244 /** A flag for opening a file and requesting synchronous data writes.
245 * This is only used when writing a meta page. It's not strictly needed;
246 * we could just do a normal write and then immediately perform a flush.
247 * But if this flag is available it saves us an extra system call.
249 * @note If O_DSYNC is undefined but exists in /usr/include,
250 * preferably set some compiler flag to get the definition.
251 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
254 # define MDB_DSYNC O_DSYNC
258 /** Function for flushing the data of a file. Define this to fsync
259 * if fdatasync() is not supported.
261 #ifndef MDB_FDATASYNC
262 # define MDB_FDATASYNC fdatasync
266 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
277 /** A page number in the database.
278 * Note that 64 bit page numbers are overkill, since pages themselves
279 * already represent 12-13 bits of addressable memory, and the OS will
280 * always limit applications to a maximum of 63 bits of address space.
282 * @note In the #MDB_node structure, we only store 48 bits of this value,
283 * which thus limits us to only 60 bits of addressable data.
285 typedef MDB_ID pgno_t;
287 /** A transaction ID.
288 * See struct MDB_txn.mt_txnid for details.
290 typedef MDB_ID txnid_t;
292 /** @defgroup debug Debug Macros
296 /** Enable debug output. Needs variable argument macros (a C99 feature).
297 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
298 * read from and written to the database (used for free space management).
304 static int mdb_debug;
305 static txnid_t mdb_debug_start;
307 /** Print a debug message with printf formatting.
308 * Requires double parenthesis around 2 or more args.
310 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
311 # define DPRINTF0(fmt, ...) \
312 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
314 # define DPRINTF(args) ((void) 0)
316 /** Print a debug string.
317 * The string is printed literally, with no format processing.
319 #define DPUTS(arg) DPRINTF(("%s", arg))
322 /** A default memory page size.
323 * The actual size is platform-dependent, but we use this for
324 * boot-strapping. We probably should not be using this any more.
325 * The #GET_PAGESIZE() macro is used to get the actual size.
327 * Note that we don't currently support Huge pages. On Linux,
328 * regular data files cannot use Huge pages, and in general
329 * Huge pages aren't actually pageable. We rely on the OS
330 * demand-pager to read our data and page it out when memory
331 * pressure from other processes is high. So until OSs have
332 * actual paging support for Huge pages, they're not viable.
334 #define MDB_PAGESIZE 4096
336 /** The minimum number of keys required in a database page.
337 * Setting this to a larger value will place a smaller bound on the
338 * maximum size of a data item. Data items larger than this size will
339 * be pushed into overflow pages instead of being stored directly in
340 * the B-tree node. This value used to default to 4. With a page size
341 * of 4096 bytes that meant that any item larger than 1024 bytes would
342 * go into an overflow page. That also meant that on average 2-3KB of
343 * each overflow page was wasted space. The value cannot be lower than
344 * 2 because then there would no longer be a tree structure. With this
345 * value, items larger than 2KB will go into overflow pages, and on
346 * average only 1KB will be wasted.
348 #define MDB_MINKEYS 2
350 /** A stamp that identifies a file as an MDB file.
351 * There's nothing special about this value other than that it is easily
352 * recognizable, and it will reflect any byte order mismatches.
354 #define MDB_MAGIC 0xBEEFC0DE
356 /** The version number for a database's datafile format. */
357 #define MDB_DATA_VERSION 1
358 /** The version number for a database's lockfile format. */
359 #define MDB_LOCK_VERSION 1
361 /** @brief The maximum size of a key in the database.
363 * The library rejects bigger keys, and cannot deal with records
364 * with bigger keys stored by a library with bigger max keysize.
366 * We require that keys all fit onto a regular page. This limit
367 * could be raised a bit further if needed; to something just
368 * under #MDB_PAGESIZE / #MDB_MINKEYS.
370 * Note that data items in an #MDB_DUPSORT database are actually keys
371 * of a subDB, so they're also limited to this size.
373 #ifndef MDB_MAXKEYSIZE
374 #define MDB_MAXKEYSIZE 511
377 /** @brief The maximum size of a data item.
379 * We only store a 32 bit value for node sizes.
381 #define MAXDATASIZE 0xffffffffUL
386 * This is used for printing a hex dump of a key's contents.
388 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
389 /** Display a key in hex.
391 * Invoke a function to display a key in hex.
393 #define DKEY(x) mdb_dkey(x, kbuf)
399 /** An invalid page number.
400 * Mainly used to denote an empty tree.
402 #define P_INVALID (~(pgno_t)0)
404 /** Test if the flags \b f are set in a flag word \b w. */
405 #define F_ISSET(w, f) (((w) & (f)) == (f))
407 /** Used for offsets within a single page.
408 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
411 typedef uint16_t indx_t;
413 /** Default size of memory map.
414 * This is certainly too small for any actual applications. Apps should always set
415 * the size explicitly using #mdb_env_set_mapsize().
417 #define DEFAULT_MAPSIZE 1048576
419 /** @defgroup readers Reader Lock Table
420 * Readers don't acquire any locks for their data access. Instead, they
421 * simply record their transaction ID in the reader table. The reader
422 * mutex is needed just to find an empty slot in the reader table. The
423 * slot's address is saved in thread-specific data so that subsequent read
424 * transactions started by the same thread need no further locking to proceed.
426 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
428 * No reader table is used if the database is on a read-only filesystem.
430 * Since the database uses multi-version concurrency control, readers don't
431 * actually need any locking. This table is used to keep track of which
432 * readers are using data from which old transactions, so that we'll know
433 * when a particular old transaction is no longer in use. Old transactions
434 * that have discarded any data pages can then have those pages reclaimed
435 * for use by a later write transaction.
437 * The lock table is constructed such that reader slots are aligned with the
438 * processor's cache line size. Any slot is only ever used by one thread.
439 * This alignment guarantees that there will be no contention or cache
440 * thrashing as threads update their own slot info, and also eliminates
441 * any need for locking when accessing a slot.
443 * A writer thread will scan every slot in the table to determine the oldest
444 * outstanding reader transaction. Any freed pages older than this will be
445 * reclaimed by the writer. The writer doesn't use any locks when scanning
446 * this table. This means that there's no guarantee that the writer will
447 * see the most up-to-date reader info, but that's not required for correct
448 * operation - all we need is to know the upper bound on the oldest reader,
449 * we don't care at all about the newest reader. So the only consequence of
450 * reading stale information here is that old pages might hang around a
451 * while longer before being reclaimed. That's actually good anyway, because
452 * the longer we delay reclaiming old pages, the more likely it is that a
453 * string of contiguous pages can be found after coalescing old pages from
454 * many old transactions together.
457 /** Number of slots in the reader table.
458 * This value was chosen somewhat arbitrarily. 126 readers plus a
459 * couple mutexes fit exactly into 8KB on my development machine.
460 * Applications should set the table size using #mdb_env_set_maxreaders().
462 #define DEFAULT_READERS 126
464 /** The size of a CPU cache line in bytes. We want our lock structures
465 * aligned to this size to avoid false cache line sharing in the
467 * This value works for most CPUs. For Itanium this should be 128.
473 /** The information we store in a single slot of the reader table.
474 * In addition to a transaction ID, we also record the process and
475 * thread ID that owns a slot, so that we can detect stale information,
476 * e.g. threads or processes that went away without cleaning up.
477 * @note We currently don't check for stale records. We simply re-init
478 * the table when we know that we're the only process opening the
481 typedef struct MDB_rxbody {
482 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
483 * Multiple readers that start at the same time will probably have the
484 * same ID here. Again, it's not important to exclude them from
485 * anything; all we need to know is which version of the DB they
486 * started from so we can avoid overwriting any data used in that
487 * particular version.
490 /** The process ID of the process owning this reader txn. */
492 /** The thread ID of the thread owning this txn. */
496 /** The actual reader record, with cacheline padding. */
497 typedef struct MDB_reader {
500 /** shorthand for mrb_txnid */
501 #define mr_txnid mru.mrx.mrb_txnid
502 #define mr_pid mru.mrx.mrb_pid
503 #define mr_tid mru.mrx.mrb_tid
504 /** cache line alignment */
505 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
509 /** The header for the reader table.
510 * The table resides in a memory-mapped file. (This is a different file
511 * than is used for the main database.)
513 * For POSIX the actual mutexes reside in the shared memory of this
514 * mapped file. On Windows, mutexes are named objects allocated by the
515 * kernel; we store the mutex names in this mapped file so that other
516 * processes can grab them. This same approach is also used on
517 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
518 * process-shared POSIX mutexes. For these cases where a named object
519 * is used, the object name is derived from a 64 bit FNV hash of the
520 * environment pathname. As such, naming collisions are extremely
521 * unlikely. If a collision occurs, the results are unpredictable.
523 typedef struct MDB_txbody {
524 /** Stamp identifying this as an MDB file. It must be set
527 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
529 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
530 char mtb_rmname[MNAME_LEN];
532 /** Mutex protecting access to this table.
533 * This is the reader lock that #LOCK_MUTEX_R acquires.
535 pthread_mutex_t mtb_mutex;
537 /** The ID of the last transaction committed to the database.
538 * This is recorded here only for convenience; the value can always
539 * be determined by reading the main database meta pages.
542 /** The number of slots that have been used in the reader table.
543 * This always records the maximum count, it is not decremented
544 * when readers release their slots.
546 unsigned mtb_numreaders;
549 /** The actual reader table definition. */
550 typedef struct MDB_txninfo {
553 #define mti_magic mt1.mtb.mtb_magic
554 #define mti_format mt1.mtb.mtb_format
555 #define mti_mutex mt1.mtb.mtb_mutex
556 #define mti_rmname mt1.mtb.mtb_rmname
557 #define mti_txnid mt1.mtb.mtb_txnid
558 #define mti_numreaders mt1.mtb.mtb_numreaders
559 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
562 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
563 char mt2_wmname[MNAME_LEN];
564 #define mti_wmname mt2.mt2_wmname
566 pthread_mutex_t mt2_wmutex;
567 #define mti_wmutex mt2.mt2_wmutex
569 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
571 MDB_reader mti_readers[1];
574 /** Lockfile format signature: version, features and field layout */
575 #define MDB_LOCK_FORMAT \
577 ((MDB_LOCK_VERSION) \
578 /* Flags which describe functionality */ \
579 + (((MDB_PIDLOCK) != 0) << 16)))
582 /** Common header for all page types.
583 * Overflow records occupy a number of contiguous pages with no
584 * headers on any page after the first.
586 typedef struct MDB_page {
587 #define mp_pgno mp_p.p_pgno
588 #define mp_next mp_p.p_next
590 pgno_t p_pgno; /**< page number */
591 void * p_next; /**< for in-memory list of freed structs */
594 /** @defgroup mdb_page Page Flags
596 * Flags for the page headers.
599 #define P_BRANCH 0x01 /**< branch page */
600 #define P_LEAF 0x02 /**< leaf page */
601 #define P_OVERFLOW 0x04 /**< overflow page */
602 #define P_META 0x08 /**< meta page */
603 #define P_DIRTY 0x10 /**< dirty page */
604 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
605 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
606 #define P_KEEP 0x8000 /**< leave this page alone during spill */
608 uint16_t mp_flags; /**< @ref mdb_page */
609 #define mp_lower mp_pb.pb.pb_lower
610 #define mp_upper mp_pb.pb.pb_upper
611 #define mp_pages mp_pb.pb_pages
614 indx_t pb_lower; /**< lower bound of free space */
615 indx_t pb_upper; /**< upper bound of free space */
617 uint32_t pb_pages; /**< number of overflow pages */
619 indx_t mp_ptrs[1]; /**< dynamic size */
622 /** Size of the page header, excluding dynamic data at the end */
623 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
625 /** Address of first usable data byte in a page, after the header */
626 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
628 /** Number of nodes on a page */
629 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
631 /** The amount of space remaining in the page */
632 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
634 /** The percentage of space used in the page, in tenths of a percent. */
635 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
636 ((env)->me_psize - PAGEHDRSZ))
637 /** The minimum page fill factor, in tenths of a percent.
638 * Pages emptier than this are candidates for merging.
640 #define FILL_THRESHOLD 250
642 /** Test if a page is a leaf page */
643 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
644 /** Test if a page is a LEAF2 page */
645 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
646 /** Test if a page is a branch page */
647 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
648 /** Test if a page is an overflow page */
649 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
650 /** Test if a page is a sub page */
651 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
653 /** The number of overflow pages needed to store the given size. */
654 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
656 /** Header for a single key/data pair within a page.
657 * We guarantee 2-byte alignment for nodes.
659 typedef struct MDB_node {
660 /** lo and hi are used for data size on leaf nodes and for
661 * child pgno on branch nodes. On 64 bit platforms, flags
662 * is also used for pgno. (Branch nodes have no flags).
663 * They are in host byte order in case that lets some
664 * accesses be optimized into a 32-bit word access.
666 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
667 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
668 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
669 /** @defgroup mdb_node Node Flags
671 * Flags for node headers.
674 #define F_BIGDATA 0x01 /**< data put on overflow page */
675 #define F_SUBDATA 0x02 /**< data is a sub-database */
676 #define F_DUPDATA 0x04 /**< data has duplicates */
678 /** valid flags for #mdb_node_add() */
679 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
682 unsigned short mn_flags; /**< @ref mdb_node */
683 unsigned short mn_ksize; /**< key size */
684 char mn_data[1]; /**< key and data are appended here */
687 /** Size of the node header, excluding dynamic data at the end */
688 #define NODESIZE offsetof(MDB_node, mn_data)
690 /** Bit position of top word in page number, for shifting mn_flags */
691 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
693 /** Size of a node in a branch page with a given key.
694 * This is just the node header plus the key, there is no data.
696 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
698 /** Size of a node in a leaf page with a given key and data.
699 * This is node header plus key plus data size.
701 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
703 /** Address of node \b i in page \b p */
704 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
706 /** Address of the key for the node */
707 #define NODEKEY(node) (void *)((node)->mn_data)
709 /** Address of the data for a node */
710 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
712 /** Get the page number pointed to by a branch node */
713 #define NODEPGNO(node) \
714 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
715 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
716 /** Set the page number in a branch node */
717 #define SETPGNO(node,pgno) do { \
718 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
719 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
721 /** Get the size of the data in a leaf node */
722 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
723 /** Set the size of the data for a leaf node */
724 #define SETDSZ(node,size) do { \
725 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
726 /** The size of a key in a node */
727 #define NODEKSZ(node) ((node)->mn_ksize)
729 /** Copy a page number from src to dst */
731 #define COPY_PGNO(dst,src) dst = src
733 #if SIZE_MAX > 4294967295UL
734 #define COPY_PGNO(dst,src) do { \
735 unsigned short *s, *d; \
736 s = (unsigned short *)&(src); \
737 d = (unsigned short *)&(dst); \
744 #define COPY_PGNO(dst,src) do { \
745 unsigned short *s, *d; \
746 s = (unsigned short *)&(src); \
747 d = (unsigned short *)&(dst); \
753 /** The address of a key in a LEAF2 page.
754 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
755 * There are no node headers, keys are stored contiguously.
757 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
759 /** Set the \b node's key into \b keyptr, if requested. */
760 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
761 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
763 /** Set the \b node's key into \b key. */
764 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
766 /** Information about a single database in the environment. */
767 typedef struct MDB_db {
768 uint32_t md_pad; /**< also ksize for LEAF2 pages */
769 uint16_t md_flags; /**< @ref mdb_dbi_open */
770 uint16_t md_depth; /**< depth of this tree */
771 pgno_t md_branch_pages; /**< number of internal pages */
772 pgno_t md_leaf_pages; /**< number of leaf pages */
773 pgno_t md_overflow_pages; /**< number of overflow pages */
774 size_t md_entries; /**< number of data items */
775 pgno_t md_root; /**< the root page of this tree */
778 /** mdb_dbi_open flags */
779 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
780 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
781 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
782 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
784 /** Handle for the DB used to track free pages. */
786 /** Handle for the default DB. */
789 /** Meta page content. */
790 typedef struct MDB_meta {
791 /** Stamp identifying this as an MDB file. It must be set
794 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
796 void *mm_address; /**< address for fixed mapping */
797 size_t mm_mapsize; /**< size of mmap region */
798 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
799 /** The size of pages used in this DB */
800 #define mm_psize mm_dbs[0].md_pad
801 /** Any persistent environment flags. @ref mdb_env */
802 #define mm_flags mm_dbs[0].md_flags
803 pgno_t mm_last_pg; /**< last used page in file */
804 txnid_t mm_txnid; /**< txnid that committed this page */
807 /** Buffer for a stack-allocated dirty page.
808 * The members define size and alignment, and silence type
809 * aliasing warnings. They are not used directly; that could
810 * mean incorrectly using several union members in parallel.
812 typedef union MDB_pagebuf {
813 char mb_raw[MDB_PAGESIZE];
816 char mm_pad[PAGEHDRSZ];
821 /** Auxiliary DB info.
822 * The information here is mostly static/read-only. There is
823 * only a single copy of this record in the environment.
825 typedef struct MDB_dbx {
826 MDB_val md_name; /**< name of the database */
827 MDB_cmp_func *md_cmp; /**< function for comparing keys */
828 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
829 MDB_rel_func *md_rel; /**< user relocate function */
830 void *md_relctx; /**< user-provided context for md_rel */
833 /** A database transaction.
834 * Every operation requires a transaction handle.
837 MDB_txn *mt_parent; /**< parent of a nested txn */
838 MDB_txn *mt_child; /**< nested txn under this txn */
839 pgno_t mt_next_pgno; /**< next unallocated page */
840 /** The ID of this transaction. IDs are integers incrementing from 1.
841 * Only committed write transactions increment the ID. If a transaction
842 * aborts, the ID may be re-used by the next writer.
845 MDB_env *mt_env; /**< the DB environment */
846 /** The list of pages that became unused during this transaction.
849 /** The sorted list of dirty pages we temporarily wrote to disk
850 * because the dirty list was full.
852 MDB_IDL mt_spill_pgs;
854 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
856 /** For read txns: This thread/txn's reader table slot, or NULL. */
859 /** Array of records for each DB known in the environment. */
861 /** Array of MDB_db records for each known DB */
863 /** @defgroup mt_dbflag Transaction DB Flags
867 #define DB_DIRTY 0x01 /**< DB was written in this txn */
868 #define DB_STALE 0x02 /**< DB record is older than txnID */
869 #define DB_NEW 0x04 /**< DB handle opened in this txn */
870 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
872 /** In write txns, array of cursors for each DB */
873 MDB_cursor **mt_cursors;
874 /** Array of flags for each DB */
875 unsigned char *mt_dbflags;
876 /** Number of DB records in use. This number only ever increments;
877 * we don't decrement it when individual DB handles are closed.
881 /** @defgroup mdb_txn Transaction Flags
885 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
886 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
887 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
888 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
890 unsigned int mt_flags; /**< @ref mdb_txn */
891 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
892 unsigned int mt_dirty_room;
893 /** Tracks which of the two meta pages was used at the start
894 * of this transaction.
896 unsigned int mt_toggle;
899 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
900 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
901 * raise this on a 64 bit machine.
903 #define CURSOR_STACK 32
907 /** Cursors are used for all DB operations */
909 /** Next cursor on this DB in this txn */
911 /** Backup of the original cursor if this cursor is a shadow */
912 MDB_cursor *mc_backup;
913 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
914 struct MDB_xcursor *mc_xcursor;
915 /** The transaction that owns this cursor */
917 /** The database handle this cursor operates on */
919 /** The database record for this cursor */
921 /** The database auxiliary record for this cursor */
923 /** The @ref mt_dbflag for this database */
924 unsigned char *mc_dbflag;
925 unsigned short mc_snum; /**< number of pushed pages */
926 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
927 /** @defgroup mdb_cursor Cursor Flags
929 * Cursor state flags.
932 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
933 #define C_EOF 0x02 /**< No more data */
934 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
935 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
936 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
938 unsigned int mc_flags; /**< @ref mdb_cursor */
939 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
940 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
943 /** Context for sorted-dup records.
944 * We could have gone to a fully recursive design, with arbitrarily
945 * deep nesting of sub-databases. But for now we only handle these
946 * levels - main DB, optional sub-DB, sorted-duplicate DB.
948 typedef struct MDB_xcursor {
949 /** A sub-cursor for traversing the Dup DB */
950 MDB_cursor mx_cursor;
951 /** The database record for this Dup DB */
953 /** The auxiliary DB record for this Dup DB */
955 /** The @ref mt_dbflag for this Dup DB */
956 unsigned char mx_dbflag;
959 /** State of FreeDB old pages, stored in the MDB_env */
960 typedef struct MDB_pgstate {
961 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
962 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
965 /** The database environment. */
967 HANDLE me_fd; /**< The main data file */
968 HANDLE me_lfd; /**< The lock file */
969 HANDLE me_mfd; /**< just for writing the meta pages */
970 /** Failed to update the meta page. Probably an I/O error. */
971 #define MDB_FATAL_ERROR 0x80000000U
972 /** Some fields are initialized. */
973 #define MDB_ENV_ACTIVE 0x20000000U
974 /** me_txkey is set */
975 #define MDB_ENV_TXKEY 0x10000000U
976 /** Have liveness lock in reader table */
977 #define MDB_LIVE_READER 0x08000000U
978 uint32_t me_flags; /**< @ref mdb_env */
979 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
980 unsigned int me_maxreaders; /**< size of the reader table */
981 unsigned int me_numreaders; /**< max numreaders set by this env */
982 MDB_dbi me_numdbs; /**< number of DBs opened */
983 MDB_dbi me_maxdbs; /**< size of the DB table */
984 pid_t me_pid; /**< process ID of this env */
985 char *me_path; /**< path to the DB files */
986 char *me_map; /**< the memory map of the data file */
987 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
988 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
989 MDB_txn *me_txn; /**< current write transaction */
990 size_t me_mapsize; /**< size of the data memory map */
991 off_t me_size; /**< current file size */
992 pgno_t me_maxpg; /**< me_mapsize / me_psize */
993 MDB_dbx *me_dbxs; /**< array of static DB info */
994 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
995 pthread_key_t me_txkey; /**< thread-key for readers */
996 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
997 # define me_pglast me_pgstate.mf_pglast
998 # define me_pghead me_pgstate.mf_pghead
999 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1000 /** IDL of pages that became unused in a write txn */
1001 MDB_IDL me_free_pgs;
1002 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1003 MDB_ID2L me_dirty_list;
1004 /** Max number of freelist items that can fit in a single overflow page */
1006 /** Max size of a node on a page */
1007 unsigned int me_nodemax;
1009 int me_pidquery; /**< Used in OpenProcess */
1010 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1012 #elif defined(MDB_USE_POSIX_SEM)
1013 sem_t *me_rmutex; /* Shared mutexes are not supported */
1018 /** Nested transaction */
1019 typedef struct MDB_ntxn {
1020 MDB_txn mnt_txn; /* the transaction */
1021 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
1024 /** max number of pages to commit in one writev() call */
1025 #define MDB_COMMIT_PAGES 64
1026 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1027 #undef MDB_COMMIT_PAGES
1028 #define MDB_COMMIT_PAGES IOV_MAX
1031 /* max bytes to write in one call */
1032 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1034 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1035 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1036 static int mdb_page_touch(MDB_cursor *mc);
1038 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1039 static int mdb_page_search_root(MDB_cursor *mc,
1040 MDB_val *key, int modify);
1041 #define MDB_PS_MODIFY 1
1042 #define MDB_PS_ROOTONLY 2
1043 static int mdb_page_search(MDB_cursor *mc,
1044 MDB_val *key, int flags);
1045 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1047 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1048 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1049 pgno_t newpgno, unsigned int nflags);
1051 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1052 static int mdb_env_pick_meta(const MDB_env *env);
1053 static int mdb_env_write_meta(MDB_txn *txn);
1054 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1055 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1057 static void mdb_env_close0(MDB_env *env, int excl);
1059 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1060 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1061 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1062 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1063 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1064 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1065 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1066 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1067 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1069 static int mdb_rebalance(MDB_cursor *mc);
1070 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1072 static void mdb_cursor_pop(MDB_cursor *mc);
1073 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1075 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1076 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1077 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1078 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1079 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1081 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1082 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1084 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1085 static void mdb_xcursor_init0(MDB_cursor *mc);
1086 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1088 static int mdb_drop0(MDB_cursor *mc, int subs);
1089 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1092 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1096 static SECURITY_DESCRIPTOR mdb_null_sd;
1097 static SECURITY_ATTRIBUTES mdb_all_sa;
1098 static int mdb_sec_inited;
1101 /** Return the library version info. */
1103 mdb_version(int *major, int *minor, int *patch)
1105 if (major) *major = MDB_VERSION_MAJOR;
1106 if (minor) *minor = MDB_VERSION_MINOR;
1107 if (patch) *patch = MDB_VERSION_PATCH;
1108 return MDB_VERSION_STRING;
1111 /** Table of descriptions for MDB @ref errors */
1112 static char *const mdb_errstr[] = {
1113 "MDB_KEYEXIST: Key/data pair already exists",
1114 "MDB_NOTFOUND: No matching key/data pair found",
1115 "MDB_PAGE_NOTFOUND: Requested page not found",
1116 "MDB_CORRUPTED: Located page was wrong type",
1117 "MDB_PANIC: Update of meta page failed",
1118 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1119 "MDB_INVALID: File is not an MDB file",
1120 "MDB_MAP_FULL: Environment mapsize limit reached",
1121 "MDB_DBS_FULL: Environment maxdbs limit reached",
1122 "MDB_READERS_FULL: Environment maxreaders limit reached",
1123 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1124 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1125 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1126 "MDB_PAGE_FULL: Internal error - page has no more space",
1127 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1128 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1129 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1130 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1131 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1135 mdb_strerror(int err)
1139 return ("Successful return: 0");
1141 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1142 i = err - MDB_KEYEXIST;
1143 return mdb_errstr[i];
1146 return strerror(err);
1150 /** Display a key in hexadecimal and return the address of the result.
1151 * @param[in] key the key to display
1152 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1153 * @return The key in hexadecimal form.
1156 mdb_dkey(MDB_val *key, char *buf)
1159 unsigned char *c = key->mv_data;
1165 if (key->mv_size > MDB_MAXKEYSIZE)
1166 return "MDB_MAXKEYSIZE";
1167 /* may want to make this a dynamic check: if the key is mostly
1168 * printable characters, print it as-is instead of converting to hex.
1172 for (i=0; i<key->mv_size; i++)
1173 ptr += sprintf(ptr, "%02x", *c++);
1175 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1180 /** Display all the keys in the page. */
1182 mdb_page_list(MDB_page *mp)
1185 unsigned int i, nkeys, nsize;
1189 nkeys = NUMKEYS(mp);
1190 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1191 for (i=0; i<nkeys; i++) {
1192 node = NODEPTR(mp, i);
1193 key.mv_size = node->mn_ksize;
1194 key.mv_data = node->mn_data;
1195 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1196 if (IS_BRANCH(mp)) {
1197 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1200 if (F_ISSET(node->mn_flags, F_BIGDATA))
1201 nsize += sizeof(pgno_t);
1203 nsize += NODEDSZ(node);
1204 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1210 mdb_cursor_chk(MDB_cursor *mc)
1216 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1217 for (i=0; i<mc->mc_top; i++) {
1219 node = NODEPTR(mp, mc->mc_ki[i]);
1220 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1223 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1229 /** Count all the pages in each DB and in the freelist
1230 * and make sure it matches the actual number of pages
1233 static void mdb_audit(MDB_txn *txn)
1237 MDB_ID freecount, count;
1242 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1243 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1244 freecount += *(MDB_ID *)data.mv_data;
1247 for (i = 0; i<txn->mt_numdbs; i++) {
1249 mdb_cursor_init(&mc, txn, i, &mx);
1250 if (txn->mt_dbs[i].md_root == P_INVALID)
1252 count += txn->mt_dbs[i].md_branch_pages +
1253 txn->mt_dbs[i].md_leaf_pages +
1254 txn->mt_dbs[i].md_overflow_pages;
1255 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1256 mdb_page_search(&mc, NULL, 0);
1260 mp = mc.mc_pg[mc.mc_top];
1261 for (j=0; j<NUMKEYS(mp); j++) {
1262 MDB_node *leaf = NODEPTR(mp, j);
1263 if (leaf->mn_flags & F_SUBDATA) {
1265 memcpy(&db, NODEDATA(leaf), sizeof(db));
1266 count += db.md_branch_pages + db.md_leaf_pages +
1267 db.md_overflow_pages;
1271 while (mdb_cursor_sibling(&mc, 1) == 0);
1274 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1275 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1276 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1282 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1284 return txn->mt_dbxs[dbi].md_cmp(a, b);
1288 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1290 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1293 /** Allocate memory for a page.
1294 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1297 mdb_page_malloc(MDB_txn *txn, unsigned num)
1299 MDB_env *env = txn->mt_env;
1300 MDB_page *ret = env->me_dpages;
1301 size_t sz = env->me_psize;
1304 VGMEMP_ALLOC(env, ret, sz);
1305 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1306 env->me_dpages = ret->mp_next;
1312 if ((ret = malloc(sz)) != NULL) {
1313 VGMEMP_ALLOC(env, ret, sz);
1318 /** Free a single page.
1319 * Saves single pages to a list, for future reuse.
1320 * (This is not used for multi-page overflow pages.)
1323 mdb_page_free(MDB_env *env, MDB_page *mp)
1325 mp->mp_next = env->me_dpages;
1326 VGMEMP_FREE(env, mp);
1327 env->me_dpages = mp;
1330 /* Free a dirty page */
1332 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1334 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1335 mdb_page_free(env, dp);
1337 /* large pages just get freed directly */
1338 VGMEMP_FREE(env, dp);
1343 /** Return all dirty pages to dpage list */
1345 mdb_dlist_free(MDB_txn *txn)
1347 MDB_env *env = txn->mt_env;
1348 MDB_ID2L dl = txn->mt_u.dirty_list;
1349 unsigned i, n = dl[0].mid;
1351 for (i = 1; i <= n; i++) {
1352 mdb_dpage_free(env, dl[i].mptr);
1357 /* Set or clear P_KEEP in non-overflow, non-sub pages in this txn's cursors.
1358 * @param[in] mc A cursor handle for the current operation.
1359 * @param[in] pflags Flags of the pages to update:
1360 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1363 mdb_cursorpages_mark(MDB_cursor *mc, unsigned pflags)
1365 MDB_txn *txn = mc->mc_txn;
1370 if (mc->mc_flags & C_UNTRACK)
1371 mc = NULL; /* will find mc in mt_cursors */
1372 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1373 for (; mc; mc=mc->mc_next) {
1374 for (m3 = mc; m3->mc_flags & C_INITIALIZED; m3 = &mx->mx_cursor) {
1375 for (j=0; j<m3->mc_snum; j++)
1376 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1378 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1379 mx = m3->mc_xcursor;
1389 static int mdb_page_flush(MDB_txn *txn);
1391 /** Spill pages from the dirty list back to disk.
1392 * This is intended to prevent running into #MDB_TXN_FULL situations,
1393 * but note that they may still occur in a few cases:
1394 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1395 * are too many of these dirtied in one txn, the txn may still get
1397 * 2) child txns may run out of space if their parents dirtied a
1398 * lot of pages and never spilled them. TODO: we probably should do
1399 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1400 * the parent's dirty_room is below a given threshold.
1401 * 3) our estimate of the txn size could be too small. At the
1402 * moment this seems unlikely.
1404 * Otherwise, if not using nested txns, it is expected that apps will
1405 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1406 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1407 * If the txn never references them again, they can be left alone.
1408 * If the txn only reads them, they can be used without any fuss.
1409 * If the txn writes them again, they can be dirtied immediately without
1410 * going thru all of the work of #mdb_page_touch(). Such references are
1411 * handled by #mdb_page_unspill().
1413 * Also note, we never spill DB root pages, nor pages of active cursors,
1414 * because we'll need these back again soon anyway. And in nested txns,
1415 * we can't spill a page in a child txn if it was already spilled in a
1416 * parent txn. That would alter the parent txns' data even though
1417 * the child hasn't committed yet, and we'd have no way to undo it if
1418 * the child aborted.
1420 * @param[in] m0 cursor A cursor handle identifying the transaction and
1421 * database for which we are checking space.
1422 * @param[in] key For a put operation, the key being stored.
1423 * @param[in] data For a put operation, the data being stored.
1424 * @return 0 on success, non-zero on failure.
1427 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1429 MDB_txn *txn = m0->mc_txn;
1431 MDB_ID2L dl = txn->mt_u.dirty_list;
1432 unsigned int i, j, k, need;
1435 if (m0->mc_flags & C_SUB)
1438 /* Estimate how much space this op will take */
1439 i = m0->mc_db->md_depth;
1440 /* Named DBs also dirty the main DB */
1441 if (m0->mc_dbi > MAIN_DBI)
1442 i += txn->mt_dbs[MAIN_DBI].md_depth;
1443 /* For puts, roughly factor in the key+data size */
1445 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1446 i += i; /* double it for good measure */
1449 if (txn->mt_dirty_room > i)
1452 if (!txn->mt_spill_pgs) {
1453 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1454 if (!txn->mt_spill_pgs)
1458 /* Mark all the dirty root pages we want to preserve */
1459 for (i=0; i<txn->mt_numdbs; i++) {
1460 if (txn->mt_dbflags[i] & DB_DIRTY) {
1461 pgno_t pgno = txn->mt_dbs[i].md_root;
1462 if (pgno == P_INVALID)
1464 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1466 if ((dp->mp_flags & P_DIRTY) && level <= 1)
1467 dp->mp_flags |= P_KEEP;
1471 /* Preserve pages used by cursors */
1472 mdb_cursorpages_mark(m0, P_DIRTY);
1474 /* Less aggressive spill - we originally spilled the entire dirty list,
1475 * with a few exceptions for cursor pages and DB root pages. But this
1476 * turns out to be a lot of wasted effort because in a large txn many
1477 * of those pages will need to be used again. So now we spill only 1/8th
1478 * of the dirty pages. Testing revealed this to be a good tradeoff,
1479 * better than 1/2, 1/4, or 1/10.
1483 if (need < MDB_IDL_UM_MAX / 8)
1484 need = MDB_IDL_UM_MAX / 8;
1486 /* Save the page IDs of all the pages we're flushing */
1487 /* flush from the tail forward, this saves a lot of shifting later on. */
1488 for (i=dl[0].mid; i>0; i--) {
1490 if (dp->mp_flags & P_KEEP)
1492 /* Can't spill twice, make sure it's not already in a parent's
1495 if (txn->mt_parent) {
1497 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1498 if (tx2->mt_spill_pgs) {
1499 j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
1500 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
1501 dp->mp_flags |= P_KEEP;
1509 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
1515 mdb_midl_sort(txn->mt_spill_pgs);
1517 /* Since we're only doing the tail 1/8th of the dirty list,
1518 * fake a dirty list to reflect this.
1522 k = dl[0].mid - i + 1;
1525 txn->mt_u.dirty_list = &dl[i-1];
1527 rc = mdb_page_flush(txn);
1529 /* reset back to the real list */
1531 dl[0].mid += dl[i-1].mid;
1533 txn->mt_u.dirty_list = dl;
1536 mdb_cursorpages_mark(m0, P_DIRTY|P_KEEP);
1538 /* Reset any dirty root pages we kept that page_flush didn't see */
1539 for (i=0; i<txn->mt_numdbs; i++) {
1540 if (txn->mt_dbflags[i] & DB_DIRTY) {
1541 pgno_t pgno = txn->mt_dbs[i].md_root;
1542 if (pgno == P_INVALID)
1544 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1546 if (dp->mp_flags & P_KEEP)
1547 dp->mp_flags ^= P_KEEP;
1553 if (txn->mt_parent) {
1555 pgno_t pgno = dl[i].mid;
1556 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1557 /* dirty pages that are dirty in an ancestor don't
1558 * count against this txn's dirty_room.
1560 for (i=1; i<=dl[0].mid; i++) {
1561 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1562 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1563 if (j <= tx2->mt_u.dirty_list[0].mid &&
1564 tx2->mt_u.dirty_list[j].mid == pgno) {
1565 txn->mt_dirty_room++;
1571 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1573 txn->mt_flags |= MDB_TXN_SPILLS;
1575 txn->mt_flags |= MDB_TXN_ERROR;
1580 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1582 mdb_find_oldest(MDB_txn *txn)
1585 txnid_t mr, oldest = txn->mt_txnid - 1;
1586 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1587 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1597 /** Add a page to the txn's dirty list */
1599 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1602 int (*insert)(MDB_ID2L, MDB_ID2 *);
1604 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1605 insert = mdb_mid2l_append;
1607 insert = mdb_mid2l_insert;
1609 mid.mid = mp->mp_pgno;
1611 insert(txn->mt_u.dirty_list, &mid);
1612 txn->mt_dirty_room--;
1615 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1616 * me_pghead and mt_next_pgno.
1618 * If there are free pages available from older transactions, they
1619 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1620 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1621 * and move me_pglast to say which records were consumed. Only this
1622 * function can create me_pghead and move me_pglast/mt_next_pgno.
1623 * @param[in] mc cursor A cursor handle identifying the transaction and
1624 * database for which we are allocating.
1625 * @param[in] num the number of pages to allocate.
1626 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1627 * will always be satisfied by a single contiguous chunk of memory.
1628 * @return 0 on success, non-zero on failure.
1631 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1633 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1634 /* Get at most <Max_retries> more freeDB records once me_pghead
1635 * has enough pages. If not enough, use new pages from the map.
1636 * If <Paranoid> and mc is updating the freeDB, only get new
1637 * records if me_pghead is empty. Then the freelist cannot play
1638 * catch-up with itself by growing while trying to save it.
1640 enum { Paranoid = 1, Max_retries = 500 };
1642 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1644 int rc, n2 = num-1, retry = Max_retries;
1645 MDB_txn *txn = mc->mc_txn;
1646 MDB_env *env = txn->mt_env;
1647 pgno_t pgno, *mop = env->me_pghead;
1648 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1650 txnid_t oldest = 0, last;
1656 /* If our dirty list is already full, we can't do anything */
1657 if (txn->mt_dirty_room == 0)
1658 return MDB_TXN_FULL;
1660 for (op = MDB_FIRST;; op = MDB_NEXT) {
1663 pgno_t *idl, old_id, new_id;
1665 /* Seek a big enough contiguous page range. Prefer
1666 * pages at the tail, just truncating the list.
1668 if (mop_len >= (unsigned)num) {
1672 if (mop[i-n2] == pgno+n2)
1674 } while (--i >= (unsigned)num);
1675 if (Max_retries < INT_MAX && --retry < 0)
1679 if (op == MDB_FIRST) { /* 1st iteration */
1680 /* Prepare to fetch more and coalesce */
1681 oldest = mdb_find_oldest(txn);
1682 last = env->me_pglast;
1683 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1686 key.mv_data = &last; /* will look up last+1 */
1687 key.mv_size = sizeof(last);
1689 if (Paranoid && mc->mc_dbi == FREE_DBI)
1692 if (Paranoid && retry < 0 && mop_len)
1696 /* Do not fetch more if the record will be too recent */
1699 rc = mdb_cursor_get(&m2, &key, NULL, op);
1701 if (rc == MDB_NOTFOUND)
1705 last = *(txnid_t*)key.mv_data;
1708 np = m2.mc_pg[m2.mc_top];
1709 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1710 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1713 idl = (MDB_ID *) data.mv_data;
1716 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1719 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1721 mop = env->me_pghead;
1723 env->me_pglast = last;
1725 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1726 last, txn->mt_dbs[FREE_DBI].md_root, i));
1728 DPRINTF(("IDL %"Z"u", idl[k]));
1730 /* Merge in descending sorted order */
1733 mop[0] = (pgno_t)-1;
1737 for (; old_id < new_id; old_id = mop[--j])
1744 /* Use new pages from the map when nothing suitable in the freeDB */
1746 pgno = txn->mt_next_pgno;
1747 if (pgno + num >= env->me_maxpg) {
1748 DPUTS("DB size maxed out");
1749 return MDB_MAP_FULL;
1753 if (env->me_flags & MDB_WRITEMAP) {
1754 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1756 if (!(np = mdb_page_malloc(txn, num)))
1760 mop[0] = mop_len -= num;
1761 /* Move any stragglers down */
1762 for (j = i-num; j < mop_len; )
1763 mop[++j] = mop[++i];
1765 txn->mt_next_pgno = pgno + num;
1768 mdb_page_dirty(txn, np);
1774 /** Copy the used portions of a non-overflow page.
1775 * @param[in] dst page to copy into
1776 * @param[in] src page to copy from
1777 * @param[in] psize size of a page
1780 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1782 enum { Align = sizeof(pgno_t) };
1783 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1785 /* If page isn't full, just copy the used portion. Adjust
1786 * alignment so memcpy may copy words instead of bytes.
1788 if ((unused &= -Align) && !IS_LEAF2(src)) {
1790 memcpy(dst, src, (lower + (Align-1)) & -Align);
1791 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1794 memcpy(dst, src, psize - unused);
1798 /** Pull a page off the txn's spill list, if present.
1799 * If a page being referenced was spilled to disk in this txn, bring
1800 * it back and make it dirty/writable again.
1801 * @param[in] tx0 the transaction handle.
1802 * @param[in] mp the page being referenced.
1803 * @param[out] ret the writable page, if any. ret is unchanged if
1804 * mp wasn't spilled.
1807 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1809 MDB_env *env = tx0->mt_env;
1812 pgno_t pgno = mp->mp_pgno;
1814 for (txn = tx0; txn; txn=txn->mt_parent) {
1815 if (!txn->mt_spill_pgs)
1817 x = mdb_midl_search(txn->mt_spill_pgs, pgno);
1818 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
1821 if (IS_OVERFLOW(mp))
1825 if (env->me_flags & MDB_WRITEMAP) {
1828 np = mdb_page_malloc(txn, num);
1832 memcpy(np, mp, num * env->me_psize);
1834 mdb_page_copy(np, mp, env->me_psize);
1837 /* If in current txn, this page is no longer spilled */
1838 for (; x < txn->mt_spill_pgs[0]; x++)
1839 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
1840 txn->mt_spill_pgs[0]--;
1841 } /* otherwise, if belonging to a parent txn, the
1842 * page remains spilled until child commits
1845 if (txn->mt_parent) {
1847 /* If this page is also in a parent's dirty list, then
1848 * it's already accounted in dirty_room, and we need to
1849 * cancel out the decrement that mdb_page_dirty does.
1851 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1852 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1853 if (x <= tx2->mt_u.dirty_list[0].mid &&
1854 tx2->mt_u.dirty_list[x].mid == pgno) {
1855 txn->mt_dirty_room++;
1860 mdb_page_dirty(tx0, np);
1861 np->mp_flags |= P_DIRTY;
1869 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1870 * @param[in] mc cursor pointing to the page to be touched
1871 * @return 0 on success, non-zero on failure.
1874 mdb_page_touch(MDB_cursor *mc)
1876 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1877 MDB_txn *txn = mc->mc_txn;
1878 MDB_cursor *m2, *m3;
1883 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1884 if (txn->mt_flags & MDB_TXN_SPILLS) {
1886 rc = mdb_page_unspill(txn, mp, &np);
1892 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1893 (rc = mdb_page_alloc(mc, 1, &np)))
1896 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1897 assert(mp->mp_pgno != pgno);
1898 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1899 /* Update the parent page, if any, to point to the new page */
1901 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1902 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1903 SETPGNO(node, pgno);
1905 mc->mc_db->md_root = pgno;
1907 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1908 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1910 /* If txn has a parent, make sure the page is in our
1914 unsigned x = mdb_mid2l_search(dl, pgno);
1915 if (x <= dl[0].mid && dl[x].mid == pgno) {
1916 if (mp != dl[x].mptr) { /* bad cursor? */
1917 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1918 return MDB_CORRUPTED;
1923 assert(dl[0].mid < MDB_IDL_UM_MAX);
1925 np = mdb_page_malloc(txn, 1);
1930 mdb_mid2l_insert(dl, &mid);
1935 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1937 np->mp_flags |= P_DIRTY;
1940 /* Adjust cursors pointing to mp */
1941 mc->mc_pg[mc->mc_top] = np;
1943 if (mc->mc_flags & C_SUB) {
1945 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1946 m3 = &m2->mc_xcursor->mx_cursor;
1947 if (m3->mc_snum < mc->mc_snum) continue;
1948 if (m3->mc_pg[mc->mc_top] == mp)
1949 m3->mc_pg[mc->mc_top] = np;
1952 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1953 if (m2->mc_snum < mc->mc_snum) continue;
1954 if (m2->mc_pg[mc->mc_top] == mp) {
1955 m2->mc_pg[mc->mc_top] = np;
1956 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1957 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1959 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1960 if (!(leaf->mn_flags & F_SUBDATA))
1961 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1970 mdb_env_sync(MDB_env *env, int force)
1973 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1974 if (env->me_flags & MDB_WRITEMAP) {
1975 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1976 ? MS_ASYNC : MS_SYNC;
1977 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1980 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1984 if (MDB_FDATASYNC(env->me_fd))
1991 /** Back up parent txn's cursors, then grab the originals for tracking */
1993 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1995 MDB_cursor *mc, *bk;
2000 for (i = src->mt_numdbs; --i >= 0; ) {
2001 if ((mc = src->mt_cursors[i]) != NULL) {
2002 size = sizeof(MDB_cursor);
2004 size += sizeof(MDB_xcursor);
2005 for (; mc; mc = bk->mc_next) {
2011 mc->mc_db = &dst->mt_dbs[i];
2012 /* Kill pointers into src - and dst to reduce abuse: The
2013 * user may not use mc until dst ends. Otherwise we'd...
2015 mc->mc_txn = NULL; /* ...set this to dst */
2016 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2017 if ((mx = mc->mc_xcursor) != NULL) {
2018 *(MDB_xcursor *)(bk+1) = *mx;
2019 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2021 mc->mc_next = dst->mt_cursors[i];
2022 dst->mt_cursors[i] = mc;
2029 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2030 * @param[in] txn the transaction handle.
2031 * @param[in] merge true to keep changes to parent cursors, false to revert.
2032 * @return 0 on success, non-zero on failure.
2035 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2037 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2041 for (i = txn->mt_numdbs; --i >= 0; ) {
2042 for (mc = cursors[i]; mc; mc = next) {
2044 if ((bk = mc->mc_backup) != NULL) {
2046 /* Commit changes to parent txn */
2047 mc->mc_next = bk->mc_next;
2048 mc->mc_backup = bk->mc_backup;
2049 mc->mc_txn = bk->mc_txn;
2050 mc->mc_db = bk->mc_db;
2051 mc->mc_dbflag = bk->mc_dbflag;
2052 if ((mx = mc->mc_xcursor) != NULL)
2053 mx->mx_cursor.mc_txn = bk->mc_txn;
2055 /* Abort nested txn */
2057 if ((mx = mc->mc_xcursor) != NULL)
2058 *mx = *(MDB_xcursor *)(bk+1);
2062 /* Only malloced cursors are permanently tracked. */
2070 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2073 mdb_txn_reset0(MDB_txn *txn, const char *act);
2075 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2081 Pidset = F_SETLK, Pidcheck = F_GETLK
2085 /** Set or check a pid lock. Set returns 0 on success.
2086 * Check returns 0 if the process is certainly dead, nonzero if it may
2087 * be alive (the lock exists or an error happened so we do not know).
2089 * On Windows Pidset is a no-op, we merely check for the existence
2090 * of the process with the given pid. On POSIX we use a single byte
2091 * lock on the lockfile, set at an offset equal to the pid.
2094 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2096 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2099 if (op == Pidcheck) {
2100 h = OpenProcess(env->me_pidquery, FALSE, pid);
2101 /* No documented "no such process" code, but other program use this: */
2103 return ErrCode() != ERROR_INVALID_PARAMETER;
2104 /* A process exists until all handles to it close. Has it exited? */
2105 ret = WaitForSingleObject(h, 0) != 0;
2112 struct flock lock_info;
2113 memset(&lock_info, 0, sizeof(lock_info));
2114 lock_info.l_type = F_WRLCK;
2115 lock_info.l_whence = SEEK_SET;
2116 lock_info.l_start = pid;
2117 lock_info.l_len = 1;
2118 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2119 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2121 } else if ((rc = ErrCode()) == EINTR) {
2129 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2130 * @param[in] txn the transaction handle to initialize
2131 * @return 0 on success, non-zero on failure.
2134 mdb_txn_renew0(MDB_txn *txn)
2136 MDB_env *env = txn->mt_env;
2139 int rc, new_notls = 0;
2142 txn->mt_numdbs = env->me_numdbs;
2143 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2145 if (txn->mt_flags & MDB_TXN_RDONLY) {
2146 if (!env->me_txns) {
2147 i = mdb_env_pick_meta(env);
2148 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2149 txn->mt_u.reader = NULL;
2151 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2152 pthread_getspecific(env->me_txkey);
2154 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2155 return MDB_BAD_RSLOT;
2157 pid_t pid = env->me_pid;
2158 pthread_t tid = pthread_self();
2160 if (!(env->me_flags & MDB_LIVE_READER)) {
2161 rc = mdb_reader_pid(env, Pidset, pid);
2163 UNLOCK_MUTEX_R(env);
2166 env->me_flags |= MDB_LIVE_READER;
2170 for (i=0; i<env->me_txns->mti_numreaders; i++)
2171 if (env->me_txns->mti_readers[i].mr_pid == 0)
2173 if (i == env->me_maxreaders) {
2174 UNLOCK_MUTEX_R(env);
2175 return MDB_READERS_FULL;
2177 env->me_txns->mti_readers[i].mr_pid = pid;
2178 env->me_txns->mti_readers[i].mr_tid = tid;
2179 if (i >= env->me_txns->mti_numreaders)
2180 env->me_txns->mti_numreaders = i+1;
2181 /* Save numreaders for un-mutexed mdb_env_close() */
2182 env->me_numreaders = env->me_txns->mti_numreaders;
2183 UNLOCK_MUTEX_R(env);
2184 r = &env->me_txns->mti_readers[i];
2185 new_notls = (env->me_flags & MDB_NOTLS);
2186 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2191 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2192 txn->mt_u.reader = r;
2194 txn->mt_toggle = txn->mt_txnid & 1;
2198 txn->mt_txnid = env->me_txns->mti_txnid;
2199 txn->mt_toggle = txn->mt_txnid & 1;
2202 if (txn->mt_txnid == mdb_debug_start)
2205 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2206 txn->mt_u.dirty_list = env->me_dirty_list;
2207 txn->mt_u.dirty_list[0].mid = 0;
2208 txn->mt_free_pgs = env->me_free_pgs;
2209 txn->mt_free_pgs[0] = 0;
2210 txn->mt_spill_pgs = NULL;
2214 /* Copy the DB info and flags */
2215 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2217 /* Moved to here to avoid a data race in read TXNs */
2218 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2220 for (i=2; i<txn->mt_numdbs; i++) {
2221 x = env->me_dbflags[i];
2222 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2223 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2225 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2227 if (env->me_maxpg < txn->mt_next_pgno) {
2228 mdb_txn_reset0(txn, "renew0-mapfail");
2230 txn->mt_u.reader->mr_pid = 0;
2231 txn->mt_u.reader = NULL;
2233 return MDB_MAP_RESIZED;
2240 mdb_txn_renew(MDB_txn *txn)
2244 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2247 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2248 DPUTS("environment had fatal error, must shutdown!");
2252 rc = mdb_txn_renew0(txn);
2253 if (rc == MDB_SUCCESS) {
2254 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2255 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2256 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2262 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2266 int rc, size, tsize = sizeof(MDB_txn);
2268 if (env->me_flags & MDB_FATAL_ERROR) {
2269 DPUTS("environment had fatal error, must shutdown!");
2272 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2275 /* Nested transactions: Max 1 child, write txns only, no writemap */
2276 if (parent->mt_child ||
2277 (flags & MDB_RDONLY) ||
2278 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2279 (env->me_flags & MDB_WRITEMAP))
2281 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2283 tsize = sizeof(MDB_ntxn);
2285 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2286 if (!(flags & MDB_RDONLY))
2287 size += env->me_maxdbs * sizeof(MDB_cursor *);
2289 if ((txn = calloc(1, size)) == NULL) {
2290 DPRINTF(("calloc: %s", strerror(ErrCode())));
2293 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2294 if (flags & MDB_RDONLY) {
2295 txn->mt_flags |= MDB_TXN_RDONLY;
2296 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2298 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2299 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2305 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2306 if (!txn->mt_u.dirty_list ||
2307 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2309 free(txn->mt_u.dirty_list);
2313 txn->mt_txnid = parent->mt_txnid;
2314 txn->mt_toggle = parent->mt_toggle;
2315 txn->mt_dirty_room = parent->mt_dirty_room;
2316 txn->mt_u.dirty_list[0].mid = 0;
2317 txn->mt_spill_pgs = NULL;
2318 txn->mt_next_pgno = parent->mt_next_pgno;
2319 parent->mt_child = txn;
2320 txn->mt_parent = parent;
2321 txn->mt_numdbs = parent->mt_numdbs;
2322 txn->mt_flags = parent->mt_flags;
2323 txn->mt_dbxs = parent->mt_dbxs;
2324 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2325 /* Copy parent's mt_dbflags, but clear DB_NEW */
2326 for (i=0; i<txn->mt_numdbs; i++)
2327 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2329 ntxn = (MDB_ntxn *)txn;
2330 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2331 if (env->me_pghead) {
2332 size = MDB_IDL_SIZEOF(env->me_pghead);
2333 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2335 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2340 rc = mdb_cursor_shadow(parent, txn);
2342 mdb_txn_reset0(txn, "beginchild-fail");
2344 rc = mdb_txn_renew0(txn);
2350 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2351 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2352 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2359 mdb_txn_env(MDB_txn *txn)
2361 if(!txn) return NULL;
2365 /** Export or close DBI handles opened in this txn. */
2367 mdb_dbis_update(MDB_txn *txn, int keep)
2370 MDB_dbi n = txn->mt_numdbs;
2371 MDB_env *env = txn->mt_env;
2372 unsigned char *tdbflags = txn->mt_dbflags;
2374 for (i = n; --i >= 2;) {
2375 if (tdbflags[i] & DB_NEW) {
2377 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2379 char *ptr = env->me_dbxs[i].md_name.mv_data;
2380 env->me_dbxs[i].md_name.mv_data = NULL;
2381 env->me_dbxs[i].md_name.mv_size = 0;
2382 env->me_dbflags[i] = 0;
2387 if (keep && env->me_numdbs < n)
2391 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2392 * May be called twice for readonly txns: First reset it, then abort.
2393 * @param[in] txn the transaction handle to reset
2394 * @param[in] act why the transaction is being reset
2397 mdb_txn_reset0(MDB_txn *txn, const char *act)
2399 MDB_env *env = txn->mt_env;
2401 /* Close any DBI handles opened in this txn */
2402 mdb_dbis_update(txn, 0);
2404 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2405 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2406 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2408 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2409 if (txn->mt_u.reader) {
2410 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2411 if (!(env->me_flags & MDB_NOTLS))
2412 txn->mt_u.reader = NULL; /* txn does not own reader */
2414 txn->mt_numdbs = 0; /* close nothing if called again */
2415 txn->mt_dbxs = NULL; /* mark txn as reset */
2417 mdb_cursors_close(txn, 0);
2419 if (!(env->me_flags & MDB_WRITEMAP)) {
2420 mdb_dlist_free(txn);
2422 mdb_midl_free(env->me_pghead);
2424 if (txn->mt_parent) {
2425 txn->mt_parent->mt_child = NULL;
2426 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2427 mdb_midl_free(txn->mt_free_pgs);
2428 mdb_midl_free(txn->mt_spill_pgs);
2429 free(txn->mt_u.dirty_list);
2433 if (mdb_midl_shrink(&txn->mt_free_pgs))
2434 env->me_free_pgs = txn->mt_free_pgs;
2435 env->me_pghead = NULL;
2439 /* The writer mutex was locked in mdb_txn_begin. */
2440 UNLOCK_MUTEX_W(env);
2445 mdb_txn_reset(MDB_txn *txn)
2450 /* This call is only valid for read-only txns */
2451 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2454 mdb_txn_reset0(txn, "reset");
2458 mdb_txn_abort(MDB_txn *txn)
2464 mdb_txn_abort(txn->mt_child);
2466 mdb_txn_reset0(txn, "abort");
2467 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2468 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2469 txn->mt_u.reader->mr_pid = 0;
2474 /** Save the freelist as of this transaction to the freeDB.
2475 * This changes the freelist. Keep trying until it stabilizes.
2478 mdb_freelist_save(MDB_txn *txn)
2480 /* env->me_pghead[] can grow and shrink during this call.
2481 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2482 * Page numbers cannot disappear from txn->mt_free_pgs[].
2485 MDB_env *env = txn->mt_env;
2486 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2487 txnid_t pglast = 0, head_id = 0;
2488 pgno_t freecnt = 0, *free_pgs, *mop;
2489 ssize_t head_room = 0, total_room = 0, mop_len;
2491 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2493 if (env->me_pghead) {
2494 /* Make sure first page of freeDB is touched and on freelist */
2495 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2496 if (rc && rc != MDB_NOTFOUND)
2501 /* Come back here after each Put() in case freelist changed */
2504 /* If using records from freeDB which we have not yet
2505 * deleted, delete them and any we reserved for me_pghead.
2507 while (pglast < env->me_pglast) {
2508 rc = mdb_cursor_first(&mc, &key, NULL);
2511 pglast = head_id = *(txnid_t *)key.mv_data;
2512 total_room = head_room = 0;
2513 assert(pglast <= env->me_pglast);
2514 rc = mdb_cursor_del(&mc, 0);
2519 /* Save the IDL of pages freed by this txn, to a single record */
2520 if (freecnt < txn->mt_free_pgs[0]) {
2522 /* Make sure last page of freeDB is touched and on freelist */
2523 key.mv_size = MDB_MAXKEYSIZE+1;
2525 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2526 if (rc && rc != MDB_NOTFOUND)
2529 free_pgs = txn->mt_free_pgs;
2530 /* Write to last page of freeDB */
2531 key.mv_size = sizeof(txn->mt_txnid);
2532 key.mv_data = &txn->mt_txnid;
2534 freecnt = free_pgs[0];
2535 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2536 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2539 /* Retry if mt_free_pgs[] grew during the Put() */
2540 free_pgs = txn->mt_free_pgs;
2541 } while (freecnt < free_pgs[0]);
2542 mdb_midl_sort(free_pgs);
2543 memcpy(data.mv_data, free_pgs, data.mv_size);
2546 unsigned int i = free_pgs[0];
2547 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2548 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2550 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2556 mop = env->me_pghead;
2557 mop_len = mop ? mop[0] : 0;
2559 /* Reserve records for me_pghead[]. Split it if multi-page,
2560 * to avoid searching freeDB for a page range. Use keys in
2561 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2563 if (total_room >= mop_len) {
2564 if (total_room == mop_len || --more < 0)
2566 } else if (head_room >= maxfree_1pg && head_id > 1) {
2567 /* Keep current record (overflow page), add a new one */
2571 /* (Re)write {key = head_id, IDL length = head_room} */
2572 total_room -= head_room;
2573 head_room = mop_len - total_room;
2574 if (head_room > maxfree_1pg && head_id > 1) {
2575 /* Overflow multi-page for part of me_pghead */
2576 head_room /= head_id; /* amortize page sizes */
2577 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2578 } else if (head_room < 0) {
2579 /* Rare case, not bothering to delete this record */
2582 key.mv_size = sizeof(head_id);
2583 key.mv_data = &head_id;
2584 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2585 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2588 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2589 total_room += head_room;
2592 /* Fill in the reserved, touched me_pghead records */
2598 rc = mdb_cursor_first(&mc, &key, &data);
2599 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2600 unsigned flags = MDB_CURRENT;
2601 txnid_t id = *(txnid_t *)key.mv_data;
2602 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2605 assert(len >= 0 && id <= env->me_pglast);
2607 if (len > mop_len) {
2609 data.mv_size = (len + 1) * sizeof(MDB_ID);
2612 data.mv_data = mop -= len;
2615 rc = mdb_cursor_put(&mc, &key, &data, flags);
2617 if (rc || !(mop_len -= len))
2624 /** Flush dirty pages to the map, after clearing their dirty flag.
2627 mdb_page_flush(MDB_txn *txn)
2629 MDB_env *env = txn->mt_env;
2630 MDB_ID2L dl = txn->mt_u.dirty_list;
2631 unsigned psize = env->me_psize, j;
2632 int i, pagecount = dl[0].mid, rc;
2633 size_t size = 0, pos = 0;
2635 MDB_page *dp = NULL;
2639 struct iovec iov[MDB_COMMIT_PAGES];
2640 ssize_t wpos = 0, wsize = 0, wres;
2641 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2646 if (env->me_flags & MDB_WRITEMAP) {
2647 /* Clear dirty flags */
2648 for (i=1; i<=pagecount; i++) {
2650 /* Don't flush this page yet */
2651 if (dp->mp_flags & P_KEEP) {
2652 dp->mp_flags ^= P_KEEP;
2656 dp->mp_flags &= ~P_DIRTY;
2662 /* Write the pages */
2664 if (i <= pagecount) {
2666 /* Don't flush this page yet */
2667 if (dp->mp_flags & P_KEEP) {
2668 dp->mp_flags ^= P_KEEP;
2673 /* clear dirty flag */
2674 dp->mp_flags &= ~P_DIRTY;
2677 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2682 /* Windows actually supports scatter/gather I/O, but only on
2683 * unbuffered file handles. Since we're relying on the OS page
2684 * cache for all our data, that's self-defeating. So we just
2685 * write pages one at a time. We use the ov structure to set
2686 * the write offset, to at least save the overhead of a Seek
2689 DPRINTF(("committing page %"Z"u", pgno));
2690 memset(&ov, 0, sizeof(ov));
2691 ov.Offset = pos & 0xffffffff;
2692 ov.OffsetHigh = pos >> 16 >> 16;
2693 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2695 DPRINTF(("WriteFile: %d", rc));
2699 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2700 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2702 /* Write previous page(s) */
2703 #ifdef MDB_USE_PWRITEV
2704 wres = pwritev(env->me_fd, iov, n, wpos);
2707 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2709 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2711 DPRINTF(("lseek: %s", strerror(rc)));
2714 wres = writev(env->me_fd, iov, n);
2717 if (wres != wsize) {
2720 DPRINTF(("Write error: %s", strerror(rc)));
2722 rc = EIO; /* TODO: Use which error code? */
2723 DPUTS("short write, filesystem full?");
2734 DPRINTF(("committing page %"Z"u", pgno));
2735 next_pos = pos + size;
2736 iov[n].iov_len = size;
2737 iov[n].iov_base = (char *)dp;
2744 for (i=1; i<=pagecount; i++) {
2746 /* This is a page we skipped above */
2749 dl[j].mid = dp->mp_pgno;
2752 mdb_dpage_free(env, dp);
2760 mdb_txn_commit(MDB_txn *txn)
2766 assert(txn != NULL);
2767 assert(txn->mt_env != NULL);
2769 if (txn->mt_child) {
2770 rc = mdb_txn_commit(txn->mt_child);
2771 txn->mt_child = NULL;
2778 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2779 mdb_dbis_update(txn, 1);
2780 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2785 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2786 DPUTS("error flag is set, can't commit");
2788 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2793 if (txn->mt_parent) {
2794 MDB_txn *parent = txn->mt_parent;
2798 /* Append our free list to parent's */
2799 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2802 mdb_midl_free(txn->mt_free_pgs);
2804 parent->mt_next_pgno = txn->mt_next_pgno;
2805 parent->mt_flags = txn->mt_flags;
2807 /* Merge our cursors into parent's and close them */
2808 mdb_cursors_close(txn, 1);
2810 /* Update parent's DB table. */
2811 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2812 parent->mt_numdbs = txn->mt_numdbs;
2813 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2814 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2815 for (i=2; i<txn->mt_numdbs; i++) {
2816 /* preserve parent's DB_NEW status */
2817 x = parent->mt_dbflags[i] & DB_NEW;
2818 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2821 dst = parent->mt_u.dirty_list;
2822 src = txn->mt_u.dirty_list;
2823 /* Remove anything in our dirty list from parent's spill list */
2824 if (parent->mt_spill_pgs) {
2825 x = parent->mt_spill_pgs[0];
2827 /* zero out our dirty pages in parent spill list */
2828 for (i=1; i<=src[0].mid; i++) {
2829 if (src[i].mid < parent->mt_spill_pgs[x])
2831 if (src[i].mid > parent->mt_spill_pgs[x]) {
2837 parent->mt_spill_pgs[x] = 0;
2840 /* OK, we had a few hits, squash zeros from the spill list */
2841 if (len < parent->mt_spill_pgs[0]) {
2843 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2844 if (parent->mt_spill_pgs[y]) {
2846 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2851 parent->mt_spill_pgs[0] = len;
2854 /* Find len = length of merging our dirty list with parent's */
2856 dst[0].mid = 0; /* simplify loops */
2857 if (parent->mt_parent) {
2858 len = x + src[0].mid;
2859 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2860 for (i = x; y && i; y--) {
2861 pgno_t yp = src[y].mid;
2862 while (yp < dst[i].mid)
2864 if (yp == dst[i].mid) {
2869 } else { /* Simplify the above for single-ancestor case */
2870 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2872 /* Merge our dirty list with parent's */
2874 for (i = len; y; dst[i--] = src[y--]) {
2875 pgno_t yp = src[y].mid;
2876 while (yp < dst[x].mid)
2877 dst[i--] = dst[x--];
2878 if (yp == dst[x].mid)
2879 free(dst[x--].mptr);
2883 free(txn->mt_u.dirty_list);
2884 parent->mt_dirty_room = txn->mt_dirty_room;
2885 if (txn->mt_spill_pgs) {
2886 if (parent->mt_spill_pgs) {
2887 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2888 mdb_midl_free(txn->mt_spill_pgs);
2889 mdb_midl_sort(parent->mt_spill_pgs);
2891 parent->mt_spill_pgs = txn->mt_spill_pgs;
2895 parent->mt_child = NULL;
2896 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2901 if (txn != env->me_txn) {
2902 DPUTS("attempt to commit unknown transaction");
2907 mdb_cursors_close(txn, 0);
2909 if (!txn->mt_u.dirty_list[0].mid &&
2910 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2913 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2914 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2916 /* Update DB root pointers */
2917 if (txn->mt_numdbs > 2) {
2921 data.mv_size = sizeof(MDB_db);
2923 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2924 for (i = 2; i < txn->mt_numdbs; i++) {
2925 if (txn->mt_dbflags[i] & DB_DIRTY) {
2926 data.mv_data = &txn->mt_dbs[i];
2927 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2934 rc = mdb_freelist_save(txn);
2938 mdb_midl_free(env->me_pghead);
2939 env->me_pghead = NULL;
2940 if (mdb_midl_shrink(&txn->mt_free_pgs))
2941 env->me_free_pgs = txn->mt_free_pgs;
2947 if ((rc = mdb_page_flush(txn)) ||
2948 (rc = mdb_env_sync(env, 0)) ||
2949 (rc = mdb_env_write_meta(txn)))
2955 mdb_dbis_update(txn, 1);
2957 UNLOCK_MUTEX_W(env);
2967 /** Read the environment parameters of a DB environment before
2968 * mapping it into memory.
2969 * @param[in] env the environment handle
2970 * @param[out] meta address of where to store the meta information
2971 * @return 0 on success, non-zero on failure.
2974 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2981 /* We don't know the page size yet, so use a minimum value.
2982 * Read both meta pages so we can use the latest one.
2985 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2989 memset(&ov, 0, sizeof(ov));
2991 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2992 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2995 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2997 if (rc != MDB_PAGESIZE) {
2998 if (rc == 0 && off == 0)
3000 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3001 DPRINTF(("read: %s", mdb_strerror(rc)));
3005 p = (MDB_page *)&pbuf;
3007 if (!F_ISSET(p->mp_flags, P_META)) {
3008 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3013 if (m->mm_magic != MDB_MAGIC) {
3014 DPUTS("meta has invalid magic");
3018 if (m->mm_version != MDB_DATA_VERSION) {
3019 DPRINTF(("database is version %u, expected version %u",
3020 m->mm_version, MDB_DATA_VERSION));
3021 return MDB_VERSION_MISMATCH;
3024 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3030 /** Write the environment parameters of a freshly created DB environment.
3031 * @param[in] env the environment handle
3032 * @param[out] meta address of where to store the meta information
3033 * @return 0 on success, non-zero on failure.
3036 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3044 memset(&ov, 0, sizeof(ov));
3045 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3047 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3050 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3051 len = pwrite(fd, ptr, size, pos); \
3052 rc = (len >= 0); } while(0)
3055 DPUTS("writing new meta page");
3057 GET_PAGESIZE(psize);
3059 meta->mm_magic = MDB_MAGIC;
3060 meta->mm_version = MDB_DATA_VERSION;
3061 meta->mm_mapsize = env->me_mapsize;
3062 meta->mm_psize = psize;
3063 meta->mm_last_pg = 1;
3064 meta->mm_flags = env->me_flags & 0xffff;
3065 meta->mm_flags |= MDB_INTEGERKEY;
3066 meta->mm_dbs[0].md_root = P_INVALID;
3067 meta->mm_dbs[1].md_root = P_INVALID;
3069 p = calloc(2, psize);
3071 p->mp_flags = P_META;
3072 *(MDB_meta *)METADATA(p) = *meta;
3074 q = (MDB_page *)((char *)p + psize);
3076 q->mp_flags = P_META;
3077 *(MDB_meta *)METADATA(q) = *meta;
3079 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3082 else if ((unsigned) len == psize * 2)
3090 /** Update the environment info to commit a transaction.
3091 * @param[in] txn the transaction that's being committed
3092 * @return 0 on success, non-zero on failure.
3095 mdb_env_write_meta(MDB_txn *txn)
3098 MDB_meta meta, metab, *mp;
3100 int rc, len, toggle;
3109 assert(txn != NULL);
3110 assert(txn->mt_env != NULL);
3112 toggle = !txn->mt_toggle;
3113 DPRINTF(("writing meta page %d for root page %"Z"u",
3114 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3117 mp = env->me_metas[toggle];
3119 if (env->me_flags & MDB_WRITEMAP) {
3120 /* Persist any increases of mapsize config */
3121 if (env->me_mapsize > mp->mm_mapsize)
3122 mp->mm_mapsize = env->me_mapsize;
3123 mp->mm_dbs[0] = txn->mt_dbs[0];
3124 mp->mm_dbs[1] = txn->mt_dbs[1];
3125 mp->mm_last_pg = txn->mt_next_pgno - 1;
3126 mp->mm_txnid = txn->mt_txnid;
3127 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3128 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3131 ptr += env->me_psize;
3132 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3139 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3140 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3142 ptr = (char *)&meta;
3143 if (env->me_mapsize > mp->mm_mapsize) {
3144 /* Persist any increases of mapsize config */
3145 meta.mm_mapsize = env->me_mapsize;
3146 off = offsetof(MDB_meta, mm_mapsize);
3148 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3150 len = sizeof(MDB_meta) - off;
3153 meta.mm_dbs[0] = txn->mt_dbs[0];
3154 meta.mm_dbs[1] = txn->mt_dbs[1];
3155 meta.mm_last_pg = txn->mt_next_pgno - 1;
3156 meta.mm_txnid = txn->mt_txnid;
3159 off += env->me_psize;
3162 /* Write to the SYNC fd */
3163 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3164 env->me_fd : env->me_mfd;
3167 memset(&ov, 0, sizeof(ov));
3169 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3173 rc = pwrite(mfd, ptr, len, off);
3176 rc = rc < 0 ? ErrCode() : EIO;
3177 DPUTS("write failed, disk error?");
3178 /* On a failure, the pagecache still contains the new data.
3179 * Write some old data back, to prevent it from being used.
3180 * Use the non-SYNC fd; we know it will fail anyway.
3182 meta.mm_last_pg = metab.mm_last_pg;
3183 meta.mm_txnid = metab.mm_txnid;
3185 memset(&ov, 0, sizeof(ov));
3187 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3189 r2 = pwrite(env->me_fd, ptr, len, off);
3190 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3193 env->me_flags |= MDB_FATAL_ERROR;
3197 /* Memory ordering issues are irrelevant; since the entire writer
3198 * is wrapped by wmutex, all of these changes will become visible
3199 * after the wmutex is unlocked. Since the DB is multi-version,
3200 * readers will get consistent data regardless of how fresh or
3201 * how stale their view of these values is.
3203 env->me_txns->mti_txnid = txn->mt_txnid;
3208 /** Check both meta pages to see which one is newer.
3209 * @param[in] env the environment handle
3210 * @return meta toggle (0 or 1).
3213 mdb_env_pick_meta(const MDB_env *env)
3215 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3219 mdb_env_create(MDB_env **env)
3223 e = calloc(1, sizeof(MDB_env));
3227 e->me_maxreaders = DEFAULT_READERS;
3228 e->me_maxdbs = e->me_numdbs = 2;
3229 e->me_fd = INVALID_HANDLE_VALUE;
3230 e->me_lfd = INVALID_HANDLE_VALUE;
3231 e->me_mfd = INVALID_HANDLE_VALUE;
3232 #ifdef MDB_USE_POSIX_SEM
3233 e->me_rmutex = SEM_FAILED;
3234 e->me_wmutex = SEM_FAILED;
3236 e->me_pid = getpid();
3237 VGMEMP_CREATE(e,0,0);
3243 mdb_env_set_mapsize(MDB_env *env, size_t size)
3247 env->me_mapsize = size;
3249 env->me_maxpg = env->me_mapsize / env->me_psize;
3254 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3258 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3263 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3265 if (env->me_map || readers < 1)
3267 env->me_maxreaders = readers;
3272 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3274 if (!env || !readers)
3276 *readers = env->me_maxreaders;
3280 /** Further setup required for opening an MDB environment
3283 mdb_env_open2(MDB_env *env)
3285 unsigned int flags = env->me_flags;
3293 memset(&meta, 0, sizeof(meta));
3295 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3298 DPUTS("new mdbenv");
3302 /* Was a mapsize configured? */
3303 if (!env->me_mapsize) {
3304 /* If this is a new environment, take the default,
3305 * else use the size recorded in the existing env.
3307 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3308 } else if (env->me_mapsize < meta.mm_mapsize) {
3309 /* If the configured size is smaller, make sure it's
3310 * still big enough. Silently round up to minimum if not.
3312 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3313 if (env->me_mapsize < minsize)
3314 env->me_mapsize = minsize;
3321 LONG sizelo, sizehi;
3322 sizelo = env->me_mapsize & 0xffffffff;
3323 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3325 /* See if we should use QueryLimited */
3327 if ((rc & 0xff) > 5)
3328 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3330 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3332 /* Windows won't create mappings for zero length files.
3333 * Just allocate the maxsize right now.
3336 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3337 || !SetEndOfFile(env->me_fd)
3338 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3341 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3342 PAGE_READWRITE : PAGE_READONLY,
3343 sizehi, sizelo, NULL);
3346 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3347 FILE_MAP_WRITE : FILE_MAP_READ,
3348 0, 0, env->me_mapsize, meta.mm_address);
3349 rc = env->me_map ? 0 : ErrCode();
3357 if (flags & MDB_WRITEMAP) {
3359 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3362 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3364 if (env->me_map == MAP_FAILED) {
3368 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3370 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3372 #ifdef POSIX_MADV_RANDOM
3373 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3374 #endif /* POSIX_MADV_RANDOM */
3375 #endif /* MADV_RANDOM */
3379 if (flags & MDB_FIXEDMAP)
3380 meta.mm_address = env->me_map;
3381 i = mdb_env_init_meta(env, &meta);
3382 if (i != MDB_SUCCESS) {
3385 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3386 /* Can happen because the address argument to mmap() is just a
3387 * hint. mmap() can pick another, e.g. if the range is in use.
3388 * The MAP_FIXED flag would prevent that, but then mmap could
3389 * instead unmap existing pages to make room for the new map.
3391 return EBUSY; /* TODO: Make a new MDB_* error code? */
3393 env->me_psize = meta.mm_psize;
3394 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3395 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3397 env->me_maxpg = env->me_mapsize / env->me_psize;
3399 p = (MDB_page *)env->me_map;
3400 env->me_metas[0] = METADATA(p);
3401 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3405 int toggle = mdb_env_pick_meta(env);
3406 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3408 DPRINTF(("opened database version %u, pagesize %u",
3409 env->me_metas[0]->mm_version, env->me_psize));
3410 DPRINTF(("using meta page %d", toggle));
3411 DPRINTF(("depth: %u", db->md_depth));
3412 DPRINTF(("entries: %"Z"u", db->md_entries));
3413 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3414 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3415 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3416 DPRINTF(("root: %"Z"u", db->md_root));
3424 /** Release a reader thread's slot in the reader lock table.
3425 * This function is called automatically when a thread exits.
3426 * @param[in] ptr This points to the slot in the reader lock table.
3429 mdb_env_reader_dest(void *ptr)
3431 MDB_reader *reader = ptr;
3437 /** Junk for arranging thread-specific callbacks on Windows. This is
3438 * necessarily platform and compiler-specific. Windows supports up
3439 * to 1088 keys. Let's assume nobody opens more than 64 environments
3440 * in a single process, for now. They can override this if needed.
3442 #ifndef MAX_TLS_KEYS
3443 #define MAX_TLS_KEYS 64
3445 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3446 static int mdb_tls_nkeys;
3448 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3452 case DLL_PROCESS_ATTACH: break;
3453 case DLL_THREAD_ATTACH: break;
3454 case DLL_THREAD_DETACH:
3455 for (i=0; i<mdb_tls_nkeys; i++) {
3456 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3457 mdb_env_reader_dest(r);
3460 case DLL_PROCESS_DETACH: break;
3465 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3467 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3471 /* Force some symbol references.
3472 * _tls_used forces the linker to create the TLS directory if not already done
3473 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3475 #pragma comment(linker, "/INCLUDE:_tls_used")
3476 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3477 #pragma const_seg(".CRT$XLB")
3478 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3479 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3482 #pragma comment(linker, "/INCLUDE:__tls_used")
3483 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3484 #pragma data_seg(".CRT$XLB")
3485 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3487 #endif /* WIN 32/64 */
3488 #endif /* !__GNUC__ */
3491 /** Downgrade the exclusive lock on the region back to shared */
3493 mdb_env_share_locks(MDB_env *env, int *excl)
3495 int rc = 0, toggle = mdb_env_pick_meta(env);
3497 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3502 /* First acquire a shared lock. The Unlock will
3503 * then release the existing exclusive lock.
3505 memset(&ov, 0, sizeof(ov));
3506 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3509 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3515 struct flock lock_info;
3516 /* The shared lock replaces the existing lock */
3517 memset((void *)&lock_info, 0, sizeof(lock_info));
3518 lock_info.l_type = F_RDLCK;
3519 lock_info.l_whence = SEEK_SET;
3520 lock_info.l_start = 0;
3521 lock_info.l_len = 1;
3522 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3523 (rc = ErrCode()) == EINTR) ;
3524 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3531 /** Try to get exlusive lock, otherwise shared.
3532 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3535 mdb_env_excl_lock(MDB_env *env, int *excl)
3539 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3543 memset(&ov, 0, sizeof(ov));
3544 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3551 struct flock lock_info;
3552 memset((void *)&lock_info, 0, sizeof(lock_info));
3553 lock_info.l_type = F_WRLCK;
3554 lock_info.l_whence = SEEK_SET;
3555 lock_info.l_start = 0;
3556 lock_info.l_len = 1;
3557 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3558 (rc = ErrCode()) == EINTR) ;
3562 # ifdef MDB_USE_POSIX_SEM
3563 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3566 lock_info.l_type = F_RDLCK;
3567 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3568 (rc = ErrCode()) == EINTR) ;
3576 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3578 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3580 * @(#) $Revision: 5.1 $
3581 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3582 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3584 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3588 * Please do not copyright this code. This code is in the public domain.
3590 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3591 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3592 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3593 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3594 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3595 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3596 * PERFORMANCE OF THIS SOFTWARE.
3599 * chongo <Landon Curt Noll> /\oo/\
3600 * http://www.isthe.com/chongo/
3602 * Share and Enjoy! :-)
3605 typedef unsigned long long mdb_hash_t;
3606 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3608 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3609 * @param[in] val value to hash
3610 * @param[in] hval initial value for hash
3611 * @return 64 bit hash
3613 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3614 * hval arg on the first call.
3617 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3619 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3620 unsigned char *end = s + val->mv_size;
3622 * FNV-1a hash each octet of the string
3625 /* xor the bottom with the current octet */
3626 hval ^= (mdb_hash_t)*s++;
3628 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3629 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3630 (hval << 7) + (hval << 8) + (hval << 40);
3632 /* return our new hash value */
3636 /** Hash the string and output the encoded hash.
3637 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3638 * very short name limits. We don't care about the encoding being reversible,
3639 * we just want to preserve as many bits of the input as possible in a
3640 * small printable string.
3641 * @param[in] str string to hash
3642 * @param[out] encbuf an array of 11 chars to hold the hash
3644 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3647 mdb_pack85(unsigned long l, char *out)
3651 for (i=0; i<5; i++) {
3652 *out++ = mdb_a85[l % 85];
3658 mdb_hash_enc(MDB_val *val, char *encbuf)
3660 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3662 mdb_pack85(h, encbuf);
3663 mdb_pack85(h>>32, encbuf+5);
3668 /** Open and/or initialize the lock region for the environment.
3669 * @param[in] env The MDB environment.
3670 * @param[in] lpath The pathname of the file used for the lock region.
3671 * @param[in] mode The Unix permissions for the file, if we create it.
3672 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3673 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3674 * @return 0 on success, non-zero on failure.
3677 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3680 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3682 # define MDB_ERRCODE_ROFS EROFS
3683 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3684 # define MDB_CLOEXEC O_CLOEXEC
3687 # define MDB_CLOEXEC 0
3694 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3695 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3696 FILE_ATTRIBUTE_NORMAL, NULL);
3698 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3700 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3702 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3707 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3708 /* Lose record locks when exec*() */
3709 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3710 fcntl(env->me_lfd, F_SETFD, fdflags);
3713 if (!(env->me_flags & MDB_NOTLS)) {
3714 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3717 env->me_flags |= MDB_ENV_TXKEY;
3719 /* Windows TLS callbacks need help finding their TLS info. */
3720 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3724 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3728 /* Try to get exclusive lock. If we succeed, then
3729 * nobody is using the lock region and we should initialize it.
3731 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3734 size = GetFileSize(env->me_lfd, NULL);
3736 size = lseek(env->me_lfd, 0, SEEK_END);
3737 if (size == -1) goto fail_errno;
3739 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3740 if (size < rsize && *excl > 0) {
3742 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3743 || !SetEndOfFile(env->me_lfd))
3746 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3750 size = rsize - sizeof(MDB_txninfo);
3751 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3756 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3758 if (!mh) goto fail_errno;
3759 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3761 if (!env->me_txns) goto fail_errno;
3763 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3765 if (m == MAP_FAILED) goto fail_errno;
3771 BY_HANDLE_FILE_INFORMATION stbuf;
3780 if (!mdb_sec_inited) {
3781 InitializeSecurityDescriptor(&mdb_null_sd,
3782 SECURITY_DESCRIPTOR_REVISION);
3783 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3784 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3785 mdb_all_sa.bInheritHandle = FALSE;
3786 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3789 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3790 idbuf.volume = stbuf.dwVolumeSerialNumber;
3791 idbuf.nhigh = stbuf.nFileIndexHigh;
3792 idbuf.nlow = stbuf.nFileIndexLow;
3793 val.mv_data = &idbuf;
3794 val.mv_size = sizeof(idbuf);
3795 mdb_hash_enc(&val, encbuf);
3796 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3797 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3798 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3799 if (!env->me_rmutex) goto fail_errno;
3800 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3801 if (!env->me_wmutex) goto fail_errno;
3802 #elif defined(MDB_USE_POSIX_SEM)
3811 #if defined(__NetBSD__)
3812 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3814 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3815 idbuf.dev = stbuf.st_dev;
3816 idbuf.ino = stbuf.st_ino;
3817 val.mv_data = &idbuf;
3818 val.mv_size = sizeof(idbuf);
3819 mdb_hash_enc(&val, encbuf);
3820 #ifdef MDB_SHORT_SEMNAMES
3821 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3823 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3824 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3825 /* Clean up after a previous run, if needed: Try to
3826 * remove both semaphores before doing anything else.
3828 sem_unlink(env->me_txns->mti_rmname);
3829 sem_unlink(env->me_txns->mti_wmname);
3830 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3831 O_CREAT|O_EXCL, mode, 1);
3832 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3833 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3834 O_CREAT|O_EXCL, mode, 1);
3835 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3836 #else /* MDB_USE_POSIX_SEM */
3837 pthread_mutexattr_t mattr;
3839 if ((rc = pthread_mutexattr_init(&mattr))
3840 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3841 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3842 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3844 pthread_mutexattr_destroy(&mattr);
3845 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3847 env->me_txns->mti_magic = MDB_MAGIC;
3848 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3849 env->me_txns->mti_txnid = 0;
3850 env->me_txns->mti_numreaders = 0;
3853 if (env->me_txns->mti_magic != MDB_MAGIC) {
3854 DPUTS("lock region has invalid magic");
3858 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3859 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3860 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3861 rc = MDB_VERSION_MISMATCH;
3865 if (rc && rc != EACCES && rc != EAGAIN) {
3869 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3870 if (!env->me_rmutex) goto fail_errno;
3871 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3872 if (!env->me_wmutex) goto fail_errno;
3873 #elif defined(MDB_USE_POSIX_SEM)
3874 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3875 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3876 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3877 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3888 /** The name of the lock file in the DB environment */
3889 #define LOCKNAME "/lock.mdb"
3890 /** The name of the data file in the DB environment */
3891 #define DATANAME "/data.mdb"
3892 /** The suffix of the lock file when no subdir is used */
3893 #define LOCKSUFF "-lock"
3894 /** Only a subset of the @ref mdb_env flags can be changed
3895 * at runtime. Changing other flags requires closing the
3896 * environment and re-opening it with the new flags.
3898 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3899 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3902 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3904 int oflags, rc, len, excl = -1;
3905 char *lpath, *dpath;
3907 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3911 if (flags & MDB_NOSUBDIR) {
3912 rc = len + sizeof(LOCKSUFF) + len + 1;
3914 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3919 if (flags & MDB_NOSUBDIR) {
3920 dpath = lpath + len + sizeof(LOCKSUFF);
3921 sprintf(lpath, "%s" LOCKSUFF, path);
3922 strcpy(dpath, path);
3924 dpath = lpath + len + sizeof(LOCKNAME);
3925 sprintf(lpath, "%s" LOCKNAME, path);
3926 sprintf(dpath, "%s" DATANAME, path);
3930 flags |= env->me_flags;
3931 if (flags & MDB_RDONLY) {
3932 /* silently ignore WRITEMAP when we're only getting read access */
3933 flags &= ~MDB_WRITEMAP;
3935 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3936 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3939 env->me_flags = flags |= MDB_ENV_ACTIVE;
3943 env->me_path = strdup(path);
3944 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3945 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3946 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3951 /* For RDONLY, get lockfile after we know datafile exists */
3952 if (!F_ISSET(flags, MDB_RDONLY)) {
3953 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3959 if (F_ISSET(flags, MDB_RDONLY)) {
3960 oflags = GENERIC_READ;
3961 len = OPEN_EXISTING;
3963 oflags = GENERIC_READ|GENERIC_WRITE;
3966 mode = FILE_ATTRIBUTE_NORMAL;
3967 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3968 NULL, len, mode, NULL);
3970 if (F_ISSET(flags, MDB_RDONLY))
3973 oflags = O_RDWR | O_CREAT;
3975 env->me_fd = open(dpath, oflags, mode);
3977 if (env->me_fd == INVALID_HANDLE_VALUE) {
3982 if (F_ISSET(flags, MDB_RDONLY)) {
3983 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3988 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3989 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3990 env->me_mfd = env->me_fd;
3992 /* Synchronous fd for meta writes. Needed even with
3993 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3996 len = OPEN_EXISTING;
3997 env->me_mfd = CreateFile(dpath, oflags,
3998 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3999 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4002 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4004 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4009 DPRINTF(("opened dbenv %p", (void *) env));
4011 rc = mdb_env_share_locks(env, &excl);
4017 mdb_env_close0(env, excl);
4023 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4025 mdb_env_close0(MDB_env *env, int excl)
4029 if (!(env->me_flags & MDB_ENV_ACTIVE))
4032 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4033 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4034 free(env->me_dbxs[i].md_name.mv_data);
4036 free(env->me_dbflags);
4039 free(env->me_dirty_list);
4040 mdb_midl_free(env->me_free_pgs);
4042 if (env->me_flags & MDB_ENV_TXKEY) {
4043 pthread_key_delete(env->me_txkey);
4045 /* Delete our key from the global list */
4046 for (i=0; i<mdb_tls_nkeys; i++)
4047 if (mdb_tls_keys[i] == env->me_txkey) {
4048 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4056 munmap(env->me_map, env->me_mapsize);
4058 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4059 (void) close(env->me_mfd);
4060 if (env->me_fd != INVALID_HANDLE_VALUE)
4061 (void) close(env->me_fd);
4063 pid_t pid = env->me_pid;
4064 /* Clearing readers is done in this function because
4065 * me_txkey with its destructor must be disabled first.
4067 for (i = env->me_numreaders; --i >= 0; )
4068 if (env->me_txns->mti_readers[i].mr_pid == pid)
4069 env->me_txns->mti_readers[i].mr_pid = 0;
4071 if (env->me_rmutex) {
4072 CloseHandle(env->me_rmutex);
4073 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4075 /* Windows automatically destroys the mutexes when
4076 * the last handle closes.
4078 #elif defined(MDB_USE_POSIX_SEM)
4079 if (env->me_rmutex != SEM_FAILED) {
4080 sem_close(env->me_rmutex);
4081 if (env->me_wmutex != SEM_FAILED)
4082 sem_close(env->me_wmutex);
4083 /* If we have the filelock: If we are the
4084 * only remaining user, clean up semaphores.
4087 mdb_env_excl_lock(env, &excl);
4089 sem_unlink(env->me_txns->mti_rmname);
4090 sem_unlink(env->me_txns->mti_wmname);
4094 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4096 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4099 /* Unlock the lockfile. Windows would have unlocked it
4100 * after closing anyway, but not necessarily at once.
4102 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4105 (void) close(env->me_lfd);
4108 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4112 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4114 MDB_txn *txn = NULL;
4120 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4124 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4127 /* Do the lock/unlock of the reader mutex before starting the
4128 * write txn. Otherwise other read txns could block writers.
4130 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4135 /* We must start the actual read txn after blocking writers */
4136 mdb_txn_reset0(txn, "reset-stage1");
4138 /* Temporarily block writers until we snapshot the meta pages */
4141 rc = mdb_txn_renew0(txn);
4143 UNLOCK_MUTEX_W(env);
4148 wsize = env->me_psize * 2;
4152 DO_WRITE(rc, fd, ptr, w2, len);
4156 } else if (len > 0) {
4162 /* Non-blocking or async handles are not supported */
4168 UNLOCK_MUTEX_W(env);
4173 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4175 if (wsize > MAX_WRITE)
4179 DO_WRITE(rc, fd, ptr, w2, len);
4183 } else if (len > 0) {
4200 mdb_env_copy(MDB_env *env, const char *path)
4204 HANDLE newfd = INVALID_HANDLE_VALUE;
4206 if (env->me_flags & MDB_NOSUBDIR) {
4207 lpath = (char *)path;
4210 len += sizeof(DATANAME);
4211 lpath = malloc(len);
4214 sprintf(lpath, "%s" DATANAME, path);
4217 /* The destination path must exist, but the destination file must not.
4218 * We don't want the OS to cache the writes, since the source data is
4219 * already in the OS cache.
4222 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4223 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4225 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4231 if (newfd == INVALID_HANDLE_VALUE) {
4236 #ifdef F_NOCACHE /* __APPLE__ */
4237 rc = fcntl(newfd, F_NOCACHE, 1);
4244 rc = mdb_env_copyfd(env, newfd);
4247 if (!(env->me_flags & MDB_NOSUBDIR))
4249 if (newfd != INVALID_HANDLE_VALUE)
4250 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4257 mdb_env_close(MDB_env *env)
4264 VGMEMP_DESTROY(env);
4265 while ((dp = env->me_dpages) != NULL) {
4266 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4267 env->me_dpages = dp->mp_next;
4271 mdb_env_close0(env, 0);
4275 /** Compare two items pointing at aligned size_t's */
4277 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4279 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4280 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4283 /** Compare two items pointing at aligned int's */
4285 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4287 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4288 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4291 /** Compare two items pointing at ints of unknown alignment.
4292 * Nodes and keys are guaranteed to be 2-byte aligned.
4295 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4297 #if BYTE_ORDER == LITTLE_ENDIAN
4298 unsigned short *u, *c;
4301 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4302 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4305 } while(!x && u > (unsigned short *)a->mv_data);
4308 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4312 /** Compare two items lexically */
4314 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4321 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4327 diff = memcmp(a->mv_data, b->mv_data, len);
4328 return diff ? diff : len_diff<0 ? -1 : len_diff;
4331 /** Compare two items in reverse byte order */
4333 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4335 const unsigned char *p1, *p2, *p1_lim;
4339 p1_lim = (const unsigned char *)a->mv_data;
4340 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4341 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4343 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4349 while (p1 > p1_lim) {
4350 diff = *--p1 - *--p2;
4354 return len_diff<0 ? -1 : len_diff;
4357 /** Search for key within a page, using binary search.
4358 * Returns the smallest entry larger or equal to the key.
4359 * If exactp is non-null, stores whether the found entry was an exact match
4360 * in *exactp (1 or 0).
4361 * Updates the cursor index with the index of the found entry.
4362 * If no entry larger or equal to the key is found, returns NULL.
4365 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4367 unsigned int i = 0, nkeys;
4370 MDB_page *mp = mc->mc_pg[mc->mc_top];
4371 MDB_node *node = NULL;
4376 nkeys = NUMKEYS(mp);
4381 COPY_PGNO(pgno, mp->mp_pgno);
4382 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4383 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4390 low = IS_LEAF(mp) ? 0 : 1;
4392 cmp = mc->mc_dbx->md_cmp;
4394 /* Branch pages have no data, so if using integer keys,
4395 * alignment is guaranteed. Use faster mdb_cmp_int.
4397 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4398 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4405 nodekey.mv_size = mc->mc_db->md_pad;
4406 node = NODEPTR(mp, 0); /* fake */
4407 while (low <= high) {
4408 i = (low + high) >> 1;
4409 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4410 rc = cmp(key, &nodekey);
4411 DPRINTF(("found leaf index %u [%s], rc = %i",
4412 i, DKEY(&nodekey), rc));
4421 while (low <= high) {
4422 i = (low + high) >> 1;
4424 node = NODEPTR(mp, i);
4425 nodekey.mv_size = NODEKSZ(node);
4426 nodekey.mv_data = NODEKEY(node);
4428 rc = cmp(key, &nodekey);
4431 DPRINTF(("found leaf index %u [%s], rc = %i",
4432 i, DKEY(&nodekey), rc));
4434 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4435 i, DKEY(&nodekey), NODEPGNO(node), rc));
4446 if (rc > 0) { /* Found entry is less than the key. */
4447 i++; /* Skip to get the smallest entry larger than key. */
4449 node = NODEPTR(mp, i);
4452 *exactp = (rc == 0);
4453 /* store the key index */
4454 mc->mc_ki[mc->mc_top] = i;
4456 /* There is no entry larger or equal to the key. */
4459 /* nodeptr is fake for LEAF2 */
4465 mdb_cursor_adjust(MDB_cursor *mc, func)
4469 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4470 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4477 /** Pop a page off the top of the cursor's stack. */
4479 mdb_cursor_pop(MDB_cursor *mc)
4483 MDB_page *top = mc->mc_pg[mc->mc_top];
4489 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4490 mc->mc_dbi, (void *) mc));
4494 /** Push a page onto the top of the cursor's stack. */
4496 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4498 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4499 mc->mc_dbi, (void *) mc));
4501 if (mc->mc_snum >= CURSOR_STACK) {
4502 assert(mc->mc_snum < CURSOR_STACK);
4503 return MDB_CURSOR_FULL;
4506 mc->mc_top = mc->mc_snum++;
4507 mc->mc_pg[mc->mc_top] = mp;
4508 mc->mc_ki[mc->mc_top] = 0;
4513 /** Find the address of the page corresponding to a given page number.
4514 * @param[in] txn the transaction for this access.
4515 * @param[in] pgno the page number for the page to retrieve.
4516 * @param[out] ret address of a pointer where the page's address will be stored.
4517 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4518 * @return 0 on success, non-zero on failure.
4521 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4523 MDB_env *env = txn->mt_env;
4527 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4531 MDB_ID2L dl = tx2->mt_u.dirty_list;
4533 /* Spilled pages were dirtied in this txn and flushed
4534 * because the dirty list got full. Bring this page
4535 * back in from the map (but don't unspill it here,
4536 * leave that unless page_touch happens again).
4538 if (tx2->mt_spill_pgs) {
4539 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4540 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4541 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4546 unsigned x = mdb_mid2l_search(dl, pgno);
4547 if (x <= dl[0].mid && dl[x].mid == pgno) {
4553 } while ((tx2 = tx2->mt_parent) != NULL);
4556 if (pgno < txn->mt_next_pgno) {
4558 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4560 DPRINTF(("page %"Z"u not found", pgno));
4562 return MDB_PAGE_NOTFOUND;
4572 /** Search for the page a given key should be in.
4573 * Pushes parent pages on the cursor stack. This function continues a
4574 * search on a cursor that has already been initialized. (Usually by
4575 * #mdb_page_search() but also by #mdb_node_move().)
4576 * @param[in,out] mc the cursor for this operation.
4577 * @param[in] key the key to search for. If NULL, search for the lowest
4578 * page. (This is used by #mdb_cursor_first().)
4579 * @param[in] modify If true, visited pages are updated with new page numbers.
4580 * @return 0 on success, non-zero on failure.
4583 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4585 MDB_page *mp = mc->mc_pg[mc->mc_top];
4589 while (IS_BRANCH(mp)) {
4593 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4594 assert(NUMKEYS(mp) > 1);
4595 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4597 if (key == NULL) /* Initialize cursor to first page. */
4599 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4600 /* cursor to last page */
4604 node = mdb_node_search(mc, key, &exact);
4606 i = NUMKEYS(mp) - 1;
4608 i = mc->mc_ki[mc->mc_top];
4617 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4618 assert(i < NUMKEYS(mp));
4619 node = NODEPTR(mp, i);
4621 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4624 mc->mc_ki[mc->mc_top] = i;
4625 if ((rc = mdb_cursor_push(mc, mp)))
4629 if ((rc = mdb_page_touch(mc)) != 0)
4631 mp = mc->mc_pg[mc->mc_top];
4636 DPRINTF(("internal error, index points to a %02X page!?",
4638 return MDB_CORRUPTED;
4641 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4642 key ? DKEY(key) : NULL));
4643 mc->mc_flags |= C_INITIALIZED;
4644 mc->mc_flags &= ~C_EOF;
4649 /** Search for the lowest key under the current branch page.
4650 * This just bypasses a NUMKEYS check in the current page
4651 * before calling mdb_page_search_root(), because the callers
4652 * are all in situations where the current page is known to
4656 mdb_page_search_lowest(MDB_cursor *mc)
4658 MDB_page *mp = mc->mc_pg[mc->mc_top];
4659 MDB_node *node = NODEPTR(mp, 0);
4662 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4665 mc->mc_ki[mc->mc_top] = 0;
4666 if ((rc = mdb_cursor_push(mc, mp)))
4668 return mdb_page_search_root(mc, NULL, 0);
4671 /** Search for the page a given key should be in.
4672 * Pushes parent pages on the cursor stack. This function just sets up
4673 * the search; it finds the root page for \b mc's database and sets this
4674 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4675 * called to complete the search.
4676 * @param[in,out] mc the cursor for this operation.
4677 * @param[in] key the key to search for. If NULL, search for the lowest
4678 * page. (This is used by #mdb_cursor_first().)
4679 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4680 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4681 * @return 0 on success, non-zero on failure.
4684 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4689 /* Make sure the txn is still viable, then find the root from
4690 * the txn's db table.
4692 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4693 DPUTS("transaction has failed, must abort");
4696 /* Make sure we're using an up-to-date root */
4697 if (mc->mc_dbi > MAIN_DBI) {
4698 if ((*mc->mc_dbflag & DB_STALE) ||
4699 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4701 unsigned char dbflag = 0;
4702 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4703 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4706 if (*mc->mc_dbflag & DB_STALE) {
4710 MDB_node *leaf = mdb_node_search(&mc2,
4711 &mc->mc_dbx->md_name, &exact);
4713 return MDB_NOTFOUND;
4714 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4717 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4719 /* The txn may not know this DBI, or another process may
4720 * have dropped and recreated the DB with other flags.
4722 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4723 return MDB_INCOMPATIBLE;
4724 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4726 if (flags & MDB_PS_MODIFY)
4728 *mc->mc_dbflag &= ~DB_STALE;
4729 *mc->mc_dbflag |= dbflag;
4732 root = mc->mc_db->md_root;
4734 if (root == P_INVALID) { /* Tree is empty. */
4735 DPUTS("tree is empty");
4736 return MDB_NOTFOUND;
4741 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4742 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4748 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4749 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4751 if (flags & MDB_PS_MODIFY) {
4752 if ((rc = mdb_page_touch(mc)))
4756 if (flags & MDB_PS_ROOTONLY)
4759 return mdb_page_search_root(mc, key, flags);
4763 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4765 MDB_txn *txn = mc->mc_txn;
4766 pgno_t pg = mp->mp_pgno;
4767 unsigned x = 0, ovpages = mp->mp_pages;
4768 MDB_env *env = txn->mt_env;
4769 MDB_IDL sl = txn->mt_spill_pgs;
4772 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4773 /* If the page is dirty or on the spill list we just acquired it,
4774 * so we should give it back to our current free list, if any.
4775 * Otherwise put it onto the list of pages we freed in this txn.
4777 * Won't create me_pghead: me_pglast must be inited along with it.
4778 * Unsupported in nested txns: They would need to hide the page
4779 * range in ancestor txns' dirty and spilled lists.
4781 if (env->me_pghead &&
4783 ((mp->mp_flags & P_DIRTY) ||
4784 (sl && (x = mdb_midl_search(sl, pg)) <= sl[0] && sl[x] == pg)))
4788 MDB_ID2 *dl, ix, iy;
4789 rc = mdb_midl_need(&env->me_pghead, ovpages);
4792 if (!(mp->mp_flags & P_DIRTY)) {
4793 /* This page is no longer spilled */
4794 for (; x < sl[0]; x++)
4799 /* Remove from dirty list */
4800 dl = txn->mt_u.dirty_list;
4802 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4810 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4811 txn->mt_flags |= MDB_TXN_ERROR;
4812 return MDB_CORRUPTED;
4815 if (!(env->me_flags & MDB_WRITEMAP))
4816 mdb_dpage_free(env, mp);
4818 /* Insert in me_pghead */
4819 mop = env->me_pghead;
4820 j = mop[0] + ovpages;
4821 for (i = mop[0]; i && mop[i] < pg; i--)
4827 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4831 mc->mc_db->md_overflow_pages -= ovpages;
4835 /** Return the data associated with a given node.
4836 * @param[in] txn The transaction for this operation.
4837 * @param[in] leaf The node being read.
4838 * @param[out] data Updated to point to the node's data.
4839 * @return 0 on success, non-zero on failure.
4842 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4844 MDB_page *omp; /* overflow page */
4848 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4849 data->mv_size = NODEDSZ(leaf);
4850 data->mv_data = NODEDATA(leaf);
4854 /* Read overflow data.
4856 data->mv_size = NODEDSZ(leaf);
4857 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4858 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4859 DPRINTF(("read overflow page %"Z"u failed", pgno));
4862 data->mv_data = METADATA(omp);
4868 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4869 MDB_val *key, MDB_val *data)
4878 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4880 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4883 if (txn->mt_flags & MDB_TXN_ERROR)
4886 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4887 return MDB_BAD_VALSIZE;
4890 mdb_cursor_init(&mc, txn, dbi, &mx);
4891 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4894 /** Find a sibling for a page.
4895 * Replaces the page at the top of the cursor's stack with the
4896 * specified sibling, if one exists.
4897 * @param[in] mc The cursor for this operation.
4898 * @param[in] move_right Non-zero if the right sibling is requested,
4899 * otherwise the left sibling.
4900 * @return 0 on success, non-zero on failure.
4903 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4909 if (mc->mc_snum < 2) {
4910 return MDB_NOTFOUND; /* root has no siblings */
4914 DPRINTF(("parent page is page %"Z"u, index %u",
4915 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4917 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4918 : (mc->mc_ki[mc->mc_top] == 0)) {
4919 DPRINTF(("no more keys left, moving to %s sibling",
4920 move_right ? "right" : "left"));
4921 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4922 /* undo cursor_pop before returning */
4929 mc->mc_ki[mc->mc_top]++;
4931 mc->mc_ki[mc->mc_top]--;
4932 DPRINTF(("just moving to %s index key %u",
4933 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4935 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4937 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4938 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4941 mdb_cursor_push(mc, mp);
4943 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4948 /** Move the cursor to the next data item. */
4950 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4956 if (mc->mc_flags & C_EOF) {
4957 return MDB_NOTFOUND;
4960 assert(mc->mc_flags & C_INITIALIZED);
4962 mp = mc->mc_pg[mc->mc_top];
4964 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4965 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4966 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4967 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4968 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4969 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4973 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4974 if (op == MDB_NEXT_DUP)
4975 return MDB_NOTFOUND;
4979 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4981 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4982 DPUTS("=====> move to next sibling page");
4983 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4984 mc->mc_flags |= C_EOF;
4987 mp = mc->mc_pg[mc->mc_top];
4988 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
4990 mc->mc_ki[mc->mc_top]++;
4992 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
4993 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
4996 key->mv_size = mc->mc_db->md_pad;
4997 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5001 assert(IS_LEAF(mp));
5002 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5004 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5005 mdb_xcursor_init1(mc, leaf);
5008 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5011 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5012 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5013 if (rc != MDB_SUCCESS)
5018 MDB_GET_KEY(leaf, key);
5022 /** Move the cursor to the previous data item. */
5024 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5030 assert(mc->mc_flags & C_INITIALIZED);
5032 mp = mc->mc_pg[mc->mc_top];
5034 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5035 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5036 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5037 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5038 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5039 if (op != MDB_PREV || rc != MDB_NOTFOUND)
5042 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5043 if (op == MDB_PREV_DUP)
5044 return MDB_NOTFOUND;
5049 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5051 if (mc->mc_ki[mc->mc_top] == 0) {
5052 DPUTS("=====> move to prev sibling page");
5053 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5056 mp = mc->mc_pg[mc->mc_top];
5057 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5058 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5060 mc->mc_ki[mc->mc_top]--;
5062 mc->mc_flags &= ~C_EOF;
5064 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5065 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5068 key->mv_size = mc->mc_db->md_pad;
5069 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5073 assert(IS_LEAF(mp));
5074 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5076 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5077 mdb_xcursor_init1(mc, leaf);
5080 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5083 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5084 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5085 if (rc != MDB_SUCCESS)
5090 MDB_GET_KEY(leaf, key);
5094 /** Set the cursor on a specific data item. */
5096 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5097 MDB_cursor_op op, int *exactp)
5101 MDB_node *leaf = NULL;
5106 assert(key->mv_size > 0);
5109 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5111 /* See if we're already on the right page */
5112 if (mc->mc_flags & C_INITIALIZED) {
5115 mp = mc->mc_pg[mc->mc_top];
5117 mc->mc_ki[mc->mc_top] = 0;
5118 return MDB_NOTFOUND;
5120 if (mp->mp_flags & P_LEAF2) {
5121 nodekey.mv_size = mc->mc_db->md_pad;
5122 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5124 leaf = NODEPTR(mp, 0);
5125 MDB_GET_KEY2(leaf, nodekey);
5127 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5129 /* Probably happens rarely, but first node on the page
5130 * was the one we wanted.
5132 mc->mc_ki[mc->mc_top] = 0;
5139 unsigned int nkeys = NUMKEYS(mp);
5141 if (mp->mp_flags & P_LEAF2) {
5142 nodekey.mv_data = LEAF2KEY(mp,
5143 nkeys-1, nodekey.mv_size);
5145 leaf = NODEPTR(mp, nkeys-1);
5146 MDB_GET_KEY2(leaf, nodekey);
5148 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5150 /* last node was the one we wanted */
5151 mc->mc_ki[mc->mc_top] = nkeys-1;
5157 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5158 /* This is definitely the right page, skip search_page */
5159 if (mp->mp_flags & P_LEAF2) {
5160 nodekey.mv_data = LEAF2KEY(mp,
5161 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5163 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5164 MDB_GET_KEY2(leaf, nodekey);
5166 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5168 /* current node was the one we wanted */
5178 /* If any parents have right-sibs, search.
5179 * Otherwise, there's nothing further.
5181 for (i=0; i<mc->mc_top; i++)
5183 NUMKEYS(mc->mc_pg[i])-1)
5185 if (i == mc->mc_top) {
5186 /* There are no other pages */
5187 mc->mc_ki[mc->mc_top] = nkeys;
5188 return MDB_NOTFOUND;
5192 /* There are no other pages */
5193 mc->mc_ki[mc->mc_top] = 0;
5194 return MDB_NOTFOUND;
5198 rc = mdb_page_search(mc, key, 0);
5199 if (rc != MDB_SUCCESS)
5202 mp = mc->mc_pg[mc->mc_top];
5203 assert(IS_LEAF(mp));
5206 leaf = mdb_node_search(mc, key, exactp);
5207 if (exactp != NULL && !*exactp) {
5208 /* MDB_SET specified and not an exact match. */
5209 return MDB_NOTFOUND;
5213 DPUTS("===> inexact leaf not found, goto sibling");
5214 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5215 return rc; /* no entries matched */
5216 mp = mc->mc_pg[mc->mc_top];
5217 assert(IS_LEAF(mp));
5218 leaf = NODEPTR(mp, 0);
5222 mc->mc_flags |= C_INITIALIZED;
5223 mc->mc_flags &= ~C_EOF;
5226 key->mv_size = mc->mc_db->md_pad;
5227 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5231 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5232 mdb_xcursor_init1(mc, leaf);
5235 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5236 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5237 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5240 if (op == MDB_GET_BOTH) {
5246 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5247 if (rc != MDB_SUCCESS)
5250 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5252 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5254 rc = mc->mc_dbx->md_dcmp(data, &d2);
5256 if (op == MDB_GET_BOTH || rc > 0)
5257 return MDB_NOTFOUND;
5262 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5263 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5268 /* The key already matches in all other cases */
5269 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5270 MDB_GET_KEY(leaf, key);
5271 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5276 /** Move the cursor to the first item in the database. */
5278 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5284 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5286 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5287 rc = mdb_page_search(mc, NULL, 0);
5288 if (rc != MDB_SUCCESS)
5291 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5293 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5294 mc->mc_flags |= C_INITIALIZED;
5295 mc->mc_flags &= ~C_EOF;
5297 mc->mc_ki[mc->mc_top] = 0;
5299 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5300 key->mv_size = mc->mc_db->md_pad;
5301 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5306 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5307 mdb_xcursor_init1(mc, leaf);
5308 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5312 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5316 MDB_GET_KEY(leaf, key);
5320 /** Move the cursor to the last item in the database. */
5322 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5328 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5330 if (!(mc->mc_flags & C_EOF)) {
5332 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5335 lkey.mv_size = MDB_MAXKEYSIZE+1;
5336 lkey.mv_data = NULL;
5337 rc = mdb_page_search(mc, &lkey, 0);
5338 if (rc != MDB_SUCCESS)
5341 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5344 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5345 mc->mc_flags |= C_INITIALIZED|C_EOF;
5346 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
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], mc->mc_ki[mc->mc_top], key->mv_size);
5355 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5356 mdb_xcursor_init1(mc, leaf);
5357 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5361 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5366 MDB_GET_KEY(leaf, key);
5371 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5376 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5380 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5384 case MDB_GET_CURRENT:
5385 if (!(mc->mc_flags & C_INITIALIZED)) {
5388 MDB_page *mp = mc->mc_pg[mc->mc_top];
5390 mc->mc_ki[mc->mc_top] = 0;
5396 key->mv_size = mc->mc_db->md_pad;
5397 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5399 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5400 MDB_GET_KEY(leaf, key);
5402 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5403 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5405 rc = mdb_node_read(mc->mc_txn, leaf, data);
5412 case MDB_GET_BOTH_RANGE:
5417 if (mc->mc_xcursor == NULL) {
5418 rc = MDB_INCOMPATIBLE;
5427 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5428 rc = MDB_BAD_VALSIZE;
5429 } else if (op == MDB_SET_RANGE)
5430 rc = mdb_cursor_set(mc, key, data, op, NULL);
5432 rc = mdb_cursor_set(mc, key, data, op, &exact);
5434 case MDB_GET_MULTIPLE:
5435 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5439 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5440 rc = MDB_INCOMPATIBLE;
5444 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5445 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5448 case MDB_NEXT_MULTIPLE:
5453 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5454 rc = MDB_INCOMPATIBLE;
5457 if (!(mc->mc_flags & C_INITIALIZED))
5458 rc = mdb_cursor_first(mc, key, data);
5460 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5461 if (rc == MDB_SUCCESS) {
5462 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5465 mx = &mc->mc_xcursor->mx_cursor;
5466 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5468 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5469 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5477 case MDB_NEXT_NODUP:
5478 if (!(mc->mc_flags & C_INITIALIZED))
5479 rc = mdb_cursor_first(mc, key, data);
5481 rc = mdb_cursor_next(mc, key, data, op);
5485 case MDB_PREV_NODUP:
5486 if (!(mc->mc_flags & C_INITIALIZED)) {
5487 rc = mdb_cursor_last(mc, key, data);
5490 mc->mc_flags |= C_INITIALIZED;
5491 mc->mc_ki[mc->mc_top]++;
5493 rc = mdb_cursor_prev(mc, key, data, op);
5496 rc = mdb_cursor_first(mc, key, data);
5499 mfunc = mdb_cursor_first;
5501 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5505 if (mc->mc_xcursor == NULL) {
5506 rc = MDB_INCOMPATIBLE;
5509 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5513 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5516 rc = mdb_cursor_last(mc, key, data);
5519 mfunc = mdb_cursor_last;
5522 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5530 /** Touch all the pages in the cursor stack.
5531 * Makes sure all the pages are writable, before attempting a write operation.
5532 * @param[in] mc The cursor to operate on.
5535 mdb_cursor_touch(MDB_cursor *mc)
5539 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5542 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5543 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5546 *mc->mc_dbflag |= DB_DIRTY;
5548 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5549 rc = mdb_page_touch(mc);
5553 mc->mc_top = mc->mc_snum-1;
5557 /** Do not spill pages to disk if txn is getting full, may fail instead */
5558 #define MDB_NOSPILL 0x8000
5561 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5564 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5565 MDB_node *leaf = NULL;
5566 MDB_val xdata, *rdata, dkey;
5569 int do_sub = 0, insert = 0;
5570 unsigned int mcount = 0, dcount = 0, nospill;
5574 char dbuf[MDB_MAXKEYSIZE+1];
5575 unsigned int nflags;
5578 /* Check this first so counter will always be zero on any
5581 if (flags & MDB_MULTIPLE) {
5582 dcount = data[1].mv_size;
5583 data[1].mv_size = 0;
5584 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5585 return MDB_INCOMPATIBLE;
5588 nospill = flags & MDB_NOSPILL;
5589 flags &= ~MDB_NOSPILL;
5591 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5592 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5594 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5595 return MDB_BAD_VALSIZE;
5597 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5598 return MDB_BAD_VALSIZE;
5600 #if SIZE_MAX > MAXDATASIZE
5601 if (data->mv_size > MAXDATASIZE)
5602 return MDB_BAD_VALSIZE;
5605 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5606 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5610 if (flags == MDB_CURRENT) {
5611 if (!(mc->mc_flags & C_INITIALIZED))
5614 } else if (mc->mc_db->md_root == P_INVALID) {
5615 /* new database, cursor has nothing to point to */
5617 mc->mc_flags &= ~C_INITIALIZED;
5622 if (flags & MDB_APPEND) {
5624 rc = mdb_cursor_last(mc, &k2, &d2);
5626 rc = mc->mc_dbx->md_cmp(key, &k2);
5629 mc->mc_ki[mc->mc_top]++;
5631 /* new key is <= last key */
5636 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5638 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5639 DPRINTF(("duplicate key [%s]", DKEY(key)));
5641 return MDB_KEYEXIST;
5643 if (rc && rc != MDB_NOTFOUND)
5647 /* Cursor is positioned, check for room in the dirty list */
5649 if (flags & MDB_MULTIPLE) {
5651 xdata.mv_size = data->mv_size * dcount;
5655 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5659 if (rc == MDB_NO_ROOT) {
5661 /* new database, write a root leaf page */
5662 DPUTS("allocating new root leaf page");
5663 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5666 mdb_cursor_push(mc, np);
5667 mc->mc_db->md_root = np->mp_pgno;
5668 mc->mc_db->md_depth++;
5669 *mc->mc_dbflag |= DB_DIRTY;
5670 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5672 np->mp_flags |= P_LEAF2;
5673 mc->mc_flags |= C_INITIALIZED;
5675 /* make sure all cursor pages are writable */
5676 rc2 = mdb_cursor_touch(mc);
5681 /* The key already exists */
5682 if (rc == MDB_SUCCESS) {
5683 /* there's only a key anyway, so this is a no-op */
5684 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5685 unsigned int ksize = mc->mc_db->md_pad;
5686 if (key->mv_size != ksize)
5687 return MDB_BAD_VALSIZE;
5688 if (flags == MDB_CURRENT) {
5689 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5690 memcpy(ptr, key->mv_data, ksize);
5695 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5698 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5699 /* Was a single item before, must convert now */
5701 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5702 /* Just overwrite the current item */
5703 if (flags == MDB_CURRENT)
5706 dkey.mv_size = NODEDSZ(leaf);
5707 dkey.mv_data = NODEDATA(leaf);
5708 #if UINT_MAX < SIZE_MAX
5709 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5710 #ifdef MISALIGNED_OK
5711 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5713 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5716 /* if data matches, skip it */
5717 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5718 if (flags & MDB_NODUPDATA)
5720 else if (flags & MDB_MULTIPLE)
5727 /* create a fake page for the dup items */
5728 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5729 dkey.mv_data = dbuf;
5730 fp = (MDB_page *)&pbuf;
5731 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5732 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5733 fp->mp_lower = PAGEHDRSZ;
5734 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5735 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5736 fp->mp_flags |= P_LEAF2;
5737 fp->mp_pad = data->mv_size;
5738 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5740 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5741 (dkey.mv_size & 1) + (data->mv_size & 1);
5743 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5746 xdata.mv_size = fp->mp_upper;
5751 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5752 /* See if we need to convert from fake page to subDB */
5754 unsigned int offset;
5758 fp = NODEDATA(leaf);
5759 if (flags == MDB_CURRENT) {
5761 fp->mp_flags |= P_DIRTY;
5762 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5763 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5767 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5768 offset = fp->mp_pad;
5769 if (SIZELEFT(fp) >= offset)
5771 offset *= 4; /* space for 4 more */
5773 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5775 offset += offset & 1;
5776 fp_flags = fp->mp_flags;
5777 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5778 offset >= mc->mc_txn->mt_env->me_nodemax) {
5779 /* yes, convert it */
5781 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5782 dummy.md_pad = fp->mp_pad;
5783 dummy.md_flags = MDB_DUPFIXED;
5784 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5785 dummy.md_flags |= MDB_INTEGERKEY;
5788 dummy.md_branch_pages = 0;
5789 dummy.md_leaf_pages = 1;
5790 dummy.md_overflow_pages = 0;
5791 dummy.md_entries = NUMKEYS(fp);
5793 xdata.mv_size = sizeof(MDB_db);
5794 xdata.mv_data = &dummy;
5795 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5797 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5798 flags |= F_DUPDATA|F_SUBDATA;
5799 dummy.md_root = mp->mp_pgno;
5800 fp_flags &= ~P_SUBP;
5802 /* no, just grow it */
5804 xdata.mv_size = NODEDSZ(leaf) + offset;
5805 xdata.mv_data = &pbuf;
5806 mp = (MDB_page *)&pbuf;
5807 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5810 mp->mp_flags = fp_flags | P_DIRTY;
5811 mp->mp_pad = fp->mp_pad;
5812 mp->mp_lower = fp->mp_lower;
5813 mp->mp_upper = fp->mp_upper + offset;
5815 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5817 nsize = NODEDSZ(leaf) - fp->mp_upper;
5818 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5819 for (i=0; i<NUMKEYS(fp); i++)
5820 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5822 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5826 /* data is on sub-DB, just store it */
5827 flags |= F_DUPDATA|F_SUBDATA;
5831 /* overflow page overwrites need special handling */
5832 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5835 unsigned psize = mc->mc_txn->mt_env->me_psize;
5836 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5838 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5839 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5841 ovpages = omp->mp_pages;
5843 /* Is the ov page large enough? */
5844 if (ovpages >= dpages) {
5845 if (!(omp->mp_flags & P_DIRTY) &&
5846 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5848 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5851 level = 0; /* dirty in this txn or clean */
5854 if (omp->mp_flags & P_DIRTY) {
5855 /* yes, overwrite it. Note in this case we don't
5856 * bother to try shrinking the page if the new data
5857 * is smaller than the overflow threshold.
5860 /* It is writable only in a parent txn */
5861 size_t sz = (size_t) psize * ovpages, off;
5862 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5868 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5869 if (!(flags & MDB_RESERVE)) {
5870 /* Copy end of page, adjusting alignment so
5871 * compiler may copy words instead of bytes.
5873 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5874 memcpy((size_t *)((char *)np + off),
5875 (size_t *)((char *)omp + off), sz - off);
5878 memcpy(np, omp, sz); /* Copy beginning of page */
5881 SETDSZ(leaf, data->mv_size);
5882 if (F_ISSET(flags, MDB_RESERVE))
5883 data->mv_data = METADATA(omp);
5885 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5889 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5891 } else if (NODEDSZ(leaf) == data->mv_size) {
5892 /* same size, just replace it. Note that we could
5893 * also reuse this node if the new data is smaller,
5894 * but instead we opt to shrink the node in that case.
5896 if (F_ISSET(flags, MDB_RESERVE))
5897 data->mv_data = NODEDATA(leaf);
5898 else if (data->mv_size)
5899 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5901 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5904 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5905 mc->mc_db->md_entries--;
5907 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5914 nflags = flags & NODE_ADD_FLAGS;
5915 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5916 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5917 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5918 nflags &= ~MDB_APPEND;
5920 nflags |= MDB_SPLIT_REPLACE;
5921 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5923 /* There is room already in this leaf page. */
5924 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5925 if (rc == 0 && !do_sub && insert) {
5926 /* Adjust other cursors pointing to mp */
5927 MDB_cursor *m2, *m3;
5928 MDB_dbi dbi = mc->mc_dbi;
5929 unsigned i = mc->mc_top;
5930 MDB_page *mp = mc->mc_pg[i];
5932 if (mc->mc_flags & C_SUB)
5935 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5936 if (mc->mc_flags & C_SUB)
5937 m3 = &m2->mc_xcursor->mx_cursor;
5940 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5941 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5948 if (rc != MDB_SUCCESS)
5949 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5951 /* Now store the actual data in the child DB. Note that we're
5952 * storing the user data in the keys field, so there are strict
5953 * size limits on dupdata. The actual data fields of the child
5954 * DB are all zero size.
5961 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5962 if (flags & MDB_CURRENT) {
5963 xflags = MDB_CURRENT|MDB_NOSPILL;
5965 mdb_xcursor_init1(mc, leaf);
5966 xflags = (flags & MDB_NODUPDATA) ?
5967 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5969 /* converted, write the original data first */
5971 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5975 /* Adjust other cursors pointing to mp */
5977 unsigned i = mc->mc_top;
5978 MDB_page *mp = mc->mc_pg[i];
5980 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5981 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5982 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5983 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5984 mdb_xcursor_init1(m2, leaf);
5988 /* we've done our job */
5991 if (flags & MDB_APPENDDUP)
5992 xflags |= MDB_APPEND;
5993 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5994 if (flags & F_SUBDATA) {
5995 void *db = NODEDATA(leaf);
5996 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5999 /* sub-writes might have failed so check rc again.
6000 * Don't increment count if we just replaced an existing item.
6002 if (!rc && !(flags & MDB_CURRENT))
6003 mc->mc_db->md_entries++;
6004 if (flags & MDB_MULTIPLE) {
6008 /* let caller know how many succeeded, if any */
6009 data[1].mv_size = mcount;
6010 if (mcount < dcount) {
6011 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6012 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6019 /* If we succeeded and the key didn't exist before, make sure
6020 * the cursor is marked valid.
6023 mc->mc_flags |= C_INITIALIZED;
6028 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6033 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6034 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6036 if (!(mc->mc_flags & C_INITIALIZED))
6039 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6041 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
6043 rc = mdb_cursor_touch(mc);
6047 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6049 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6050 if (!(flags & MDB_NODUPDATA)) {
6051 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6052 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6054 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6055 /* If sub-DB still has entries, we're done */
6056 if (mc->mc_xcursor->mx_db.md_entries) {
6057 if (leaf->mn_flags & F_SUBDATA) {
6058 /* update subDB info */
6059 void *db = NODEDATA(leaf);
6060 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6063 /* shrink fake page */
6064 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6065 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6066 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6067 /* fix other sub-DB cursors pointed at this fake page */
6068 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6069 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6070 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6071 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6072 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6075 mc->mc_db->md_entries--;
6078 /* otherwise fall thru and delete the sub-DB */
6081 if (leaf->mn_flags & F_SUBDATA) {
6082 /* add all the child DB's pages to the free list */
6083 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6084 if (rc == MDB_SUCCESS) {
6085 mc->mc_db->md_entries -=
6086 mc->mc_xcursor->mx_db.md_entries;
6091 return mdb_cursor_del0(mc, leaf);
6094 /** Allocate and initialize new pages for a database.
6095 * @param[in] mc a cursor on the database being added to.
6096 * @param[in] flags flags defining what type of page is being allocated.
6097 * @param[in] num the number of pages to allocate. This is usually 1,
6098 * unless allocating overflow pages for a large record.
6099 * @param[out] mp Address of a page, or NULL on failure.
6100 * @return 0 on success, non-zero on failure.
6103 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6108 if ((rc = mdb_page_alloc(mc, num, &np)))
6110 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6111 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6112 np->mp_flags = flags | P_DIRTY;
6113 np->mp_lower = PAGEHDRSZ;
6114 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6117 mc->mc_db->md_branch_pages++;
6118 else if (IS_LEAF(np))
6119 mc->mc_db->md_leaf_pages++;
6120 else if (IS_OVERFLOW(np)) {
6121 mc->mc_db->md_overflow_pages += num;
6129 /** Calculate the size of a leaf node.
6130 * The size depends on the environment's page size; if a data item
6131 * is too large it will be put onto an overflow page and the node
6132 * size will only include the key and not the data. Sizes are always
6133 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6134 * of the #MDB_node headers.
6135 * @param[in] env The environment handle.
6136 * @param[in] key The key for the node.
6137 * @param[in] data The data for the node.
6138 * @return The number of bytes needed to store the node.
6141 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6145 sz = LEAFSIZE(key, data);
6146 if (sz >= env->me_nodemax) {
6147 /* put on overflow page */
6148 sz -= data->mv_size - sizeof(pgno_t);
6152 return sz + sizeof(indx_t);
6155 /** Calculate the size of a branch node.
6156 * The size should depend on the environment's page size but since
6157 * we currently don't support spilling large keys onto overflow
6158 * pages, it's simply the size of the #MDB_node header plus the
6159 * size of the key. Sizes are always rounded up to an even number
6160 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6161 * @param[in] env The environment handle.
6162 * @param[in] key The key for the node.
6163 * @return The number of bytes needed to store the node.
6166 mdb_branch_size(MDB_env *env, MDB_val *key)
6171 if (sz >= env->me_nodemax) {
6172 /* put on overflow page */
6173 /* not implemented */
6174 /* sz -= key->size - sizeof(pgno_t); */
6177 return sz + sizeof(indx_t);
6180 /** Add a node to the page pointed to by the cursor.
6181 * @param[in] mc The cursor for this operation.
6182 * @param[in] indx The index on the page where the new node should be added.
6183 * @param[in] key The key for the new node.
6184 * @param[in] data The data for the new node, if any.
6185 * @param[in] pgno The page number, if adding a branch node.
6186 * @param[in] flags Flags for the node.
6187 * @return 0 on success, non-zero on failure. Possible errors are:
6189 * <li>ENOMEM - failed to allocate overflow pages for the node.
6190 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6191 * should never happen since all callers already calculate the
6192 * page's free space before calling this function.
6196 mdb_node_add(MDB_cursor *mc, indx_t indx,
6197 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6200 size_t node_size = NODESIZE;
6203 MDB_page *mp = mc->mc_pg[mc->mc_top];
6204 MDB_page *ofp = NULL; /* overflow page */
6207 assert(mp->mp_upper >= mp->mp_lower);
6209 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6210 IS_LEAF(mp) ? "leaf" : "branch",
6211 IS_SUBP(mp) ? "sub-" : "",
6212 mp->mp_pgno, indx, data ? data->mv_size : 0,
6213 key ? key->mv_size : 0, key ? DKEY(key) : NULL));
6216 /* Move higher keys up one slot. */
6217 int ksize = mc->mc_db->md_pad, dif;
6218 char *ptr = LEAF2KEY(mp, indx, ksize);
6219 dif = NUMKEYS(mp) - indx;
6221 memmove(ptr+ksize, ptr, dif*ksize);
6222 /* insert new key */
6223 memcpy(ptr, key->mv_data, ksize);
6225 /* Just using these for counting */
6226 mp->mp_lower += sizeof(indx_t);
6227 mp->mp_upper -= ksize - sizeof(indx_t);
6232 node_size += key->mv_size;
6236 if (F_ISSET(flags, F_BIGDATA)) {
6237 /* Data already on overflow page. */
6238 node_size += sizeof(pgno_t);
6239 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6240 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6242 /* Put data on overflow page. */
6243 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6244 data->mv_size, node_size+data->mv_size));
6245 node_size += sizeof(pgno_t);
6246 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6248 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6251 node_size += data->mv_size;
6254 node_size += node_size & 1;
6256 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6257 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6258 mp->mp_pgno, NUMKEYS(mp)));
6259 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6260 mp->mp_upper - mp->mp_lower));
6261 DPRINTF(("node size = %"Z"u", node_size));
6262 return MDB_PAGE_FULL;
6265 /* Move higher pointers up one slot. */
6266 for (i = NUMKEYS(mp); i > indx; i--)
6267 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6269 /* Adjust free space offsets. */
6270 ofs = mp->mp_upper - node_size;
6271 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6272 mp->mp_ptrs[indx] = ofs;
6274 mp->mp_lower += sizeof(indx_t);
6276 /* Write the node data. */
6277 node = NODEPTR(mp, indx);
6278 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6279 node->mn_flags = flags;
6281 SETDSZ(node,data->mv_size);
6286 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6291 if (F_ISSET(flags, F_BIGDATA))
6292 memcpy(node->mn_data + key->mv_size, data->mv_data,
6294 else if (F_ISSET(flags, MDB_RESERVE))
6295 data->mv_data = node->mn_data + key->mv_size;
6297 memcpy(node->mn_data + key->mv_size, data->mv_data,
6300 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6302 if (F_ISSET(flags, MDB_RESERVE))
6303 data->mv_data = METADATA(ofp);
6305 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6312 /** Delete the specified node from a page.
6313 * @param[in] mp The page to operate on.
6314 * @param[in] indx The index of the node to delete.
6315 * @param[in] ksize The size of a node. Only used if the page is
6316 * part of a #MDB_DUPFIXED database.
6319 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6322 indx_t i, j, numkeys, ptr;
6329 COPY_PGNO(pgno, mp->mp_pgno);
6330 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6331 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6334 assert(indx < NUMKEYS(mp));
6337 int x = NUMKEYS(mp) - 1 - indx;
6338 base = LEAF2KEY(mp, indx, ksize);
6340 memmove(base, base + ksize, x * ksize);
6341 mp->mp_lower -= sizeof(indx_t);
6342 mp->mp_upper += ksize - sizeof(indx_t);
6346 node = NODEPTR(mp, indx);
6347 sz = NODESIZE + node->mn_ksize;
6349 if (F_ISSET(node->mn_flags, F_BIGDATA))
6350 sz += sizeof(pgno_t);
6352 sz += NODEDSZ(node);
6356 ptr = mp->mp_ptrs[indx];
6357 numkeys = NUMKEYS(mp);
6358 for (i = j = 0; i < numkeys; i++) {
6360 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6361 if (mp->mp_ptrs[i] < ptr)
6362 mp->mp_ptrs[j] += sz;
6367 base = (char *)mp + mp->mp_upper;
6368 memmove(base + sz, base, ptr - mp->mp_upper);
6370 mp->mp_lower -= sizeof(indx_t);
6374 /** Compact the main page after deleting a node on a subpage.
6375 * @param[in] mp The main page to operate on.
6376 * @param[in] indx The index of the subpage on the main page.
6379 mdb_node_shrink(MDB_page *mp, indx_t indx)
6386 indx_t i, numkeys, ptr;
6388 node = NODEPTR(mp, indx);
6389 sp = (MDB_page *)NODEDATA(node);
6390 osize = NODEDSZ(node);
6392 delta = sp->mp_upper - sp->mp_lower;
6393 SETDSZ(node, osize - delta);
6394 xp = (MDB_page *)((char *)sp + delta);
6396 /* shift subpage upward */
6398 nsize = NUMKEYS(sp) * sp->mp_pad;
6399 memmove(METADATA(xp), METADATA(sp), nsize);
6402 nsize = osize - sp->mp_upper;
6403 numkeys = NUMKEYS(sp);
6404 for (i=numkeys-1; i>=0; i--)
6405 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6407 xp->mp_upper = sp->mp_lower;
6408 xp->mp_lower = sp->mp_lower;
6409 xp->mp_flags = sp->mp_flags;
6410 xp->mp_pad = sp->mp_pad;
6411 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6413 /* shift lower nodes upward */
6414 ptr = mp->mp_ptrs[indx];
6415 numkeys = NUMKEYS(mp);
6416 for (i = 0; i < numkeys; i++) {
6417 if (mp->mp_ptrs[i] <= ptr)
6418 mp->mp_ptrs[i] += delta;
6421 base = (char *)mp + mp->mp_upper;
6422 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6423 mp->mp_upper += delta;
6426 /** Initial setup of a sorted-dups cursor.
6427 * Sorted duplicates are implemented as a sub-database for the given key.
6428 * The duplicate data items are actually keys of the sub-database.
6429 * Operations on the duplicate data items are performed using a sub-cursor
6430 * initialized when the sub-database is first accessed. This function does
6431 * the preliminary setup of the sub-cursor, filling in the fields that
6432 * depend only on the parent DB.
6433 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6436 mdb_xcursor_init0(MDB_cursor *mc)
6438 MDB_xcursor *mx = mc->mc_xcursor;
6440 mx->mx_cursor.mc_xcursor = NULL;
6441 mx->mx_cursor.mc_txn = mc->mc_txn;
6442 mx->mx_cursor.mc_db = &mx->mx_db;
6443 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6444 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6445 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6446 mx->mx_cursor.mc_snum = 0;
6447 mx->mx_cursor.mc_top = 0;
6448 mx->mx_cursor.mc_flags = C_SUB;
6449 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6450 mx->mx_dbx.md_dcmp = NULL;
6451 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6454 /** Final setup of a sorted-dups cursor.
6455 * Sets up the fields that depend on the data from the main cursor.
6456 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6457 * @param[in] node The data containing the #MDB_db record for the
6458 * sorted-dup database.
6461 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6463 MDB_xcursor *mx = mc->mc_xcursor;
6465 if (node->mn_flags & F_SUBDATA) {
6466 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6467 mx->mx_cursor.mc_pg[0] = 0;
6468 mx->mx_cursor.mc_snum = 0;
6469 mx->mx_cursor.mc_flags = C_SUB;
6471 MDB_page *fp = NODEDATA(node);
6472 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6473 mx->mx_db.md_flags = 0;
6474 mx->mx_db.md_depth = 1;
6475 mx->mx_db.md_branch_pages = 0;
6476 mx->mx_db.md_leaf_pages = 1;
6477 mx->mx_db.md_overflow_pages = 0;
6478 mx->mx_db.md_entries = NUMKEYS(fp);
6479 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6480 mx->mx_cursor.mc_snum = 1;
6481 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6482 mx->mx_cursor.mc_top = 0;
6483 mx->mx_cursor.mc_pg[0] = fp;
6484 mx->mx_cursor.mc_ki[0] = 0;
6485 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6486 mx->mx_db.md_flags = MDB_DUPFIXED;
6487 mx->mx_db.md_pad = fp->mp_pad;
6488 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6489 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6492 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6493 mx->mx_db.md_root));
6494 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6496 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6497 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6498 #if UINT_MAX < SIZE_MAX
6499 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6500 #ifdef MISALIGNED_OK
6501 mx->mx_dbx.md_cmp = mdb_cmp_long;
6503 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6508 /** Initialize a cursor for a given transaction and database. */
6510 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6513 mc->mc_backup = NULL;
6516 mc->mc_db = &txn->mt_dbs[dbi];
6517 mc->mc_dbx = &txn->mt_dbxs[dbi];
6518 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6523 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6525 mc->mc_xcursor = mx;
6526 mdb_xcursor_init0(mc);
6528 mc->mc_xcursor = NULL;
6530 if (*mc->mc_dbflag & DB_STALE) {
6531 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6536 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6539 size_t size = sizeof(MDB_cursor);
6541 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6544 if (txn->mt_flags & MDB_TXN_ERROR)
6547 /* Allow read access to the freelist */
6548 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6551 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6552 size += sizeof(MDB_xcursor);
6554 if ((mc = malloc(size)) != NULL) {
6555 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6556 if (txn->mt_cursors) {
6557 mc->mc_next = txn->mt_cursors[dbi];
6558 txn->mt_cursors[dbi] = mc;
6559 mc->mc_flags |= C_UNTRACK;
6571 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6573 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6576 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6579 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6583 /* Return the count of duplicate data items for the current key */
6585 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6589 if (mc == NULL || countp == NULL)
6592 if (mc->mc_xcursor == NULL)
6593 return MDB_INCOMPATIBLE;
6595 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6596 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6599 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6602 *countp = mc->mc_xcursor->mx_db.md_entries;
6608 mdb_cursor_close(MDB_cursor *mc)
6610 if (mc && !mc->mc_backup) {
6611 /* remove from txn, if tracked */
6612 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6613 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6614 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6616 *prev = mc->mc_next;
6623 mdb_cursor_txn(MDB_cursor *mc)
6625 if (!mc) return NULL;
6630 mdb_cursor_dbi(MDB_cursor *mc)
6636 /** Replace the key for a node with a new key.
6637 * @param[in] mc Cursor pointing to the node to operate on.
6638 * @param[in] key The new key to use.
6639 * @return 0 on success, non-zero on failure.
6642 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6649 indx_t ptr, i, numkeys, indx;
6652 indx = mc->mc_ki[mc->mc_top];
6653 mp = mc->mc_pg[mc->mc_top];
6654 node = NODEPTR(mp, indx);
6655 ptr = mp->mp_ptrs[indx];
6659 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6660 k2.mv_data = NODEKEY(node);
6661 k2.mv_size = node->mn_ksize;
6662 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6664 mdb_dkey(&k2, kbuf2),
6670 delta0 = delta = key->mv_size - node->mn_ksize;
6672 /* Must be 2-byte aligned. If new key is
6673 * shorter by 1, the shift will be skipped.
6675 delta += (delta & 1);
6677 if (delta > 0 && SIZELEFT(mp) < delta) {
6679 /* not enough space left, do a delete and split */
6680 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6681 pgno = NODEPGNO(node);
6682 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6683 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6686 numkeys = NUMKEYS(mp);
6687 for (i = 0; i < numkeys; i++) {
6688 if (mp->mp_ptrs[i] <= ptr)
6689 mp->mp_ptrs[i] -= delta;
6692 base = (char *)mp + mp->mp_upper;
6693 len = ptr - mp->mp_upper + NODESIZE;
6694 memmove(base - delta, base, len);
6695 mp->mp_upper -= delta;
6697 node = NODEPTR(mp, indx);
6700 /* But even if no shift was needed, update ksize */
6702 node->mn_ksize = key->mv_size;
6705 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6711 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6713 /** Move a node from csrc to cdst.
6716 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6723 unsigned short flags;
6727 /* Mark src and dst as dirty. */
6728 if ((rc = mdb_page_touch(csrc)) ||
6729 (rc = mdb_page_touch(cdst)))
6732 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6733 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6734 key.mv_size = csrc->mc_db->md_pad;
6735 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6737 data.mv_data = NULL;
6741 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6742 assert(!((long)srcnode&1));
6743 srcpg = NODEPGNO(srcnode);
6744 flags = srcnode->mn_flags;
6745 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6746 unsigned int snum = csrc->mc_snum;
6748 /* must find the lowest key below src */
6749 mdb_page_search_lowest(csrc);
6750 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6751 key.mv_size = csrc->mc_db->md_pad;
6752 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6754 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6755 key.mv_size = NODEKSZ(s2);
6756 key.mv_data = NODEKEY(s2);
6758 csrc->mc_snum = snum--;
6759 csrc->mc_top = snum;
6761 key.mv_size = NODEKSZ(srcnode);
6762 key.mv_data = NODEKEY(srcnode);
6764 data.mv_size = NODEDSZ(srcnode);
6765 data.mv_data = NODEDATA(srcnode);
6767 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6768 unsigned int snum = cdst->mc_snum;
6771 /* must find the lowest key below dst */
6772 mdb_page_search_lowest(cdst);
6773 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6774 bkey.mv_size = cdst->mc_db->md_pad;
6775 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6777 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6778 bkey.mv_size = NODEKSZ(s2);
6779 bkey.mv_data = NODEKEY(s2);
6781 cdst->mc_snum = snum--;
6782 cdst->mc_top = snum;
6783 mdb_cursor_copy(cdst, &mn);
6785 rc = mdb_update_key(&mn, &bkey);
6790 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6791 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6792 csrc->mc_ki[csrc->mc_top],
6794 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6795 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6797 /* Add the node to the destination page.
6799 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6800 if (rc != MDB_SUCCESS)
6803 /* Delete the node from the source page.
6805 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6808 /* Adjust other cursors pointing to mp */
6809 MDB_cursor *m2, *m3;
6810 MDB_dbi dbi = csrc->mc_dbi;
6811 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6813 if (csrc->mc_flags & C_SUB)
6816 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6817 if (csrc->mc_flags & C_SUB)
6818 m3 = &m2->mc_xcursor->mx_cursor;
6821 if (m3 == csrc) continue;
6822 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6823 csrc->mc_ki[csrc->mc_top]) {
6824 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6825 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6830 /* Update the parent separators.
6832 if (csrc->mc_ki[csrc->mc_top] == 0) {
6833 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6834 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6835 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6837 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6838 key.mv_size = NODEKSZ(srcnode);
6839 key.mv_data = NODEKEY(srcnode);
6841 DPRINTF(("update separator for source page %"Z"u to [%s]",
6842 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6843 mdb_cursor_copy(csrc, &mn);
6846 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6849 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6851 indx_t ix = csrc->mc_ki[csrc->mc_top];
6852 nullkey.mv_size = 0;
6853 csrc->mc_ki[csrc->mc_top] = 0;
6854 rc = mdb_update_key(csrc, &nullkey);
6855 csrc->mc_ki[csrc->mc_top] = ix;
6856 assert(rc == MDB_SUCCESS);
6860 if (cdst->mc_ki[cdst->mc_top] == 0) {
6861 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6862 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6863 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6865 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6866 key.mv_size = NODEKSZ(srcnode);
6867 key.mv_data = NODEKEY(srcnode);
6869 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6870 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6871 mdb_cursor_copy(cdst, &mn);
6874 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6877 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6879 indx_t ix = cdst->mc_ki[cdst->mc_top];
6880 nullkey.mv_size = 0;
6881 cdst->mc_ki[cdst->mc_top] = 0;
6882 rc = mdb_update_key(cdst, &nullkey);
6883 cdst->mc_ki[cdst->mc_top] = ix;
6884 assert(rc == MDB_SUCCESS);
6891 /** Merge one page into another.
6892 * The nodes from the page pointed to by \b csrc will
6893 * be copied to the page pointed to by \b cdst and then
6894 * the \b csrc page will be freed.
6895 * @param[in] csrc Cursor pointing to the source page.
6896 * @param[in] cdst Cursor pointing to the destination page.
6899 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6907 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6908 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6910 assert(csrc->mc_snum > 1); /* can't merge root page */
6911 assert(cdst->mc_snum > 1);
6913 /* Mark dst as dirty. */
6914 if ((rc = mdb_page_touch(cdst)))
6917 /* Move all nodes from src to dst.
6919 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6920 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6921 key.mv_size = csrc->mc_db->md_pad;
6922 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6923 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6924 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6925 if (rc != MDB_SUCCESS)
6927 key.mv_data = (char *)key.mv_data + key.mv_size;
6930 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6931 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6932 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6933 unsigned int snum = csrc->mc_snum;
6935 /* must find the lowest key below src */
6936 mdb_page_search_lowest(csrc);
6937 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6938 key.mv_size = csrc->mc_db->md_pad;
6939 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6941 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6942 key.mv_size = NODEKSZ(s2);
6943 key.mv_data = NODEKEY(s2);
6945 csrc->mc_snum = snum--;
6946 csrc->mc_top = snum;
6948 key.mv_size = srcnode->mn_ksize;
6949 key.mv_data = NODEKEY(srcnode);
6952 data.mv_size = NODEDSZ(srcnode);
6953 data.mv_data = NODEDATA(srcnode);
6954 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6955 if (rc != MDB_SUCCESS)
6960 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6961 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
6962 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
6964 /* Unlink the src page from parent and add to free list.
6966 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6967 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6970 rc = mdb_update_key(csrc, &key);
6976 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6977 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6980 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6981 csrc->mc_db->md_leaf_pages--;
6983 csrc->mc_db->md_branch_pages--;
6985 /* Adjust other cursors pointing to mp */
6986 MDB_cursor *m2, *m3;
6987 MDB_dbi dbi = csrc->mc_dbi;
6988 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6990 if (csrc->mc_flags & C_SUB)
6993 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6994 if (csrc->mc_flags & C_SUB)
6995 m3 = &m2->mc_xcursor->mx_cursor;
6998 if (m3 == csrc) continue;
6999 if (m3->mc_snum < csrc->mc_snum) continue;
7000 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7001 m3->mc_pg[csrc->mc_top] = mp;
7002 m3->mc_ki[csrc->mc_top] += nkeys;
7006 mdb_cursor_pop(csrc);
7008 return mdb_rebalance(csrc);
7011 /** Copy the contents of a cursor.
7012 * @param[in] csrc The cursor to copy from.
7013 * @param[out] cdst The cursor to copy to.
7016 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7020 cdst->mc_txn = csrc->mc_txn;
7021 cdst->mc_dbi = csrc->mc_dbi;
7022 cdst->mc_db = csrc->mc_db;
7023 cdst->mc_dbx = csrc->mc_dbx;
7024 cdst->mc_snum = csrc->mc_snum;
7025 cdst->mc_top = csrc->mc_top;
7026 cdst->mc_flags = csrc->mc_flags;
7028 for (i=0; i<csrc->mc_snum; i++) {
7029 cdst->mc_pg[i] = csrc->mc_pg[i];
7030 cdst->mc_ki[i] = csrc->mc_ki[i];
7034 /** Rebalance the tree after a delete operation.
7035 * @param[in] mc Cursor pointing to the page where rebalancing
7037 * @return 0 on success, non-zero on failure.
7040 mdb_rebalance(MDB_cursor *mc)
7044 unsigned int ptop, minkeys;
7047 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7051 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7052 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7053 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7054 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7055 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7059 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7060 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7063 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7064 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7070 if (mc->mc_snum < 2) {
7071 MDB_page *mp = mc->mc_pg[0];
7073 DPUTS("Can't rebalance a subpage, ignoring");
7076 if (NUMKEYS(mp) == 0) {
7077 DPUTS("tree is completely empty");
7078 mc->mc_db->md_root = P_INVALID;
7079 mc->mc_db->md_depth = 0;
7080 mc->mc_db->md_leaf_pages = 0;
7081 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7084 /* Adjust cursors pointing to mp */
7088 MDB_cursor *m2, *m3;
7089 MDB_dbi dbi = mc->mc_dbi;
7091 if (mc->mc_flags & C_SUB)
7094 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7095 if (mc->mc_flags & C_SUB)
7096 m3 = &m2->mc_xcursor->mx_cursor;
7099 if (m3->mc_snum < mc->mc_snum) continue;
7100 if (m3->mc_pg[0] == mp) {
7106 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7107 DPUTS("collapsing root page!");
7108 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7111 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7112 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7115 mc->mc_db->md_depth--;
7116 mc->mc_db->md_branch_pages--;
7117 mc->mc_ki[0] = mc->mc_ki[1];
7119 /* Adjust other cursors pointing to mp */
7120 MDB_cursor *m2, *m3;
7121 MDB_dbi dbi = mc->mc_dbi;
7123 if (mc->mc_flags & C_SUB)
7126 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7127 if (mc->mc_flags & C_SUB)
7128 m3 = &m2->mc_xcursor->mx_cursor;
7131 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7132 if (m3->mc_pg[0] == mp) {
7133 m3->mc_pg[0] = mc->mc_pg[0];
7136 m3->mc_ki[0] = m3->mc_ki[1];
7141 DPUTS("root page doesn't need rebalancing");
7145 /* The parent (branch page) must have at least 2 pointers,
7146 * otherwise the tree is invalid.
7148 ptop = mc->mc_top-1;
7149 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7151 /* Leaf page fill factor is below the threshold.
7152 * Try to move keys from left or right neighbor, or
7153 * merge with a neighbor page.
7158 mdb_cursor_copy(mc, &mn);
7159 mn.mc_xcursor = NULL;
7161 if (mc->mc_ki[ptop] == 0) {
7162 /* We're the leftmost leaf in our parent.
7164 DPUTS("reading right neighbor");
7166 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7167 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7170 mn.mc_ki[mn.mc_top] = 0;
7171 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7173 /* There is at least one neighbor to the left.
7175 DPUTS("reading left neighbor");
7177 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7178 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7181 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7182 mc->mc_ki[mc->mc_top] = 0;
7185 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7186 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7187 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7189 /* If the neighbor page is above threshold and has enough keys,
7190 * move one key from it. Otherwise we should try to merge them.
7191 * (A branch page must never have less than 2 keys.)
7193 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7194 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7195 return mdb_node_move(&mn, mc);
7197 if (mc->mc_ki[ptop] == 0)
7198 rc = mdb_page_merge(&mn, mc);
7200 rc = mdb_page_merge(mc, &mn);
7201 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7206 /** Complete a delete operation started by #mdb_cursor_del(). */
7208 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7214 mp = mc->mc_pg[mc->mc_top];
7215 ki = mc->mc_ki[mc->mc_top];
7217 /* add overflow pages to free list */
7218 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7222 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7223 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7224 (rc = mdb_ovpage_free(mc, omp)))
7227 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7228 mc->mc_db->md_entries--;
7229 rc = mdb_rebalance(mc);
7230 if (rc != MDB_SUCCESS)
7231 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7232 /* if mc points past last node in page, invalidate */
7233 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7234 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7237 /* Adjust other cursors pointing to mp */
7240 MDB_dbi dbi = mc->mc_dbi;
7242 mp = mc->mc_pg[mc->mc_top];
7243 nkeys = NUMKEYS(mp);
7244 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7247 if (!(m2->mc_flags & C_INITIALIZED))
7249 if (m2->mc_pg[mc->mc_top] == mp) {
7250 if (m2->mc_ki[mc->mc_top] > ki)
7251 m2->mc_ki[mc->mc_top]--;
7252 if (m2->mc_ki[mc->mc_top] >= nkeys)
7253 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7262 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7263 MDB_val *key, MDB_val *data)
7268 MDB_val rdata, *xdata;
7272 assert(key != NULL);
7274 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7276 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7279 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7280 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7282 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7283 return MDB_BAD_VALSIZE;
7286 mdb_cursor_init(&mc, txn, dbi, &mx);
7289 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7290 /* must ignore any data */
7301 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7303 /* let mdb_page_split know about this cursor if needed:
7304 * delete will trigger a rebalance; if it needs to move
7305 * a node from one page to another, it will have to
7306 * update the parent's separator key(s). If the new sepkey
7307 * is larger than the current one, the parent page may
7308 * run out of space, triggering a split. We need this
7309 * cursor to be consistent until the end of the rebalance.
7311 mc.mc_flags |= C_UNTRACK;
7312 mc.mc_next = txn->mt_cursors[dbi];
7313 txn->mt_cursors[dbi] = &mc;
7314 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7315 txn->mt_cursors[dbi] = mc.mc_next;
7320 /** Split a page and insert a new node.
7321 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7322 * The cursor will be updated to point to the actual page and index where
7323 * the node got inserted after the split.
7324 * @param[in] newkey The key for the newly inserted node.
7325 * @param[in] newdata The data for the newly inserted node.
7326 * @param[in] newpgno The page number, if the new node is a branch node.
7327 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7328 * @return 0 on success, non-zero on failure.
7331 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7332 unsigned int nflags)
7335 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7338 unsigned int i, j, split_indx, nkeys, pmax;
7340 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7342 MDB_page *mp, *rp, *pp;
7347 mp = mc->mc_pg[mc->mc_top];
7348 newindx = mc->mc_ki[mc->mc_top];
7350 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7351 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7352 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7354 /* Create a right sibling. */
7355 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7357 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7359 if (mc->mc_snum < 2) {
7360 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7362 /* shift current top to make room for new parent */
7363 mc->mc_pg[1] = mc->mc_pg[0];
7364 mc->mc_ki[1] = mc->mc_ki[0];
7367 mc->mc_db->md_root = pp->mp_pgno;
7368 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7369 mc->mc_db->md_depth++;
7372 /* Add left (implicit) pointer. */
7373 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7374 /* undo the pre-push */
7375 mc->mc_pg[0] = mc->mc_pg[1];
7376 mc->mc_ki[0] = mc->mc_ki[1];
7377 mc->mc_db->md_root = mp->mp_pgno;
7378 mc->mc_db->md_depth--;
7385 ptop = mc->mc_top-1;
7386 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7389 mc->mc_flags |= C_SPLITTING;
7390 mdb_cursor_copy(mc, &mn);
7391 mn.mc_pg[mn.mc_top] = rp;
7392 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7394 if (nflags & MDB_APPEND) {
7395 mn.mc_ki[mn.mc_top] = 0;
7397 split_indx = newindx;
7402 nkeys = NUMKEYS(mp);
7403 split_indx = nkeys / 2;
7404 if (newindx < split_indx)
7410 unsigned int lsize, rsize, ksize;
7411 /* Move half of the keys to the right sibling */
7413 x = mc->mc_ki[mc->mc_top] - split_indx;
7414 ksize = mc->mc_db->md_pad;
7415 split = LEAF2KEY(mp, split_indx, ksize);
7416 rsize = (nkeys - split_indx) * ksize;
7417 lsize = (nkeys - split_indx) * sizeof(indx_t);
7418 mp->mp_lower -= lsize;
7419 rp->mp_lower += lsize;
7420 mp->mp_upper += rsize - lsize;
7421 rp->mp_upper -= rsize - lsize;
7422 sepkey.mv_size = ksize;
7423 if (newindx == split_indx) {
7424 sepkey.mv_data = newkey->mv_data;
7426 sepkey.mv_data = split;
7429 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7430 memcpy(rp->mp_ptrs, split, rsize);
7431 sepkey.mv_data = rp->mp_ptrs;
7432 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7433 memcpy(ins, newkey->mv_data, ksize);
7434 mp->mp_lower += sizeof(indx_t);
7435 mp->mp_upper -= ksize - sizeof(indx_t);
7438 memcpy(rp->mp_ptrs, split, x * ksize);
7439 ins = LEAF2KEY(rp, x, ksize);
7440 memcpy(ins, newkey->mv_data, ksize);
7441 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7442 rp->mp_lower += sizeof(indx_t);
7443 rp->mp_upper -= ksize - sizeof(indx_t);
7444 mc->mc_ki[mc->mc_top] = x;
7445 mc->mc_pg[mc->mc_top] = rp;
7450 /* For leaf pages, check the split point based on what
7451 * fits where, since otherwise mdb_node_add can fail.
7453 * This check is only needed when the data items are
7454 * relatively large, such that being off by one will
7455 * make the difference between success or failure.
7457 * It's also relevant if a page happens to be laid out
7458 * such that one half of its nodes are all "small" and
7459 * the other half of its nodes are "large." If the new
7460 * item is also "large" and falls on the half with
7461 * "large" nodes, it also may not fit.
7464 unsigned int psize, nsize;
7465 /* Maximum free space in an empty page */
7466 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7467 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7468 if ((nkeys < 20) || (nsize > pmax/16)) {
7469 if (newindx <= split_indx) {
7472 for (i=0; i<split_indx; i++) {
7473 node = NODEPTR(mp, i);
7474 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7475 if (F_ISSET(node->mn_flags, F_BIGDATA))
7476 psize += sizeof(pgno_t);
7478 psize += NODEDSZ(node);
7482 split_indx = newindx;
7493 for (i=nkeys-1; i>=split_indx; i--) {
7494 node = NODEPTR(mp, i);
7495 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7496 if (F_ISSET(node->mn_flags, F_BIGDATA))
7497 psize += sizeof(pgno_t);
7499 psize += NODEDSZ(node);
7503 split_indx = newindx;
7514 /* First find the separating key between the split pages.
7515 * The case where newindx == split_indx is ambiguous; the
7516 * new item could go to the new page or stay on the original
7517 * page. If newpos == 1 it goes to the new page.
7519 if (newindx == split_indx && newpos) {
7520 sepkey.mv_size = newkey->mv_size;
7521 sepkey.mv_data = newkey->mv_data;
7523 node = NODEPTR(mp, split_indx);
7524 sepkey.mv_size = node->mn_ksize;
7525 sepkey.mv_data = NODEKEY(node);
7529 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7531 /* Copy separator key to the parent.
7533 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7537 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7540 if (mn.mc_snum == mc->mc_snum) {
7541 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7542 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7543 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7544 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7549 /* Right page might now have changed parent.
7550 * Check if left page also changed parent.
7552 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7553 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7554 for (i=0; i<ptop; i++) {
7555 mc->mc_pg[i] = mn.mc_pg[i];
7556 mc->mc_ki[i] = mn.mc_ki[i];
7558 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7559 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7563 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7566 mc->mc_flags ^= C_SPLITTING;
7567 if (rc != MDB_SUCCESS) {
7570 if (nflags & MDB_APPEND) {
7571 mc->mc_pg[mc->mc_top] = rp;
7572 mc->mc_ki[mc->mc_top] = 0;
7573 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7576 for (i=0; i<mc->mc_top; i++)
7577 mc->mc_ki[i] = mn.mc_ki[i];
7584 /* Move half of the keys to the right sibling. */
7586 /* grab a page to hold a temporary copy */
7587 copy = mdb_page_malloc(mc->mc_txn, 1);
7591 copy->mp_pgno = mp->mp_pgno;
7592 copy->mp_flags = mp->mp_flags;
7593 copy->mp_lower = PAGEHDRSZ;
7594 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7595 mc->mc_pg[mc->mc_top] = copy;
7596 for (i = j = 0; i <= nkeys; j++) {
7597 if (i == split_indx) {
7598 /* Insert in right sibling. */
7599 /* Reset insert index for right sibling. */
7600 if (i != newindx || (newpos ^ ins_new)) {
7602 mc->mc_pg[mc->mc_top] = rp;
7606 if (i == newindx && !ins_new) {
7607 /* Insert the original entry that caused the split. */
7608 rkey.mv_data = newkey->mv_data;
7609 rkey.mv_size = newkey->mv_size;
7618 /* Update index for the new key. */
7619 mc->mc_ki[mc->mc_top] = j;
7620 } else if (i == nkeys) {
7623 node = NODEPTR(mp, i);
7624 rkey.mv_data = NODEKEY(node);
7625 rkey.mv_size = node->mn_ksize;
7627 xdata.mv_data = NODEDATA(node);
7628 xdata.mv_size = NODEDSZ(node);
7631 pgno = NODEPGNO(node);
7632 flags = node->mn_flags;
7637 if (!IS_LEAF(mp) && j == 0) {
7638 /* First branch index doesn't need key data. */
7642 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7646 nkeys = NUMKEYS(copy);
7647 for (i=0; i<nkeys; i++)
7648 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7649 mp->mp_lower = copy->mp_lower;
7650 mp->mp_upper = copy->mp_upper;
7651 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7652 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7654 /* reset back to original page */
7655 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7656 mc->mc_pg[mc->mc_top] = mp;
7657 if (nflags & MDB_RESERVE) {
7658 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7659 if (!(node->mn_flags & F_BIGDATA))
7660 newdata->mv_data = NODEDATA(node);
7664 /* Make sure mc_ki is still valid.
7666 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7667 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7668 for (i=0; i<ptop; i++) {
7669 mc->mc_pg[i] = mn.mc_pg[i];
7670 mc->mc_ki[i] = mn.mc_ki[i];
7672 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7673 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7677 /* return tmp page to freelist */
7678 mdb_page_free(mc->mc_txn->mt_env, copy);
7681 /* Adjust other cursors pointing to mp */
7682 MDB_cursor *m2, *m3;
7683 MDB_dbi dbi = mc->mc_dbi;
7684 int fixup = NUMKEYS(mp);
7686 if (mc->mc_flags & C_SUB)
7689 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7690 if (mc->mc_flags & C_SUB)
7691 m3 = &m2->mc_xcursor->mx_cursor;
7696 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7698 if (m3->mc_flags & C_SPLITTING)
7703 for (k=m3->mc_top; k>=0; k--) {
7704 m3->mc_ki[k+1] = m3->mc_ki[k];
7705 m3->mc_pg[k+1] = m3->mc_pg[k];
7707 if (m3->mc_ki[0] >= split_indx) {
7712 m3->mc_pg[0] = mc->mc_pg[0];
7716 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7717 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7718 m3->mc_ki[mc->mc_top]++;
7719 if (m3->mc_ki[mc->mc_top] >= fixup) {
7720 m3->mc_pg[mc->mc_top] = rp;
7721 m3->mc_ki[mc->mc_top] -= fixup;
7722 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7724 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7725 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7734 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7735 MDB_val *key, MDB_val *data, unsigned int flags)
7740 assert(key != NULL);
7741 assert(data != NULL);
7743 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7746 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7747 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7749 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7750 return MDB_BAD_VALSIZE;
7753 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7756 mdb_cursor_init(&mc, txn, dbi, &mx);
7757 return mdb_cursor_put(&mc, key, data, flags);
7761 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7763 if ((flag & CHANGEABLE) != flag)
7766 env->me_flags |= flag;
7768 env->me_flags &= ~flag;
7773 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7778 *arg = env->me_flags;
7783 mdb_env_get_path(MDB_env *env, const char **arg)
7788 *arg = env->me_path;
7792 /** Common code for #mdb_stat() and #mdb_env_stat().
7793 * @param[in] env the environment to operate in.
7794 * @param[in] db the #MDB_db record containing the stats to return.
7795 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7796 * @return 0, this function always succeeds.
7799 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7801 arg->ms_psize = env->me_psize;
7802 arg->ms_depth = db->md_depth;
7803 arg->ms_branch_pages = db->md_branch_pages;
7804 arg->ms_leaf_pages = db->md_leaf_pages;
7805 arg->ms_overflow_pages = db->md_overflow_pages;
7806 arg->ms_entries = db->md_entries;
7811 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7815 if (env == NULL || arg == NULL)
7818 toggle = mdb_env_pick_meta(env);
7820 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7824 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7828 if (env == NULL || arg == NULL)
7831 toggle = mdb_env_pick_meta(env);
7832 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7833 arg->me_mapsize = env->me_mapsize;
7834 arg->me_maxreaders = env->me_maxreaders;
7836 /* me_numreaders may be zero if this process never used any readers. Use
7837 * the shared numreader count if it exists.
7839 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7841 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7842 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7846 /** Set the default comparison functions for a database.
7847 * Called immediately after a database is opened to set the defaults.
7848 * The user can then override them with #mdb_set_compare() or
7849 * #mdb_set_dupsort().
7850 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7851 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7854 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7856 uint16_t f = txn->mt_dbs[dbi].md_flags;
7858 txn->mt_dbxs[dbi].md_cmp =
7859 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7860 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7862 txn->mt_dbxs[dbi].md_dcmp =
7863 !(f & MDB_DUPSORT) ? 0 :
7864 ((f & MDB_INTEGERDUP)
7865 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7866 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7869 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7874 int rc, dbflag, exact;
7875 unsigned int unused = 0;
7878 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7879 mdb_default_cmp(txn, FREE_DBI);
7882 if ((flags & VALID_FLAGS) != flags)
7884 if (txn->mt_flags & MDB_TXN_ERROR)
7890 if (flags & PERSISTENT_FLAGS) {
7891 uint16_t f2 = flags & PERSISTENT_FLAGS;
7892 /* make sure flag changes get committed */
7893 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7894 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7895 txn->mt_flags |= MDB_TXN_DIRTY;
7898 mdb_default_cmp(txn, MAIN_DBI);
7902 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7903 mdb_default_cmp(txn, MAIN_DBI);
7906 /* Is the DB already open? */
7908 for (i=2; i<txn->mt_numdbs; i++) {
7909 if (!txn->mt_dbxs[i].md_name.mv_size) {
7910 /* Remember this free slot */
7911 if (!unused) unused = i;
7914 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7915 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7921 /* If no free slot and max hit, fail */
7922 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7923 return MDB_DBS_FULL;
7925 /* Cannot mix named databases with some mainDB flags */
7926 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7927 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7929 /* Find the DB info */
7930 dbflag = DB_NEW|DB_VALID;
7933 key.mv_data = (void *)name;
7934 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7935 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7936 if (rc == MDB_SUCCESS) {
7937 /* make sure this is actually a DB */
7938 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7939 if (!(node->mn_flags & F_SUBDATA))
7940 return MDB_INCOMPATIBLE;
7941 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7942 /* Create if requested */
7944 data.mv_size = sizeof(MDB_db);
7945 data.mv_data = &dummy;
7946 memset(&dummy, 0, sizeof(dummy));
7947 dummy.md_root = P_INVALID;
7948 dummy.md_flags = flags & PERSISTENT_FLAGS;
7949 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7953 /* OK, got info, add to table */
7954 if (rc == MDB_SUCCESS) {
7955 unsigned int slot = unused ? unused : txn->mt_numdbs;
7956 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7957 txn->mt_dbxs[slot].md_name.mv_size = len;
7958 txn->mt_dbxs[slot].md_rel = NULL;
7959 txn->mt_dbflags[slot] = dbflag;
7960 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7962 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7963 mdb_default_cmp(txn, slot);
7972 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7974 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7977 if (txn->mt_dbflags[dbi] & DB_STALE) {
7980 /* Stale, must read the DB's root. cursor_init does it for us. */
7981 mdb_cursor_init(&mc, txn, dbi, &mx);
7983 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7986 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7989 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7991 ptr = env->me_dbxs[dbi].md_name.mv_data;
7992 env->me_dbxs[dbi].md_name.mv_data = NULL;
7993 env->me_dbxs[dbi].md_name.mv_size = 0;
7994 env->me_dbflags[dbi] = 0;
7998 int mdb_dbi_flags(MDB_env *env, MDB_dbi dbi, unsigned int *flags)
8000 /* We could return the flags for the FREE_DBI too but what's the point? */
8001 if (dbi < MAIN_DBI || dbi >= env->me_numdbs)
8003 *flags = env->me_dbflags[dbi];
8007 /** Add all the DB's pages to the free list.
8008 * @param[in] mc Cursor on the DB to free.
8009 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8010 * @return 0 on success, non-zero on failure.
8013 mdb_drop0(MDB_cursor *mc, int subs)
8017 rc = mdb_page_search(mc, NULL, 0);
8018 if (rc == MDB_SUCCESS) {
8019 MDB_txn *txn = mc->mc_txn;
8024 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8025 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8028 mdb_cursor_copy(mc, &mx);
8029 while (mc->mc_snum > 0) {
8030 MDB_page *mp = mc->mc_pg[mc->mc_top];
8031 unsigned n = NUMKEYS(mp);
8033 for (i=0; i<n; i++) {
8034 ni = NODEPTR(mp, i);
8035 if (ni->mn_flags & F_BIGDATA) {
8038 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8039 rc = mdb_page_get(txn, pg, &omp, NULL);
8042 assert(IS_OVERFLOW(omp));
8043 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8047 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8048 mdb_xcursor_init1(mc, ni);
8049 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8055 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8057 for (i=0; i<n; i++) {
8059 ni = NODEPTR(mp, i);
8062 mdb_midl_xappend(txn->mt_free_pgs, pg);
8067 mc->mc_ki[mc->mc_top] = i;
8068 rc = mdb_cursor_sibling(mc, 1);
8070 /* no more siblings, go back to beginning
8071 * of previous level.
8075 for (i=1; i<mc->mc_snum; i++) {
8077 mc->mc_pg[i] = mx.mc_pg[i];
8082 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8083 } else if (rc == MDB_NOTFOUND) {
8089 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8091 MDB_cursor *mc, *m2;
8094 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8097 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8100 rc = mdb_cursor_open(txn, dbi, &mc);
8104 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8105 /* Invalidate the dropped DB's cursors */
8106 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8107 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8111 /* Can't delete the main DB */
8112 if (del && dbi > MAIN_DBI) {
8113 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8115 txn->mt_dbflags[dbi] = DB_STALE;
8116 mdb_dbi_close(txn->mt_env, dbi);
8119 /* reset the DB record, mark it dirty */
8120 txn->mt_dbflags[dbi] |= DB_DIRTY;
8121 txn->mt_dbs[dbi].md_depth = 0;
8122 txn->mt_dbs[dbi].md_branch_pages = 0;
8123 txn->mt_dbs[dbi].md_leaf_pages = 0;
8124 txn->mt_dbs[dbi].md_overflow_pages = 0;
8125 txn->mt_dbs[dbi].md_entries = 0;
8126 txn->mt_dbs[dbi].md_root = P_INVALID;
8128 txn->mt_flags |= MDB_TXN_DIRTY;
8131 mdb_cursor_close(mc);
8135 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8137 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8140 txn->mt_dbxs[dbi].md_cmp = cmp;
8144 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8146 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8149 txn->mt_dbxs[dbi].md_dcmp = cmp;
8153 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8155 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8158 txn->mt_dbxs[dbi].md_rel = rel;
8162 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8164 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8167 txn->mt_dbxs[dbi].md_relctx = ctx;
8171 int mdb_env_get_maxkeysize(MDB_env *env)
8173 return MDB_MAXKEYSIZE;
8176 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8178 unsigned int i, rdrs;
8185 if (!env->me_txns) {
8186 return func("(no reader locks)\n", ctx);
8188 rdrs = env->me_txns->mti_numreaders;
8189 mr = env->me_txns->mti_readers;
8190 for (i=0; i<rdrs; i++) {
8195 if (mr[i].mr_txnid == (txnid_t)-1) {
8196 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8198 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8202 func(" pid thread txnid\n", ctx);
8204 rc = func(buf, ctx);
8210 func("(no active readers)\n", ctx);
8215 /* insert pid into list if not already present.
8216 * return -1 if already present.
8218 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8220 /* binary search of pid in list */
8222 unsigned cursor = 1;
8224 unsigned n = ids[0];
8227 unsigned pivot = n >> 1;
8228 cursor = base + pivot + 1;
8229 val = pid - ids[cursor];
8234 } else if ( val > 0 ) {
8239 /* found, so it's a duplicate */
8248 for (n = ids[0]; n > cursor; n--)
8254 int mdb_reader_check(MDB_env *env, int *dead)
8256 unsigned int i, j, rdrs;
8267 rdrs = env->me_txns->mti_numreaders;
8268 pids = malloc((rdrs+1) * sizeof(pid_t));
8272 mr = env->me_txns->mti_readers;
8274 for (i=0; i<rdrs; i++) {
8275 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8277 if (mdb_pid_insert(pids, pid) == 0) {
8278 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8280 /* Recheck, a new process may have reused pid */
8281 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8282 for (j=i; j<rdrs; j++)
8283 if (mr[j].mr_pid == pid) {
8288 UNLOCK_MUTEX_R(env);