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 dirty, non-overflow, non-sub pages watched by txn.
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.
1361 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1362 * @return 0 on success, non-zero on failure.
1365 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1367 MDB_txn *txn = mc->mc_txn;
1372 int rc = MDB_SUCCESS, level;
1374 /* Mark pages seen by cursors */
1375 if (mc->mc_flags & C_UNTRACK)
1376 mc = NULL; /* will find mc in mt_cursors */
1377 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1378 for (; mc; mc=mc->mc_next) {
1379 for (m3 = mc; m3->mc_flags & C_INITIALIZED; m3 = &mx->mx_cursor) {
1380 for (j=0; j<m3->mc_snum; j++)
1381 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1383 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1384 mx = m3->mc_xcursor;
1394 /* Mark dirty root pages */
1395 for (i=0; i<txn->mt_numdbs; i++) {
1396 if (txn->mt_dbflags[i] & DB_DIRTY) {
1397 pgno_t pgno = txn->mt_dbs[i].md_root;
1398 if (pgno == P_INVALID)
1400 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1402 if ((dp->mp_flags & (P_DIRTY|P_KEEP)) == pflags && level <= 1)
1403 dp->mp_flags ^= P_KEEP;
1411 static int mdb_page_flush(MDB_txn *txn, int keep);
1413 /** Spill pages from the dirty list back to disk.
1414 * This is intended to prevent running into #MDB_TXN_FULL situations,
1415 * but note that they may still occur in a few cases:
1416 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1417 * are too many of these dirtied in one txn, the txn may still get
1419 * 2) child txns may run out of space if their parents dirtied a
1420 * lot of pages and never spilled them. TODO: we probably should do
1421 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1422 * the parent's dirty_room is below a given threshold.
1423 * 3) our estimate of the txn size could be too small. At the
1424 * moment this seems unlikely.
1426 * Otherwise, if not using nested txns, it is expected that apps will
1427 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1428 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1429 * If the txn never references them again, they can be left alone.
1430 * If the txn only reads them, they can be used without any fuss.
1431 * If the txn writes them again, they can be dirtied immediately without
1432 * going thru all of the work of #mdb_page_touch(). Such references are
1433 * handled by #mdb_page_unspill().
1435 * Also note, we never spill DB root pages, nor pages of active cursors,
1436 * because we'll need these back again soon anyway. And in nested txns,
1437 * we can't spill a page in a child txn if it was already spilled in a
1438 * parent txn. That would alter the parent txns' data even though
1439 * the child hasn't committed yet, and we'd have no way to undo it if
1440 * the child aborted.
1442 * @param[in] m0 cursor A cursor handle identifying the transaction and
1443 * database for which we are checking space.
1444 * @param[in] key For a put operation, the key being stored.
1445 * @param[in] data For a put operation, the data being stored.
1446 * @return 0 on success, non-zero on failure.
1449 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1451 MDB_txn *txn = m0->mc_txn;
1453 MDB_ID2L dl = txn->mt_u.dirty_list;
1454 unsigned int i, j, need;
1457 if (m0->mc_flags & C_SUB)
1460 /* Estimate how much space this op will take */
1461 i = m0->mc_db->md_depth;
1462 /* Named DBs also dirty the main DB */
1463 if (m0->mc_dbi > MAIN_DBI)
1464 i += txn->mt_dbs[MAIN_DBI].md_depth;
1465 /* For puts, roughly factor in the key+data size */
1467 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1468 i += i; /* double it for good measure */
1471 if (txn->mt_dirty_room > i)
1474 if (!txn->mt_spill_pgs) {
1475 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1476 if (!txn->mt_spill_pgs)
1479 /* strip any dups */
1480 MDB_IDL sl = txn->mt_spill_pgs;
1481 unsigned int num = sl[0];
1483 for (i=1; i<=num; i++) {
1491 /* Preserve pages which may soon be dirtied again */
1492 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1495 /* Less aggressive spill - we originally spilled the entire dirty list,
1496 * with a few exceptions for cursor pages and DB root pages. But this
1497 * turns out to be a lot of wasted effort because in a large txn many
1498 * of those pages will need to be used again. So now we spill only 1/8th
1499 * of the dirty pages. Testing revealed this to be a good tradeoff,
1500 * better than 1/2, 1/4, or 1/10.
1502 if (need < MDB_IDL_UM_MAX / 8)
1503 need = MDB_IDL_UM_MAX / 8;
1505 /* Save the page IDs of all the pages we're flushing */
1506 /* flush from the tail forward, this saves a lot of shifting later on. */
1507 for (i=dl[0].mid; i && need; i--) {
1509 if (dp->mp_flags & P_KEEP)
1511 /* Can't spill twice, make sure it's not already in a parent's
1514 if (txn->mt_parent) {
1516 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1517 if (tx2->mt_spill_pgs) {
1518 j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
1519 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
1520 dp->mp_flags |= P_KEEP;
1528 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
1532 mdb_midl_sort(txn->mt_spill_pgs);
1534 /* Flush the spilled part of dirty list */
1535 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1538 /* Reset any dirty pages we kept that page_flush didn't see */
1539 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1543 if (txn->mt_parent) {
1544 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1545 /* dirty pages that are dirty in an ancestor don't
1546 * count against this txn's dirty_room.
1548 for (i=1; i<=dl[0].mid; i++) {
1549 pgno_t pgno = dl[i].mid;
1551 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1552 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1553 if (j <= tx2->mt_u.dirty_list[0].mid &&
1554 tx2->mt_u.dirty_list[j].mid == pgno) {
1555 txn->mt_dirty_room++;
1561 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1563 txn->mt_flags |= MDB_TXN_SPILLS;
1565 txn->mt_flags |= MDB_TXN_ERROR;
1570 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1572 mdb_find_oldest(MDB_txn *txn)
1575 txnid_t mr, oldest = txn->mt_txnid - 1;
1576 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1577 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1587 /** Add a page to the txn's dirty list */
1589 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1592 int (*insert)(MDB_ID2L, MDB_ID2 *);
1594 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1595 insert = mdb_mid2l_append;
1597 insert = mdb_mid2l_insert;
1599 mid.mid = mp->mp_pgno;
1601 insert(txn->mt_u.dirty_list, &mid);
1602 txn->mt_dirty_room--;
1605 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1606 * me_pghead and mt_next_pgno.
1608 * If there are free pages available from older transactions, they
1609 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1610 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1611 * and move me_pglast to say which records were consumed. Only this
1612 * function can create me_pghead and move me_pglast/mt_next_pgno.
1613 * @param[in] mc cursor A cursor handle identifying the transaction and
1614 * database for which we are allocating.
1615 * @param[in] num the number of pages to allocate.
1616 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1617 * will always be satisfied by a single contiguous chunk of memory.
1618 * @return 0 on success, non-zero on failure.
1621 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1623 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1624 /* Get at most <Max_retries> more freeDB records once me_pghead
1625 * has enough pages. If not enough, use new pages from the map.
1626 * If <Paranoid> and mc is updating the freeDB, only get new
1627 * records if me_pghead is empty. Then the freelist cannot play
1628 * catch-up with itself by growing while trying to save it.
1630 enum { Paranoid = 1, Max_retries = 500 };
1632 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1634 int rc, n2 = num-1, retry = Max_retries;
1635 MDB_txn *txn = mc->mc_txn;
1636 MDB_env *env = txn->mt_env;
1637 pgno_t pgno, *mop = env->me_pghead;
1638 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1640 txnid_t oldest = 0, last;
1646 /* If our dirty list is already full, we can't do anything */
1647 if (txn->mt_dirty_room == 0)
1648 return MDB_TXN_FULL;
1650 for (op = MDB_FIRST;; op = MDB_NEXT) {
1653 pgno_t *idl, old_id, new_id;
1655 /* Seek a big enough contiguous page range. Prefer
1656 * pages at the tail, just truncating the list.
1658 if (mop_len >= (unsigned)num) {
1662 if (mop[i-n2] == pgno+n2)
1664 } while (--i >= (unsigned)num);
1665 if (Max_retries < INT_MAX && --retry < 0)
1669 if (op == MDB_FIRST) { /* 1st iteration */
1670 /* Prepare to fetch more and coalesce */
1671 oldest = mdb_find_oldest(txn);
1672 last = env->me_pglast;
1673 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1676 key.mv_data = &last; /* will look up last+1 */
1677 key.mv_size = sizeof(last);
1679 if (Paranoid && mc->mc_dbi == FREE_DBI)
1682 if (Paranoid && retry < 0 && mop_len)
1686 /* Do not fetch more if the record will be too recent */
1689 rc = mdb_cursor_get(&m2, &key, NULL, op);
1691 if (rc == MDB_NOTFOUND)
1695 last = *(txnid_t*)key.mv_data;
1698 np = m2.mc_pg[m2.mc_top];
1699 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1700 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1703 idl = (MDB_ID *) data.mv_data;
1706 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1709 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1711 mop = env->me_pghead;
1713 env->me_pglast = last;
1715 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1716 last, txn->mt_dbs[FREE_DBI].md_root, i));
1718 DPRINTF(("IDL %"Z"u", idl[k]));
1720 /* Merge in descending sorted order */
1723 mop[0] = (pgno_t)-1;
1727 for (; old_id < new_id; old_id = mop[--j])
1734 /* Use new pages from the map when nothing suitable in the freeDB */
1736 pgno = txn->mt_next_pgno;
1737 if (pgno + num >= env->me_maxpg) {
1738 DPUTS("DB size maxed out");
1739 return MDB_MAP_FULL;
1743 if (env->me_flags & MDB_WRITEMAP) {
1744 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1746 if (!(np = mdb_page_malloc(txn, num)))
1750 mop[0] = mop_len -= num;
1751 /* Move any stragglers down */
1752 for (j = i-num; j < mop_len; )
1753 mop[++j] = mop[++i];
1755 txn->mt_next_pgno = pgno + num;
1758 mdb_page_dirty(txn, np);
1764 /** Copy the used portions of a non-overflow page.
1765 * @param[in] dst page to copy into
1766 * @param[in] src page to copy from
1767 * @param[in] psize size of a page
1770 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1772 enum { Align = sizeof(pgno_t) };
1773 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1775 /* If page isn't full, just copy the used portion. Adjust
1776 * alignment so memcpy may copy words instead of bytes.
1778 if ((unused &= -Align) && !IS_LEAF2(src)) {
1780 memcpy(dst, src, (lower + (Align-1)) & -Align);
1781 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1784 memcpy(dst, src, psize - unused);
1788 /** Pull a page off the txn's spill list, if present.
1789 * If a page being referenced was spilled to disk in this txn, bring
1790 * it back and make it dirty/writable again.
1791 * @param[in] tx0 the transaction handle.
1792 * @param[in] mp the page being referenced.
1793 * @param[out] ret the writable page, if any. ret is unchanged if
1794 * mp wasn't spilled.
1797 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1799 MDB_env *env = tx0->mt_env;
1802 pgno_t pgno = mp->mp_pgno;
1804 for (txn = tx0; txn; txn=txn->mt_parent) {
1805 if (!txn->mt_spill_pgs)
1807 x = mdb_midl_search(txn->mt_spill_pgs, pgno);
1808 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
1811 if (IS_OVERFLOW(mp))
1815 if (env->me_flags & MDB_WRITEMAP) {
1818 np = mdb_page_malloc(txn, num);
1822 memcpy(np, mp, num * env->me_psize);
1824 mdb_page_copy(np, mp, env->me_psize);
1827 /* If in current txn, this page is no longer spilled.
1828 * If it happens to be the last page, truncate the spill list.
1829 * Otherwise temporarily dup its neighbor over it. Dups will
1830 * be stripped out later by the next mdb_page_spill run.
1832 if (x == txn->mt_spill_pgs[0])
1833 txn->mt_spill_pgs[0]--;
1835 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
1836 } /* otherwise, if belonging to a parent txn, the
1837 * page remains spilled until child commits
1840 if (txn->mt_parent) {
1842 /* If this page is also in a parent's dirty list, then
1843 * it's already accounted in dirty_room, and we need to
1844 * cancel out the decrement that mdb_page_dirty does.
1846 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1847 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1848 if (x <= tx2->mt_u.dirty_list[0].mid &&
1849 tx2->mt_u.dirty_list[x].mid == pgno) {
1850 txn->mt_dirty_room++;
1855 mdb_page_dirty(tx0, np);
1856 np->mp_flags |= P_DIRTY;
1864 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1865 * @param[in] mc cursor pointing to the page to be touched
1866 * @return 0 on success, non-zero on failure.
1869 mdb_page_touch(MDB_cursor *mc)
1871 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1872 MDB_txn *txn = mc->mc_txn;
1873 MDB_cursor *m2, *m3;
1878 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1879 if (txn->mt_flags & MDB_TXN_SPILLS) {
1881 rc = mdb_page_unspill(txn, mp, &np);
1887 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1888 (rc = mdb_page_alloc(mc, 1, &np)))
1891 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1892 assert(mp->mp_pgno != pgno);
1893 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1894 /* Update the parent page, if any, to point to the new page */
1896 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1897 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1898 SETPGNO(node, pgno);
1900 mc->mc_db->md_root = pgno;
1902 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1903 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1905 /* If txn has a parent, make sure the page is in our
1909 unsigned x = mdb_mid2l_search(dl, pgno);
1910 if (x <= dl[0].mid && dl[x].mid == pgno) {
1911 if (mp != dl[x].mptr) { /* bad cursor? */
1912 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1913 return MDB_CORRUPTED;
1918 assert(dl[0].mid < MDB_IDL_UM_MAX);
1920 np = mdb_page_malloc(txn, 1);
1925 mdb_mid2l_insert(dl, &mid);
1930 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1932 np->mp_flags |= P_DIRTY;
1935 /* Adjust cursors pointing to mp */
1936 mc->mc_pg[mc->mc_top] = np;
1938 if (mc->mc_flags & C_SUB) {
1940 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1941 m3 = &m2->mc_xcursor->mx_cursor;
1942 if (m3->mc_snum < mc->mc_snum) continue;
1943 if (m3->mc_pg[mc->mc_top] == mp)
1944 m3->mc_pg[mc->mc_top] = np;
1947 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1948 if (m2->mc_snum < mc->mc_snum) continue;
1949 if (m2->mc_pg[mc->mc_top] == mp) {
1950 m2->mc_pg[mc->mc_top] = np;
1951 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1952 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1954 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1955 if (!(leaf->mn_flags & F_SUBDATA))
1956 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1965 mdb_env_sync(MDB_env *env, int force)
1968 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1969 if (env->me_flags & MDB_WRITEMAP) {
1970 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1971 ? MS_ASYNC : MS_SYNC;
1972 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1975 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1979 if (MDB_FDATASYNC(env->me_fd))
1986 /** Back up parent txn's cursors, then grab the originals for tracking */
1988 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1990 MDB_cursor *mc, *bk;
1995 for (i = src->mt_numdbs; --i >= 0; ) {
1996 if ((mc = src->mt_cursors[i]) != NULL) {
1997 size = sizeof(MDB_cursor);
1999 size += sizeof(MDB_xcursor);
2000 for (; mc; mc = bk->mc_next) {
2006 mc->mc_db = &dst->mt_dbs[i];
2007 /* Kill pointers into src - and dst to reduce abuse: The
2008 * user may not use mc until dst ends. Otherwise we'd...
2010 mc->mc_txn = NULL; /* ...set this to dst */
2011 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2012 if ((mx = mc->mc_xcursor) != NULL) {
2013 *(MDB_xcursor *)(bk+1) = *mx;
2014 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2016 mc->mc_next = dst->mt_cursors[i];
2017 dst->mt_cursors[i] = mc;
2024 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2025 * @param[in] txn the transaction handle.
2026 * @param[in] merge true to keep changes to parent cursors, false to revert.
2027 * @return 0 on success, non-zero on failure.
2030 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2032 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2036 for (i = txn->mt_numdbs; --i >= 0; ) {
2037 for (mc = cursors[i]; mc; mc = next) {
2039 if ((bk = mc->mc_backup) != NULL) {
2041 /* Commit changes to parent txn */
2042 mc->mc_next = bk->mc_next;
2043 mc->mc_backup = bk->mc_backup;
2044 mc->mc_txn = bk->mc_txn;
2045 mc->mc_db = bk->mc_db;
2046 mc->mc_dbflag = bk->mc_dbflag;
2047 if ((mx = mc->mc_xcursor) != NULL)
2048 mx->mx_cursor.mc_txn = bk->mc_txn;
2050 /* Abort nested txn */
2052 if ((mx = mc->mc_xcursor) != NULL)
2053 *mx = *(MDB_xcursor *)(bk+1);
2057 /* Only malloced cursors are permanently tracked. */
2065 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2068 mdb_txn_reset0(MDB_txn *txn, const char *act);
2070 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2076 Pidset = F_SETLK, Pidcheck = F_GETLK
2080 /** Set or check a pid lock. Set returns 0 on success.
2081 * Check returns 0 if the process is certainly dead, nonzero if it may
2082 * be alive (the lock exists or an error happened so we do not know).
2084 * On Windows Pidset is a no-op, we merely check for the existence
2085 * of the process with the given pid. On POSIX we use a single byte
2086 * lock on the lockfile, set at an offset equal to the pid.
2089 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2091 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2094 if (op == Pidcheck) {
2095 h = OpenProcess(env->me_pidquery, FALSE, pid);
2096 /* No documented "no such process" code, but other program use this: */
2098 return ErrCode() != ERROR_INVALID_PARAMETER;
2099 /* A process exists until all handles to it close. Has it exited? */
2100 ret = WaitForSingleObject(h, 0) != 0;
2107 struct flock lock_info;
2108 memset(&lock_info, 0, sizeof(lock_info));
2109 lock_info.l_type = F_WRLCK;
2110 lock_info.l_whence = SEEK_SET;
2111 lock_info.l_start = pid;
2112 lock_info.l_len = 1;
2113 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2114 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2116 } else if ((rc = ErrCode()) == EINTR) {
2124 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2125 * @param[in] txn the transaction handle to initialize
2126 * @return 0 on success, non-zero on failure.
2129 mdb_txn_renew0(MDB_txn *txn)
2131 MDB_env *env = txn->mt_env;
2134 int rc, new_notls = 0;
2137 txn->mt_numdbs = env->me_numdbs;
2138 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2140 if (txn->mt_flags & MDB_TXN_RDONLY) {
2141 if (!env->me_txns) {
2142 i = mdb_env_pick_meta(env);
2143 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2144 txn->mt_u.reader = NULL;
2146 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2147 pthread_getspecific(env->me_txkey);
2149 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2150 return MDB_BAD_RSLOT;
2152 pid_t pid = env->me_pid;
2153 pthread_t tid = pthread_self();
2155 if (!(env->me_flags & MDB_LIVE_READER)) {
2156 rc = mdb_reader_pid(env, Pidset, pid);
2158 UNLOCK_MUTEX_R(env);
2161 env->me_flags |= MDB_LIVE_READER;
2165 for (i=0; i<env->me_txns->mti_numreaders; i++)
2166 if (env->me_txns->mti_readers[i].mr_pid == 0)
2168 if (i == env->me_maxreaders) {
2169 UNLOCK_MUTEX_R(env);
2170 return MDB_READERS_FULL;
2172 env->me_txns->mti_readers[i].mr_pid = pid;
2173 env->me_txns->mti_readers[i].mr_tid = tid;
2174 if (i >= env->me_txns->mti_numreaders)
2175 env->me_txns->mti_numreaders = i+1;
2176 /* Save numreaders for un-mutexed mdb_env_close() */
2177 env->me_numreaders = env->me_txns->mti_numreaders;
2178 UNLOCK_MUTEX_R(env);
2179 r = &env->me_txns->mti_readers[i];
2180 new_notls = (env->me_flags & MDB_NOTLS);
2181 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2186 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2187 txn->mt_u.reader = r;
2189 txn->mt_toggle = txn->mt_txnid & 1;
2193 txn->mt_txnid = env->me_txns->mti_txnid;
2194 txn->mt_toggle = txn->mt_txnid & 1;
2197 if (txn->mt_txnid == mdb_debug_start)
2200 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2201 txn->mt_u.dirty_list = env->me_dirty_list;
2202 txn->mt_u.dirty_list[0].mid = 0;
2203 txn->mt_free_pgs = env->me_free_pgs;
2204 txn->mt_free_pgs[0] = 0;
2205 txn->mt_spill_pgs = NULL;
2209 /* Copy the DB info and flags */
2210 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2212 /* Moved to here to avoid a data race in read TXNs */
2213 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2215 for (i=2; i<txn->mt_numdbs; i++) {
2216 x = env->me_dbflags[i];
2217 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2218 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2220 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2222 if (env->me_maxpg < txn->mt_next_pgno) {
2223 mdb_txn_reset0(txn, "renew0-mapfail");
2225 txn->mt_u.reader->mr_pid = 0;
2226 txn->mt_u.reader = NULL;
2228 return MDB_MAP_RESIZED;
2235 mdb_txn_renew(MDB_txn *txn)
2239 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2242 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2243 DPUTS("environment had fatal error, must shutdown!");
2247 rc = mdb_txn_renew0(txn);
2248 if (rc == MDB_SUCCESS) {
2249 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2250 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2251 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2257 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2261 int rc, size, tsize = sizeof(MDB_txn);
2263 if (env->me_flags & MDB_FATAL_ERROR) {
2264 DPUTS("environment had fatal error, must shutdown!");
2267 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2270 /* Nested transactions: Max 1 child, write txns only, no writemap */
2271 if (parent->mt_child ||
2272 (flags & MDB_RDONLY) ||
2273 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2274 (env->me_flags & MDB_WRITEMAP))
2276 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2278 tsize = sizeof(MDB_ntxn);
2280 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2281 if (!(flags & MDB_RDONLY))
2282 size += env->me_maxdbs * sizeof(MDB_cursor *);
2284 if ((txn = calloc(1, size)) == NULL) {
2285 DPRINTF(("calloc: %s", strerror(ErrCode())));
2288 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2289 if (flags & MDB_RDONLY) {
2290 txn->mt_flags |= MDB_TXN_RDONLY;
2291 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2293 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2294 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2300 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2301 if (!txn->mt_u.dirty_list ||
2302 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2304 free(txn->mt_u.dirty_list);
2308 txn->mt_txnid = parent->mt_txnid;
2309 txn->mt_toggle = parent->mt_toggle;
2310 txn->mt_dirty_room = parent->mt_dirty_room;
2311 txn->mt_u.dirty_list[0].mid = 0;
2312 txn->mt_spill_pgs = NULL;
2313 txn->mt_next_pgno = parent->mt_next_pgno;
2314 parent->mt_child = txn;
2315 txn->mt_parent = parent;
2316 txn->mt_numdbs = parent->mt_numdbs;
2317 txn->mt_flags = parent->mt_flags;
2318 txn->mt_dbxs = parent->mt_dbxs;
2319 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2320 /* Copy parent's mt_dbflags, but clear DB_NEW */
2321 for (i=0; i<txn->mt_numdbs; i++)
2322 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2324 ntxn = (MDB_ntxn *)txn;
2325 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2326 if (env->me_pghead) {
2327 size = MDB_IDL_SIZEOF(env->me_pghead);
2328 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2330 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2335 rc = mdb_cursor_shadow(parent, txn);
2337 mdb_txn_reset0(txn, "beginchild-fail");
2339 rc = mdb_txn_renew0(txn);
2345 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2346 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2347 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2354 mdb_txn_env(MDB_txn *txn)
2356 if(!txn) return NULL;
2360 /** Export or close DBI handles opened in this txn. */
2362 mdb_dbis_update(MDB_txn *txn, int keep)
2365 MDB_dbi n = txn->mt_numdbs;
2366 MDB_env *env = txn->mt_env;
2367 unsigned char *tdbflags = txn->mt_dbflags;
2369 for (i = n; --i >= 2;) {
2370 if (tdbflags[i] & DB_NEW) {
2372 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2374 char *ptr = env->me_dbxs[i].md_name.mv_data;
2375 env->me_dbxs[i].md_name.mv_data = NULL;
2376 env->me_dbxs[i].md_name.mv_size = 0;
2377 env->me_dbflags[i] = 0;
2382 if (keep && env->me_numdbs < n)
2386 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2387 * May be called twice for readonly txns: First reset it, then abort.
2388 * @param[in] txn the transaction handle to reset
2389 * @param[in] act why the transaction is being reset
2392 mdb_txn_reset0(MDB_txn *txn, const char *act)
2394 MDB_env *env = txn->mt_env;
2396 /* Close any DBI handles opened in this txn */
2397 mdb_dbis_update(txn, 0);
2399 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2400 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2401 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2403 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2404 if (txn->mt_u.reader) {
2405 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2406 if (!(env->me_flags & MDB_NOTLS))
2407 txn->mt_u.reader = NULL; /* txn does not own reader */
2409 txn->mt_numdbs = 0; /* close nothing if called again */
2410 txn->mt_dbxs = NULL; /* mark txn as reset */
2412 mdb_cursors_close(txn, 0);
2414 if (!(env->me_flags & MDB_WRITEMAP)) {
2415 mdb_dlist_free(txn);
2417 mdb_midl_free(env->me_pghead);
2419 if (txn->mt_parent) {
2420 txn->mt_parent->mt_child = NULL;
2421 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2422 mdb_midl_free(txn->mt_free_pgs);
2423 mdb_midl_free(txn->mt_spill_pgs);
2424 free(txn->mt_u.dirty_list);
2428 if (mdb_midl_shrink(&txn->mt_free_pgs))
2429 env->me_free_pgs = txn->mt_free_pgs;
2430 env->me_pghead = NULL;
2434 /* The writer mutex was locked in mdb_txn_begin. */
2435 UNLOCK_MUTEX_W(env);
2440 mdb_txn_reset(MDB_txn *txn)
2445 /* This call is only valid for read-only txns */
2446 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2449 mdb_txn_reset0(txn, "reset");
2453 mdb_txn_abort(MDB_txn *txn)
2459 mdb_txn_abort(txn->mt_child);
2461 mdb_txn_reset0(txn, "abort");
2462 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2463 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2464 txn->mt_u.reader->mr_pid = 0;
2469 /** Save the freelist as of this transaction to the freeDB.
2470 * This changes the freelist. Keep trying until it stabilizes.
2473 mdb_freelist_save(MDB_txn *txn)
2475 /* env->me_pghead[] can grow and shrink during this call.
2476 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2477 * Page numbers cannot disappear from txn->mt_free_pgs[].
2480 MDB_env *env = txn->mt_env;
2481 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2482 txnid_t pglast = 0, head_id = 0;
2483 pgno_t freecnt = 0, *free_pgs, *mop;
2484 ssize_t head_room = 0, total_room = 0, mop_len;
2486 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2488 if (env->me_pghead) {
2489 /* Make sure first page of freeDB is touched and on freelist */
2490 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2491 if (rc && rc != MDB_NOTFOUND)
2496 /* Come back here after each Put() in case freelist changed */
2499 /* If using records from freeDB which we have not yet
2500 * deleted, delete them and any we reserved for me_pghead.
2502 while (pglast < env->me_pglast) {
2503 rc = mdb_cursor_first(&mc, &key, NULL);
2506 pglast = head_id = *(txnid_t *)key.mv_data;
2507 total_room = head_room = 0;
2508 assert(pglast <= env->me_pglast);
2509 rc = mdb_cursor_del(&mc, 0);
2514 /* Save the IDL of pages freed by this txn, to a single record */
2515 if (freecnt < txn->mt_free_pgs[0]) {
2517 /* Make sure last page of freeDB is touched and on freelist */
2518 key.mv_size = MDB_MAXKEYSIZE+1;
2520 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2521 if (rc && rc != MDB_NOTFOUND)
2524 free_pgs = txn->mt_free_pgs;
2525 /* Write to last page of freeDB */
2526 key.mv_size = sizeof(txn->mt_txnid);
2527 key.mv_data = &txn->mt_txnid;
2529 freecnt = free_pgs[0];
2530 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2531 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2534 /* Retry if mt_free_pgs[] grew during the Put() */
2535 free_pgs = txn->mt_free_pgs;
2536 } while (freecnt < free_pgs[0]);
2537 mdb_midl_sort(free_pgs);
2538 memcpy(data.mv_data, free_pgs, data.mv_size);
2541 unsigned int i = free_pgs[0];
2542 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2543 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2545 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2551 mop = env->me_pghead;
2552 mop_len = mop ? mop[0] : 0;
2554 /* Reserve records for me_pghead[]. Split it if multi-page,
2555 * to avoid searching freeDB for a page range. Use keys in
2556 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2558 if (total_room >= mop_len) {
2559 if (total_room == mop_len || --more < 0)
2561 } else if (head_room >= maxfree_1pg && head_id > 1) {
2562 /* Keep current record (overflow page), add a new one */
2566 /* (Re)write {key = head_id, IDL length = head_room} */
2567 total_room -= head_room;
2568 head_room = mop_len - total_room;
2569 if (head_room > maxfree_1pg && head_id > 1) {
2570 /* Overflow multi-page for part of me_pghead */
2571 head_room /= head_id; /* amortize page sizes */
2572 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2573 } else if (head_room < 0) {
2574 /* Rare case, not bothering to delete this record */
2577 key.mv_size = sizeof(head_id);
2578 key.mv_data = &head_id;
2579 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2580 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2583 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2584 total_room += head_room;
2587 /* Fill in the reserved, touched me_pghead records */
2593 rc = mdb_cursor_first(&mc, &key, &data);
2594 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2595 unsigned flags = MDB_CURRENT;
2596 txnid_t id = *(txnid_t *)key.mv_data;
2597 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2600 assert(len >= 0 && id <= env->me_pglast);
2602 if (len > mop_len) {
2604 data.mv_size = (len + 1) * sizeof(MDB_ID);
2607 data.mv_data = mop -= len;
2610 rc = mdb_cursor_put(&mc, &key, &data, flags);
2612 if (rc || !(mop_len -= len))
2619 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2620 * @param[in] txn the transaction that's being committed
2621 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2622 * @return 0 on success, non-zero on failure.
2625 mdb_page_flush(MDB_txn *txn, int keep)
2627 MDB_env *env = txn->mt_env;
2628 MDB_ID2L dl = txn->mt_u.dirty_list;
2629 unsigned psize = env->me_psize, j;
2630 int i, pagecount = dl[0].mid, rc;
2631 size_t size = 0, pos = 0;
2633 MDB_page *dp = NULL;
2637 struct iovec iov[MDB_COMMIT_PAGES];
2638 ssize_t wpos = 0, wsize = 0, wres;
2639 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2645 if (env->me_flags & MDB_WRITEMAP) {
2646 /* Clear dirty flags */
2647 while (++i <= pagecount) {
2649 /* Don't flush this page yet */
2650 if (dp->mp_flags & P_KEEP) {
2651 dp->mp_flags ^= P_KEEP;
2655 dp->mp_flags &= ~P_DIRTY;
2661 /* Write the pages */
2663 if (++i <= pagecount) {
2665 /* Don't flush this page yet */
2666 if (dp->mp_flags & P_KEEP) {
2667 dp->mp_flags ^= P_KEEP;
2672 /* clear dirty flag */
2673 dp->mp_flags &= ~P_DIRTY;
2676 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2681 /* Windows actually supports scatter/gather I/O, but only on
2682 * unbuffered file handles. Since we're relying on the OS page
2683 * cache for all our data, that's self-defeating. So we just
2684 * write pages one at a time. We use the ov structure to set
2685 * the write offset, to at least save the overhead of a Seek
2688 DPRINTF(("committing page %"Z"u", pgno));
2689 memset(&ov, 0, sizeof(ov));
2690 ov.Offset = pos & 0xffffffff;
2691 ov.OffsetHigh = pos >> 16 >> 16;
2692 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2694 DPRINTF(("WriteFile: %d", rc));
2698 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2699 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2701 /* Write previous page(s) */
2702 #ifdef MDB_USE_PWRITEV
2703 wres = pwritev(env->me_fd, iov, n, wpos);
2706 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2708 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2710 DPRINTF(("lseek: %s", strerror(rc)));
2713 wres = writev(env->me_fd, iov, n);
2716 if (wres != wsize) {
2719 DPRINTF(("Write error: %s", strerror(rc)));
2721 rc = EIO; /* TODO: Use which error code? */
2722 DPUTS("short write, filesystem full?");
2733 DPRINTF(("committing page %"Z"u", pgno));
2734 next_pos = pos + size;
2735 iov[n].iov_len = size;
2736 iov[n].iov_base = (char *)dp;
2742 for (i = keep; ++i <= pagecount; ) {
2744 /* This is a page we skipped above */
2747 dl[j].mid = dp->mp_pgno;
2750 mdb_dpage_free(env, dp);
2758 mdb_txn_commit(MDB_txn *txn)
2764 assert(txn != NULL);
2765 assert(txn->mt_env != NULL);
2767 if (txn->mt_child) {
2768 rc = mdb_txn_commit(txn->mt_child);
2769 txn->mt_child = NULL;
2776 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2777 mdb_dbis_update(txn, 1);
2778 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2783 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2784 DPUTS("error flag is set, can't commit");
2786 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2791 if (txn->mt_parent) {
2792 MDB_txn *parent = txn->mt_parent;
2796 /* Append our free list to parent's */
2797 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2800 mdb_midl_free(txn->mt_free_pgs);
2802 parent->mt_next_pgno = txn->mt_next_pgno;
2803 parent->mt_flags = txn->mt_flags;
2805 /* Merge our cursors into parent's and close them */
2806 mdb_cursors_close(txn, 1);
2808 /* Update parent's DB table. */
2809 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2810 parent->mt_numdbs = txn->mt_numdbs;
2811 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2812 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2813 for (i=2; i<txn->mt_numdbs; i++) {
2814 /* preserve parent's DB_NEW status */
2815 x = parent->mt_dbflags[i] & DB_NEW;
2816 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2819 dst = parent->mt_u.dirty_list;
2820 src = txn->mt_u.dirty_list;
2821 /* Remove anything in our dirty list from parent's spill list */
2822 if (parent->mt_spill_pgs) {
2823 x = parent->mt_spill_pgs[0];
2825 /* zero out our dirty pages in parent spill list */
2826 for (i=1; i<=src[0].mid; i++) {
2827 if (src[i].mid < parent->mt_spill_pgs[x])
2829 if (src[i].mid > parent->mt_spill_pgs[x]) {
2835 parent->mt_spill_pgs[x] = 0;
2838 /* OK, we had a few hits, squash zeros from the spill list */
2839 if (len < parent->mt_spill_pgs[0]) {
2841 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2842 if (parent->mt_spill_pgs[y]) {
2844 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2849 parent->mt_spill_pgs[0] = len;
2852 /* Find len = length of merging our dirty list with parent's */
2854 dst[0].mid = 0; /* simplify loops */
2855 if (parent->mt_parent) {
2856 len = x + src[0].mid;
2857 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2858 for (i = x; y && i; y--) {
2859 pgno_t yp = src[y].mid;
2860 while (yp < dst[i].mid)
2862 if (yp == dst[i].mid) {
2867 } else { /* Simplify the above for single-ancestor case */
2868 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2870 /* Merge our dirty list with parent's */
2872 for (i = len; y; dst[i--] = src[y--]) {
2873 pgno_t yp = src[y].mid;
2874 while (yp < dst[x].mid)
2875 dst[i--] = dst[x--];
2876 if (yp == dst[x].mid)
2877 free(dst[x--].mptr);
2881 free(txn->mt_u.dirty_list);
2882 parent->mt_dirty_room = txn->mt_dirty_room;
2883 if (txn->mt_spill_pgs) {
2884 if (parent->mt_spill_pgs) {
2885 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2886 mdb_midl_free(txn->mt_spill_pgs);
2887 mdb_midl_sort(parent->mt_spill_pgs);
2889 parent->mt_spill_pgs = txn->mt_spill_pgs;
2893 parent->mt_child = NULL;
2894 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2899 if (txn != env->me_txn) {
2900 DPUTS("attempt to commit unknown transaction");
2905 mdb_cursors_close(txn, 0);
2907 if (!txn->mt_u.dirty_list[0].mid &&
2908 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2911 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2912 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2914 /* Update DB root pointers */
2915 if (txn->mt_numdbs > 2) {
2919 data.mv_size = sizeof(MDB_db);
2921 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2922 for (i = 2; i < txn->mt_numdbs; i++) {
2923 if (txn->mt_dbflags[i] & DB_DIRTY) {
2924 data.mv_data = &txn->mt_dbs[i];
2925 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2932 rc = mdb_freelist_save(txn);
2936 mdb_midl_free(env->me_pghead);
2937 env->me_pghead = NULL;
2938 if (mdb_midl_shrink(&txn->mt_free_pgs))
2939 env->me_free_pgs = txn->mt_free_pgs;
2945 if ((rc = mdb_page_flush(txn, 0)) ||
2946 (rc = mdb_env_sync(env, 0)) ||
2947 (rc = mdb_env_write_meta(txn)))
2953 mdb_dbis_update(txn, 1);
2955 UNLOCK_MUTEX_W(env);
2965 /** Read the environment parameters of a DB environment before
2966 * mapping it into memory.
2967 * @param[in] env the environment handle
2968 * @param[out] meta address of where to store the meta information
2969 * @return 0 on success, non-zero on failure.
2972 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2979 /* We don't know the page size yet, so use a minimum value.
2980 * Read both meta pages so we can use the latest one.
2983 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2987 memset(&ov, 0, sizeof(ov));
2989 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2990 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2993 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2995 if (rc != MDB_PAGESIZE) {
2996 if (rc == 0 && off == 0)
2998 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2999 DPRINTF(("read: %s", mdb_strerror(rc)));
3003 p = (MDB_page *)&pbuf;
3005 if (!F_ISSET(p->mp_flags, P_META)) {
3006 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3011 if (m->mm_magic != MDB_MAGIC) {
3012 DPUTS("meta has invalid magic");
3016 if (m->mm_version != MDB_DATA_VERSION) {
3017 DPRINTF(("database is version %u, expected version %u",
3018 m->mm_version, MDB_DATA_VERSION));
3019 return MDB_VERSION_MISMATCH;
3022 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3028 /** Write the environment parameters of a freshly created DB environment.
3029 * @param[in] env the environment handle
3030 * @param[out] meta address of where to store the meta information
3031 * @return 0 on success, non-zero on failure.
3034 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3042 memset(&ov, 0, sizeof(ov));
3043 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3045 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3048 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3049 len = pwrite(fd, ptr, size, pos); \
3050 rc = (len >= 0); } while(0)
3053 DPUTS("writing new meta page");
3055 GET_PAGESIZE(psize);
3057 meta->mm_magic = MDB_MAGIC;
3058 meta->mm_version = MDB_DATA_VERSION;
3059 meta->mm_mapsize = env->me_mapsize;
3060 meta->mm_psize = psize;
3061 meta->mm_last_pg = 1;
3062 meta->mm_flags = env->me_flags & 0xffff;
3063 meta->mm_flags |= MDB_INTEGERKEY;
3064 meta->mm_dbs[0].md_root = P_INVALID;
3065 meta->mm_dbs[1].md_root = P_INVALID;
3067 p = calloc(2, psize);
3069 p->mp_flags = P_META;
3070 *(MDB_meta *)METADATA(p) = *meta;
3072 q = (MDB_page *)((char *)p + psize);
3074 q->mp_flags = P_META;
3075 *(MDB_meta *)METADATA(q) = *meta;
3077 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3080 else if ((unsigned) len == psize * 2)
3088 /** Update the environment info to commit a transaction.
3089 * @param[in] txn the transaction that's being committed
3090 * @return 0 on success, non-zero on failure.
3093 mdb_env_write_meta(MDB_txn *txn)
3096 MDB_meta meta, metab, *mp;
3098 int rc, len, toggle;
3107 assert(txn != NULL);
3108 assert(txn->mt_env != NULL);
3110 toggle = !txn->mt_toggle;
3111 DPRINTF(("writing meta page %d for root page %"Z"u",
3112 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3115 mp = env->me_metas[toggle];
3117 if (env->me_flags & MDB_WRITEMAP) {
3118 /* Persist any increases of mapsize config */
3119 if (env->me_mapsize > mp->mm_mapsize)
3120 mp->mm_mapsize = env->me_mapsize;
3121 mp->mm_dbs[0] = txn->mt_dbs[0];
3122 mp->mm_dbs[1] = txn->mt_dbs[1];
3123 mp->mm_last_pg = txn->mt_next_pgno - 1;
3124 mp->mm_txnid = txn->mt_txnid;
3125 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3126 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3129 ptr += env->me_psize;
3130 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3137 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3138 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3140 ptr = (char *)&meta;
3141 if (env->me_mapsize > mp->mm_mapsize) {
3142 /* Persist any increases of mapsize config */
3143 meta.mm_mapsize = env->me_mapsize;
3144 off = offsetof(MDB_meta, mm_mapsize);
3146 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3148 len = sizeof(MDB_meta) - off;
3151 meta.mm_dbs[0] = txn->mt_dbs[0];
3152 meta.mm_dbs[1] = txn->mt_dbs[1];
3153 meta.mm_last_pg = txn->mt_next_pgno - 1;
3154 meta.mm_txnid = txn->mt_txnid;
3157 off += env->me_psize;
3160 /* Write to the SYNC fd */
3161 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3162 env->me_fd : env->me_mfd;
3165 memset(&ov, 0, sizeof(ov));
3167 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3171 rc = pwrite(mfd, ptr, len, off);
3174 rc = rc < 0 ? ErrCode() : EIO;
3175 DPUTS("write failed, disk error?");
3176 /* On a failure, the pagecache still contains the new data.
3177 * Write some old data back, to prevent it from being used.
3178 * Use the non-SYNC fd; we know it will fail anyway.
3180 meta.mm_last_pg = metab.mm_last_pg;
3181 meta.mm_txnid = metab.mm_txnid;
3183 memset(&ov, 0, sizeof(ov));
3185 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3187 r2 = pwrite(env->me_fd, ptr, len, off);
3188 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3191 env->me_flags |= MDB_FATAL_ERROR;
3195 /* Memory ordering issues are irrelevant; since the entire writer
3196 * is wrapped by wmutex, all of these changes will become visible
3197 * after the wmutex is unlocked. Since the DB is multi-version,
3198 * readers will get consistent data regardless of how fresh or
3199 * how stale their view of these values is.
3201 env->me_txns->mti_txnid = txn->mt_txnid;
3206 /** Check both meta pages to see which one is newer.
3207 * @param[in] env the environment handle
3208 * @return meta toggle (0 or 1).
3211 mdb_env_pick_meta(const MDB_env *env)
3213 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3217 mdb_env_create(MDB_env **env)
3221 e = calloc(1, sizeof(MDB_env));
3225 e->me_maxreaders = DEFAULT_READERS;
3226 e->me_maxdbs = e->me_numdbs = 2;
3227 e->me_fd = INVALID_HANDLE_VALUE;
3228 e->me_lfd = INVALID_HANDLE_VALUE;
3229 e->me_mfd = INVALID_HANDLE_VALUE;
3230 #ifdef MDB_USE_POSIX_SEM
3231 e->me_rmutex = SEM_FAILED;
3232 e->me_wmutex = SEM_FAILED;
3234 e->me_pid = getpid();
3235 VGMEMP_CREATE(e,0,0);
3241 mdb_env_set_mapsize(MDB_env *env, size_t size)
3245 env->me_mapsize = size;
3247 env->me_maxpg = env->me_mapsize / env->me_psize;
3252 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3256 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3261 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3263 if (env->me_map || readers < 1)
3265 env->me_maxreaders = readers;
3270 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3272 if (!env || !readers)
3274 *readers = env->me_maxreaders;
3278 /** Further setup required for opening an MDB environment
3281 mdb_env_open2(MDB_env *env)
3283 unsigned int flags = env->me_flags;
3291 memset(&meta, 0, sizeof(meta));
3293 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3296 DPUTS("new mdbenv");
3300 /* Was a mapsize configured? */
3301 if (!env->me_mapsize) {
3302 /* If this is a new environment, take the default,
3303 * else use the size recorded in the existing env.
3305 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3306 } else if (env->me_mapsize < meta.mm_mapsize) {
3307 /* If the configured size is smaller, make sure it's
3308 * still big enough. Silently round up to minimum if not.
3310 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3311 if (env->me_mapsize < minsize)
3312 env->me_mapsize = minsize;
3319 LONG sizelo, sizehi;
3320 sizelo = env->me_mapsize & 0xffffffff;
3321 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3323 /* See if we should use QueryLimited */
3325 if ((rc & 0xff) > 5)
3326 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3328 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3330 /* Windows won't create mappings for zero length files.
3331 * Just allocate the maxsize right now.
3334 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3335 || !SetEndOfFile(env->me_fd)
3336 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3339 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3340 PAGE_READWRITE : PAGE_READONLY,
3341 sizehi, sizelo, NULL);
3344 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3345 FILE_MAP_WRITE : FILE_MAP_READ,
3346 0, 0, env->me_mapsize, meta.mm_address);
3347 rc = env->me_map ? 0 : ErrCode();
3355 if (flags & MDB_WRITEMAP) {
3357 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3360 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3362 if (env->me_map == MAP_FAILED) {
3366 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3368 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3370 #ifdef POSIX_MADV_RANDOM
3371 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3372 #endif /* POSIX_MADV_RANDOM */
3373 #endif /* MADV_RANDOM */
3377 if (flags & MDB_FIXEDMAP)
3378 meta.mm_address = env->me_map;
3379 i = mdb_env_init_meta(env, &meta);
3380 if (i != MDB_SUCCESS) {
3383 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3384 /* Can happen because the address argument to mmap() is just a
3385 * hint. mmap() can pick another, e.g. if the range is in use.
3386 * The MAP_FIXED flag would prevent that, but then mmap could
3387 * instead unmap existing pages to make room for the new map.
3389 return EBUSY; /* TODO: Make a new MDB_* error code? */
3391 env->me_psize = meta.mm_psize;
3392 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3393 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3395 env->me_maxpg = env->me_mapsize / env->me_psize;
3397 p = (MDB_page *)env->me_map;
3398 env->me_metas[0] = METADATA(p);
3399 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3403 int toggle = mdb_env_pick_meta(env);
3404 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3406 DPRINTF(("opened database version %u, pagesize %u",
3407 env->me_metas[0]->mm_version, env->me_psize));
3408 DPRINTF(("using meta page %d", toggle));
3409 DPRINTF(("depth: %u", db->md_depth));
3410 DPRINTF(("entries: %"Z"u", db->md_entries));
3411 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3412 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3413 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3414 DPRINTF(("root: %"Z"u", db->md_root));
3422 /** Release a reader thread's slot in the reader lock table.
3423 * This function is called automatically when a thread exits.
3424 * @param[in] ptr This points to the slot in the reader lock table.
3427 mdb_env_reader_dest(void *ptr)
3429 MDB_reader *reader = ptr;
3435 /** Junk for arranging thread-specific callbacks on Windows. This is
3436 * necessarily platform and compiler-specific. Windows supports up
3437 * to 1088 keys. Let's assume nobody opens more than 64 environments
3438 * in a single process, for now. They can override this if needed.
3440 #ifndef MAX_TLS_KEYS
3441 #define MAX_TLS_KEYS 64
3443 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3444 static int mdb_tls_nkeys;
3446 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3450 case DLL_PROCESS_ATTACH: break;
3451 case DLL_THREAD_ATTACH: break;
3452 case DLL_THREAD_DETACH:
3453 for (i=0; i<mdb_tls_nkeys; i++) {
3454 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3455 mdb_env_reader_dest(r);
3458 case DLL_PROCESS_DETACH: break;
3463 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3465 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3469 /* Force some symbol references.
3470 * _tls_used forces the linker to create the TLS directory if not already done
3471 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3473 #pragma comment(linker, "/INCLUDE:_tls_used")
3474 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3475 #pragma const_seg(".CRT$XLB")
3476 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3477 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3480 #pragma comment(linker, "/INCLUDE:__tls_used")
3481 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3482 #pragma data_seg(".CRT$XLB")
3483 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3485 #endif /* WIN 32/64 */
3486 #endif /* !__GNUC__ */
3489 /** Downgrade the exclusive lock on the region back to shared */
3491 mdb_env_share_locks(MDB_env *env, int *excl)
3493 int rc = 0, toggle = mdb_env_pick_meta(env);
3495 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3500 /* First acquire a shared lock. The Unlock will
3501 * then release the existing exclusive lock.
3503 memset(&ov, 0, sizeof(ov));
3504 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3507 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3513 struct flock lock_info;
3514 /* The shared lock replaces the existing lock */
3515 memset((void *)&lock_info, 0, sizeof(lock_info));
3516 lock_info.l_type = F_RDLCK;
3517 lock_info.l_whence = SEEK_SET;
3518 lock_info.l_start = 0;
3519 lock_info.l_len = 1;
3520 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3521 (rc = ErrCode()) == EINTR) ;
3522 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3529 /** Try to get exlusive lock, otherwise shared.
3530 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3533 mdb_env_excl_lock(MDB_env *env, int *excl)
3537 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3541 memset(&ov, 0, sizeof(ov));
3542 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3549 struct flock lock_info;
3550 memset((void *)&lock_info, 0, sizeof(lock_info));
3551 lock_info.l_type = F_WRLCK;
3552 lock_info.l_whence = SEEK_SET;
3553 lock_info.l_start = 0;
3554 lock_info.l_len = 1;
3555 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3556 (rc = ErrCode()) == EINTR) ;
3560 # ifdef MDB_USE_POSIX_SEM
3561 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3564 lock_info.l_type = F_RDLCK;
3565 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3566 (rc = ErrCode()) == EINTR) ;
3574 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3576 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3578 * @(#) $Revision: 5.1 $
3579 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3580 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3582 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3586 * Please do not copyright this code. This code is in the public domain.
3588 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3589 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3590 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3591 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3592 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3593 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3594 * PERFORMANCE OF THIS SOFTWARE.
3597 * chongo <Landon Curt Noll> /\oo/\
3598 * http://www.isthe.com/chongo/
3600 * Share and Enjoy! :-)
3603 typedef unsigned long long mdb_hash_t;
3604 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3606 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3607 * @param[in] val value to hash
3608 * @param[in] hval initial value for hash
3609 * @return 64 bit hash
3611 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3612 * hval arg on the first call.
3615 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3617 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3618 unsigned char *end = s + val->mv_size;
3620 * FNV-1a hash each octet of the string
3623 /* xor the bottom with the current octet */
3624 hval ^= (mdb_hash_t)*s++;
3626 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3627 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3628 (hval << 7) + (hval << 8) + (hval << 40);
3630 /* return our new hash value */
3634 /** Hash the string and output the encoded hash.
3635 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3636 * very short name limits. We don't care about the encoding being reversible,
3637 * we just want to preserve as many bits of the input as possible in a
3638 * small printable string.
3639 * @param[in] str string to hash
3640 * @param[out] encbuf an array of 11 chars to hold the hash
3642 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3645 mdb_pack85(unsigned long l, char *out)
3649 for (i=0; i<5; i++) {
3650 *out++ = mdb_a85[l % 85];
3656 mdb_hash_enc(MDB_val *val, char *encbuf)
3658 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3660 mdb_pack85(h, encbuf);
3661 mdb_pack85(h>>32, encbuf+5);
3666 /** Open and/or initialize the lock region for the environment.
3667 * @param[in] env The MDB environment.
3668 * @param[in] lpath The pathname of the file used for the lock region.
3669 * @param[in] mode The Unix permissions for the file, if we create it.
3670 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3671 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3672 * @return 0 on success, non-zero on failure.
3675 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3678 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3680 # define MDB_ERRCODE_ROFS EROFS
3681 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3682 # define MDB_CLOEXEC O_CLOEXEC
3685 # define MDB_CLOEXEC 0
3692 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3693 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3694 FILE_ATTRIBUTE_NORMAL, NULL);
3696 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3698 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3700 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3705 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3706 /* Lose record locks when exec*() */
3707 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3708 fcntl(env->me_lfd, F_SETFD, fdflags);
3711 if (!(env->me_flags & MDB_NOTLS)) {
3712 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3715 env->me_flags |= MDB_ENV_TXKEY;
3717 /* Windows TLS callbacks need help finding their TLS info. */
3718 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3722 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3726 /* Try to get exclusive lock. If we succeed, then
3727 * nobody is using the lock region and we should initialize it.
3729 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3732 size = GetFileSize(env->me_lfd, NULL);
3734 size = lseek(env->me_lfd, 0, SEEK_END);
3735 if (size == -1) goto fail_errno;
3737 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3738 if (size < rsize && *excl > 0) {
3740 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3741 || !SetEndOfFile(env->me_lfd))
3744 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3748 size = rsize - sizeof(MDB_txninfo);
3749 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3754 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3756 if (!mh) goto fail_errno;
3757 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3759 if (!env->me_txns) goto fail_errno;
3761 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3763 if (m == MAP_FAILED) goto fail_errno;
3769 BY_HANDLE_FILE_INFORMATION stbuf;
3778 if (!mdb_sec_inited) {
3779 InitializeSecurityDescriptor(&mdb_null_sd,
3780 SECURITY_DESCRIPTOR_REVISION);
3781 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3782 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3783 mdb_all_sa.bInheritHandle = FALSE;
3784 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3787 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3788 idbuf.volume = stbuf.dwVolumeSerialNumber;
3789 idbuf.nhigh = stbuf.nFileIndexHigh;
3790 idbuf.nlow = stbuf.nFileIndexLow;
3791 val.mv_data = &idbuf;
3792 val.mv_size = sizeof(idbuf);
3793 mdb_hash_enc(&val, encbuf);
3794 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3795 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3796 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3797 if (!env->me_rmutex) goto fail_errno;
3798 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3799 if (!env->me_wmutex) goto fail_errno;
3800 #elif defined(MDB_USE_POSIX_SEM)
3809 #if defined(__NetBSD__)
3810 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3812 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3813 idbuf.dev = stbuf.st_dev;
3814 idbuf.ino = stbuf.st_ino;
3815 val.mv_data = &idbuf;
3816 val.mv_size = sizeof(idbuf);
3817 mdb_hash_enc(&val, encbuf);
3818 #ifdef MDB_SHORT_SEMNAMES
3819 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3821 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3822 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3823 /* Clean up after a previous run, if needed: Try to
3824 * remove both semaphores before doing anything else.
3826 sem_unlink(env->me_txns->mti_rmname);
3827 sem_unlink(env->me_txns->mti_wmname);
3828 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3829 O_CREAT|O_EXCL, mode, 1);
3830 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3831 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3832 O_CREAT|O_EXCL, mode, 1);
3833 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3834 #else /* MDB_USE_POSIX_SEM */
3835 pthread_mutexattr_t mattr;
3837 if ((rc = pthread_mutexattr_init(&mattr))
3838 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3839 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3840 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3842 pthread_mutexattr_destroy(&mattr);
3843 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3845 env->me_txns->mti_magic = MDB_MAGIC;
3846 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3847 env->me_txns->mti_txnid = 0;
3848 env->me_txns->mti_numreaders = 0;
3851 if (env->me_txns->mti_magic != MDB_MAGIC) {
3852 DPUTS("lock region has invalid magic");
3856 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3857 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3858 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3859 rc = MDB_VERSION_MISMATCH;
3863 if (rc && rc != EACCES && rc != EAGAIN) {
3867 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3868 if (!env->me_rmutex) goto fail_errno;
3869 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3870 if (!env->me_wmutex) goto fail_errno;
3871 #elif defined(MDB_USE_POSIX_SEM)
3872 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3873 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3874 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3875 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3886 /** The name of the lock file in the DB environment */
3887 #define LOCKNAME "/lock.mdb"
3888 /** The name of the data file in the DB environment */
3889 #define DATANAME "/data.mdb"
3890 /** The suffix of the lock file when no subdir is used */
3891 #define LOCKSUFF "-lock"
3892 /** Only a subset of the @ref mdb_env flags can be changed
3893 * at runtime. Changing other flags requires closing the
3894 * environment and re-opening it with the new flags.
3896 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3897 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3900 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3902 int oflags, rc, len, excl = -1;
3903 char *lpath, *dpath;
3905 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3909 if (flags & MDB_NOSUBDIR) {
3910 rc = len + sizeof(LOCKSUFF) + len + 1;
3912 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3917 if (flags & MDB_NOSUBDIR) {
3918 dpath = lpath + len + sizeof(LOCKSUFF);
3919 sprintf(lpath, "%s" LOCKSUFF, path);
3920 strcpy(dpath, path);
3922 dpath = lpath + len + sizeof(LOCKNAME);
3923 sprintf(lpath, "%s" LOCKNAME, path);
3924 sprintf(dpath, "%s" DATANAME, path);
3928 flags |= env->me_flags;
3929 if (flags & MDB_RDONLY) {
3930 /* silently ignore WRITEMAP when we're only getting read access */
3931 flags &= ~MDB_WRITEMAP;
3933 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3934 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3937 env->me_flags = flags |= MDB_ENV_ACTIVE;
3941 env->me_path = strdup(path);
3942 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3943 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3944 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3949 /* For RDONLY, get lockfile after we know datafile exists */
3950 if (!F_ISSET(flags, MDB_RDONLY)) {
3951 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3957 if (F_ISSET(flags, MDB_RDONLY)) {
3958 oflags = GENERIC_READ;
3959 len = OPEN_EXISTING;
3961 oflags = GENERIC_READ|GENERIC_WRITE;
3964 mode = FILE_ATTRIBUTE_NORMAL;
3965 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3966 NULL, len, mode, NULL);
3968 if (F_ISSET(flags, MDB_RDONLY))
3971 oflags = O_RDWR | O_CREAT;
3973 env->me_fd = open(dpath, oflags, mode);
3975 if (env->me_fd == INVALID_HANDLE_VALUE) {
3980 if (F_ISSET(flags, MDB_RDONLY)) {
3981 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3986 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3987 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3988 env->me_mfd = env->me_fd;
3990 /* Synchronous fd for meta writes. Needed even with
3991 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3994 len = OPEN_EXISTING;
3995 env->me_mfd = CreateFile(dpath, oflags,
3996 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3997 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4000 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4002 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4007 DPRINTF(("opened dbenv %p", (void *) env));
4009 rc = mdb_env_share_locks(env, &excl);
4015 mdb_env_close0(env, excl);
4021 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4023 mdb_env_close0(MDB_env *env, int excl)
4027 if (!(env->me_flags & MDB_ENV_ACTIVE))
4030 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4031 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4032 free(env->me_dbxs[i].md_name.mv_data);
4034 free(env->me_dbflags);
4037 free(env->me_dirty_list);
4038 mdb_midl_free(env->me_free_pgs);
4040 if (env->me_flags & MDB_ENV_TXKEY) {
4041 pthread_key_delete(env->me_txkey);
4043 /* Delete our key from the global list */
4044 for (i=0; i<mdb_tls_nkeys; i++)
4045 if (mdb_tls_keys[i] == env->me_txkey) {
4046 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4054 munmap(env->me_map, env->me_mapsize);
4056 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4057 (void) close(env->me_mfd);
4058 if (env->me_fd != INVALID_HANDLE_VALUE)
4059 (void) close(env->me_fd);
4061 pid_t pid = env->me_pid;
4062 /* Clearing readers is done in this function because
4063 * me_txkey with its destructor must be disabled first.
4065 for (i = env->me_numreaders; --i >= 0; )
4066 if (env->me_txns->mti_readers[i].mr_pid == pid)
4067 env->me_txns->mti_readers[i].mr_pid = 0;
4069 if (env->me_rmutex) {
4070 CloseHandle(env->me_rmutex);
4071 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4073 /* Windows automatically destroys the mutexes when
4074 * the last handle closes.
4076 #elif defined(MDB_USE_POSIX_SEM)
4077 if (env->me_rmutex != SEM_FAILED) {
4078 sem_close(env->me_rmutex);
4079 if (env->me_wmutex != SEM_FAILED)
4080 sem_close(env->me_wmutex);
4081 /* If we have the filelock: If we are the
4082 * only remaining user, clean up semaphores.
4085 mdb_env_excl_lock(env, &excl);
4087 sem_unlink(env->me_txns->mti_rmname);
4088 sem_unlink(env->me_txns->mti_wmname);
4092 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4094 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4097 /* Unlock the lockfile. Windows would have unlocked it
4098 * after closing anyway, but not necessarily at once.
4100 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4103 (void) close(env->me_lfd);
4106 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4110 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4112 MDB_txn *txn = NULL;
4118 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4122 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4125 /* Do the lock/unlock of the reader mutex before starting the
4126 * write txn. Otherwise other read txns could block writers.
4128 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4133 /* We must start the actual read txn after blocking writers */
4134 mdb_txn_reset0(txn, "reset-stage1");
4136 /* Temporarily block writers until we snapshot the meta pages */
4139 rc = mdb_txn_renew0(txn);
4141 UNLOCK_MUTEX_W(env);
4146 wsize = env->me_psize * 2;
4150 DO_WRITE(rc, fd, ptr, w2, len);
4154 } else if (len > 0) {
4160 /* Non-blocking or async handles are not supported */
4166 UNLOCK_MUTEX_W(env);
4171 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4173 if (wsize > MAX_WRITE)
4177 DO_WRITE(rc, fd, ptr, w2, len);
4181 } else if (len > 0) {
4198 mdb_env_copy(MDB_env *env, const char *path)
4202 HANDLE newfd = INVALID_HANDLE_VALUE;
4204 if (env->me_flags & MDB_NOSUBDIR) {
4205 lpath = (char *)path;
4208 len += sizeof(DATANAME);
4209 lpath = malloc(len);
4212 sprintf(lpath, "%s" DATANAME, path);
4215 /* The destination path must exist, but the destination file must not.
4216 * We don't want the OS to cache the writes, since the source data is
4217 * already in the OS cache.
4220 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4221 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4223 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4229 if (newfd == INVALID_HANDLE_VALUE) {
4234 #ifdef F_NOCACHE /* __APPLE__ */
4235 rc = fcntl(newfd, F_NOCACHE, 1);
4242 rc = mdb_env_copyfd(env, newfd);
4245 if (!(env->me_flags & MDB_NOSUBDIR))
4247 if (newfd != INVALID_HANDLE_VALUE)
4248 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4255 mdb_env_close(MDB_env *env)
4262 VGMEMP_DESTROY(env);
4263 while ((dp = env->me_dpages) != NULL) {
4264 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4265 env->me_dpages = dp->mp_next;
4269 mdb_env_close0(env, 0);
4273 /** Compare two items pointing at aligned size_t's */
4275 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4277 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4278 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4281 /** Compare two items pointing at aligned int's */
4283 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4285 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4286 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4289 /** Compare two items pointing at ints of unknown alignment.
4290 * Nodes and keys are guaranteed to be 2-byte aligned.
4293 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4295 #if BYTE_ORDER == LITTLE_ENDIAN
4296 unsigned short *u, *c;
4299 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4300 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4303 } while(!x && u > (unsigned short *)a->mv_data);
4306 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4310 /** Compare two items lexically */
4312 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4319 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4325 diff = memcmp(a->mv_data, b->mv_data, len);
4326 return diff ? diff : len_diff<0 ? -1 : len_diff;
4329 /** Compare two items in reverse byte order */
4331 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4333 const unsigned char *p1, *p2, *p1_lim;
4337 p1_lim = (const unsigned char *)a->mv_data;
4338 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4339 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4341 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4347 while (p1 > p1_lim) {
4348 diff = *--p1 - *--p2;
4352 return len_diff<0 ? -1 : len_diff;
4355 /** Search for key within a page, using binary search.
4356 * Returns the smallest entry larger or equal to the key.
4357 * If exactp is non-null, stores whether the found entry was an exact match
4358 * in *exactp (1 or 0).
4359 * Updates the cursor index with the index of the found entry.
4360 * If no entry larger or equal to the key is found, returns NULL.
4363 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4365 unsigned int i = 0, nkeys;
4368 MDB_page *mp = mc->mc_pg[mc->mc_top];
4369 MDB_node *node = NULL;
4374 nkeys = NUMKEYS(mp);
4379 COPY_PGNO(pgno, mp->mp_pgno);
4380 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4381 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4388 low = IS_LEAF(mp) ? 0 : 1;
4390 cmp = mc->mc_dbx->md_cmp;
4392 /* Branch pages have no data, so if using integer keys,
4393 * alignment is guaranteed. Use faster mdb_cmp_int.
4395 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4396 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4403 nodekey.mv_size = mc->mc_db->md_pad;
4404 node = NODEPTR(mp, 0); /* fake */
4405 while (low <= high) {
4406 i = (low + high) >> 1;
4407 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4408 rc = cmp(key, &nodekey);
4409 DPRINTF(("found leaf index %u [%s], rc = %i",
4410 i, DKEY(&nodekey), rc));
4419 while (low <= high) {
4420 i = (low + high) >> 1;
4422 node = NODEPTR(mp, i);
4423 nodekey.mv_size = NODEKSZ(node);
4424 nodekey.mv_data = NODEKEY(node);
4426 rc = cmp(key, &nodekey);
4429 DPRINTF(("found leaf index %u [%s], rc = %i",
4430 i, DKEY(&nodekey), rc));
4432 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4433 i, DKEY(&nodekey), NODEPGNO(node), rc));
4444 if (rc > 0) { /* Found entry is less than the key. */
4445 i++; /* Skip to get the smallest entry larger than key. */
4447 node = NODEPTR(mp, i);
4450 *exactp = (rc == 0);
4451 /* store the key index */
4452 mc->mc_ki[mc->mc_top] = i;
4454 /* There is no entry larger or equal to the key. */
4457 /* nodeptr is fake for LEAF2 */
4463 mdb_cursor_adjust(MDB_cursor *mc, func)
4467 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4468 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4475 /** Pop a page off the top of the cursor's stack. */
4477 mdb_cursor_pop(MDB_cursor *mc)
4481 MDB_page *top = mc->mc_pg[mc->mc_top];
4487 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4488 mc->mc_dbi, (void *) mc));
4492 /** Push a page onto the top of the cursor's stack. */
4494 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4496 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4497 mc->mc_dbi, (void *) mc));
4499 if (mc->mc_snum >= CURSOR_STACK) {
4500 assert(mc->mc_snum < CURSOR_STACK);
4501 return MDB_CURSOR_FULL;
4504 mc->mc_top = mc->mc_snum++;
4505 mc->mc_pg[mc->mc_top] = mp;
4506 mc->mc_ki[mc->mc_top] = 0;
4511 /** Find the address of the page corresponding to a given page number.
4512 * @param[in] txn the transaction for this access.
4513 * @param[in] pgno the page number for the page to retrieve.
4514 * @param[out] ret address of a pointer where the page's address will be stored.
4515 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4516 * @return 0 on success, non-zero on failure.
4519 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4521 MDB_env *env = txn->mt_env;
4525 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4529 MDB_ID2L dl = tx2->mt_u.dirty_list;
4531 /* Spilled pages were dirtied in this txn and flushed
4532 * because the dirty list got full. Bring this page
4533 * back in from the map (but don't unspill it here,
4534 * leave that unless page_touch happens again).
4536 if (tx2->mt_spill_pgs) {
4537 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4538 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4539 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4544 unsigned x = mdb_mid2l_search(dl, pgno);
4545 if (x <= dl[0].mid && dl[x].mid == pgno) {
4551 } while ((tx2 = tx2->mt_parent) != NULL);
4554 if (pgno < txn->mt_next_pgno) {
4556 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4558 DPRINTF(("page %"Z"u not found", pgno));
4560 return MDB_PAGE_NOTFOUND;
4570 /** Search for the page a given key should be in.
4571 * Pushes parent pages on the cursor stack. This function continues a
4572 * search on a cursor that has already been initialized. (Usually by
4573 * #mdb_page_search() but also by #mdb_node_move().)
4574 * @param[in,out] mc the cursor for this operation.
4575 * @param[in] key the key to search for. If NULL, search for the lowest
4576 * page. (This is used by #mdb_cursor_first().)
4577 * @param[in] modify If true, visited pages are updated with new page numbers.
4578 * @return 0 on success, non-zero on failure.
4581 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4583 MDB_page *mp = mc->mc_pg[mc->mc_top];
4587 while (IS_BRANCH(mp)) {
4591 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4592 assert(NUMKEYS(mp) > 1);
4593 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4595 if (key == NULL) /* Initialize cursor to first page. */
4597 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4598 /* cursor to last page */
4602 node = mdb_node_search(mc, key, &exact);
4604 i = NUMKEYS(mp) - 1;
4606 i = mc->mc_ki[mc->mc_top];
4615 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4616 assert(i < NUMKEYS(mp));
4617 node = NODEPTR(mp, i);
4619 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4622 mc->mc_ki[mc->mc_top] = i;
4623 if ((rc = mdb_cursor_push(mc, mp)))
4627 if ((rc = mdb_page_touch(mc)) != 0)
4629 mp = mc->mc_pg[mc->mc_top];
4634 DPRINTF(("internal error, index points to a %02X page!?",
4636 return MDB_CORRUPTED;
4639 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4640 key ? DKEY(key) : NULL));
4641 mc->mc_flags |= C_INITIALIZED;
4642 mc->mc_flags &= ~C_EOF;
4647 /** Search for the lowest key under the current branch page.
4648 * This just bypasses a NUMKEYS check in the current page
4649 * before calling mdb_page_search_root(), because the callers
4650 * are all in situations where the current page is known to
4654 mdb_page_search_lowest(MDB_cursor *mc)
4656 MDB_page *mp = mc->mc_pg[mc->mc_top];
4657 MDB_node *node = NODEPTR(mp, 0);
4660 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4663 mc->mc_ki[mc->mc_top] = 0;
4664 if ((rc = mdb_cursor_push(mc, mp)))
4666 return mdb_page_search_root(mc, NULL, 0);
4669 /** Search for the page a given key should be in.
4670 * Pushes parent pages on the cursor stack. This function just sets up
4671 * the search; it finds the root page for \b mc's database and sets this
4672 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4673 * called to complete the search.
4674 * @param[in,out] mc the cursor for this operation.
4675 * @param[in] key the key to search for. If NULL, search for the lowest
4676 * page. (This is used by #mdb_cursor_first().)
4677 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4678 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4679 * @return 0 on success, non-zero on failure.
4682 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4687 /* Make sure the txn is still viable, then find the root from
4688 * the txn's db table.
4690 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4691 DPUTS("transaction has failed, must abort");
4694 /* Make sure we're using an up-to-date root */
4695 if (mc->mc_dbi > MAIN_DBI) {
4696 if ((*mc->mc_dbflag & DB_STALE) ||
4697 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4699 unsigned char dbflag = 0;
4700 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4701 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4704 if (*mc->mc_dbflag & DB_STALE) {
4708 MDB_node *leaf = mdb_node_search(&mc2,
4709 &mc->mc_dbx->md_name, &exact);
4711 return MDB_NOTFOUND;
4712 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4715 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4717 /* The txn may not know this DBI, or another process may
4718 * have dropped and recreated the DB with other flags.
4720 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4721 return MDB_INCOMPATIBLE;
4722 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4724 if (flags & MDB_PS_MODIFY)
4726 *mc->mc_dbflag &= ~DB_STALE;
4727 *mc->mc_dbflag |= dbflag;
4730 root = mc->mc_db->md_root;
4732 if (root == P_INVALID) { /* Tree is empty. */
4733 DPUTS("tree is empty");
4734 return MDB_NOTFOUND;
4739 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4740 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4746 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4747 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4749 if (flags & MDB_PS_MODIFY) {
4750 if ((rc = mdb_page_touch(mc)))
4754 if (flags & MDB_PS_ROOTONLY)
4757 return mdb_page_search_root(mc, key, flags);
4761 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4763 MDB_txn *txn = mc->mc_txn;
4764 pgno_t pg = mp->mp_pgno;
4765 unsigned x = 0, ovpages = mp->mp_pages;
4766 MDB_env *env = txn->mt_env;
4767 MDB_IDL sl = txn->mt_spill_pgs;
4770 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4771 /* If the page is dirty or on the spill list we just acquired it,
4772 * so we should give it back to our current free list, if any.
4773 * Otherwise put it onto the list of pages we freed in this txn.
4775 * Won't create me_pghead: me_pglast must be inited along with it.
4776 * Unsupported in nested txns: They would need to hide the page
4777 * range in ancestor txns' dirty and spilled lists.
4779 if (env->me_pghead &&
4781 ((mp->mp_flags & P_DIRTY) ||
4782 (sl && (x = mdb_midl_search(sl, pg)) <= sl[0] && sl[x] == pg)))
4786 MDB_ID2 *dl, ix, iy;
4787 rc = mdb_midl_need(&env->me_pghead, ovpages);
4790 if (!(mp->mp_flags & P_DIRTY)) {
4791 /* This page is no longer spilled */
4792 for (; x < sl[0]; x++)
4797 /* Remove from dirty list */
4798 dl = txn->mt_u.dirty_list;
4800 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4808 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4809 txn->mt_flags |= MDB_TXN_ERROR;
4810 return MDB_CORRUPTED;
4813 if (!(env->me_flags & MDB_WRITEMAP))
4814 mdb_dpage_free(env, mp);
4816 /* Insert in me_pghead */
4817 mop = env->me_pghead;
4818 j = mop[0] + ovpages;
4819 for (i = mop[0]; i && mop[i] < pg; i--)
4825 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4829 mc->mc_db->md_overflow_pages -= ovpages;
4833 /** Return the data associated with a given node.
4834 * @param[in] txn The transaction for this operation.
4835 * @param[in] leaf The node being read.
4836 * @param[out] data Updated to point to the node's data.
4837 * @return 0 on success, non-zero on failure.
4840 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4842 MDB_page *omp; /* overflow page */
4846 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4847 data->mv_size = NODEDSZ(leaf);
4848 data->mv_data = NODEDATA(leaf);
4852 /* Read overflow data.
4854 data->mv_size = NODEDSZ(leaf);
4855 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4856 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4857 DPRINTF(("read overflow page %"Z"u failed", pgno));
4860 data->mv_data = METADATA(omp);
4866 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4867 MDB_val *key, MDB_val *data)
4876 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4878 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4881 if (txn->mt_flags & MDB_TXN_ERROR)
4884 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4885 return MDB_BAD_VALSIZE;
4888 mdb_cursor_init(&mc, txn, dbi, &mx);
4889 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4892 /** Find a sibling for a page.
4893 * Replaces the page at the top of the cursor's stack with the
4894 * specified sibling, if one exists.
4895 * @param[in] mc The cursor for this operation.
4896 * @param[in] move_right Non-zero if the right sibling is requested,
4897 * otherwise the left sibling.
4898 * @return 0 on success, non-zero on failure.
4901 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4907 if (mc->mc_snum < 2) {
4908 return MDB_NOTFOUND; /* root has no siblings */
4912 DPRINTF(("parent page is page %"Z"u, index %u",
4913 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4915 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4916 : (mc->mc_ki[mc->mc_top] == 0)) {
4917 DPRINTF(("no more keys left, moving to %s sibling",
4918 move_right ? "right" : "left"));
4919 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4920 /* undo cursor_pop before returning */
4927 mc->mc_ki[mc->mc_top]++;
4929 mc->mc_ki[mc->mc_top]--;
4930 DPRINTF(("just moving to %s index key %u",
4931 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4933 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4935 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4936 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4939 mdb_cursor_push(mc, mp);
4941 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4946 /** Move the cursor to the next data item. */
4948 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4954 if (mc->mc_flags & C_EOF) {
4955 return MDB_NOTFOUND;
4958 assert(mc->mc_flags & C_INITIALIZED);
4960 mp = mc->mc_pg[mc->mc_top];
4962 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4963 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4964 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4965 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4966 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4967 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
4968 if (rc == MDB_SUCCESS)
4969 MDB_GET_KEY(leaf, key);
4974 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4975 if (op == MDB_NEXT_DUP)
4976 return MDB_NOTFOUND;
4980 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4982 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4983 DPUTS("=====> move to next sibling page");
4984 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4985 mc->mc_flags |= C_EOF;
4988 mp = mc->mc_pg[mc->mc_top];
4989 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
4991 mc->mc_ki[mc->mc_top]++;
4993 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
4994 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
4997 key->mv_size = mc->mc_db->md_pad;
4998 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5002 assert(IS_LEAF(mp));
5003 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5005 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5006 mdb_xcursor_init1(mc, leaf);
5009 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5012 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5013 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5014 if (rc != MDB_SUCCESS)
5019 MDB_GET_KEY(leaf, key);
5023 /** Move the cursor to the previous data item. */
5025 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5031 assert(mc->mc_flags & C_INITIALIZED);
5033 mp = mc->mc_pg[mc->mc_top];
5035 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5036 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5037 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5038 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5039 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5040 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5041 if (rc == MDB_SUCCESS)
5042 MDB_GET_KEY(leaf, key);
5046 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5047 if (op == MDB_PREV_DUP)
5048 return MDB_NOTFOUND;
5053 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5055 if (mc->mc_ki[mc->mc_top] == 0) {
5056 DPUTS("=====> move to prev sibling page");
5057 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5060 mp = mc->mc_pg[mc->mc_top];
5061 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5062 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5064 mc->mc_ki[mc->mc_top]--;
5066 mc->mc_flags &= ~C_EOF;
5068 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5069 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5072 key->mv_size = mc->mc_db->md_pad;
5073 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5077 assert(IS_LEAF(mp));
5078 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5080 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5081 mdb_xcursor_init1(mc, leaf);
5084 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5087 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5088 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5089 if (rc != MDB_SUCCESS)
5094 MDB_GET_KEY(leaf, key);
5098 /** Set the cursor on a specific data item. */
5100 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5101 MDB_cursor_op op, int *exactp)
5105 MDB_node *leaf = NULL;
5110 assert(key->mv_size > 0);
5113 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5115 /* See if we're already on the right page */
5116 if (mc->mc_flags & C_INITIALIZED) {
5119 mp = mc->mc_pg[mc->mc_top];
5121 mc->mc_ki[mc->mc_top] = 0;
5122 return MDB_NOTFOUND;
5124 if (mp->mp_flags & P_LEAF2) {
5125 nodekey.mv_size = mc->mc_db->md_pad;
5126 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5128 leaf = NODEPTR(mp, 0);
5129 MDB_GET_KEY2(leaf, nodekey);
5131 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5133 /* Probably happens rarely, but first node on the page
5134 * was the one we wanted.
5136 mc->mc_ki[mc->mc_top] = 0;
5143 unsigned int nkeys = NUMKEYS(mp);
5145 if (mp->mp_flags & P_LEAF2) {
5146 nodekey.mv_data = LEAF2KEY(mp,
5147 nkeys-1, nodekey.mv_size);
5149 leaf = NODEPTR(mp, nkeys-1);
5150 MDB_GET_KEY2(leaf, nodekey);
5152 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5154 /* last node was the one we wanted */
5155 mc->mc_ki[mc->mc_top] = nkeys-1;
5161 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5162 /* This is definitely the right page, skip search_page */
5163 if (mp->mp_flags & P_LEAF2) {
5164 nodekey.mv_data = LEAF2KEY(mp,
5165 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5167 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5168 MDB_GET_KEY2(leaf, nodekey);
5170 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5172 /* current node was the one we wanted */
5182 /* If any parents have right-sibs, search.
5183 * Otherwise, there's nothing further.
5185 for (i=0; i<mc->mc_top; i++)
5187 NUMKEYS(mc->mc_pg[i])-1)
5189 if (i == mc->mc_top) {
5190 /* There are no other pages */
5191 mc->mc_ki[mc->mc_top] = nkeys;
5192 return MDB_NOTFOUND;
5196 /* There are no other pages */
5197 mc->mc_ki[mc->mc_top] = 0;
5198 return MDB_NOTFOUND;
5202 rc = mdb_page_search(mc, key, 0);
5203 if (rc != MDB_SUCCESS)
5206 mp = mc->mc_pg[mc->mc_top];
5207 assert(IS_LEAF(mp));
5210 leaf = mdb_node_search(mc, key, exactp);
5211 if (exactp != NULL && !*exactp) {
5212 /* MDB_SET specified and not an exact match. */
5213 return MDB_NOTFOUND;
5217 DPUTS("===> inexact leaf not found, goto sibling");
5218 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5219 return rc; /* no entries matched */
5220 mp = mc->mc_pg[mc->mc_top];
5221 assert(IS_LEAF(mp));
5222 leaf = NODEPTR(mp, 0);
5226 mc->mc_flags |= C_INITIALIZED;
5227 mc->mc_flags &= ~C_EOF;
5230 key->mv_size = mc->mc_db->md_pad;
5231 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5235 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5236 mdb_xcursor_init1(mc, leaf);
5239 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5240 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5241 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5244 if (op == MDB_GET_BOTH) {
5250 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5251 if (rc != MDB_SUCCESS)
5254 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5256 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5258 rc = mc->mc_dbx->md_dcmp(data, &d2);
5260 if (op == MDB_GET_BOTH || rc > 0)
5261 return MDB_NOTFOUND;
5266 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5267 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5272 /* The key already matches in all other cases */
5273 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5274 MDB_GET_KEY(leaf, key);
5275 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5280 /** Move the cursor to the first item in the database. */
5282 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5288 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5290 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5291 rc = mdb_page_search(mc, NULL, 0);
5292 if (rc != MDB_SUCCESS)
5295 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5297 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5298 mc->mc_flags |= C_INITIALIZED;
5299 mc->mc_flags &= ~C_EOF;
5301 mc->mc_ki[mc->mc_top] = 0;
5303 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5304 key->mv_size = mc->mc_db->md_pad;
5305 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5310 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5311 mdb_xcursor_init1(mc, leaf);
5312 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5316 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5320 MDB_GET_KEY(leaf, key);
5324 /** Move the cursor to the last item in the database. */
5326 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5332 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5334 if (!(mc->mc_flags & C_EOF)) {
5336 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5339 lkey.mv_size = MDB_MAXKEYSIZE+1;
5340 lkey.mv_data = NULL;
5341 rc = mdb_page_search(mc, &lkey, 0);
5342 if (rc != MDB_SUCCESS)
5345 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5348 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5349 mc->mc_flags |= C_INITIALIZED|C_EOF;
5350 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5352 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5353 key->mv_size = mc->mc_db->md_pad;
5354 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5359 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5360 mdb_xcursor_init1(mc, leaf);
5361 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5365 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5370 MDB_GET_KEY(leaf, key);
5375 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5380 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5384 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5388 case MDB_GET_CURRENT:
5389 if (!(mc->mc_flags & C_INITIALIZED)) {
5392 MDB_page *mp = mc->mc_pg[mc->mc_top];
5394 mc->mc_ki[mc->mc_top] = 0;
5400 key->mv_size = mc->mc_db->md_pad;
5401 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5403 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5404 MDB_GET_KEY(leaf, key);
5406 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5407 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5409 rc = mdb_node_read(mc->mc_txn, leaf, data);
5416 case MDB_GET_BOTH_RANGE:
5421 if (mc->mc_xcursor == NULL) {
5422 rc = MDB_INCOMPATIBLE;
5431 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5432 rc = MDB_BAD_VALSIZE;
5433 } else if (op == MDB_SET_RANGE)
5434 rc = mdb_cursor_set(mc, key, data, op, NULL);
5436 rc = mdb_cursor_set(mc, key, data, op, &exact);
5438 case MDB_GET_MULTIPLE:
5439 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5443 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5444 rc = MDB_INCOMPATIBLE;
5448 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5449 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5452 case MDB_NEXT_MULTIPLE:
5457 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5458 rc = MDB_INCOMPATIBLE;
5461 if (!(mc->mc_flags & C_INITIALIZED))
5462 rc = mdb_cursor_first(mc, key, data);
5464 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5465 if (rc == MDB_SUCCESS) {
5466 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5469 mx = &mc->mc_xcursor->mx_cursor;
5470 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5472 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5473 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5481 case MDB_NEXT_NODUP:
5482 if (!(mc->mc_flags & C_INITIALIZED))
5483 rc = mdb_cursor_first(mc, key, data);
5485 rc = mdb_cursor_next(mc, key, data, op);
5489 case MDB_PREV_NODUP:
5490 if (!(mc->mc_flags & C_INITIALIZED)) {
5491 rc = mdb_cursor_last(mc, key, data);
5494 mc->mc_flags |= C_INITIALIZED;
5495 mc->mc_ki[mc->mc_top]++;
5497 rc = mdb_cursor_prev(mc, key, data, op);
5500 rc = mdb_cursor_first(mc, key, data);
5503 mfunc = mdb_cursor_first;
5505 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5509 if (mc->mc_xcursor == NULL) {
5510 rc = MDB_INCOMPATIBLE;
5513 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5517 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5520 rc = mdb_cursor_last(mc, key, data);
5523 mfunc = mdb_cursor_last;
5526 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5534 /** Touch all the pages in the cursor stack.
5535 * Makes sure all the pages are writable, before attempting a write operation.
5536 * @param[in] mc The cursor to operate on.
5539 mdb_cursor_touch(MDB_cursor *mc)
5543 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5546 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5547 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5550 *mc->mc_dbflag |= DB_DIRTY;
5552 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5553 rc = mdb_page_touch(mc);
5557 mc->mc_top = mc->mc_snum-1;
5561 /** Do not spill pages to disk if txn is getting full, may fail instead */
5562 #define MDB_NOSPILL 0x8000
5565 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5568 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5569 MDB_node *leaf = NULL;
5570 MDB_val xdata, *rdata, dkey;
5573 int do_sub = 0, insert = 0;
5574 unsigned int mcount = 0, dcount = 0, nospill;
5578 char dbuf[MDB_MAXKEYSIZE+1];
5579 unsigned int nflags;
5582 /* Check this first so counter will always be zero on any
5585 if (flags & MDB_MULTIPLE) {
5586 dcount = data[1].mv_size;
5587 data[1].mv_size = 0;
5588 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5589 return MDB_INCOMPATIBLE;
5592 nospill = flags & MDB_NOSPILL;
5593 flags &= ~MDB_NOSPILL;
5595 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5596 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5598 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5599 return MDB_BAD_VALSIZE;
5601 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5602 return MDB_BAD_VALSIZE;
5604 #if SIZE_MAX > MAXDATASIZE
5605 if (data->mv_size > MAXDATASIZE)
5606 return MDB_BAD_VALSIZE;
5609 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5610 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5614 if (flags == MDB_CURRENT) {
5615 if (!(mc->mc_flags & C_INITIALIZED))
5618 } else if (mc->mc_db->md_root == P_INVALID) {
5619 /* new database, cursor has nothing to point to */
5621 mc->mc_flags &= ~C_INITIALIZED;
5626 if (flags & MDB_APPEND) {
5628 rc = mdb_cursor_last(mc, &k2, &d2);
5630 rc = mc->mc_dbx->md_cmp(key, &k2);
5633 mc->mc_ki[mc->mc_top]++;
5635 /* new key is <= last key */
5640 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5642 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5643 DPRINTF(("duplicate key [%s]", DKEY(key)));
5645 return MDB_KEYEXIST;
5647 if (rc && rc != MDB_NOTFOUND)
5651 /* Cursor is positioned, check for room in the dirty list */
5653 if (flags & MDB_MULTIPLE) {
5655 xdata.mv_size = data->mv_size * dcount;
5659 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5663 if (rc == MDB_NO_ROOT) {
5665 /* new database, write a root leaf page */
5666 DPUTS("allocating new root leaf page");
5667 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5670 mdb_cursor_push(mc, np);
5671 mc->mc_db->md_root = np->mp_pgno;
5672 mc->mc_db->md_depth++;
5673 *mc->mc_dbflag |= DB_DIRTY;
5674 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5676 np->mp_flags |= P_LEAF2;
5677 mc->mc_flags |= C_INITIALIZED;
5679 /* make sure all cursor pages are writable */
5680 rc2 = mdb_cursor_touch(mc);
5685 /* The key already exists */
5686 if (rc == MDB_SUCCESS) {
5687 /* there's only a key anyway, so this is a no-op */
5688 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5689 unsigned int ksize = mc->mc_db->md_pad;
5690 if (key->mv_size != ksize)
5691 return MDB_BAD_VALSIZE;
5692 if (flags == MDB_CURRENT) {
5693 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5694 memcpy(ptr, key->mv_data, ksize);
5699 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5702 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5703 /* Was a single item before, must convert now */
5705 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5706 /* Just overwrite the current item */
5707 if (flags == MDB_CURRENT)
5710 dkey.mv_size = NODEDSZ(leaf);
5711 dkey.mv_data = NODEDATA(leaf);
5712 #if UINT_MAX < SIZE_MAX
5713 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5714 #ifdef MISALIGNED_OK
5715 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5717 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5720 /* if data matches, skip it */
5721 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5722 if (flags & MDB_NODUPDATA)
5724 else if (flags & MDB_MULTIPLE)
5731 /* create a fake page for the dup items */
5732 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5733 dkey.mv_data = dbuf;
5734 fp = (MDB_page *)&pbuf;
5735 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5736 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5737 fp->mp_lower = PAGEHDRSZ;
5738 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5739 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5740 fp->mp_flags |= P_LEAF2;
5741 fp->mp_pad = data->mv_size;
5742 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5744 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5745 (dkey.mv_size & 1) + (data->mv_size & 1);
5747 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5750 xdata.mv_size = fp->mp_upper;
5755 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5756 /* See if we need to convert from fake page to subDB */
5758 unsigned int offset;
5762 fp = NODEDATA(leaf);
5763 if (flags == MDB_CURRENT) {
5765 fp->mp_flags |= P_DIRTY;
5766 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5767 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5771 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5772 offset = fp->mp_pad;
5773 if (SIZELEFT(fp) >= offset)
5775 offset *= 4; /* space for 4 more */
5777 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5779 offset += offset & 1;
5780 fp_flags = fp->mp_flags;
5781 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5782 offset >= mc->mc_txn->mt_env->me_nodemax) {
5783 /* yes, convert it */
5785 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5786 dummy.md_pad = fp->mp_pad;
5787 dummy.md_flags = MDB_DUPFIXED;
5788 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5789 dummy.md_flags |= MDB_INTEGERKEY;
5792 dummy.md_branch_pages = 0;
5793 dummy.md_leaf_pages = 1;
5794 dummy.md_overflow_pages = 0;
5795 dummy.md_entries = NUMKEYS(fp);
5797 xdata.mv_size = sizeof(MDB_db);
5798 xdata.mv_data = &dummy;
5799 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5801 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5802 flags |= F_DUPDATA|F_SUBDATA;
5803 dummy.md_root = mp->mp_pgno;
5804 fp_flags &= ~P_SUBP;
5806 /* no, just grow it */
5808 xdata.mv_size = NODEDSZ(leaf) + offset;
5809 xdata.mv_data = &pbuf;
5810 mp = (MDB_page *)&pbuf;
5811 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5814 mp->mp_flags = fp_flags | P_DIRTY;
5815 mp->mp_pad = fp->mp_pad;
5816 mp->mp_lower = fp->mp_lower;
5817 mp->mp_upper = fp->mp_upper + offset;
5819 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5821 nsize = NODEDSZ(leaf) - fp->mp_upper;
5822 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5823 for (i=0; i<NUMKEYS(fp); i++)
5824 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5826 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5830 /* data is on sub-DB, just store it */
5831 flags |= F_DUPDATA|F_SUBDATA;
5835 /* overflow page overwrites need special handling */
5836 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5839 unsigned psize = mc->mc_txn->mt_env->me_psize;
5840 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5842 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5843 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5845 ovpages = omp->mp_pages;
5847 /* Is the ov page large enough? */
5848 if (ovpages >= dpages) {
5849 if (!(omp->mp_flags & P_DIRTY) &&
5850 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5852 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5855 level = 0; /* dirty in this txn or clean */
5858 if (omp->mp_flags & P_DIRTY) {
5859 /* yes, overwrite it. Note in this case we don't
5860 * bother to try shrinking the page if the new data
5861 * is smaller than the overflow threshold.
5864 /* It is writable only in a parent txn */
5865 size_t sz = (size_t) psize * ovpages, off;
5866 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5872 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5873 if (!(flags & MDB_RESERVE)) {
5874 /* Copy end of page, adjusting alignment so
5875 * compiler may copy words instead of bytes.
5877 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5878 memcpy((size_t *)((char *)np + off),
5879 (size_t *)((char *)omp + off), sz - off);
5882 memcpy(np, omp, sz); /* Copy beginning of page */
5885 SETDSZ(leaf, data->mv_size);
5886 if (F_ISSET(flags, MDB_RESERVE))
5887 data->mv_data = METADATA(omp);
5889 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5893 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5895 } else if (NODEDSZ(leaf) == data->mv_size) {
5896 /* same size, just replace it. Note that we could
5897 * also reuse this node if the new data is smaller,
5898 * but instead we opt to shrink the node in that case.
5900 if (F_ISSET(flags, MDB_RESERVE))
5901 data->mv_data = NODEDATA(leaf);
5902 else if (data->mv_size)
5903 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5905 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5908 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5909 mc->mc_db->md_entries--;
5911 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5918 nflags = flags & NODE_ADD_FLAGS;
5919 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5920 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5921 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5922 nflags &= ~MDB_APPEND;
5924 nflags |= MDB_SPLIT_REPLACE;
5925 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5927 /* There is room already in this leaf page. */
5928 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5929 if (rc == 0 && !do_sub && insert) {
5930 /* Adjust other cursors pointing to mp */
5931 MDB_cursor *m2, *m3;
5932 MDB_dbi dbi = mc->mc_dbi;
5933 unsigned i = mc->mc_top;
5934 MDB_page *mp = mc->mc_pg[i];
5936 if (mc->mc_flags & C_SUB)
5939 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5940 if (mc->mc_flags & C_SUB)
5941 m3 = &m2->mc_xcursor->mx_cursor;
5944 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5945 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5952 if (rc != MDB_SUCCESS)
5953 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5955 /* Now store the actual data in the child DB. Note that we're
5956 * storing the user data in the keys field, so there are strict
5957 * size limits on dupdata. The actual data fields of the child
5958 * DB are all zero size.
5965 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5966 if (flags & MDB_CURRENT) {
5967 xflags = MDB_CURRENT|MDB_NOSPILL;
5969 mdb_xcursor_init1(mc, leaf);
5970 xflags = (flags & MDB_NODUPDATA) ?
5971 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5973 /* converted, write the original data first */
5975 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5979 /* Adjust other cursors pointing to mp */
5981 unsigned i = mc->mc_top;
5982 MDB_page *mp = mc->mc_pg[i];
5984 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5985 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5986 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5987 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5988 mdb_xcursor_init1(m2, leaf);
5992 /* we've done our job */
5995 if (flags & MDB_APPENDDUP)
5996 xflags |= MDB_APPEND;
5997 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5998 if (flags & F_SUBDATA) {
5999 void *db = NODEDATA(leaf);
6000 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6003 /* sub-writes might have failed so check rc again.
6004 * Don't increment count if we just replaced an existing item.
6006 if (!rc && !(flags & MDB_CURRENT))
6007 mc->mc_db->md_entries++;
6008 if (flags & MDB_MULTIPLE) {
6012 /* let caller know how many succeeded, if any */
6013 data[1].mv_size = mcount;
6014 if (mcount < dcount) {
6015 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6016 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6023 /* If we succeeded and the key didn't exist before, make sure
6024 * the cursor is marked valid.
6027 mc->mc_flags |= C_INITIALIZED;
6032 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6037 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6038 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6040 if (!(mc->mc_flags & C_INITIALIZED))
6043 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6045 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
6047 rc = mdb_cursor_touch(mc);
6051 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6053 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6054 if (!(flags & MDB_NODUPDATA)) {
6055 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6056 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6058 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6059 /* If sub-DB still has entries, we're done */
6060 if (mc->mc_xcursor->mx_db.md_entries) {
6061 if (leaf->mn_flags & F_SUBDATA) {
6062 /* update subDB info */
6063 void *db = NODEDATA(leaf);
6064 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6067 /* shrink fake page */
6068 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6069 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6070 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6071 /* fix other sub-DB cursors pointed at this fake page */
6072 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6073 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6074 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6075 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6076 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6079 mc->mc_db->md_entries--;
6082 /* otherwise fall thru and delete the sub-DB */
6085 if (leaf->mn_flags & F_SUBDATA) {
6086 /* add all the child DB's pages to the free list */
6087 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6088 if (rc == MDB_SUCCESS) {
6089 mc->mc_db->md_entries -=
6090 mc->mc_xcursor->mx_db.md_entries;
6095 return mdb_cursor_del0(mc, leaf);
6098 /** Allocate and initialize new pages for a database.
6099 * @param[in] mc a cursor on the database being added to.
6100 * @param[in] flags flags defining what type of page is being allocated.
6101 * @param[in] num the number of pages to allocate. This is usually 1,
6102 * unless allocating overflow pages for a large record.
6103 * @param[out] mp Address of a page, or NULL on failure.
6104 * @return 0 on success, non-zero on failure.
6107 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6112 if ((rc = mdb_page_alloc(mc, num, &np)))
6114 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6115 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6116 np->mp_flags = flags | P_DIRTY;
6117 np->mp_lower = PAGEHDRSZ;
6118 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6121 mc->mc_db->md_branch_pages++;
6122 else if (IS_LEAF(np))
6123 mc->mc_db->md_leaf_pages++;
6124 else if (IS_OVERFLOW(np)) {
6125 mc->mc_db->md_overflow_pages += num;
6133 /** Calculate the size of a leaf node.
6134 * The size depends on the environment's page size; if a data item
6135 * is too large it will be put onto an overflow page and the node
6136 * size will only include the key and not the data. Sizes are always
6137 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6138 * of the #MDB_node headers.
6139 * @param[in] env The environment handle.
6140 * @param[in] key The key for the node.
6141 * @param[in] data The data for the node.
6142 * @return The number of bytes needed to store the node.
6145 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6149 sz = LEAFSIZE(key, data);
6150 if (sz >= env->me_nodemax) {
6151 /* put on overflow page */
6152 sz -= data->mv_size - sizeof(pgno_t);
6156 return sz + sizeof(indx_t);
6159 /** Calculate the size of a branch node.
6160 * The size should depend on the environment's page size but since
6161 * we currently don't support spilling large keys onto overflow
6162 * pages, it's simply the size of the #MDB_node header plus the
6163 * size of the key. Sizes are always rounded up to an even number
6164 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6165 * @param[in] env The environment handle.
6166 * @param[in] key The key for the node.
6167 * @return The number of bytes needed to store the node.
6170 mdb_branch_size(MDB_env *env, MDB_val *key)
6175 if (sz >= env->me_nodemax) {
6176 /* put on overflow page */
6177 /* not implemented */
6178 /* sz -= key->size - sizeof(pgno_t); */
6181 return sz + sizeof(indx_t);
6184 /** Add a node to the page pointed to by the cursor.
6185 * @param[in] mc The cursor for this operation.
6186 * @param[in] indx The index on the page where the new node should be added.
6187 * @param[in] key The key for the new node.
6188 * @param[in] data The data for the new node, if any.
6189 * @param[in] pgno The page number, if adding a branch node.
6190 * @param[in] flags Flags for the node.
6191 * @return 0 on success, non-zero on failure. Possible errors are:
6193 * <li>ENOMEM - failed to allocate overflow pages for the node.
6194 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6195 * should never happen since all callers already calculate the
6196 * page's free space before calling this function.
6200 mdb_node_add(MDB_cursor *mc, indx_t indx,
6201 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6204 size_t node_size = NODESIZE;
6207 MDB_page *mp = mc->mc_pg[mc->mc_top];
6208 MDB_page *ofp = NULL; /* overflow page */
6211 assert(mp->mp_upper >= mp->mp_lower);
6213 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6214 IS_LEAF(mp) ? "leaf" : "branch",
6215 IS_SUBP(mp) ? "sub-" : "",
6216 mp->mp_pgno, indx, data ? data->mv_size : 0,
6217 key ? key->mv_size : 0, key ? DKEY(key) : NULL));
6220 /* Move higher keys up one slot. */
6221 int ksize = mc->mc_db->md_pad, dif;
6222 char *ptr = LEAF2KEY(mp, indx, ksize);
6223 dif = NUMKEYS(mp) - indx;
6225 memmove(ptr+ksize, ptr, dif*ksize);
6226 /* insert new key */
6227 memcpy(ptr, key->mv_data, ksize);
6229 /* Just using these for counting */
6230 mp->mp_lower += sizeof(indx_t);
6231 mp->mp_upper -= ksize - sizeof(indx_t);
6236 node_size += key->mv_size;
6240 if (F_ISSET(flags, F_BIGDATA)) {
6241 /* Data already on overflow page. */
6242 node_size += sizeof(pgno_t);
6243 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6244 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6246 /* Put data on overflow page. */
6247 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6248 data->mv_size, node_size+data->mv_size));
6249 node_size += sizeof(pgno_t);
6250 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6252 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6255 node_size += data->mv_size;
6258 node_size += node_size & 1;
6260 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6261 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6262 mp->mp_pgno, NUMKEYS(mp)));
6263 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6264 mp->mp_upper - mp->mp_lower));
6265 DPRINTF(("node size = %"Z"u", node_size));
6266 return MDB_PAGE_FULL;
6269 /* Move higher pointers up one slot. */
6270 for (i = NUMKEYS(mp); i > indx; i--)
6271 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6273 /* Adjust free space offsets. */
6274 ofs = mp->mp_upper - node_size;
6275 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6276 mp->mp_ptrs[indx] = ofs;
6278 mp->mp_lower += sizeof(indx_t);
6280 /* Write the node data. */
6281 node = NODEPTR(mp, indx);
6282 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6283 node->mn_flags = flags;
6285 SETDSZ(node,data->mv_size);
6290 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6295 if (F_ISSET(flags, F_BIGDATA))
6296 memcpy(node->mn_data + key->mv_size, data->mv_data,
6298 else if (F_ISSET(flags, MDB_RESERVE))
6299 data->mv_data = node->mn_data + key->mv_size;
6301 memcpy(node->mn_data + key->mv_size, data->mv_data,
6304 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6306 if (F_ISSET(flags, MDB_RESERVE))
6307 data->mv_data = METADATA(ofp);
6309 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6316 /** Delete the specified node from a page.
6317 * @param[in] mp The page to operate on.
6318 * @param[in] indx The index of the node to delete.
6319 * @param[in] ksize The size of a node. Only used if the page is
6320 * part of a #MDB_DUPFIXED database.
6323 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6326 indx_t i, j, numkeys, ptr;
6333 COPY_PGNO(pgno, mp->mp_pgno);
6334 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6335 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6338 assert(indx < NUMKEYS(mp));
6341 int x = NUMKEYS(mp) - 1 - indx;
6342 base = LEAF2KEY(mp, indx, ksize);
6344 memmove(base, base + ksize, x * ksize);
6345 mp->mp_lower -= sizeof(indx_t);
6346 mp->mp_upper += ksize - sizeof(indx_t);
6350 node = NODEPTR(mp, indx);
6351 sz = NODESIZE + node->mn_ksize;
6353 if (F_ISSET(node->mn_flags, F_BIGDATA))
6354 sz += sizeof(pgno_t);
6356 sz += NODEDSZ(node);
6360 ptr = mp->mp_ptrs[indx];
6361 numkeys = NUMKEYS(mp);
6362 for (i = j = 0; i < numkeys; i++) {
6364 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6365 if (mp->mp_ptrs[i] < ptr)
6366 mp->mp_ptrs[j] += sz;
6371 base = (char *)mp + mp->mp_upper;
6372 memmove(base + sz, base, ptr - mp->mp_upper);
6374 mp->mp_lower -= sizeof(indx_t);
6378 /** Compact the main page after deleting a node on a subpage.
6379 * @param[in] mp The main page to operate on.
6380 * @param[in] indx The index of the subpage on the main page.
6383 mdb_node_shrink(MDB_page *mp, indx_t indx)
6390 indx_t i, numkeys, ptr;
6392 node = NODEPTR(mp, indx);
6393 sp = (MDB_page *)NODEDATA(node);
6394 osize = NODEDSZ(node);
6396 delta = sp->mp_upper - sp->mp_lower;
6397 SETDSZ(node, osize - delta);
6398 xp = (MDB_page *)((char *)sp + delta);
6400 /* shift subpage upward */
6402 nsize = NUMKEYS(sp) * sp->mp_pad;
6403 memmove(METADATA(xp), METADATA(sp), nsize);
6406 nsize = osize - sp->mp_upper;
6407 numkeys = NUMKEYS(sp);
6408 for (i=numkeys-1; i>=0; i--)
6409 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6411 xp->mp_upper = sp->mp_lower;
6412 xp->mp_lower = sp->mp_lower;
6413 xp->mp_flags = sp->mp_flags;
6414 xp->mp_pad = sp->mp_pad;
6415 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6417 /* shift lower nodes upward */
6418 ptr = mp->mp_ptrs[indx];
6419 numkeys = NUMKEYS(mp);
6420 for (i = 0; i < numkeys; i++) {
6421 if (mp->mp_ptrs[i] <= ptr)
6422 mp->mp_ptrs[i] += delta;
6425 base = (char *)mp + mp->mp_upper;
6426 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6427 mp->mp_upper += delta;
6430 /** Initial setup of a sorted-dups cursor.
6431 * Sorted duplicates are implemented as a sub-database for the given key.
6432 * The duplicate data items are actually keys of the sub-database.
6433 * Operations on the duplicate data items are performed using a sub-cursor
6434 * initialized when the sub-database is first accessed. This function does
6435 * the preliminary setup of the sub-cursor, filling in the fields that
6436 * depend only on the parent DB.
6437 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6440 mdb_xcursor_init0(MDB_cursor *mc)
6442 MDB_xcursor *mx = mc->mc_xcursor;
6444 mx->mx_cursor.mc_xcursor = NULL;
6445 mx->mx_cursor.mc_txn = mc->mc_txn;
6446 mx->mx_cursor.mc_db = &mx->mx_db;
6447 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6448 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6449 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6450 mx->mx_cursor.mc_snum = 0;
6451 mx->mx_cursor.mc_top = 0;
6452 mx->mx_cursor.mc_flags = C_SUB;
6453 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6454 mx->mx_dbx.md_dcmp = NULL;
6455 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6458 /** Final setup of a sorted-dups cursor.
6459 * Sets up the fields that depend on the data from the main cursor.
6460 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6461 * @param[in] node The data containing the #MDB_db record for the
6462 * sorted-dup database.
6465 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6467 MDB_xcursor *mx = mc->mc_xcursor;
6469 if (node->mn_flags & F_SUBDATA) {
6470 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6471 mx->mx_cursor.mc_pg[0] = 0;
6472 mx->mx_cursor.mc_snum = 0;
6473 mx->mx_cursor.mc_flags = C_SUB;
6475 MDB_page *fp = NODEDATA(node);
6476 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6477 mx->mx_db.md_flags = 0;
6478 mx->mx_db.md_depth = 1;
6479 mx->mx_db.md_branch_pages = 0;
6480 mx->mx_db.md_leaf_pages = 1;
6481 mx->mx_db.md_overflow_pages = 0;
6482 mx->mx_db.md_entries = NUMKEYS(fp);
6483 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6484 mx->mx_cursor.mc_snum = 1;
6485 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6486 mx->mx_cursor.mc_top = 0;
6487 mx->mx_cursor.mc_pg[0] = fp;
6488 mx->mx_cursor.mc_ki[0] = 0;
6489 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6490 mx->mx_db.md_flags = MDB_DUPFIXED;
6491 mx->mx_db.md_pad = fp->mp_pad;
6492 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6493 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6496 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6497 mx->mx_db.md_root));
6498 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6500 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6501 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6502 #if UINT_MAX < SIZE_MAX
6503 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6504 #ifdef MISALIGNED_OK
6505 mx->mx_dbx.md_cmp = mdb_cmp_long;
6507 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6512 /** Initialize a cursor for a given transaction and database. */
6514 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6517 mc->mc_backup = NULL;
6520 mc->mc_db = &txn->mt_dbs[dbi];
6521 mc->mc_dbx = &txn->mt_dbxs[dbi];
6522 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6527 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6529 mc->mc_xcursor = mx;
6530 mdb_xcursor_init0(mc);
6532 mc->mc_xcursor = NULL;
6534 if (*mc->mc_dbflag & DB_STALE) {
6535 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6540 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6543 size_t size = sizeof(MDB_cursor);
6545 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6548 if (txn->mt_flags & MDB_TXN_ERROR)
6551 /* Allow read access to the freelist */
6552 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6555 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6556 size += sizeof(MDB_xcursor);
6558 if ((mc = malloc(size)) != NULL) {
6559 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6560 if (txn->mt_cursors) {
6561 mc->mc_next = txn->mt_cursors[dbi];
6562 txn->mt_cursors[dbi] = mc;
6563 mc->mc_flags |= C_UNTRACK;
6575 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6577 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6580 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6583 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6587 /* Return the count of duplicate data items for the current key */
6589 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6593 if (mc == NULL || countp == NULL)
6596 if (mc->mc_xcursor == NULL)
6597 return MDB_INCOMPATIBLE;
6599 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6600 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6603 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6606 *countp = mc->mc_xcursor->mx_db.md_entries;
6612 mdb_cursor_close(MDB_cursor *mc)
6614 if (mc && !mc->mc_backup) {
6615 /* remove from txn, if tracked */
6616 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6617 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6618 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6620 *prev = mc->mc_next;
6627 mdb_cursor_txn(MDB_cursor *mc)
6629 if (!mc) return NULL;
6634 mdb_cursor_dbi(MDB_cursor *mc)
6640 /** Replace the key for a node with a new key.
6641 * @param[in] mc Cursor pointing to the node to operate on.
6642 * @param[in] key The new key to use.
6643 * @return 0 on success, non-zero on failure.
6646 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6653 indx_t ptr, i, numkeys, indx;
6656 indx = mc->mc_ki[mc->mc_top];
6657 mp = mc->mc_pg[mc->mc_top];
6658 node = NODEPTR(mp, indx);
6659 ptr = mp->mp_ptrs[indx];
6663 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6664 k2.mv_data = NODEKEY(node);
6665 k2.mv_size = node->mn_ksize;
6666 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6668 mdb_dkey(&k2, kbuf2),
6674 delta0 = delta = key->mv_size - node->mn_ksize;
6676 /* Must be 2-byte aligned. If new key is
6677 * shorter by 1, the shift will be skipped.
6679 delta += (delta & 1);
6681 if (delta > 0 && SIZELEFT(mp) < delta) {
6683 /* not enough space left, do a delete and split */
6684 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6685 pgno = NODEPGNO(node);
6686 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6687 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6690 numkeys = NUMKEYS(mp);
6691 for (i = 0; i < numkeys; i++) {
6692 if (mp->mp_ptrs[i] <= ptr)
6693 mp->mp_ptrs[i] -= delta;
6696 base = (char *)mp + mp->mp_upper;
6697 len = ptr - mp->mp_upper + NODESIZE;
6698 memmove(base - delta, base, len);
6699 mp->mp_upper -= delta;
6701 node = NODEPTR(mp, indx);
6704 /* But even if no shift was needed, update ksize */
6706 node->mn_ksize = key->mv_size;
6709 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6715 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6717 /** Move a node from csrc to cdst.
6720 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6727 unsigned short flags;
6731 /* Mark src and dst as dirty. */
6732 if ((rc = mdb_page_touch(csrc)) ||
6733 (rc = mdb_page_touch(cdst)))
6736 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6737 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6738 key.mv_size = csrc->mc_db->md_pad;
6739 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6741 data.mv_data = NULL;
6745 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6746 assert(!((long)srcnode&1));
6747 srcpg = NODEPGNO(srcnode);
6748 flags = srcnode->mn_flags;
6749 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6750 unsigned int snum = csrc->mc_snum;
6752 /* must find the lowest key below src */
6753 mdb_page_search_lowest(csrc);
6754 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6755 key.mv_size = csrc->mc_db->md_pad;
6756 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6758 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6759 key.mv_size = NODEKSZ(s2);
6760 key.mv_data = NODEKEY(s2);
6762 csrc->mc_snum = snum--;
6763 csrc->mc_top = snum;
6765 key.mv_size = NODEKSZ(srcnode);
6766 key.mv_data = NODEKEY(srcnode);
6768 data.mv_size = NODEDSZ(srcnode);
6769 data.mv_data = NODEDATA(srcnode);
6771 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6772 unsigned int snum = cdst->mc_snum;
6775 /* must find the lowest key below dst */
6776 mdb_page_search_lowest(cdst);
6777 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6778 bkey.mv_size = cdst->mc_db->md_pad;
6779 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6781 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6782 bkey.mv_size = NODEKSZ(s2);
6783 bkey.mv_data = NODEKEY(s2);
6785 cdst->mc_snum = snum--;
6786 cdst->mc_top = snum;
6787 mdb_cursor_copy(cdst, &mn);
6789 rc = mdb_update_key(&mn, &bkey);
6794 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6795 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6796 csrc->mc_ki[csrc->mc_top],
6798 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6799 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6801 /* Add the node to the destination page.
6803 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6804 if (rc != MDB_SUCCESS)
6807 /* Delete the node from the source page.
6809 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6812 /* Adjust other cursors pointing to mp */
6813 MDB_cursor *m2, *m3;
6814 MDB_dbi dbi = csrc->mc_dbi;
6815 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6817 if (csrc->mc_flags & C_SUB)
6820 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6821 if (csrc->mc_flags & C_SUB)
6822 m3 = &m2->mc_xcursor->mx_cursor;
6825 if (m3 == csrc) continue;
6826 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6827 csrc->mc_ki[csrc->mc_top]) {
6828 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6829 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6834 /* Update the parent separators.
6836 if (csrc->mc_ki[csrc->mc_top] == 0) {
6837 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6838 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6839 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6841 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6842 key.mv_size = NODEKSZ(srcnode);
6843 key.mv_data = NODEKEY(srcnode);
6845 DPRINTF(("update separator for source page %"Z"u to [%s]",
6846 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6847 mdb_cursor_copy(csrc, &mn);
6850 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6853 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6855 indx_t ix = csrc->mc_ki[csrc->mc_top];
6856 nullkey.mv_size = 0;
6857 csrc->mc_ki[csrc->mc_top] = 0;
6858 rc = mdb_update_key(csrc, &nullkey);
6859 csrc->mc_ki[csrc->mc_top] = ix;
6860 assert(rc == MDB_SUCCESS);
6864 if (cdst->mc_ki[cdst->mc_top] == 0) {
6865 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6866 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6867 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6869 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6870 key.mv_size = NODEKSZ(srcnode);
6871 key.mv_data = NODEKEY(srcnode);
6873 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6874 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6875 mdb_cursor_copy(cdst, &mn);
6878 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6881 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6883 indx_t ix = cdst->mc_ki[cdst->mc_top];
6884 nullkey.mv_size = 0;
6885 cdst->mc_ki[cdst->mc_top] = 0;
6886 rc = mdb_update_key(cdst, &nullkey);
6887 cdst->mc_ki[cdst->mc_top] = ix;
6888 assert(rc == MDB_SUCCESS);
6895 /** Merge one page into another.
6896 * The nodes from the page pointed to by \b csrc will
6897 * be copied to the page pointed to by \b cdst and then
6898 * the \b csrc page will be freed.
6899 * @param[in] csrc Cursor pointing to the source page.
6900 * @param[in] cdst Cursor pointing to the destination page.
6903 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6911 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6912 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6914 assert(csrc->mc_snum > 1); /* can't merge root page */
6915 assert(cdst->mc_snum > 1);
6917 /* Mark dst as dirty. */
6918 if ((rc = mdb_page_touch(cdst)))
6921 /* Move all nodes from src to dst.
6923 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6924 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6925 key.mv_size = csrc->mc_db->md_pad;
6926 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6927 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6928 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6929 if (rc != MDB_SUCCESS)
6931 key.mv_data = (char *)key.mv_data + key.mv_size;
6934 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6935 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6936 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6937 unsigned int snum = csrc->mc_snum;
6939 /* must find the lowest key below src */
6940 mdb_page_search_lowest(csrc);
6941 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6942 key.mv_size = csrc->mc_db->md_pad;
6943 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6945 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6946 key.mv_size = NODEKSZ(s2);
6947 key.mv_data = NODEKEY(s2);
6949 csrc->mc_snum = snum--;
6950 csrc->mc_top = snum;
6952 key.mv_size = srcnode->mn_ksize;
6953 key.mv_data = NODEKEY(srcnode);
6956 data.mv_size = NODEDSZ(srcnode);
6957 data.mv_data = NODEDATA(srcnode);
6958 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6959 if (rc != MDB_SUCCESS)
6964 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6965 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
6966 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
6968 /* Unlink the src page from parent and add to free list.
6970 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6971 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6974 rc = mdb_update_key(csrc, &key);
6980 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6981 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6984 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6985 csrc->mc_db->md_leaf_pages--;
6987 csrc->mc_db->md_branch_pages--;
6989 /* Adjust other cursors pointing to mp */
6990 MDB_cursor *m2, *m3;
6991 MDB_dbi dbi = csrc->mc_dbi;
6992 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6994 if (csrc->mc_flags & C_SUB)
6997 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6998 if (csrc->mc_flags & C_SUB)
6999 m3 = &m2->mc_xcursor->mx_cursor;
7002 if (m3 == csrc) continue;
7003 if (m3->mc_snum < csrc->mc_snum) continue;
7004 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7005 m3->mc_pg[csrc->mc_top] = mp;
7006 m3->mc_ki[csrc->mc_top] += nkeys;
7010 mdb_cursor_pop(csrc);
7012 return mdb_rebalance(csrc);
7015 /** Copy the contents of a cursor.
7016 * @param[in] csrc The cursor to copy from.
7017 * @param[out] cdst The cursor to copy to.
7020 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7024 cdst->mc_txn = csrc->mc_txn;
7025 cdst->mc_dbi = csrc->mc_dbi;
7026 cdst->mc_db = csrc->mc_db;
7027 cdst->mc_dbx = csrc->mc_dbx;
7028 cdst->mc_snum = csrc->mc_snum;
7029 cdst->mc_top = csrc->mc_top;
7030 cdst->mc_flags = csrc->mc_flags;
7032 for (i=0; i<csrc->mc_snum; i++) {
7033 cdst->mc_pg[i] = csrc->mc_pg[i];
7034 cdst->mc_ki[i] = csrc->mc_ki[i];
7038 /** Rebalance the tree after a delete operation.
7039 * @param[in] mc Cursor pointing to the page where rebalancing
7041 * @return 0 on success, non-zero on failure.
7044 mdb_rebalance(MDB_cursor *mc)
7048 unsigned int ptop, minkeys;
7051 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7055 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7056 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7057 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7058 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7059 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7063 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7064 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7067 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7068 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7074 if (mc->mc_snum < 2) {
7075 MDB_page *mp = mc->mc_pg[0];
7077 DPUTS("Can't rebalance a subpage, ignoring");
7080 if (NUMKEYS(mp) == 0) {
7081 DPUTS("tree is completely empty");
7082 mc->mc_db->md_root = P_INVALID;
7083 mc->mc_db->md_depth = 0;
7084 mc->mc_db->md_leaf_pages = 0;
7085 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7088 /* Adjust cursors pointing to mp */
7092 MDB_cursor *m2, *m3;
7093 MDB_dbi dbi = mc->mc_dbi;
7095 if (mc->mc_flags & C_SUB)
7098 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7099 if (mc->mc_flags & C_SUB)
7100 m3 = &m2->mc_xcursor->mx_cursor;
7103 if (m3->mc_snum < mc->mc_snum) continue;
7104 if (m3->mc_pg[0] == mp) {
7110 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7111 DPUTS("collapsing root page!");
7112 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7115 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7116 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7119 mc->mc_db->md_depth--;
7120 mc->mc_db->md_branch_pages--;
7121 mc->mc_ki[0] = mc->mc_ki[1];
7123 /* Adjust other cursors pointing to mp */
7124 MDB_cursor *m2, *m3;
7125 MDB_dbi dbi = mc->mc_dbi;
7127 if (mc->mc_flags & C_SUB)
7130 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7131 if (mc->mc_flags & C_SUB)
7132 m3 = &m2->mc_xcursor->mx_cursor;
7135 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7136 if (m3->mc_pg[0] == mp) {
7137 m3->mc_pg[0] = mc->mc_pg[0];
7140 m3->mc_ki[0] = m3->mc_ki[1];
7145 DPUTS("root page doesn't need rebalancing");
7149 /* The parent (branch page) must have at least 2 pointers,
7150 * otherwise the tree is invalid.
7152 ptop = mc->mc_top-1;
7153 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7155 /* Leaf page fill factor is below the threshold.
7156 * Try to move keys from left or right neighbor, or
7157 * merge with a neighbor page.
7162 mdb_cursor_copy(mc, &mn);
7163 mn.mc_xcursor = NULL;
7165 if (mc->mc_ki[ptop] == 0) {
7166 /* We're the leftmost leaf in our parent.
7168 DPUTS("reading right neighbor");
7170 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7171 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7174 mn.mc_ki[mn.mc_top] = 0;
7175 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7177 /* There is at least one neighbor to the left.
7179 DPUTS("reading left neighbor");
7181 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7182 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7185 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7186 mc->mc_ki[mc->mc_top] = 0;
7189 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7190 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7191 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7193 /* If the neighbor page is above threshold and has enough keys,
7194 * move one key from it. Otherwise we should try to merge them.
7195 * (A branch page must never have less than 2 keys.)
7197 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7198 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7199 return mdb_node_move(&mn, mc);
7201 if (mc->mc_ki[ptop] == 0)
7202 rc = mdb_page_merge(&mn, mc);
7204 rc = mdb_page_merge(mc, &mn);
7205 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7210 /** Complete a delete operation started by #mdb_cursor_del(). */
7212 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7219 mp = mc->mc_pg[mc->mc_top];
7220 ki = mc->mc_ki[mc->mc_top];
7222 /* add overflow pages to free list */
7223 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7227 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7228 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7229 (rc = mdb_ovpage_free(mc, omp)))
7232 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7233 mc->mc_db->md_entries--;
7234 rc = mdb_rebalance(mc);
7235 if (rc != MDB_SUCCESS)
7236 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7239 MDB_dbi dbi = mc->mc_dbi;
7241 mp = mc->mc_pg[mc->mc_top];
7242 nkeys = NUMKEYS(mp);
7244 /* if mc points past last node in page, find next sibling */
7245 if (mc->mc_ki[mc->mc_top] >= nkeys)
7246 mdb_cursor_sibling(mc, 1);
7248 /* Adjust other cursors pointing to mp */
7249 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7252 if (!(m2->mc_flags & C_INITIALIZED))
7254 if (m2->mc_pg[mc->mc_top] == mp) {
7255 if (m2->mc_ki[mc->mc_top] > ki)
7256 m2->mc_ki[mc->mc_top]--;
7257 if (m2->mc_ki[mc->mc_top] >= nkeys)
7258 mdb_cursor_sibling(m2, 1);
7267 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7268 MDB_val *key, MDB_val *data)
7273 MDB_val rdata, *xdata;
7277 assert(key != NULL);
7279 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7281 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7284 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7285 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7287 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7288 return MDB_BAD_VALSIZE;
7291 mdb_cursor_init(&mc, txn, dbi, &mx);
7294 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7295 /* must ignore any data */
7306 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7308 /* let mdb_page_split know about this cursor if needed:
7309 * delete will trigger a rebalance; if it needs to move
7310 * a node from one page to another, it will have to
7311 * update the parent's separator key(s). If the new sepkey
7312 * is larger than the current one, the parent page may
7313 * run out of space, triggering a split. We need this
7314 * cursor to be consistent until the end of the rebalance.
7316 mc.mc_flags |= C_UNTRACK;
7317 mc.mc_next = txn->mt_cursors[dbi];
7318 txn->mt_cursors[dbi] = &mc;
7319 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7320 txn->mt_cursors[dbi] = mc.mc_next;
7325 /** Split a page and insert a new node.
7326 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7327 * The cursor will be updated to point to the actual page and index where
7328 * the node got inserted after the split.
7329 * @param[in] newkey The key for the newly inserted node.
7330 * @param[in] newdata The data for the newly inserted node.
7331 * @param[in] newpgno The page number, if the new node is a branch node.
7332 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7333 * @return 0 on success, non-zero on failure.
7336 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7337 unsigned int nflags)
7340 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7343 unsigned int i, j, split_indx, nkeys, pmax;
7345 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7347 MDB_page *mp, *rp, *pp;
7352 mp = mc->mc_pg[mc->mc_top];
7353 newindx = mc->mc_ki[mc->mc_top];
7355 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7356 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7357 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7359 /* Create a right sibling. */
7360 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7362 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7364 if (mc->mc_snum < 2) {
7365 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7367 /* shift current top to make room for new parent */
7368 mc->mc_pg[1] = mc->mc_pg[0];
7369 mc->mc_ki[1] = mc->mc_ki[0];
7372 mc->mc_db->md_root = pp->mp_pgno;
7373 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7374 mc->mc_db->md_depth++;
7377 /* Add left (implicit) pointer. */
7378 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7379 /* undo the pre-push */
7380 mc->mc_pg[0] = mc->mc_pg[1];
7381 mc->mc_ki[0] = mc->mc_ki[1];
7382 mc->mc_db->md_root = mp->mp_pgno;
7383 mc->mc_db->md_depth--;
7390 ptop = mc->mc_top-1;
7391 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7394 mc->mc_flags |= C_SPLITTING;
7395 mdb_cursor_copy(mc, &mn);
7396 mn.mc_pg[mn.mc_top] = rp;
7397 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7399 if (nflags & MDB_APPEND) {
7400 mn.mc_ki[mn.mc_top] = 0;
7402 split_indx = newindx;
7407 nkeys = NUMKEYS(mp);
7408 split_indx = nkeys / 2;
7409 if (newindx < split_indx)
7415 unsigned int lsize, rsize, ksize;
7416 /* Move half of the keys to the right sibling */
7418 x = mc->mc_ki[mc->mc_top] - split_indx;
7419 ksize = mc->mc_db->md_pad;
7420 split = LEAF2KEY(mp, split_indx, ksize);
7421 rsize = (nkeys - split_indx) * ksize;
7422 lsize = (nkeys - split_indx) * sizeof(indx_t);
7423 mp->mp_lower -= lsize;
7424 rp->mp_lower += lsize;
7425 mp->mp_upper += rsize - lsize;
7426 rp->mp_upper -= rsize - lsize;
7427 sepkey.mv_size = ksize;
7428 if (newindx == split_indx) {
7429 sepkey.mv_data = newkey->mv_data;
7431 sepkey.mv_data = split;
7434 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7435 memcpy(rp->mp_ptrs, split, rsize);
7436 sepkey.mv_data = rp->mp_ptrs;
7437 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7438 memcpy(ins, newkey->mv_data, ksize);
7439 mp->mp_lower += sizeof(indx_t);
7440 mp->mp_upper -= ksize - sizeof(indx_t);
7443 memcpy(rp->mp_ptrs, split, x * ksize);
7444 ins = LEAF2KEY(rp, x, ksize);
7445 memcpy(ins, newkey->mv_data, ksize);
7446 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7447 rp->mp_lower += sizeof(indx_t);
7448 rp->mp_upper -= ksize - sizeof(indx_t);
7449 mc->mc_ki[mc->mc_top] = x;
7450 mc->mc_pg[mc->mc_top] = rp;
7455 /* For leaf pages, check the split point based on what
7456 * fits where, since otherwise mdb_node_add can fail.
7458 * This check is only needed when the data items are
7459 * relatively large, such that being off by one will
7460 * make the difference between success or failure.
7462 * It's also relevant if a page happens to be laid out
7463 * such that one half of its nodes are all "small" and
7464 * the other half of its nodes are "large." If the new
7465 * item is also "large" and falls on the half with
7466 * "large" nodes, it also may not fit.
7469 unsigned int psize, nsize;
7470 /* Maximum free space in an empty page */
7471 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7472 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7473 if ((nkeys < 20) || (nsize > pmax/16)) {
7474 if (newindx <= split_indx) {
7477 for (i=0; i<split_indx; i++) {
7478 node = NODEPTR(mp, i);
7479 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7480 if (F_ISSET(node->mn_flags, F_BIGDATA))
7481 psize += sizeof(pgno_t);
7483 psize += NODEDSZ(node);
7487 split_indx = newindx;
7498 for (i=nkeys-1; i>=split_indx; i--) {
7499 node = NODEPTR(mp, i);
7500 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7501 if (F_ISSET(node->mn_flags, F_BIGDATA))
7502 psize += sizeof(pgno_t);
7504 psize += NODEDSZ(node);
7508 split_indx = newindx;
7519 /* First find the separating key between the split pages.
7520 * The case where newindx == split_indx is ambiguous; the
7521 * new item could go to the new page or stay on the original
7522 * page. If newpos == 1 it goes to the new page.
7524 if (newindx == split_indx && newpos) {
7525 sepkey.mv_size = newkey->mv_size;
7526 sepkey.mv_data = newkey->mv_data;
7528 node = NODEPTR(mp, split_indx);
7529 sepkey.mv_size = node->mn_ksize;
7530 sepkey.mv_data = NODEKEY(node);
7534 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7536 /* Copy separator key to the parent.
7538 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7542 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7545 if (mn.mc_snum == mc->mc_snum) {
7546 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7547 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7548 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7549 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7554 /* Right page might now have changed parent.
7555 * Check if left page also changed parent.
7557 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7558 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7559 for (i=0; i<ptop; i++) {
7560 mc->mc_pg[i] = mn.mc_pg[i];
7561 mc->mc_ki[i] = mn.mc_ki[i];
7563 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7564 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7568 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7571 mc->mc_flags ^= C_SPLITTING;
7572 if (rc != MDB_SUCCESS) {
7575 if (nflags & MDB_APPEND) {
7576 mc->mc_pg[mc->mc_top] = rp;
7577 mc->mc_ki[mc->mc_top] = 0;
7578 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7581 for (i=0; i<mc->mc_top; i++)
7582 mc->mc_ki[i] = mn.mc_ki[i];
7589 /* Move half of the keys to the right sibling. */
7591 /* grab a page to hold a temporary copy */
7592 copy = mdb_page_malloc(mc->mc_txn, 1);
7596 copy->mp_pgno = mp->mp_pgno;
7597 copy->mp_flags = mp->mp_flags;
7598 copy->mp_lower = PAGEHDRSZ;
7599 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7600 mc->mc_pg[mc->mc_top] = copy;
7601 for (i = j = 0; i <= nkeys; j++) {
7602 if (i == split_indx) {
7603 /* Insert in right sibling. */
7604 /* Reset insert index for right sibling. */
7605 if (i != newindx || (newpos ^ ins_new)) {
7607 mc->mc_pg[mc->mc_top] = rp;
7611 if (i == newindx && !ins_new) {
7612 /* Insert the original entry that caused the split. */
7613 rkey.mv_data = newkey->mv_data;
7614 rkey.mv_size = newkey->mv_size;
7623 /* Update index for the new key. */
7624 mc->mc_ki[mc->mc_top] = j;
7625 } else if (i == nkeys) {
7628 node = NODEPTR(mp, i);
7629 rkey.mv_data = NODEKEY(node);
7630 rkey.mv_size = node->mn_ksize;
7632 xdata.mv_data = NODEDATA(node);
7633 xdata.mv_size = NODEDSZ(node);
7636 pgno = NODEPGNO(node);
7637 flags = node->mn_flags;
7642 if (!IS_LEAF(mp) && j == 0) {
7643 /* First branch index doesn't need key data. */
7647 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7651 nkeys = NUMKEYS(copy);
7652 for (i=0; i<nkeys; i++)
7653 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7654 mp->mp_lower = copy->mp_lower;
7655 mp->mp_upper = copy->mp_upper;
7656 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7657 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7659 /* reset back to original page */
7660 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7661 mc->mc_pg[mc->mc_top] = mp;
7662 if (nflags & MDB_RESERVE) {
7663 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7664 if (!(node->mn_flags & F_BIGDATA))
7665 newdata->mv_data = NODEDATA(node);
7669 /* Make sure mc_ki is still valid.
7671 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7672 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7673 for (i=0; i<ptop; i++) {
7674 mc->mc_pg[i] = mn.mc_pg[i];
7675 mc->mc_ki[i] = mn.mc_ki[i];
7677 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7678 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7682 /* return tmp page to freelist */
7683 mdb_page_free(mc->mc_txn->mt_env, copy);
7686 /* Adjust other cursors pointing to mp */
7687 MDB_cursor *m2, *m3;
7688 MDB_dbi dbi = mc->mc_dbi;
7689 int fixup = NUMKEYS(mp);
7691 if (mc->mc_flags & C_SUB)
7694 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7695 if (mc->mc_flags & C_SUB)
7696 m3 = &m2->mc_xcursor->mx_cursor;
7701 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7703 if (m3->mc_flags & C_SPLITTING)
7708 for (k=m3->mc_top; k>=0; k--) {
7709 m3->mc_ki[k+1] = m3->mc_ki[k];
7710 m3->mc_pg[k+1] = m3->mc_pg[k];
7712 if (m3->mc_ki[0] >= split_indx) {
7717 m3->mc_pg[0] = mc->mc_pg[0];
7721 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7722 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7723 m3->mc_ki[mc->mc_top]++;
7724 if (m3->mc_ki[mc->mc_top] >= fixup) {
7725 m3->mc_pg[mc->mc_top] = rp;
7726 m3->mc_ki[mc->mc_top] -= fixup;
7727 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7729 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7730 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7739 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7740 MDB_val *key, MDB_val *data, unsigned int flags)
7745 assert(key != NULL);
7746 assert(data != NULL);
7748 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7751 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7752 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7754 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7755 return MDB_BAD_VALSIZE;
7758 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7761 mdb_cursor_init(&mc, txn, dbi, &mx);
7762 return mdb_cursor_put(&mc, key, data, flags);
7766 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7768 if ((flag & CHANGEABLE) != flag)
7771 env->me_flags |= flag;
7773 env->me_flags &= ~flag;
7778 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7783 *arg = env->me_flags;
7788 mdb_env_get_path(MDB_env *env, const char **arg)
7793 *arg = env->me_path;
7797 /** Common code for #mdb_stat() and #mdb_env_stat().
7798 * @param[in] env the environment to operate in.
7799 * @param[in] db the #MDB_db record containing the stats to return.
7800 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7801 * @return 0, this function always succeeds.
7804 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7806 arg->ms_psize = env->me_psize;
7807 arg->ms_depth = db->md_depth;
7808 arg->ms_branch_pages = db->md_branch_pages;
7809 arg->ms_leaf_pages = db->md_leaf_pages;
7810 arg->ms_overflow_pages = db->md_overflow_pages;
7811 arg->ms_entries = db->md_entries;
7816 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7820 if (env == NULL || arg == NULL)
7823 toggle = mdb_env_pick_meta(env);
7825 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7829 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7833 if (env == NULL || arg == NULL)
7836 toggle = mdb_env_pick_meta(env);
7837 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7838 arg->me_mapsize = env->me_mapsize;
7839 arg->me_maxreaders = env->me_maxreaders;
7841 /* me_numreaders may be zero if this process never used any readers. Use
7842 * the shared numreader count if it exists.
7844 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7846 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7847 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7851 /** Set the default comparison functions for a database.
7852 * Called immediately after a database is opened to set the defaults.
7853 * The user can then override them with #mdb_set_compare() or
7854 * #mdb_set_dupsort().
7855 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7856 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7859 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7861 uint16_t f = txn->mt_dbs[dbi].md_flags;
7863 txn->mt_dbxs[dbi].md_cmp =
7864 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7865 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7867 txn->mt_dbxs[dbi].md_dcmp =
7868 !(f & MDB_DUPSORT) ? 0 :
7869 ((f & MDB_INTEGERDUP)
7870 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7871 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7874 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7879 int rc, dbflag, exact;
7880 unsigned int unused = 0;
7883 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7884 mdb_default_cmp(txn, FREE_DBI);
7887 if ((flags & VALID_FLAGS) != flags)
7889 if (txn->mt_flags & MDB_TXN_ERROR)
7895 if (flags & PERSISTENT_FLAGS) {
7896 uint16_t f2 = flags & PERSISTENT_FLAGS;
7897 /* make sure flag changes get committed */
7898 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7899 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7900 txn->mt_flags |= MDB_TXN_DIRTY;
7903 mdb_default_cmp(txn, MAIN_DBI);
7907 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7908 mdb_default_cmp(txn, MAIN_DBI);
7911 /* Is the DB already open? */
7913 for (i=2; i<txn->mt_numdbs; i++) {
7914 if (!txn->mt_dbxs[i].md_name.mv_size) {
7915 /* Remember this free slot */
7916 if (!unused) unused = i;
7919 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7920 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7926 /* If no free slot and max hit, fail */
7927 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7928 return MDB_DBS_FULL;
7930 /* Cannot mix named databases with some mainDB flags */
7931 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7932 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7934 /* Find the DB info */
7935 dbflag = DB_NEW|DB_VALID;
7938 key.mv_data = (void *)name;
7939 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7940 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7941 if (rc == MDB_SUCCESS) {
7942 /* make sure this is actually a DB */
7943 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7944 if (!(node->mn_flags & F_SUBDATA))
7945 return MDB_INCOMPATIBLE;
7946 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7947 /* Create if requested */
7949 data.mv_size = sizeof(MDB_db);
7950 data.mv_data = &dummy;
7951 memset(&dummy, 0, sizeof(dummy));
7952 dummy.md_root = P_INVALID;
7953 dummy.md_flags = flags & PERSISTENT_FLAGS;
7954 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7958 /* OK, got info, add to table */
7959 if (rc == MDB_SUCCESS) {
7960 unsigned int slot = unused ? unused : txn->mt_numdbs;
7961 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7962 txn->mt_dbxs[slot].md_name.mv_size = len;
7963 txn->mt_dbxs[slot].md_rel = NULL;
7964 txn->mt_dbflags[slot] = dbflag;
7965 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7967 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7968 mdb_default_cmp(txn, slot);
7977 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7979 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7982 if (txn->mt_dbflags[dbi] & DB_STALE) {
7985 /* Stale, must read the DB's root. cursor_init does it for us. */
7986 mdb_cursor_init(&mc, txn, dbi, &mx);
7988 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7991 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7994 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7996 ptr = env->me_dbxs[dbi].md_name.mv_data;
7997 env->me_dbxs[dbi].md_name.mv_data = NULL;
7998 env->me_dbxs[dbi].md_name.mv_size = 0;
7999 env->me_dbflags[dbi] = 0;
8003 int mdb_dbi_flags(MDB_env *env, MDB_dbi dbi, unsigned int *flags)
8005 /* We could return the flags for the FREE_DBI too but what's the point? */
8006 if (dbi < MAIN_DBI || dbi >= env->me_numdbs)
8008 *flags = env->me_dbflags[dbi];
8012 /** Add all the DB's pages to the free list.
8013 * @param[in] mc Cursor on the DB to free.
8014 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8015 * @return 0 on success, non-zero on failure.
8018 mdb_drop0(MDB_cursor *mc, int subs)
8022 rc = mdb_page_search(mc, NULL, 0);
8023 if (rc == MDB_SUCCESS) {
8024 MDB_txn *txn = mc->mc_txn;
8029 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8030 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8033 mdb_cursor_copy(mc, &mx);
8034 while (mc->mc_snum > 0) {
8035 MDB_page *mp = mc->mc_pg[mc->mc_top];
8036 unsigned n = NUMKEYS(mp);
8038 for (i=0; i<n; i++) {
8039 ni = NODEPTR(mp, i);
8040 if (ni->mn_flags & F_BIGDATA) {
8043 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8044 rc = mdb_page_get(txn, pg, &omp, NULL);
8047 assert(IS_OVERFLOW(omp));
8048 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8052 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8053 mdb_xcursor_init1(mc, ni);
8054 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8060 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8062 for (i=0; i<n; i++) {
8064 ni = NODEPTR(mp, i);
8067 mdb_midl_xappend(txn->mt_free_pgs, pg);
8072 mc->mc_ki[mc->mc_top] = i;
8073 rc = mdb_cursor_sibling(mc, 1);
8075 /* no more siblings, go back to beginning
8076 * of previous level.
8080 for (i=1; i<mc->mc_snum; i++) {
8082 mc->mc_pg[i] = mx.mc_pg[i];
8087 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8088 } else if (rc == MDB_NOTFOUND) {
8094 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8096 MDB_cursor *mc, *m2;
8099 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8102 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8105 rc = mdb_cursor_open(txn, dbi, &mc);
8109 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8110 /* Invalidate the dropped DB's cursors */
8111 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8112 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8116 /* Can't delete the main DB */
8117 if (del && dbi > MAIN_DBI) {
8118 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8120 txn->mt_dbflags[dbi] = DB_STALE;
8121 mdb_dbi_close(txn->mt_env, dbi);
8124 /* reset the DB record, mark it dirty */
8125 txn->mt_dbflags[dbi] |= DB_DIRTY;
8126 txn->mt_dbs[dbi].md_depth = 0;
8127 txn->mt_dbs[dbi].md_branch_pages = 0;
8128 txn->mt_dbs[dbi].md_leaf_pages = 0;
8129 txn->mt_dbs[dbi].md_overflow_pages = 0;
8130 txn->mt_dbs[dbi].md_entries = 0;
8131 txn->mt_dbs[dbi].md_root = P_INVALID;
8133 txn->mt_flags |= MDB_TXN_DIRTY;
8136 mdb_cursor_close(mc);
8140 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8142 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8145 txn->mt_dbxs[dbi].md_cmp = cmp;
8149 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8151 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8154 txn->mt_dbxs[dbi].md_dcmp = cmp;
8158 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8160 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8163 txn->mt_dbxs[dbi].md_rel = rel;
8167 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8169 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8172 txn->mt_dbxs[dbi].md_relctx = ctx;
8176 int mdb_env_get_maxkeysize(MDB_env *env)
8178 return MDB_MAXKEYSIZE;
8181 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8183 unsigned int i, rdrs;
8190 if (!env->me_txns) {
8191 return func("(no reader locks)\n", ctx);
8193 rdrs = env->me_txns->mti_numreaders;
8194 mr = env->me_txns->mti_readers;
8195 for (i=0; i<rdrs; i++) {
8200 if (mr[i].mr_txnid == (txnid_t)-1) {
8201 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8203 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8207 func(" pid thread txnid\n", ctx);
8209 rc = func(buf, ctx);
8215 func("(no active readers)\n", ctx);
8220 /* insert pid into list if not already present.
8221 * return -1 if already present.
8223 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8225 /* binary search of pid in list */
8227 unsigned cursor = 1;
8229 unsigned n = ids[0];
8232 unsigned pivot = n >> 1;
8233 cursor = base + pivot + 1;
8234 val = pid - ids[cursor];
8239 } else if ( val > 0 ) {
8244 /* found, so it's a duplicate */
8253 for (n = ids[0]; n > cursor; n--)
8259 int mdb_reader_check(MDB_env *env, int *dead)
8261 unsigned int i, j, rdrs;
8272 rdrs = env->me_txns->mti_numreaders;
8273 pids = malloc((rdrs+1) * sizeof(pid_t));
8277 mr = env->me_txns->mti_readers;
8279 for (i=0; i<rdrs; i++) {
8280 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8282 if (mdb_pid_insert(pids, pid) == 0) {
8283 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8285 /* Recheck, a new process may have reused pid */
8286 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8287 for (j=i; j<rdrs; j++)
8288 if (mr[j].mr_pid == pid) {
8293 UNLOCK_MUTEX_R(env);