2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define MDB_PIDLOCK 0
144 #define pthread_t DWORD
145 #define pthread_mutex_t HANDLE
146 #define pthread_key_t DWORD
147 #define pthread_self() GetCurrentThreadId()
148 #define pthread_key_create(x,y) \
149 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
150 #define pthread_key_delete(x) TlsFree(x)
151 #define pthread_getspecific(x) TlsGetValue(x)
152 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
153 #define pthread_mutex_unlock(x) ReleaseMutex(x)
154 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
155 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
156 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
157 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
158 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
159 #define getpid() GetCurrentProcessId()
160 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
161 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
162 #define ErrCode() GetLastError()
163 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
164 #define close(fd) (CloseHandle(fd) ? 0 : -1)
165 #define munmap(ptr,len) UnmapViewOfFile(ptr)
166 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
167 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
169 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
174 #define Z "z" /**< printf format modifier for size_t */
176 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
177 #define MDB_PIDLOCK 1
179 #ifdef MDB_USE_POSIX_SEM
181 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
182 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
183 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
184 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
187 mdb_sem_wait(sem_t *sem)
190 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
195 /** Lock the reader mutex.
197 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
198 /** Unlock the reader mutex.
200 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
202 /** Lock the writer mutex.
203 * Only a single write transaction is allowed at a time. Other writers
204 * will block waiting for this mutex.
206 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
207 /** Unlock the writer mutex.
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
210 #endif /* MDB_USE_POSIX_SEM */
212 /** Get the error code for the last failed system function.
214 #define ErrCode() errno
216 /** An abstraction for a file handle.
217 * On POSIX systems file handles are small integers. On Windows
218 * they're opaque pointers.
222 /** A value for an invalid file handle.
223 * Mainly used to initialize file variables and signify that they are
226 #define INVALID_HANDLE_VALUE (-1)
228 /** Get the size of a memory page for the system.
229 * This is the basic size that the platform's memory manager uses, and is
230 * fundamental to the use of memory-mapped files.
232 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
235 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
238 #define MNAME_LEN (sizeof(pthread_mutex_t))
244 /** A flag for opening a file and requesting synchronous data writes.
245 * This is only used when writing a meta page. It's not strictly needed;
246 * we could just do a normal write and then immediately perform a flush.
247 * But if this flag is available it saves us an extra system call.
249 * @note If O_DSYNC is undefined but exists in /usr/include,
250 * preferably set some compiler flag to get the definition.
251 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
254 # define MDB_DSYNC O_DSYNC
258 /** Function for flushing the data of a file. Define this to fsync
259 * if fdatasync() is not supported.
261 #ifndef MDB_FDATASYNC
262 # define MDB_FDATASYNC fdatasync
266 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
277 /** A page number in the database.
278 * Note that 64 bit page numbers are overkill, since pages themselves
279 * already represent 12-13 bits of addressable memory, and the OS will
280 * always limit applications to a maximum of 63 bits of address space.
282 * @note In the #MDB_node structure, we only store 48 bits of this value,
283 * which thus limits us to only 60 bits of addressable data.
285 typedef MDB_ID pgno_t;
287 /** A transaction ID.
288 * See struct MDB_txn.mt_txnid for details.
290 typedef MDB_ID txnid_t;
292 /** @defgroup debug Debug Macros
296 /** Enable debug output. Needs variable argument macros (a C99 feature).
297 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
298 * read from and written to the database (used for free space management).
304 static int mdb_debug;
305 static txnid_t mdb_debug_start;
307 /** Print a debug message with printf formatting.
308 * Requires double parenthesis around 2 or more args.
310 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
311 # define DPRINTF0(fmt, ...) \
312 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
314 # define DPRINTF(args) ((void) 0)
316 /** Print a debug string.
317 * The string is printed literally, with no format processing.
319 #define DPUTS(arg) DPRINTF(("%s", arg))
320 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
322 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
325 /** A default memory page size.
326 * The actual size is platform-dependent, but we use this for
327 * boot-strapping. We probably should not be using this any more.
328 * The #GET_PAGESIZE() macro is used to get the actual size.
330 * Note that we don't currently support Huge pages. On Linux,
331 * regular data files cannot use Huge pages, and in general
332 * Huge pages aren't actually pageable. We rely on the OS
333 * demand-pager to read our data and page it out when memory
334 * pressure from other processes is high. So until OSs have
335 * actual paging support for Huge pages, they're not viable.
337 #define MDB_PAGESIZE 4096
339 /** The minimum number of keys required in a database page.
340 * Setting this to a larger value will place a smaller bound on the
341 * maximum size of a data item. Data items larger than this size will
342 * be pushed into overflow pages instead of being stored directly in
343 * the B-tree node. This value used to default to 4. With a page size
344 * of 4096 bytes that meant that any item larger than 1024 bytes would
345 * go into an overflow page. That also meant that on average 2-3KB of
346 * each overflow page was wasted space. The value cannot be lower than
347 * 2 because then there would no longer be a tree structure. With this
348 * value, items larger than 2KB will go into overflow pages, and on
349 * average only 1KB will be wasted.
351 #define MDB_MINKEYS 2
353 /** A stamp that identifies a file as an MDB file.
354 * There's nothing special about this value other than that it is easily
355 * recognizable, and it will reflect any byte order mismatches.
357 #define MDB_MAGIC 0xBEEFC0DE
359 /** The version number for a database's datafile format. */
360 #define MDB_DATA_VERSION 1
361 /** The version number for a database's lockfile format. */
362 #define MDB_LOCK_VERSION 1
364 /** @brief The maximum size of a key in the database.
366 * The library rejects bigger keys, and cannot deal with records
367 * with bigger keys stored by a library with bigger max keysize.
369 * We require that keys all fit onto a regular page. This limit
370 * could be raised a bit further if needed; to something just
371 * under #MDB_PAGESIZE / #MDB_MINKEYS.
373 * Note that data items in an #MDB_DUPSORT database are actually keys
374 * of a subDB, so they're also limited to this size.
376 #ifndef MDB_MAXKEYSIZE
377 #define MDB_MAXKEYSIZE 511
380 /** @brief The maximum size of a data item.
382 * We only store a 32 bit value for node sizes.
384 #define MAXDATASIZE 0xffffffffUL
389 * This is used for printing a hex dump of a key's contents.
391 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
392 /** Display a key in hex.
394 * Invoke a function to display a key in hex.
396 #define DKEY(x) mdb_dkey(x, kbuf)
402 /** An invalid page number.
403 * Mainly used to denote an empty tree.
405 #define P_INVALID (~(pgno_t)0)
407 /** Test if the flags \b f are set in a flag word \b w. */
408 #define F_ISSET(w, f) (((w) & (f)) == (f))
410 /** Used for offsets within a single page.
411 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
414 typedef uint16_t indx_t;
416 /** Default size of memory map.
417 * This is certainly too small for any actual applications. Apps should always set
418 * the size explicitly using #mdb_env_set_mapsize().
420 #define DEFAULT_MAPSIZE 1048576
422 /** @defgroup readers Reader Lock Table
423 * Readers don't acquire any locks for their data access. Instead, they
424 * simply record their transaction ID in the reader table. The reader
425 * mutex is needed just to find an empty slot in the reader table. The
426 * slot's address is saved in thread-specific data so that subsequent read
427 * transactions started by the same thread need no further locking to proceed.
429 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
431 * No reader table is used if the database is on a read-only filesystem, or
432 * if #MDB_NOLOCK is set.
434 * Since the database uses multi-version concurrency control, readers don't
435 * actually need any locking. This table is used to keep track of which
436 * readers are using data from which old transactions, so that we'll know
437 * when a particular old transaction is no longer in use. Old transactions
438 * that have discarded any data pages can then have those pages reclaimed
439 * for use by a later write transaction.
441 * The lock table is constructed such that reader slots are aligned with the
442 * processor's cache line size. Any slot is only ever used by one thread.
443 * This alignment guarantees that there will be no contention or cache
444 * thrashing as threads update their own slot info, and also eliminates
445 * any need for locking when accessing a slot.
447 * A writer thread will scan every slot in the table to determine the oldest
448 * outstanding reader transaction. Any freed pages older than this will be
449 * reclaimed by the writer. The writer doesn't use any locks when scanning
450 * this table. This means that there's no guarantee that the writer will
451 * see the most up-to-date reader info, but that's not required for correct
452 * operation - all we need is to know the upper bound on the oldest reader,
453 * we don't care at all about the newest reader. So the only consequence of
454 * reading stale information here is that old pages might hang around a
455 * while longer before being reclaimed. That's actually good anyway, because
456 * the longer we delay reclaiming old pages, the more likely it is that a
457 * string of contiguous pages can be found after coalescing old pages from
458 * many old transactions together.
461 /** Number of slots in the reader table.
462 * This value was chosen somewhat arbitrarily. 126 readers plus a
463 * couple mutexes fit exactly into 8KB on my development machine.
464 * Applications should set the table size using #mdb_env_set_maxreaders().
466 #define DEFAULT_READERS 126
468 /** The size of a CPU cache line in bytes. We want our lock structures
469 * aligned to this size to avoid false cache line sharing in the
471 * This value works for most CPUs. For Itanium this should be 128.
477 /** The information we store in a single slot of the reader table.
478 * In addition to a transaction ID, we also record the process and
479 * thread ID that owns a slot, so that we can detect stale information,
480 * e.g. threads or processes that went away without cleaning up.
481 * @note We currently don't check for stale records. We simply re-init
482 * the table when we know that we're the only process opening the
485 typedef struct MDB_rxbody {
486 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
487 * Multiple readers that start at the same time will probably have the
488 * same ID here. Again, it's not important to exclude them from
489 * anything; all we need to know is which version of the DB they
490 * started from so we can avoid overwriting any data used in that
491 * particular version.
494 /** The process ID of the process owning this reader txn. */
496 /** The thread ID of the thread owning this txn. */
500 /** The actual reader record, with cacheline padding. */
501 typedef struct MDB_reader {
504 /** shorthand for mrb_txnid */
505 #define mr_txnid mru.mrx.mrb_txnid
506 #define mr_pid mru.mrx.mrb_pid
507 #define mr_tid mru.mrx.mrb_tid
508 /** cache line alignment */
509 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
513 /** The header for the reader table.
514 * The table resides in a memory-mapped file. (This is a different file
515 * than is used for the main database.)
517 * For POSIX the actual mutexes reside in the shared memory of this
518 * mapped file. On Windows, mutexes are named objects allocated by the
519 * kernel; we store the mutex names in this mapped file so that other
520 * processes can grab them. This same approach is also used on
521 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
522 * process-shared POSIX mutexes. For these cases where a named object
523 * is used, the object name is derived from a 64 bit FNV hash of the
524 * environment pathname. As such, naming collisions are extremely
525 * unlikely. If a collision occurs, the results are unpredictable.
527 typedef struct MDB_txbody {
528 /** Stamp identifying this as an MDB file. It must be set
531 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
533 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
534 char mtb_rmname[MNAME_LEN];
536 /** Mutex protecting access to this table.
537 * This is the reader lock that #LOCK_MUTEX_R acquires.
539 pthread_mutex_t mtb_mutex;
541 /** The ID of the last transaction committed to the database.
542 * This is recorded here only for convenience; the value can always
543 * be determined by reading the main database meta pages.
546 /** The number of slots that have been used in the reader table.
547 * This always records the maximum count, it is not decremented
548 * when readers release their slots.
550 unsigned mtb_numreaders;
553 /** The actual reader table definition. */
554 typedef struct MDB_txninfo {
557 #define mti_magic mt1.mtb.mtb_magic
558 #define mti_format mt1.mtb.mtb_format
559 #define mti_mutex mt1.mtb.mtb_mutex
560 #define mti_rmname mt1.mtb.mtb_rmname
561 #define mti_txnid mt1.mtb.mtb_txnid
562 #define mti_numreaders mt1.mtb.mtb_numreaders
563 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
566 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
567 char mt2_wmname[MNAME_LEN];
568 #define mti_wmname mt2.mt2_wmname
570 pthread_mutex_t mt2_wmutex;
571 #define mti_wmutex mt2.mt2_wmutex
573 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
575 MDB_reader mti_readers[1];
578 /** Lockfile format signature: version, features and field layout */
579 #define MDB_LOCK_FORMAT \
581 ((MDB_LOCK_VERSION) \
582 /* Flags which describe functionality */ \
583 + (((MDB_PIDLOCK) != 0) << 16)))
586 /** Common header for all page types.
587 * Overflow records occupy a number of contiguous pages with no
588 * headers on any page after the first.
590 typedef struct MDB_page {
591 #define mp_pgno mp_p.p_pgno
592 #define mp_next mp_p.p_next
594 pgno_t p_pgno; /**< page number */
595 void * p_next; /**< for in-memory list of freed structs */
598 /** @defgroup mdb_page Page Flags
600 * Flags for the page headers.
603 #define P_BRANCH 0x01 /**< branch page */
604 #define P_LEAF 0x02 /**< leaf page */
605 #define P_OVERFLOW 0x04 /**< overflow page */
606 #define P_META 0x08 /**< meta page */
607 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
608 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
609 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
610 #define P_KEEP 0x8000 /**< leave this page alone during spill */
612 uint16_t mp_flags; /**< @ref mdb_page */
613 #define mp_lower mp_pb.pb.pb_lower
614 #define mp_upper mp_pb.pb.pb_upper
615 #define mp_pages mp_pb.pb_pages
618 indx_t pb_lower; /**< lower bound of free space */
619 indx_t pb_upper; /**< upper bound of free space */
621 uint32_t pb_pages; /**< number of overflow pages */
623 indx_t mp_ptrs[1]; /**< dynamic size */
626 /** Size of the page header, excluding dynamic data at the end */
627 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
629 /** Address of first usable data byte in a page, after the header */
630 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
632 /** Number of nodes on a page */
633 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
635 /** The amount of space remaining in the page */
636 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
638 /** The percentage of space used in the page, in tenths of a percent. */
639 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
640 ((env)->me_psize - PAGEHDRSZ))
641 /** The minimum page fill factor, in tenths of a percent.
642 * Pages emptier than this are candidates for merging.
644 #define FILL_THRESHOLD 250
646 /** Test if a page is a leaf page */
647 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
648 /** Test if a page is a LEAF2 page */
649 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
650 /** Test if a page is a branch page */
651 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
652 /** Test if a page is an overflow page */
653 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
654 /** Test if a page is a sub page */
655 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
657 /** The number of overflow pages needed to store the given size. */
658 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
660 /** Header for a single key/data pair within a page.
661 * We guarantee 2-byte alignment for nodes.
663 typedef struct MDB_node {
664 /** lo and hi are used for data size on leaf nodes and for
665 * child pgno on branch nodes. On 64 bit platforms, flags
666 * is also used for pgno. (Branch nodes have no flags).
667 * They are in host byte order in case that lets some
668 * accesses be optimized into a 32-bit word access.
670 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
671 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
672 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
673 /** @defgroup mdb_node Node Flags
675 * Flags for node headers.
678 #define F_BIGDATA 0x01 /**< data put on overflow page */
679 #define F_SUBDATA 0x02 /**< data is a sub-database */
680 #define F_DUPDATA 0x04 /**< data has duplicates */
682 /** valid flags for #mdb_node_add() */
683 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
686 unsigned short mn_flags; /**< @ref mdb_node */
687 unsigned short mn_ksize; /**< key size */
688 char mn_data[1]; /**< key and data are appended here */
691 /** Size of the node header, excluding dynamic data at the end */
692 #define NODESIZE offsetof(MDB_node, mn_data)
694 /** Bit position of top word in page number, for shifting mn_flags */
695 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
697 /** Size of a node in a branch page with a given key.
698 * This is just the node header plus the key, there is no data.
700 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
702 /** Size of a node in a leaf page with a given key and data.
703 * This is node header plus key plus data size.
705 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
707 /** Address of node \b i in page \b p */
708 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
710 /** Address of the key for the node */
711 #define NODEKEY(node) (void *)((node)->mn_data)
713 /** Address of the data for a node */
714 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
716 /** Get the page number pointed to by a branch node */
717 #define NODEPGNO(node) \
718 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
719 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
720 /** Set the page number in a branch node */
721 #define SETPGNO(node,pgno) do { \
722 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
723 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
725 /** Get the size of the data in a leaf node */
726 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
727 /** Set the size of the data for a leaf node */
728 #define SETDSZ(node,size) do { \
729 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
730 /** The size of a key in a node */
731 #define NODEKSZ(node) ((node)->mn_ksize)
733 /** Copy a page number from src to dst */
735 #define COPY_PGNO(dst,src) dst = src
737 #if SIZE_MAX > 4294967295UL
738 #define COPY_PGNO(dst,src) do { \
739 unsigned short *s, *d; \
740 s = (unsigned short *)&(src); \
741 d = (unsigned short *)&(dst); \
748 #define COPY_PGNO(dst,src) do { \
749 unsigned short *s, *d; \
750 s = (unsigned short *)&(src); \
751 d = (unsigned short *)&(dst); \
757 /** The address of a key in a LEAF2 page.
758 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
759 * There are no node headers, keys are stored contiguously.
761 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
763 /** Set the \b node's key into \b keyptr, if requested. */
764 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
765 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
767 /** Set the \b node's key into \b key. */
768 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
770 /** Information about a single database in the environment. */
771 typedef struct MDB_db {
772 uint32_t md_pad; /**< also ksize for LEAF2 pages */
773 uint16_t md_flags; /**< @ref mdb_dbi_open */
774 uint16_t md_depth; /**< depth of this tree */
775 pgno_t md_branch_pages; /**< number of internal pages */
776 pgno_t md_leaf_pages; /**< number of leaf pages */
777 pgno_t md_overflow_pages; /**< number of overflow pages */
778 size_t md_entries; /**< number of data items */
779 pgno_t md_root; /**< the root page of this tree */
782 /** mdb_dbi_open flags */
783 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
784 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
785 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
786 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
788 /** Handle for the DB used to track free pages. */
790 /** Handle for the default DB. */
793 /** Meta page content.
794 * A meta page is the start point for accessing a database snapshot.
795 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
797 typedef struct MDB_meta {
798 /** Stamp identifying this as an MDB file. It must be set
801 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
803 void *mm_address; /**< address for fixed mapping */
804 size_t mm_mapsize; /**< size of mmap region */
805 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
806 /** The size of pages used in this DB */
807 #define mm_psize mm_dbs[0].md_pad
808 /** Any persistent environment flags. @ref mdb_env */
809 #define mm_flags mm_dbs[0].md_flags
810 pgno_t mm_last_pg; /**< last used page in file */
811 txnid_t mm_txnid; /**< txnid that committed this page */
814 /** Buffer for a stack-allocated dirty page.
815 * The members define size and alignment, and silence type
816 * aliasing warnings. They are not used directly; that could
817 * mean incorrectly using several union members in parallel.
819 typedef union MDB_pagebuf {
820 char mb_raw[MDB_PAGESIZE];
823 char mm_pad[PAGEHDRSZ];
828 /** Auxiliary DB info.
829 * The information here is mostly static/read-only. There is
830 * only a single copy of this record in the environment.
832 typedef struct MDB_dbx {
833 MDB_val md_name; /**< name of the database */
834 MDB_cmp_func *md_cmp; /**< function for comparing keys */
835 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
836 MDB_rel_func *md_rel; /**< user relocate function */
837 void *md_relctx; /**< user-provided context for md_rel */
840 /** A database transaction.
841 * Every operation requires a transaction handle.
844 MDB_txn *mt_parent; /**< parent of a nested txn */
845 MDB_txn *mt_child; /**< nested txn under this txn */
846 pgno_t mt_next_pgno; /**< next unallocated page */
847 /** The ID of this transaction. IDs are integers incrementing from 1.
848 * Only committed write transactions increment the ID. If a transaction
849 * aborts, the ID may be re-used by the next writer.
852 MDB_env *mt_env; /**< the DB environment */
853 /** The list of pages that became unused during this transaction.
856 /** The sorted list of dirty pages we temporarily wrote to disk
857 * because the dirty list was full. page numbers in here are
858 * shifted left by 1, deleted slots have the LSB set.
860 MDB_IDL mt_spill_pgs;
862 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
864 /** For read txns: This thread/txn's reader table slot, or NULL. */
867 /** Array of records for each DB known in the environment. */
869 /** Array of MDB_db records for each known DB */
871 /** @defgroup mt_dbflag Transaction DB Flags
875 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
876 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
877 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
878 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
880 /** In write txns, array of cursors for each DB */
881 MDB_cursor **mt_cursors;
882 /** Array of flags for each DB */
883 unsigned char *mt_dbflags;
884 /** Number of DB records in use. This number only ever increments;
885 * we don't decrement it when individual DB handles are closed.
889 /** @defgroup mdb_txn Transaction Flags
893 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
894 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
895 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
896 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
898 unsigned int mt_flags; /**< @ref mdb_txn */
899 /** dirty_list room: Array size - #dirty pages visible to this txn.
900 * Includes ancestor txns' dirty pages not hidden by other txns'
901 * dirty/spilled pages. Thus commit(nested txn) has room to merge
902 * dirty_list into mt_parent after freeing hidden mt_parent pages.
904 unsigned int mt_dirty_room;
907 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
908 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
909 * raise this on a 64 bit machine.
911 #define CURSOR_STACK 32
915 /** Cursors are used for all DB operations.
916 * A cursor holds a path of (page pointer, key index) from the DB
917 * root to a position in the DB, plus other state. #MDB_DUPSORT
918 * cursors include an xcursor to the current data item. Write txns
919 * track their cursors and keep them up to date when data moves.
920 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
921 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
924 /** Next cursor on this DB in this txn */
926 /** Backup of the original cursor if this cursor is a shadow */
927 MDB_cursor *mc_backup;
928 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
929 struct MDB_xcursor *mc_xcursor;
930 /** The transaction that owns this cursor */
932 /** The database handle this cursor operates on */
934 /** The database record for this cursor */
936 /** The database auxiliary record for this cursor */
938 /** The @ref mt_dbflag for this database */
939 unsigned char *mc_dbflag;
940 unsigned short mc_snum; /**< number of pushed pages */
941 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
942 /** @defgroup mdb_cursor Cursor Flags
944 * Cursor state flags.
947 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
948 #define C_EOF 0x02 /**< No more data */
949 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
950 #define C_DEL 0x08 /**< last op was a cursor_del */
951 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
952 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
954 unsigned int mc_flags; /**< @ref mdb_cursor */
955 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
956 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
959 /** Context for sorted-dup records.
960 * We could have gone to a fully recursive design, with arbitrarily
961 * deep nesting of sub-databases. But for now we only handle these
962 * levels - main DB, optional sub-DB, sorted-duplicate DB.
964 typedef struct MDB_xcursor {
965 /** A sub-cursor for traversing the Dup DB */
966 MDB_cursor mx_cursor;
967 /** The database record for this Dup DB */
969 /** The auxiliary DB record for this Dup DB */
971 /** The @ref mt_dbflag for this Dup DB */
972 unsigned char mx_dbflag;
975 /** State of FreeDB old pages, stored in the MDB_env */
976 typedef struct MDB_pgstate {
977 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
978 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
981 /** The database environment. */
983 HANDLE me_fd; /**< The main data file */
984 HANDLE me_lfd; /**< The lock file */
985 HANDLE me_mfd; /**< just for writing the meta pages */
986 /** Failed to update the meta page. Probably an I/O error. */
987 #define MDB_FATAL_ERROR 0x80000000U
988 /** Some fields are initialized. */
989 #define MDB_ENV_ACTIVE 0x20000000U
990 /** me_txkey is set */
991 #define MDB_ENV_TXKEY 0x10000000U
992 /** Have liveness lock in reader table */
993 #define MDB_LIVE_READER 0x08000000U
994 uint32_t me_flags; /**< @ref mdb_env */
995 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
996 unsigned int me_maxreaders; /**< size of the reader table */
997 unsigned int me_numreaders; /**< max numreaders set by this env */
998 MDB_dbi me_numdbs; /**< number of DBs opened */
999 MDB_dbi me_maxdbs; /**< size of the DB table */
1000 pid_t me_pid; /**< process ID of this env */
1001 char *me_path; /**< path to the DB files */
1002 char *me_map; /**< the memory map of the data file */
1003 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1004 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1005 MDB_txn *me_txn; /**< current write transaction */
1006 size_t me_mapsize; /**< size of the data memory map */
1007 off_t me_size; /**< current file size */
1008 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1009 MDB_dbx *me_dbxs; /**< array of static DB info */
1010 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1011 pthread_key_t me_txkey; /**< thread-key for readers */
1012 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1013 # define me_pglast me_pgstate.mf_pglast
1014 # define me_pghead me_pgstate.mf_pghead
1015 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1016 /** IDL of pages that became unused in a write txn */
1017 MDB_IDL me_free_pgs;
1018 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1019 MDB_ID2L me_dirty_list;
1020 /** Max number of freelist items that can fit in a single overflow page */
1022 /** Max size of a node on a page */
1023 unsigned int me_nodemax;
1025 int me_pidquery; /**< Used in OpenProcess */
1026 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1028 #elif defined(MDB_USE_POSIX_SEM)
1029 sem_t *me_rmutex; /* Shared mutexes are not supported */
1034 /** Nested transaction */
1035 typedef struct MDB_ntxn {
1036 MDB_txn mnt_txn; /**< the transaction */
1037 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1040 /** max number of pages to commit in one writev() call */
1041 #define MDB_COMMIT_PAGES 64
1042 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1043 #undef MDB_COMMIT_PAGES
1044 #define MDB_COMMIT_PAGES IOV_MAX
1047 /* max bytes to write in one call */
1048 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1050 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1051 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1052 static int mdb_page_touch(MDB_cursor *mc);
1054 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1055 static int mdb_page_search_root(MDB_cursor *mc,
1056 MDB_val *key, int modify);
1057 #define MDB_PS_MODIFY 1
1058 #define MDB_PS_ROOTONLY 2
1059 #define MDB_PS_FIRST 4
1060 #define MDB_PS_LAST 8
1061 static int mdb_page_search(MDB_cursor *mc,
1062 MDB_val *key, int flags);
1063 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1065 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1066 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1067 pgno_t newpgno, unsigned int nflags);
1069 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1070 static int mdb_env_pick_meta(const MDB_env *env);
1071 static int mdb_env_write_meta(MDB_txn *txn);
1072 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1073 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1075 static void mdb_env_close0(MDB_env *env, int excl);
1077 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1078 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1079 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1080 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1081 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1082 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1083 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1084 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1085 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1087 static int mdb_rebalance(MDB_cursor *mc);
1088 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1090 static void mdb_cursor_pop(MDB_cursor *mc);
1091 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1093 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1094 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1095 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1096 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1097 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1099 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1100 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1102 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1103 static void mdb_xcursor_init0(MDB_cursor *mc);
1104 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1106 static int mdb_drop0(MDB_cursor *mc, int subs);
1107 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1110 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1114 static SECURITY_DESCRIPTOR mdb_null_sd;
1115 static SECURITY_ATTRIBUTES mdb_all_sa;
1116 static int mdb_sec_inited;
1119 /** Return the library version info. */
1121 mdb_version(int *major, int *minor, int *patch)
1123 if (major) *major = MDB_VERSION_MAJOR;
1124 if (minor) *minor = MDB_VERSION_MINOR;
1125 if (patch) *patch = MDB_VERSION_PATCH;
1126 return MDB_VERSION_STRING;
1129 /** Table of descriptions for MDB @ref errors */
1130 static char *const mdb_errstr[] = {
1131 "MDB_KEYEXIST: Key/data pair already exists",
1132 "MDB_NOTFOUND: No matching key/data pair found",
1133 "MDB_PAGE_NOTFOUND: Requested page not found",
1134 "MDB_CORRUPTED: Located page was wrong type",
1135 "MDB_PANIC: Update of meta page failed",
1136 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1137 "MDB_INVALID: File is not an MDB file",
1138 "MDB_MAP_FULL: Environment mapsize limit reached",
1139 "MDB_DBS_FULL: Environment maxdbs limit reached",
1140 "MDB_READERS_FULL: Environment maxreaders limit reached",
1141 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1142 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1143 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1144 "MDB_PAGE_FULL: Internal error - page has no more space",
1145 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1146 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1147 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1148 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1149 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1153 mdb_strerror(int err)
1157 return ("Successful return: 0");
1159 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1160 i = err - MDB_KEYEXIST;
1161 return mdb_errstr[i];
1164 return strerror(err);
1168 /** Display a key in hexadecimal and return the address of the result.
1169 * @param[in] key the key to display
1170 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1171 * @return The key in hexadecimal form.
1174 mdb_dkey(MDB_val *key, char *buf)
1177 unsigned char *c = key->mv_data;
1183 if (key->mv_size > MDB_MAXKEYSIZE)
1184 return "MDB_MAXKEYSIZE";
1185 /* may want to make this a dynamic check: if the key is mostly
1186 * printable characters, print it as-is instead of converting to hex.
1190 for (i=0; i<key->mv_size; i++)
1191 ptr += sprintf(ptr, "%02x", *c++);
1193 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1198 /** Display all the keys in the page. */
1200 mdb_page_list(MDB_page *mp)
1203 unsigned int i, nkeys, nsize;
1207 nkeys = NUMKEYS(mp);
1208 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1209 for (i=0; i<nkeys; i++) {
1210 node = NODEPTR(mp, i);
1211 key.mv_size = node->mn_ksize;
1212 key.mv_data = node->mn_data;
1213 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1214 if (IS_BRANCH(mp)) {
1215 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1218 if (F_ISSET(node->mn_flags, F_BIGDATA))
1219 nsize += sizeof(pgno_t);
1221 nsize += NODEDSZ(node);
1222 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1228 mdb_cursor_chk(MDB_cursor *mc)
1234 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1235 for (i=0; i<mc->mc_top; i++) {
1237 node = NODEPTR(mp, mc->mc_ki[i]);
1238 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1241 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1247 /** Count all the pages in each DB and in the freelist
1248 * and make sure it matches the actual number of pages
1251 static void mdb_audit(MDB_txn *txn)
1255 MDB_ID freecount, count;
1260 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1261 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1262 freecount += *(MDB_ID *)data.mv_data;
1265 for (i = 0; i<txn->mt_numdbs; i++) {
1267 mdb_cursor_init(&mc, txn, i, &mx);
1268 if (txn->mt_dbs[i].md_root == P_INVALID)
1270 count += txn->mt_dbs[i].md_branch_pages +
1271 txn->mt_dbs[i].md_leaf_pages +
1272 txn->mt_dbs[i].md_overflow_pages;
1273 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1274 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1278 mp = mc.mc_pg[mc.mc_top];
1279 for (j=0; j<NUMKEYS(mp); j++) {
1280 MDB_node *leaf = NODEPTR(mp, j);
1281 if (leaf->mn_flags & F_SUBDATA) {
1283 memcpy(&db, NODEDATA(leaf), sizeof(db));
1284 count += db.md_branch_pages + db.md_leaf_pages +
1285 db.md_overflow_pages;
1289 while (mdb_cursor_sibling(&mc, 1) == 0);
1292 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1293 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1294 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1300 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1302 return txn->mt_dbxs[dbi].md_cmp(a, b);
1306 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1308 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1311 /** Allocate memory for a page.
1312 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1315 mdb_page_malloc(MDB_txn *txn, unsigned num)
1317 MDB_env *env = txn->mt_env;
1318 MDB_page *ret = env->me_dpages;
1319 size_t sz = env->me_psize;
1322 VGMEMP_ALLOC(env, ret, sz);
1323 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1324 env->me_dpages = ret->mp_next;
1330 if ((ret = malloc(sz)) != NULL) {
1331 VGMEMP_ALLOC(env, ret, sz);
1336 /** Free a single page.
1337 * Saves single pages to a list, for future reuse.
1338 * (This is not used for multi-page overflow pages.)
1341 mdb_page_free(MDB_env *env, MDB_page *mp)
1343 mp->mp_next = env->me_dpages;
1344 VGMEMP_FREE(env, mp);
1345 env->me_dpages = mp;
1348 /** Free a dirty page */
1350 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1352 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1353 mdb_page_free(env, dp);
1355 /* large pages just get freed directly */
1356 VGMEMP_FREE(env, dp);
1361 /** Return all dirty pages to dpage list */
1363 mdb_dlist_free(MDB_txn *txn)
1365 MDB_env *env = txn->mt_env;
1366 MDB_ID2L dl = txn->mt_u.dirty_list;
1367 unsigned i, n = dl[0].mid;
1369 for (i = 1; i <= n; i++) {
1370 mdb_dpage_free(env, dl[i].mptr);
1375 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1376 * @param[in] mc A cursor handle for the current operation.
1377 * @param[in] pflags Flags of the pages to update:
1378 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1379 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1380 * @return 0 on success, non-zero on failure.
1383 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1385 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1386 MDB_txn *txn = mc->mc_txn;
1392 int rc = MDB_SUCCESS, level;
1394 /* Mark pages seen by cursors */
1395 if (mc->mc_flags & C_UNTRACK)
1396 mc = NULL; /* will find mc in mt_cursors */
1397 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1398 for (; mc; mc=mc->mc_next) {
1399 if (!(mc->mc_flags & C_INITIALIZED))
1401 for (m3 = mc;; m3 = &mx->mx_cursor) {
1403 for (j=0; j<m3->mc_snum; j++) {
1405 if ((mp->mp_flags & Mask) == pflags)
1406 mp->mp_flags ^= P_KEEP;
1408 mx = m3->mc_xcursor;
1409 /* Proceed to mx if it is at a sub-database */
1410 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1412 if (! (mp && (mp->mp_flags & P_LEAF)))
1414 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1415 if (!(leaf->mn_flags & F_SUBDATA))
1424 /* Mark dirty root pages */
1425 for (i=0; i<txn->mt_numdbs; i++) {
1426 if (txn->mt_dbflags[i] & DB_DIRTY) {
1427 pgno_t pgno = txn->mt_dbs[i].md_root;
1428 if (pgno == P_INVALID)
1430 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1432 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1433 dp->mp_flags ^= P_KEEP;
1441 static int mdb_page_flush(MDB_txn *txn, int keep);
1443 /** Spill pages from the dirty list back to disk.
1444 * This is intended to prevent running into #MDB_TXN_FULL situations,
1445 * but note that they may still occur in a few cases:
1446 * 1) our estimate of the txn size could be too small. Currently this
1447 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1448 * 2) child txns may run out of space if their parents dirtied a
1449 * lot of pages and never spilled them. TODO: we probably should do
1450 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1451 * the parent's dirty_room is below a given threshold.
1453 * Otherwise, if not using nested txns, it is expected that apps will
1454 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1455 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1456 * If the txn never references them again, they can be left alone.
1457 * If the txn only reads them, they can be used without any fuss.
1458 * If the txn writes them again, they can be dirtied immediately without
1459 * going thru all of the work of #mdb_page_touch(). Such references are
1460 * handled by #mdb_page_unspill().
1462 * Also note, we never spill DB root pages, nor pages of active cursors,
1463 * because we'll need these back again soon anyway. And in nested txns,
1464 * we can't spill a page in a child txn if it was already spilled in a
1465 * parent txn. That would alter the parent txns' data even though
1466 * the child hasn't committed yet, and we'd have no way to undo it if
1467 * the child aborted.
1469 * @param[in] m0 cursor A cursor handle identifying the transaction and
1470 * database for which we are checking space.
1471 * @param[in] key For a put operation, the key being stored.
1472 * @param[in] data For a put operation, the data being stored.
1473 * @return 0 on success, non-zero on failure.
1476 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1478 MDB_txn *txn = m0->mc_txn;
1480 MDB_ID2L dl = txn->mt_u.dirty_list;
1481 unsigned int i, j, need;
1484 if (m0->mc_flags & C_SUB)
1487 /* Estimate how much space this op will take */
1488 i = m0->mc_db->md_depth;
1489 /* Named DBs also dirty the main DB */
1490 if (m0->mc_dbi > MAIN_DBI)
1491 i += txn->mt_dbs[MAIN_DBI].md_depth;
1492 /* For puts, roughly factor in the key+data size */
1494 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1495 i += i; /* double it for good measure */
1498 if (txn->mt_dirty_room > i)
1501 if (!txn->mt_spill_pgs) {
1502 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1503 if (!txn->mt_spill_pgs)
1506 /* purge deleted slots */
1507 MDB_IDL sl = txn->mt_spill_pgs;
1508 unsigned int num = sl[0];
1510 for (i=1; i<=num; i++) {
1517 /* Preserve pages which may soon be dirtied again */
1518 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1521 /* Less aggressive spill - we originally spilled the entire dirty list,
1522 * with a few exceptions for cursor pages and DB root pages. But this
1523 * turns out to be a lot of wasted effort because in a large txn many
1524 * of those pages will need to be used again. So now we spill only 1/8th
1525 * of the dirty pages. Testing revealed this to be a good tradeoff,
1526 * better than 1/2, 1/4, or 1/10.
1528 if (need < MDB_IDL_UM_MAX / 8)
1529 need = MDB_IDL_UM_MAX / 8;
1531 /* Save the page IDs of all the pages we're flushing */
1532 /* flush from the tail forward, this saves a lot of shifting later on. */
1533 for (i=dl[0].mid; i && need; i--) {
1534 MDB_ID pn = dl[i].mid << 1;
1536 if (dp->mp_flags & P_KEEP)
1538 /* Can't spill twice, make sure it's not already in a parent's
1541 if (txn->mt_parent) {
1543 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1544 if (tx2->mt_spill_pgs) {
1545 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1546 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1547 dp->mp_flags |= P_KEEP;
1555 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1559 mdb_midl_sort(txn->mt_spill_pgs);
1561 /* Flush the spilled part of dirty list */
1562 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1565 /* Reset any dirty pages we kept that page_flush didn't see */
1566 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1569 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1573 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1575 mdb_find_oldest(MDB_txn *txn)
1578 txnid_t mr, oldest = txn->mt_txnid - 1;
1579 if (txn->mt_env->me_txns) {
1580 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1581 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1592 /** Add a page to the txn's dirty list */
1594 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1597 int (*insert)(MDB_ID2L, MDB_ID2 *);
1599 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1600 insert = mdb_mid2l_append;
1602 insert = mdb_mid2l_insert;
1604 mid.mid = mp->mp_pgno;
1606 insert(txn->mt_u.dirty_list, &mid);
1607 txn->mt_dirty_room--;
1610 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1611 * me_pghead and mt_next_pgno.
1613 * If there are free pages available from older transactions, they
1614 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1615 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1616 * and move me_pglast to say which records were consumed. Only this
1617 * function can create me_pghead and move me_pglast/mt_next_pgno.
1618 * @param[in] mc cursor A cursor handle identifying the transaction and
1619 * database for which we are allocating.
1620 * @param[in] num the number of pages to allocate.
1621 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1622 * will always be satisfied by a single contiguous chunk of memory.
1623 * @return 0 on success, non-zero on failure.
1626 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1628 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1629 /* Get at most <Max_retries> more freeDB records once me_pghead
1630 * has enough pages. If not enough, use new pages from the map.
1631 * If <Paranoid> and mc is updating the freeDB, only get new
1632 * records if me_pghead is empty. Then the freelist cannot play
1633 * catch-up with itself by growing while trying to save it.
1635 enum { Paranoid = 1, Max_retries = 500 };
1637 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1639 int rc, n2 = num-1, retry = Max_retries;
1640 MDB_txn *txn = mc->mc_txn;
1641 MDB_env *env = txn->mt_env;
1642 pgno_t pgno, *mop = env->me_pghead;
1643 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1645 txnid_t oldest = 0, last;
1651 /* If our dirty list is already full, we can't do anything */
1652 if (txn->mt_dirty_room == 0)
1653 return MDB_TXN_FULL;
1655 for (op = MDB_FIRST;; op = MDB_NEXT) {
1658 pgno_t *idl, old_id, new_id;
1660 /* Seek a big enough contiguous page range. Prefer
1661 * pages at the tail, just truncating the list.
1663 if (mop_len >= (unsigned)num) {
1667 if (mop[i-n2] == pgno+n2)
1669 } while (--i >= (unsigned)num);
1670 if (Max_retries < INT_MAX && --retry < 0)
1674 if (op == MDB_FIRST) { /* 1st iteration */
1675 /* Prepare to fetch more and coalesce */
1676 oldest = mdb_find_oldest(txn);
1677 last = env->me_pglast;
1678 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1681 key.mv_data = &last; /* will look up last+1 */
1682 key.mv_size = sizeof(last);
1684 if (Paranoid && mc->mc_dbi == FREE_DBI)
1687 if (Paranoid && retry < 0 && mop_len)
1691 /* Do not fetch more if the record will be too recent */
1694 rc = mdb_cursor_get(&m2, &key, NULL, op);
1696 if (rc == MDB_NOTFOUND)
1700 last = *(txnid_t*)key.mv_data;
1703 np = m2.mc_pg[m2.mc_top];
1704 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1705 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1708 idl = (MDB_ID *) data.mv_data;
1711 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1714 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1716 mop = env->me_pghead;
1718 env->me_pglast = last;
1720 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1721 last, txn->mt_dbs[FREE_DBI].md_root, i));
1723 DPRINTF(("IDL %"Z"u", idl[k]));
1725 /* Merge in descending sorted order */
1728 mop[0] = (pgno_t)-1;
1732 for (; old_id < new_id; old_id = mop[--j])
1739 /* Use new pages from the map when nothing suitable in the freeDB */
1741 pgno = txn->mt_next_pgno;
1742 if (pgno + num >= env->me_maxpg) {
1743 DPUTS("DB size maxed out");
1744 return MDB_MAP_FULL;
1748 if (env->me_flags & MDB_WRITEMAP) {
1749 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1751 if (!(np = mdb_page_malloc(txn, num)))
1755 mop[0] = mop_len -= num;
1756 /* Move any stragglers down */
1757 for (j = i-num; j < mop_len; )
1758 mop[++j] = mop[++i];
1760 txn->mt_next_pgno = pgno + num;
1763 mdb_page_dirty(txn, np);
1769 /** Copy the used portions of a non-overflow page.
1770 * @param[in] dst page to copy into
1771 * @param[in] src page to copy from
1772 * @param[in] psize size of a page
1775 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1777 enum { Align = sizeof(pgno_t) };
1778 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1780 /* If page isn't full, just copy the used portion. Adjust
1781 * alignment so memcpy may copy words instead of bytes.
1783 if ((unused &= -Align) && !IS_LEAF2(src)) {
1785 memcpy(dst, src, (lower + (Align-1)) & -Align);
1786 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1789 memcpy(dst, src, psize - unused);
1793 /** Pull a page off the txn's spill list, if present.
1794 * If a page being referenced was spilled to disk in this txn, bring
1795 * it back and make it dirty/writable again.
1796 * @param[in] txn the transaction handle.
1797 * @param[in] mp the page being referenced.
1798 * @param[out] ret the writable page, if any. ret is unchanged if
1799 * mp wasn't spilled.
1802 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1804 MDB_env *env = txn->mt_env;
1807 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1809 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1810 if (!tx2->mt_spill_pgs)
1812 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1813 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1816 if (txn->mt_dirty_room == 0)
1817 return MDB_TXN_FULL;
1818 if (IS_OVERFLOW(mp))
1822 if (env->me_flags & MDB_WRITEMAP) {
1825 np = mdb_page_malloc(txn, num);
1829 memcpy(np, mp, num * env->me_psize);
1831 mdb_page_copy(np, mp, env->me_psize);
1834 /* If in current txn, this page is no longer spilled.
1835 * If it happens to be the last page, truncate the spill list.
1836 * Otherwise mark it as deleted by setting the LSB.
1838 if (x == txn->mt_spill_pgs[0])
1839 txn->mt_spill_pgs[0]--;
1841 txn->mt_spill_pgs[x] |= 1;
1842 } /* otherwise, if belonging to a parent txn, the
1843 * page remains spilled until child commits
1846 mdb_page_dirty(txn, np);
1847 np->mp_flags |= P_DIRTY;
1855 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1856 * @param[in] mc cursor pointing to the page to be touched
1857 * @return 0 on success, non-zero on failure.
1860 mdb_page_touch(MDB_cursor *mc)
1862 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1863 MDB_txn *txn = mc->mc_txn;
1864 MDB_cursor *m2, *m3;
1868 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1869 if (txn->mt_flags & MDB_TXN_SPILLS) {
1871 rc = mdb_page_unspill(txn, mp, &np);
1877 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1878 (rc = mdb_page_alloc(mc, 1, &np)))
1881 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
1882 mp->mp_pgno, pgno));
1883 assert(mp->mp_pgno != pgno);
1884 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1885 /* Update the parent page, if any, to point to the new page */
1887 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1888 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1889 SETPGNO(node, pgno);
1891 mc->mc_db->md_root = pgno;
1893 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1894 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1896 /* If txn has a parent, make sure the page is in our
1900 unsigned x = mdb_mid2l_search(dl, pgno);
1901 if (x <= dl[0].mid && dl[x].mid == pgno) {
1902 if (mp != dl[x].mptr) { /* bad cursor? */
1903 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1904 return MDB_CORRUPTED;
1909 assert(dl[0].mid < MDB_IDL_UM_MAX);
1911 np = mdb_page_malloc(txn, 1);
1916 mdb_mid2l_insert(dl, &mid);
1921 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1923 np->mp_flags |= P_DIRTY;
1926 /* Adjust cursors pointing to mp */
1927 mc->mc_pg[mc->mc_top] = np;
1928 m2 = txn->mt_cursors[mc->mc_dbi];
1929 if (mc->mc_flags & C_SUB) {
1930 for (; m2; m2=m2->mc_next) {
1931 m3 = &m2->mc_xcursor->mx_cursor;
1932 if (m3->mc_snum < mc->mc_snum) continue;
1933 if (m3->mc_pg[mc->mc_top] == mp)
1934 m3->mc_pg[mc->mc_top] = np;
1937 for (; m2; m2=m2->mc_next) {
1938 if (m2->mc_snum < mc->mc_snum) continue;
1939 if (m2->mc_pg[mc->mc_top] == mp) {
1940 m2->mc_pg[mc->mc_top] = np;
1941 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1942 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1944 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1945 if (!(leaf->mn_flags & F_SUBDATA))
1946 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1955 mdb_env_sync(MDB_env *env, int force)
1958 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1959 if (env->me_flags & MDB_WRITEMAP) {
1960 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1961 ? MS_ASYNC : MS_SYNC;
1962 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1965 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1969 if (MDB_FDATASYNC(env->me_fd))
1976 /** Back up parent txn's cursors, then grab the originals for tracking */
1978 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1980 MDB_cursor *mc, *bk;
1985 for (i = src->mt_numdbs; --i >= 0; ) {
1986 if ((mc = src->mt_cursors[i]) != NULL) {
1987 size = sizeof(MDB_cursor);
1989 size += sizeof(MDB_xcursor);
1990 for (; mc; mc = bk->mc_next) {
1996 mc->mc_db = &dst->mt_dbs[i];
1997 /* Kill pointers into src - and dst to reduce abuse: The
1998 * user may not use mc until dst ends. Otherwise we'd...
2000 mc->mc_txn = NULL; /* ...set this to dst */
2001 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2002 if ((mx = mc->mc_xcursor) != NULL) {
2003 *(MDB_xcursor *)(bk+1) = *mx;
2004 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2006 mc->mc_next = dst->mt_cursors[i];
2007 dst->mt_cursors[i] = mc;
2014 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2015 * @param[in] txn the transaction handle.
2016 * @param[in] merge true to keep changes to parent cursors, false to revert.
2017 * @return 0 on success, non-zero on failure.
2020 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2022 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2026 for (i = txn->mt_numdbs; --i >= 0; ) {
2027 for (mc = cursors[i]; mc; mc = next) {
2029 if ((bk = mc->mc_backup) != NULL) {
2031 /* Commit changes to parent txn */
2032 mc->mc_next = bk->mc_next;
2033 mc->mc_backup = bk->mc_backup;
2034 mc->mc_txn = bk->mc_txn;
2035 mc->mc_db = bk->mc_db;
2036 mc->mc_dbflag = bk->mc_dbflag;
2037 if ((mx = mc->mc_xcursor) != NULL)
2038 mx->mx_cursor.mc_txn = bk->mc_txn;
2040 /* Abort nested txn */
2042 if ((mx = mc->mc_xcursor) != NULL)
2043 *mx = *(MDB_xcursor *)(bk+1);
2047 /* Only malloced cursors are permanently tracked. */
2055 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2058 mdb_txn_reset0(MDB_txn *txn, const char *act);
2060 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2066 Pidset = F_SETLK, Pidcheck = F_GETLK
2070 /** Set or check a pid lock. Set returns 0 on success.
2071 * Check returns 0 if the process is certainly dead, nonzero if it may
2072 * be alive (the lock exists or an error happened so we do not know).
2074 * On Windows Pidset is a no-op, we merely check for the existence
2075 * of the process with the given pid. On POSIX we use a single byte
2076 * lock on the lockfile, set at an offset equal to the pid.
2079 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2081 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2084 if (op == Pidcheck) {
2085 h = OpenProcess(env->me_pidquery, FALSE, pid);
2086 /* No documented "no such process" code, but other program use this: */
2088 return ErrCode() != ERROR_INVALID_PARAMETER;
2089 /* A process exists until all handles to it close. Has it exited? */
2090 ret = WaitForSingleObject(h, 0) != 0;
2097 struct flock lock_info;
2098 memset(&lock_info, 0, sizeof(lock_info));
2099 lock_info.l_type = F_WRLCK;
2100 lock_info.l_whence = SEEK_SET;
2101 lock_info.l_start = pid;
2102 lock_info.l_len = 1;
2103 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2104 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2106 } else if ((rc = ErrCode()) == EINTR) {
2114 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2115 * @param[in] txn the transaction handle to initialize
2116 * @return 0 on success, non-zero on failure.
2119 mdb_txn_renew0(MDB_txn *txn)
2121 MDB_env *env = txn->mt_env;
2125 int rc, new_notls = 0;
2128 txn->mt_numdbs = env->me_numdbs;
2129 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2131 if (txn->mt_flags & MDB_TXN_RDONLY) {
2132 if (!env->me_txns) {
2133 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2134 txn->mt_txnid = meta->mm_txnid;
2135 txn->mt_u.reader = NULL;
2137 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2138 pthread_getspecific(env->me_txkey);
2140 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2141 return MDB_BAD_RSLOT;
2143 pid_t pid = env->me_pid;
2144 pthread_t tid = pthread_self();
2146 if (!(env->me_flags & MDB_LIVE_READER)) {
2147 rc = mdb_reader_pid(env, Pidset, pid);
2149 UNLOCK_MUTEX_R(env);
2152 env->me_flags |= MDB_LIVE_READER;
2156 for (i=0; i<env->me_txns->mti_numreaders; i++)
2157 if (env->me_txns->mti_readers[i].mr_pid == 0)
2159 if (i == env->me_maxreaders) {
2160 UNLOCK_MUTEX_R(env);
2161 return MDB_READERS_FULL;
2163 env->me_txns->mti_readers[i].mr_pid = pid;
2164 env->me_txns->mti_readers[i].mr_tid = tid;
2165 if (i >= env->me_txns->mti_numreaders)
2166 env->me_txns->mti_numreaders = i+1;
2167 /* Save numreaders for un-mutexed mdb_env_close() */
2168 env->me_numreaders = env->me_txns->mti_numreaders;
2169 UNLOCK_MUTEX_R(env);
2170 r = &env->me_txns->mti_readers[i];
2171 new_notls = (env->me_flags & MDB_NOTLS);
2172 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2177 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2178 txn->mt_u.reader = r;
2179 meta = env->me_metas[txn->mt_txnid & 1];
2185 txn->mt_txnid = env->me_txns->mti_txnid;
2186 meta = env->me_metas[txn->mt_txnid & 1];
2188 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2189 txn->mt_txnid = meta->mm_txnid;
2193 if (txn->mt_txnid == mdb_debug_start)
2196 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2197 txn->mt_u.dirty_list = env->me_dirty_list;
2198 txn->mt_u.dirty_list[0].mid = 0;
2199 txn->mt_free_pgs = env->me_free_pgs;
2200 txn->mt_free_pgs[0] = 0;
2201 txn->mt_spill_pgs = NULL;
2205 /* Copy the DB info and flags */
2206 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2208 /* Moved to here to avoid a data race in read TXNs */
2209 txn->mt_next_pgno = meta->mm_last_pg+1;
2211 for (i=2; i<txn->mt_numdbs; i++) {
2212 x = env->me_dbflags[i];
2213 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2214 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2216 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2218 if (env->me_maxpg < txn->mt_next_pgno) {
2219 mdb_txn_reset0(txn, "renew0-mapfail");
2221 txn->mt_u.reader->mr_pid = 0;
2222 txn->mt_u.reader = NULL;
2224 return MDB_MAP_RESIZED;
2231 mdb_txn_renew(MDB_txn *txn)
2235 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2238 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2239 DPUTS("environment had fatal error, must shutdown!");
2243 rc = mdb_txn_renew0(txn);
2244 if (rc == MDB_SUCCESS) {
2245 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2246 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2247 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2253 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2257 int rc, size, tsize = sizeof(MDB_txn);
2259 if (env->me_flags & MDB_FATAL_ERROR) {
2260 DPUTS("environment had fatal error, must shutdown!");
2263 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2266 /* Nested transactions: Max 1 child, write txns only, no writemap */
2267 if (parent->mt_child ||
2268 (flags & MDB_RDONLY) ||
2269 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2270 (env->me_flags & MDB_WRITEMAP))
2272 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2274 tsize = sizeof(MDB_ntxn);
2276 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2277 if (!(flags & MDB_RDONLY))
2278 size += env->me_maxdbs * sizeof(MDB_cursor *);
2280 if ((txn = calloc(1, size)) == NULL) {
2281 DPRINTF(("calloc: %s", strerror(ErrCode())));
2284 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2285 if (flags & MDB_RDONLY) {
2286 txn->mt_flags |= MDB_TXN_RDONLY;
2287 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2289 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2290 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2296 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2297 if (!txn->mt_u.dirty_list ||
2298 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2300 free(txn->mt_u.dirty_list);
2304 txn->mt_txnid = parent->mt_txnid;
2305 txn->mt_dirty_room = parent->mt_dirty_room;
2306 txn->mt_u.dirty_list[0].mid = 0;
2307 txn->mt_spill_pgs = NULL;
2308 txn->mt_next_pgno = parent->mt_next_pgno;
2309 parent->mt_child = txn;
2310 txn->mt_parent = parent;
2311 txn->mt_numdbs = parent->mt_numdbs;
2312 txn->mt_flags = parent->mt_flags;
2313 txn->mt_dbxs = parent->mt_dbxs;
2314 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2315 /* Copy parent's mt_dbflags, but clear DB_NEW */
2316 for (i=0; i<txn->mt_numdbs; i++)
2317 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2319 ntxn = (MDB_ntxn *)txn;
2320 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2321 if (env->me_pghead) {
2322 size = MDB_IDL_SIZEOF(env->me_pghead);
2323 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2325 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2330 rc = mdb_cursor_shadow(parent, txn);
2332 mdb_txn_reset0(txn, "beginchild-fail");
2334 rc = mdb_txn_renew0(txn);
2340 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2341 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2342 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2349 mdb_txn_env(MDB_txn *txn)
2351 if(!txn) return NULL;
2355 /** Export or close DBI handles opened in this txn. */
2357 mdb_dbis_update(MDB_txn *txn, int keep)
2360 MDB_dbi n = txn->mt_numdbs;
2361 MDB_env *env = txn->mt_env;
2362 unsigned char *tdbflags = txn->mt_dbflags;
2364 for (i = n; --i >= 2;) {
2365 if (tdbflags[i] & DB_NEW) {
2367 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2369 char *ptr = env->me_dbxs[i].md_name.mv_data;
2370 env->me_dbxs[i].md_name.mv_data = NULL;
2371 env->me_dbxs[i].md_name.mv_size = 0;
2372 env->me_dbflags[i] = 0;
2377 if (keep && env->me_numdbs < n)
2381 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2382 * May be called twice for readonly txns: First reset it, then abort.
2383 * @param[in] txn the transaction handle to reset
2384 * @param[in] act why the transaction is being reset
2387 mdb_txn_reset0(MDB_txn *txn, const char *act)
2389 MDB_env *env = txn->mt_env;
2391 /* Close any DBI handles opened in this txn */
2392 mdb_dbis_update(txn, 0);
2394 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2395 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2396 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2398 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2399 if (txn->mt_u.reader) {
2400 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2401 if (!(env->me_flags & MDB_NOTLS))
2402 txn->mt_u.reader = NULL; /* txn does not own reader */
2404 txn->mt_numdbs = 0; /* close nothing if called again */
2405 txn->mt_dbxs = NULL; /* mark txn as reset */
2407 mdb_cursors_close(txn, 0);
2409 if (!(env->me_flags & MDB_WRITEMAP)) {
2410 mdb_dlist_free(txn);
2412 mdb_midl_free(env->me_pghead);
2414 if (txn->mt_parent) {
2415 txn->mt_parent->mt_child = NULL;
2416 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2417 mdb_midl_free(txn->mt_free_pgs);
2418 mdb_midl_free(txn->mt_spill_pgs);
2419 free(txn->mt_u.dirty_list);
2423 if (mdb_midl_shrink(&txn->mt_free_pgs))
2424 env->me_free_pgs = txn->mt_free_pgs;
2425 env->me_pghead = NULL;
2429 /* The writer mutex was locked in mdb_txn_begin. */
2431 UNLOCK_MUTEX_W(env);
2436 mdb_txn_reset(MDB_txn *txn)
2441 /* This call is only valid for read-only txns */
2442 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2445 mdb_txn_reset0(txn, "reset");
2449 mdb_txn_abort(MDB_txn *txn)
2455 mdb_txn_abort(txn->mt_child);
2457 mdb_txn_reset0(txn, "abort");
2458 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2459 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2460 txn->mt_u.reader->mr_pid = 0;
2465 /** Save the freelist as of this transaction to the freeDB.
2466 * This changes the freelist. Keep trying until it stabilizes.
2469 mdb_freelist_save(MDB_txn *txn)
2471 /* env->me_pghead[] can grow and shrink during this call.
2472 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2473 * Page numbers cannot disappear from txn->mt_free_pgs[].
2476 MDB_env *env = txn->mt_env;
2477 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2478 txnid_t pglast = 0, head_id = 0;
2479 pgno_t freecnt = 0, *free_pgs, *mop;
2480 ssize_t head_room = 0, total_room = 0, mop_len;
2482 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2484 if (env->me_pghead) {
2485 /* Make sure first page of freeDB is touched and on freelist */
2486 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2487 if (rc && rc != MDB_NOTFOUND)
2492 /* Come back here after each Put() in case freelist changed */
2495 /* If using records from freeDB which we have not yet
2496 * deleted, delete them and any we reserved for me_pghead.
2498 while (pglast < env->me_pglast) {
2499 rc = mdb_cursor_first(&mc, &key, NULL);
2502 pglast = head_id = *(txnid_t *)key.mv_data;
2503 total_room = head_room = 0;
2504 assert(pglast <= env->me_pglast);
2505 rc = mdb_cursor_del(&mc, 0);
2510 /* Save the IDL of pages freed by this txn, to a single record */
2511 if (freecnt < txn->mt_free_pgs[0]) {
2513 /* Make sure last page of freeDB is touched and on freelist */
2514 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2515 if (rc && rc != MDB_NOTFOUND)
2518 free_pgs = txn->mt_free_pgs;
2519 /* Write to last page of freeDB */
2520 key.mv_size = sizeof(txn->mt_txnid);
2521 key.mv_data = &txn->mt_txnid;
2523 freecnt = free_pgs[0];
2524 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2525 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2528 /* Retry if mt_free_pgs[] grew during the Put() */
2529 free_pgs = txn->mt_free_pgs;
2530 } while (freecnt < free_pgs[0]);
2531 mdb_midl_sort(free_pgs);
2532 memcpy(data.mv_data, free_pgs, data.mv_size);
2535 unsigned int i = free_pgs[0];
2536 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2537 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2539 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2545 mop = env->me_pghead;
2546 mop_len = mop ? mop[0] : 0;
2548 /* Reserve records for me_pghead[]. Split it if multi-page,
2549 * to avoid searching freeDB for a page range. Use keys in
2550 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2552 if (total_room >= mop_len) {
2553 if (total_room == mop_len || --more < 0)
2555 } else if (head_room >= maxfree_1pg && head_id > 1) {
2556 /* Keep current record (overflow page), add a new one */
2560 /* (Re)write {key = head_id, IDL length = head_room} */
2561 total_room -= head_room;
2562 head_room = mop_len - total_room;
2563 if (head_room > maxfree_1pg && head_id > 1) {
2564 /* Overflow multi-page for part of me_pghead */
2565 head_room /= head_id; /* amortize page sizes */
2566 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2567 } else if (head_room < 0) {
2568 /* Rare case, not bothering to delete this record */
2571 key.mv_size = sizeof(head_id);
2572 key.mv_data = &head_id;
2573 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2574 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2577 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2578 total_room += head_room;
2581 /* Fill in the reserved me_pghead records */
2587 rc = mdb_cursor_first(&mc, &key, &data);
2588 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2589 unsigned flags = MDB_CURRENT;
2590 txnid_t id = *(txnid_t *)key.mv_data;
2591 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2594 assert(len >= 0 && id <= env->me_pglast);
2596 if (len > mop_len) {
2598 data.mv_size = (len + 1) * sizeof(MDB_ID);
2601 data.mv_data = mop -= len;
2604 rc = mdb_cursor_put(&mc, &key, &data, flags);
2606 if (rc || !(mop_len -= len))
2613 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2614 * @param[in] txn the transaction that's being committed
2615 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2616 * @return 0 on success, non-zero on failure.
2619 mdb_page_flush(MDB_txn *txn, int keep)
2621 MDB_env *env = txn->mt_env;
2622 MDB_ID2L dl = txn->mt_u.dirty_list;
2623 unsigned psize = env->me_psize, j;
2624 int i, pagecount = dl[0].mid, rc;
2625 size_t size = 0, pos = 0;
2627 MDB_page *dp = NULL;
2631 struct iovec iov[MDB_COMMIT_PAGES];
2632 ssize_t wpos = 0, wsize = 0, wres;
2633 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2639 if (env->me_flags & MDB_WRITEMAP) {
2640 /* Clear dirty flags */
2641 while (++i <= pagecount) {
2643 /* Don't flush this page yet */
2644 if (dp->mp_flags & P_KEEP) {
2645 dp->mp_flags ^= P_KEEP;
2649 dp->mp_flags &= ~P_DIRTY;
2654 /* Write the pages */
2656 if (++i <= pagecount) {
2658 /* Don't flush this page yet */
2659 if (dp->mp_flags & P_KEEP) {
2660 dp->mp_flags ^= P_KEEP;
2665 /* clear dirty flag */
2666 dp->mp_flags &= ~P_DIRTY;
2669 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2674 /* Windows actually supports scatter/gather I/O, but only on
2675 * unbuffered file handles. Since we're relying on the OS page
2676 * cache for all our data, that's self-defeating. So we just
2677 * write pages one at a time. We use the ov structure to set
2678 * the write offset, to at least save the overhead of a Seek
2681 DPRINTF(("committing page %"Z"u", pgno));
2682 memset(&ov, 0, sizeof(ov));
2683 ov.Offset = pos & 0xffffffff;
2684 ov.OffsetHigh = pos >> 16 >> 16;
2685 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2687 DPRINTF(("WriteFile: %d", rc));
2691 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2692 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2694 /* Write previous page(s) */
2695 #ifdef MDB_USE_PWRITEV
2696 wres = pwritev(env->me_fd, iov, n, wpos);
2699 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2701 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2703 DPRINTF(("lseek: %s", strerror(rc)));
2706 wres = writev(env->me_fd, iov, n);
2709 if (wres != wsize) {
2712 DPRINTF(("Write error: %s", strerror(rc)));
2714 rc = EIO; /* TODO: Use which error code? */
2715 DPUTS("short write, filesystem full?");
2726 DPRINTF(("committing page %"Z"u", pgno));
2727 next_pos = pos + size;
2728 iov[n].iov_len = size;
2729 iov[n].iov_base = (char *)dp;
2735 for (i = keep; ++i <= pagecount; ) {
2737 /* This is a page we skipped above */
2740 dl[j].mid = dp->mp_pgno;
2743 mdb_dpage_free(env, dp);
2748 txn->mt_dirty_room += i - j;
2754 mdb_txn_commit(MDB_txn *txn)
2760 assert(txn != NULL);
2761 assert(txn->mt_env != NULL);
2763 if (txn->mt_child) {
2764 rc = mdb_txn_commit(txn->mt_child);
2765 txn->mt_child = NULL;
2772 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2773 mdb_dbis_update(txn, 1);
2774 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2779 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2780 DPUTS("error flag is set, can't commit");
2782 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2787 if (txn->mt_parent) {
2788 MDB_txn *parent = txn->mt_parent;
2791 unsigned x, y, len, ps_len;
2793 /* Append our free list to parent's */
2794 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2797 mdb_midl_free(txn->mt_free_pgs);
2798 /* Failures after this must either undo the changes
2799 * to the parent or set MDB_TXN_ERROR in the parent.
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 ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2824 pspill[0] = (pgno_t)-1;
2825 /* Mark our dirty pages as deleted in parent spill list */
2826 for (i=0, len=src[0].mid; ++i <= len; ) {
2827 MDB_ID pn = src[i].mid << 1;
2828 while (pn > pspill[x])
2830 if (pn == pspill[x]) {
2835 /* Squash deleted pagenums if we deleted any */
2836 for (x=y; ++x <= ps_len; )
2837 if (!(pspill[x] & 1))
2838 pspill[++y] = pspill[x];
2842 /* Find len = length of merging our dirty list with parent's */
2844 dst[0].mid = 0; /* simplify loops */
2845 if (parent->mt_parent) {
2846 len = x + src[0].mid;
2847 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2848 for (i = x; y && i; y--) {
2849 pgno_t yp = src[y].mid;
2850 while (yp < dst[i].mid)
2852 if (yp == dst[i].mid) {
2857 } else { /* Simplify the above for single-ancestor case */
2858 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2860 /* Merge our dirty list with parent's */
2862 for (i = len; y; dst[i--] = src[y--]) {
2863 pgno_t yp = src[y].mid;
2864 while (yp < dst[x].mid)
2865 dst[i--] = dst[x--];
2866 if (yp == dst[x].mid)
2867 free(dst[x--].mptr);
2871 free(txn->mt_u.dirty_list);
2872 parent->mt_dirty_room = txn->mt_dirty_room;
2873 if (txn->mt_spill_pgs) {
2874 if (parent->mt_spill_pgs) {
2875 /* TODO: Prevent failure here, so parent does not fail */
2876 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2878 parent->mt_flags |= MDB_TXN_ERROR;
2879 mdb_midl_free(txn->mt_spill_pgs);
2880 mdb_midl_sort(parent->mt_spill_pgs);
2882 parent->mt_spill_pgs = txn->mt_spill_pgs;
2886 parent->mt_child = NULL;
2887 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2892 if (txn != env->me_txn) {
2893 DPUTS("attempt to commit unknown transaction");
2898 mdb_cursors_close(txn, 0);
2900 if (!txn->mt_u.dirty_list[0].mid &&
2901 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2904 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2905 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2907 /* Update DB root pointers */
2908 if (txn->mt_numdbs > 2) {
2912 data.mv_size = sizeof(MDB_db);
2914 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2915 for (i = 2; i < txn->mt_numdbs; i++) {
2916 if (txn->mt_dbflags[i] & DB_DIRTY) {
2917 data.mv_data = &txn->mt_dbs[i];
2918 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2925 rc = mdb_freelist_save(txn);
2929 mdb_midl_free(env->me_pghead);
2930 env->me_pghead = NULL;
2931 if (mdb_midl_shrink(&txn->mt_free_pgs))
2932 env->me_free_pgs = txn->mt_free_pgs;
2938 if ((rc = mdb_page_flush(txn, 0)) ||
2939 (rc = mdb_env_sync(env, 0)) ||
2940 (rc = mdb_env_write_meta(txn)))
2946 mdb_dbis_update(txn, 1);
2949 UNLOCK_MUTEX_W(env);
2959 /** Read the environment parameters of a DB environment before
2960 * mapping it into memory.
2961 * @param[in] env the environment handle
2962 * @param[out] meta address of where to store the meta information
2963 * @return 0 on success, non-zero on failure.
2966 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2973 /* We don't know the page size yet, so use a minimum value.
2974 * Read both meta pages so we can use the latest one.
2977 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2981 memset(&ov, 0, sizeof(ov));
2983 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2984 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2987 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2989 if (rc != MDB_PAGESIZE) {
2990 if (rc == 0 && off == 0)
2992 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2993 DPRINTF(("read: %s", mdb_strerror(rc)));
2997 p = (MDB_page *)&pbuf;
2999 if (!F_ISSET(p->mp_flags, P_META)) {
3000 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3005 if (m->mm_magic != MDB_MAGIC) {
3006 DPUTS("meta has invalid magic");
3010 if (m->mm_version != MDB_DATA_VERSION) {
3011 DPRINTF(("database is version %u, expected version %u",
3012 m->mm_version, MDB_DATA_VERSION));
3013 return MDB_VERSION_MISMATCH;
3016 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3022 /** Write the environment parameters of a freshly created DB environment.
3023 * @param[in] env the environment handle
3024 * @param[out] meta address of where to store the meta information
3025 * @return 0 on success, non-zero on failure.
3028 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3036 memset(&ov, 0, sizeof(ov));
3037 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3039 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3042 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3043 len = pwrite(fd, ptr, size, pos); \
3044 rc = (len >= 0); } while(0)
3047 DPUTS("writing new meta page");
3049 psize = env->me_psize;
3051 meta->mm_magic = MDB_MAGIC;
3052 meta->mm_version = MDB_DATA_VERSION;
3053 meta->mm_mapsize = env->me_mapsize;
3054 meta->mm_psize = psize;
3055 meta->mm_last_pg = 1;
3056 meta->mm_flags = env->me_flags & 0xffff;
3057 meta->mm_flags |= MDB_INTEGERKEY;
3058 meta->mm_dbs[0].md_root = P_INVALID;
3059 meta->mm_dbs[1].md_root = P_INVALID;
3061 p = calloc(2, psize);
3063 p->mp_flags = P_META;
3064 *(MDB_meta *)METADATA(p) = *meta;
3066 q = (MDB_page *)((char *)p + psize);
3068 q->mp_flags = P_META;
3069 *(MDB_meta *)METADATA(q) = *meta;
3071 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3074 else if ((unsigned) len == psize * 2)
3082 /** Update the environment info to commit a transaction.
3083 * @param[in] txn the transaction that's being committed
3084 * @return 0 on success, non-zero on failure.
3087 mdb_env_write_meta(MDB_txn *txn)
3090 MDB_meta meta, metab, *mp;
3092 int rc, len, toggle;
3101 assert(txn != NULL);
3102 assert(txn->mt_env != NULL);
3104 toggle = txn->mt_txnid & 1;
3105 DPRINTF(("writing meta page %d for root page %"Z"u",
3106 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3109 mp = env->me_metas[toggle];
3111 if (env->me_flags & MDB_WRITEMAP) {
3112 /* Persist any increases of mapsize config */
3113 if (env->me_mapsize > mp->mm_mapsize)
3114 mp->mm_mapsize = env->me_mapsize;
3115 mp->mm_dbs[0] = txn->mt_dbs[0];
3116 mp->mm_dbs[1] = txn->mt_dbs[1];
3117 mp->mm_last_pg = txn->mt_next_pgno - 1;
3118 mp->mm_txnid = txn->mt_txnid;
3119 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3120 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3123 ptr += env->me_psize;
3124 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3131 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3132 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3134 ptr = (char *)&meta;
3135 if (env->me_mapsize > mp->mm_mapsize) {
3136 /* Persist any increases of mapsize config */
3137 meta.mm_mapsize = env->me_mapsize;
3138 off = offsetof(MDB_meta, mm_mapsize);
3140 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3142 len = sizeof(MDB_meta) - off;
3145 meta.mm_dbs[0] = txn->mt_dbs[0];
3146 meta.mm_dbs[1] = txn->mt_dbs[1];
3147 meta.mm_last_pg = txn->mt_next_pgno - 1;
3148 meta.mm_txnid = txn->mt_txnid;
3151 off += env->me_psize;
3154 /* Write to the SYNC fd */
3155 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3156 env->me_fd : env->me_mfd;
3159 memset(&ov, 0, sizeof(ov));
3161 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3165 rc = pwrite(mfd, ptr, len, off);
3168 rc = rc < 0 ? ErrCode() : EIO;
3169 DPUTS("write failed, disk error?");
3170 /* On a failure, the pagecache still contains the new data.
3171 * Write some old data back, to prevent it from being used.
3172 * Use the non-SYNC fd; we know it will fail anyway.
3174 meta.mm_last_pg = metab.mm_last_pg;
3175 meta.mm_txnid = metab.mm_txnid;
3177 memset(&ov, 0, sizeof(ov));
3179 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3181 r2 = pwrite(env->me_fd, ptr, len, off);
3182 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3185 env->me_flags |= MDB_FATAL_ERROR;
3189 /* Memory ordering issues are irrelevant; since the entire writer
3190 * is wrapped by wmutex, all of these changes will become visible
3191 * after the wmutex is unlocked. Since the DB is multi-version,
3192 * readers will get consistent data regardless of how fresh or
3193 * how stale their view of these values is.
3196 env->me_txns->mti_txnid = txn->mt_txnid;
3201 /** Check both meta pages to see which one is newer.
3202 * @param[in] env the environment handle
3203 * @return meta toggle (0 or 1).
3206 mdb_env_pick_meta(const MDB_env *env)
3208 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3212 mdb_env_create(MDB_env **env)
3216 e = calloc(1, sizeof(MDB_env));
3220 e->me_maxreaders = DEFAULT_READERS;
3221 e->me_maxdbs = e->me_numdbs = 2;
3222 e->me_fd = INVALID_HANDLE_VALUE;
3223 e->me_lfd = INVALID_HANDLE_VALUE;
3224 e->me_mfd = INVALID_HANDLE_VALUE;
3225 #ifdef MDB_USE_POSIX_SEM
3226 e->me_rmutex = SEM_FAILED;
3227 e->me_wmutex = SEM_FAILED;
3229 e->me_pid = getpid();
3230 VGMEMP_CREATE(e,0,0);
3236 mdb_env_map(MDB_env *env, void *addr, int newsize)
3239 unsigned int flags = env->me_flags;
3243 LONG sizelo, sizehi;
3244 sizelo = env->me_mapsize & 0xffffffff;
3245 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3247 /* Windows won't create mappings for zero length files.
3248 * Just allocate the maxsize right now.
3251 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3252 || !SetEndOfFile(env->me_fd)
3253 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3256 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3257 PAGE_READWRITE : PAGE_READONLY,
3258 sizehi, sizelo, NULL);
3261 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3262 FILE_MAP_WRITE : FILE_MAP_READ,
3263 0, 0, env->me_mapsize, addr);
3264 rc = env->me_map ? 0 : ErrCode();
3269 int prot = PROT_READ;
3270 if (flags & MDB_WRITEMAP) {
3272 if (newsize && ftruncate(env->me_fd, env->me_mapsize) < 0)
3275 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3277 if (env->me_map == MAP_FAILED) {
3281 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3283 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3285 #ifdef POSIX_MADV_RANDOM
3286 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3287 #endif /* POSIX_MADV_RANDOM */
3288 #endif /* MADV_RANDOM */
3291 /* Can happen because the address argument to mmap() is just a
3292 * hint. mmap() can pick another, e.g. if the range is in use.
3293 * The MAP_FIXED flag would prevent that, but then mmap could
3294 * instead unmap existing pages to make room for the new map.
3296 if (addr && env->me_map != addr)
3297 return EBUSY; /* TODO: Make a new MDB_* error code? */
3299 p = (MDB_page *)env->me_map;
3300 env->me_metas[0] = METADATA(p);
3301 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3307 mdb_env_set_mapsize(MDB_env *env, size_t size)
3309 /* If env is already open, caller is responsible for making
3310 * sure there are no active txns.
3318 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3319 munmap(env->me_map, env->me_mapsize);
3320 env->me_mapsize = size;
3321 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3322 rc = mdb_env_map(env, old, 1);
3326 env->me_mapsize = size;
3328 env->me_maxpg = env->me_mapsize / env->me_psize;
3333 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3337 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3342 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3344 if (env->me_map || readers < 1)
3346 env->me_maxreaders = readers;
3351 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3353 if (!env || !readers)
3355 *readers = env->me_maxreaders;
3359 /** Further setup required for opening an MDB environment
3362 mdb_env_open2(MDB_env *env)
3364 unsigned int flags = env->me_flags;
3365 int i, newenv = 0, rc;
3369 /* See if we should use QueryLimited */
3371 if ((rc & 0xff) > 5)
3372 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3374 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3377 memset(&meta, 0, sizeof(meta));
3379 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3382 DPUTS("new mdbenv");
3384 GET_PAGESIZE(env->me_psize);
3386 env->me_psize = meta.mm_psize;
3389 /* Was a mapsize configured? */
3390 if (!env->me_mapsize) {
3391 /* If this is a new environment, take the default,
3392 * else use the size recorded in the existing env.
3394 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3395 } else if (env->me_mapsize < meta.mm_mapsize) {
3396 /* If the configured size is smaller, make sure it's
3397 * still big enough. Silently round up to minimum if not.
3399 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3400 if (env->me_mapsize < minsize)
3401 env->me_mapsize = minsize;
3404 rc = mdb_env_map(env, meta.mm_address, newenv);
3409 if (flags & MDB_FIXEDMAP)
3410 meta.mm_address = env->me_map;
3411 i = mdb_env_init_meta(env, &meta);
3412 if (i != MDB_SUCCESS) {
3416 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3417 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3419 env->me_maxpg = env->me_mapsize / env->me_psize;
3422 int toggle = mdb_env_pick_meta(env);
3423 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3425 DPRINTF(("opened database version %u, pagesize %u",
3426 env->me_metas[0]->mm_version, env->me_psize));
3427 DPRINTF(("using meta page %d", toggle));
3428 DPRINTF(("depth: %u", db->md_depth));
3429 DPRINTF(("entries: %"Z"u", db->md_entries));
3430 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3431 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3432 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3433 DPRINTF(("root: %"Z"u", db->md_root));
3441 /** Release a reader thread's slot in the reader lock table.
3442 * This function is called automatically when a thread exits.
3443 * @param[in] ptr This points to the slot in the reader lock table.
3446 mdb_env_reader_dest(void *ptr)
3448 MDB_reader *reader = ptr;
3454 /** Junk for arranging thread-specific callbacks on Windows. This is
3455 * necessarily platform and compiler-specific. Windows supports up
3456 * to 1088 keys. Let's assume nobody opens more than 64 environments
3457 * in a single process, for now. They can override this if needed.
3459 #ifndef MAX_TLS_KEYS
3460 #define MAX_TLS_KEYS 64
3462 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3463 static int mdb_tls_nkeys;
3465 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3469 case DLL_PROCESS_ATTACH: break;
3470 case DLL_THREAD_ATTACH: break;
3471 case DLL_THREAD_DETACH:
3472 for (i=0; i<mdb_tls_nkeys; i++) {
3473 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3474 mdb_env_reader_dest(r);
3477 case DLL_PROCESS_DETACH: break;
3482 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3484 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3488 /* Force some symbol references.
3489 * _tls_used forces the linker to create the TLS directory if not already done
3490 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3492 #pragma comment(linker, "/INCLUDE:_tls_used")
3493 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3494 #pragma const_seg(".CRT$XLB")
3495 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3496 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3499 #pragma comment(linker, "/INCLUDE:__tls_used")
3500 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3501 #pragma data_seg(".CRT$XLB")
3502 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3504 #endif /* WIN 32/64 */
3505 #endif /* !__GNUC__ */
3508 /** Downgrade the exclusive lock on the region back to shared */
3510 mdb_env_share_locks(MDB_env *env, int *excl)
3512 int rc = 0, toggle = mdb_env_pick_meta(env);
3514 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3519 /* First acquire a shared lock. The Unlock will
3520 * then release the existing exclusive lock.
3522 memset(&ov, 0, sizeof(ov));
3523 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3526 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3532 struct flock lock_info;
3533 /* The shared lock replaces the existing lock */
3534 memset((void *)&lock_info, 0, sizeof(lock_info));
3535 lock_info.l_type = F_RDLCK;
3536 lock_info.l_whence = SEEK_SET;
3537 lock_info.l_start = 0;
3538 lock_info.l_len = 1;
3539 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3540 (rc = ErrCode()) == EINTR) ;
3541 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3548 /** Try to get exlusive lock, otherwise shared.
3549 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3552 mdb_env_excl_lock(MDB_env *env, int *excl)
3556 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3560 memset(&ov, 0, sizeof(ov));
3561 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3568 struct flock lock_info;
3569 memset((void *)&lock_info, 0, sizeof(lock_info));
3570 lock_info.l_type = F_WRLCK;
3571 lock_info.l_whence = SEEK_SET;
3572 lock_info.l_start = 0;
3573 lock_info.l_len = 1;
3574 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3575 (rc = ErrCode()) == EINTR) ;
3579 # ifdef MDB_USE_POSIX_SEM
3580 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3583 lock_info.l_type = F_RDLCK;
3584 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3585 (rc = ErrCode()) == EINTR) ;
3593 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3595 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3597 * @(#) $Revision: 5.1 $
3598 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3599 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3601 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3605 * Please do not copyright this code. This code is in the public domain.
3607 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3608 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3609 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3610 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3611 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3612 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3613 * PERFORMANCE OF THIS SOFTWARE.
3616 * chongo <Landon Curt Noll> /\oo/\
3617 * http://www.isthe.com/chongo/
3619 * Share and Enjoy! :-)
3622 typedef unsigned long long mdb_hash_t;
3623 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3625 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3626 * @param[in] val value to hash
3627 * @param[in] hval initial value for hash
3628 * @return 64 bit hash
3630 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3631 * hval arg on the first call.
3634 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3636 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3637 unsigned char *end = s + val->mv_size;
3639 * FNV-1a hash each octet of the string
3642 /* xor the bottom with the current octet */
3643 hval ^= (mdb_hash_t)*s++;
3645 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3646 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3647 (hval << 7) + (hval << 8) + (hval << 40);
3649 /* return our new hash value */
3653 /** Hash the string and output the encoded hash.
3654 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3655 * very short name limits. We don't care about the encoding being reversible,
3656 * we just want to preserve as many bits of the input as possible in a
3657 * small printable string.
3658 * @param[in] str string to hash
3659 * @param[out] encbuf an array of 11 chars to hold the hash
3661 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3664 mdb_pack85(unsigned long l, char *out)
3668 for (i=0; i<5; i++) {
3669 *out++ = mdb_a85[l % 85];
3675 mdb_hash_enc(MDB_val *val, char *encbuf)
3677 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3679 mdb_pack85(h, encbuf);
3680 mdb_pack85(h>>32, encbuf+5);
3685 /** Open and/or initialize the lock region for the environment.
3686 * @param[in] env The MDB environment.
3687 * @param[in] lpath The pathname of the file used for the lock region.
3688 * @param[in] mode The Unix permissions for the file, if we create it.
3689 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3690 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3691 * @return 0 on success, non-zero on failure.
3694 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3697 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3699 # define MDB_ERRCODE_ROFS EROFS
3700 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3701 # define MDB_CLOEXEC O_CLOEXEC
3704 # define MDB_CLOEXEC 0
3711 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3712 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3713 FILE_ATTRIBUTE_NORMAL, NULL);
3715 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3717 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3719 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3724 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3725 /* Lose record locks when exec*() */
3726 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3727 fcntl(env->me_lfd, F_SETFD, fdflags);
3730 if (!(env->me_flags & MDB_NOTLS)) {
3731 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3734 env->me_flags |= MDB_ENV_TXKEY;
3736 /* Windows TLS callbacks need help finding their TLS info. */
3737 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3741 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3745 /* Try to get exclusive lock. If we succeed, then
3746 * nobody is using the lock region and we should initialize it.
3748 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3751 size = GetFileSize(env->me_lfd, NULL);
3753 size = lseek(env->me_lfd, 0, SEEK_END);
3754 if (size == -1) goto fail_errno;
3756 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3757 if (size < rsize && *excl > 0) {
3759 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3760 || !SetEndOfFile(env->me_lfd))
3763 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3767 size = rsize - sizeof(MDB_txninfo);
3768 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3773 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3775 if (!mh) goto fail_errno;
3776 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3778 if (!env->me_txns) goto fail_errno;
3780 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3782 if (m == MAP_FAILED) goto fail_errno;
3788 BY_HANDLE_FILE_INFORMATION stbuf;
3797 if (!mdb_sec_inited) {
3798 InitializeSecurityDescriptor(&mdb_null_sd,
3799 SECURITY_DESCRIPTOR_REVISION);
3800 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3801 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3802 mdb_all_sa.bInheritHandle = FALSE;
3803 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3806 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3807 idbuf.volume = stbuf.dwVolumeSerialNumber;
3808 idbuf.nhigh = stbuf.nFileIndexHigh;
3809 idbuf.nlow = stbuf.nFileIndexLow;
3810 val.mv_data = &idbuf;
3811 val.mv_size = sizeof(idbuf);
3812 mdb_hash_enc(&val, encbuf);
3813 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3814 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3815 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3816 if (!env->me_rmutex) goto fail_errno;
3817 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3818 if (!env->me_wmutex) goto fail_errno;
3819 #elif defined(MDB_USE_POSIX_SEM)
3828 #if defined(__NetBSD__)
3829 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3831 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3832 idbuf.dev = stbuf.st_dev;
3833 idbuf.ino = stbuf.st_ino;
3834 val.mv_data = &idbuf;
3835 val.mv_size = sizeof(idbuf);
3836 mdb_hash_enc(&val, encbuf);
3837 #ifdef MDB_SHORT_SEMNAMES
3838 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3840 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3841 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3842 /* Clean up after a previous run, if needed: Try to
3843 * remove both semaphores before doing anything else.
3845 sem_unlink(env->me_txns->mti_rmname);
3846 sem_unlink(env->me_txns->mti_wmname);
3847 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3848 O_CREAT|O_EXCL, mode, 1);
3849 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3850 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3851 O_CREAT|O_EXCL, mode, 1);
3852 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3853 #else /* MDB_USE_POSIX_SEM */
3854 pthread_mutexattr_t mattr;
3856 if ((rc = pthread_mutexattr_init(&mattr))
3857 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3858 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3859 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3861 pthread_mutexattr_destroy(&mattr);
3862 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3864 env->me_txns->mti_magic = MDB_MAGIC;
3865 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3866 env->me_txns->mti_txnid = 0;
3867 env->me_txns->mti_numreaders = 0;
3870 if (env->me_txns->mti_magic != MDB_MAGIC) {
3871 DPUTS("lock region has invalid magic");
3875 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3876 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3877 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3878 rc = MDB_VERSION_MISMATCH;
3882 if (rc && rc != EACCES && rc != EAGAIN) {
3886 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3887 if (!env->me_rmutex) goto fail_errno;
3888 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3889 if (!env->me_wmutex) goto fail_errno;
3890 #elif defined(MDB_USE_POSIX_SEM)
3891 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3892 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3893 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3894 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3905 /** The name of the lock file in the DB environment */
3906 #define LOCKNAME "/lock.mdb"
3907 /** The name of the data file in the DB environment */
3908 #define DATANAME "/data.mdb"
3909 /** The suffix of the lock file when no subdir is used */
3910 #define LOCKSUFF "-lock"
3911 /** Only a subset of the @ref mdb_env flags can be changed
3912 * at runtime. Changing other flags requires closing the
3913 * environment and re-opening it with the new flags.
3915 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3916 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK)
3919 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3921 int oflags, rc, len, excl = -1;
3922 char *lpath, *dpath;
3924 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3928 if (flags & MDB_NOSUBDIR) {
3929 rc = len + sizeof(LOCKSUFF) + len + 1;
3931 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3936 if (flags & MDB_NOSUBDIR) {
3937 dpath = lpath + len + sizeof(LOCKSUFF);
3938 sprintf(lpath, "%s" LOCKSUFF, path);
3939 strcpy(dpath, path);
3941 dpath = lpath + len + sizeof(LOCKNAME);
3942 sprintf(lpath, "%s" LOCKNAME, path);
3943 sprintf(dpath, "%s" DATANAME, path);
3947 flags |= env->me_flags;
3948 if (flags & MDB_RDONLY) {
3949 /* silently ignore WRITEMAP when we're only getting read access */
3950 flags &= ~MDB_WRITEMAP;
3952 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3953 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3956 env->me_flags = flags |= MDB_ENV_ACTIVE;
3960 env->me_path = strdup(path);
3961 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3962 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3963 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3968 /* For RDONLY, get lockfile after we know datafile exists */
3969 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
3970 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3976 if (F_ISSET(flags, MDB_RDONLY)) {
3977 oflags = GENERIC_READ;
3978 len = OPEN_EXISTING;
3980 oflags = GENERIC_READ|GENERIC_WRITE;
3983 mode = FILE_ATTRIBUTE_NORMAL;
3984 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3985 NULL, len, mode, NULL);
3987 if (F_ISSET(flags, MDB_RDONLY))
3990 oflags = O_RDWR | O_CREAT;
3992 env->me_fd = open(dpath, oflags, mode);
3994 if (env->me_fd == INVALID_HANDLE_VALUE) {
3999 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4000 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4005 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4006 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4007 env->me_mfd = env->me_fd;
4009 /* Synchronous fd for meta writes. Needed even with
4010 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4013 len = OPEN_EXISTING;
4014 env->me_mfd = CreateFile(dpath, oflags,
4015 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4016 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4019 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4021 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4026 DPRINTF(("opened dbenv %p", (void *) env));
4028 rc = mdb_env_share_locks(env, &excl);
4034 mdb_env_close0(env, excl);
4040 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4042 mdb_env_close0(MDB_env *env, int excl)
4046 if (!(env->me_flags & MDB_ENV_ACTIVE))
4049 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4050 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4051 free(env->me_dbxs[i].md_name.mv_data);
4053 free(env->me_dbflags);
4056 free(env->me_dirty_list);
4057 mdb_midl_free(env->me_free_pgs);
4059 if (env->me_flags & MDB_ENV_TXKEY) {
4060 pthread_key_delete(env->me_txkey);
4062 /* Delete our key from the global list */
4063 for (i=0; i<mdb_tls_nkeys; i++)
4064 if (mdb_tls_keys[i] == env->me_txkey) {
4065 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4073 munmap(env->me_map, env->me_mapsize);
4075 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4076 (void) close(env->me_mfd);
4077 if (env->me_fd != INVALID_HANDLE_VALUE)
4078 (void) close(env->me_fd);
4080 pid_t pid = env->me_pid;
4081 /* Clearing readers is done in this function because
4082 * me_txkey with its destructor must be disabled first.
4084 for (i = env->me_numreaders; --i >= 0; )
4085 if (env->me_txns->mti_readers[i].mr_pid == pid)
4086 env->me_txns->mti_readers[i].mr_pid = 0;
4088 if (env->me_rmutex) {
4089 CloseHandle(env->me_rmutex);
4090 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4092 /* Windows automatically destroys the mutexes when
4093 * the last handle closes.
4095 #elif defined(MDB_USE_POSIX_SEM)
4096 if (env->me_rmutex != SEM_FAILED) {
4097 sem_close(env->me_rmutex);
4098 if (env->me_wmutex != SEM_FAILED)
4099 sem_close(env->me_wmutex);
4100 /* If we have the filelock: If we are the
4101 * only remaining user, clean up semaphores.
4104 mdb_env_excl_lock(env, &excl);
4106 sem_unlink(env->me_txns->mti_rmname);
4107 sem_unlink(env->me_txns->mti_wmname);
4111 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4113 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4116 /* Unlock the lockfile. Windows would have unlocked it
4117 * after closing anyway, but not necessarily at once.
4119 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4122 (void) close(env->me_lfd);
4125 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4129 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4131 MDB_txn *txn = NULL;
4137 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4141 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4144 /* Do the lock/unlock of the reader mutex before starting the
4145 * write txn. Otherwise other read txns could block writers.
4147 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4152 /* We must start the actual read txn after blocking writers */
4153 mdb_txn_reset0(txn, "reset-stage1");
4155 /* Temporarily block writers until we snapshot the meta pages */
4158 rc = mdb_txn_renew0(txn);
4160 UNLOCK_MUTEX_W(env);
4165 wsize = env->me_psize * 2;
4169 DO_WRITE(rc, fd, ptr, w2, len);
4173 } else if (len > 0) {
4179 /* Non-blocking or async handles are not supported */
4185 UNLOCK_MUTEX_W(env);
4190 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4192 if (wsize > MAX_WRITE)
4196 DO_WRITE(rc, fd, ptr, w2, len);
4200 } else if (len > 0) {
4217 mdb_env_copy(MDB_env *env, const char *path)
4221 HANDLE newfd = INVALID_HANDLE_VALUE;
4223 if (env->me_flags & MDB_NOSUBDIR) {
4224 lpath = (char *)path;
4227 len += sizeof(DATANAME);
4228 lpath = malloc(len);
4231 sprintf(lpath, "%s" DATANAME, path);
4234 /* The destination path must exist, but the destination file must not.
4235 * We don't want the OS to cache the writes, since the source data is
4236 * already in the OS cache.
4239 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4240 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4242 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4244 if (newfd == INVALID_HANDLE_VALUE) {
4250 /* Set O_DIRECT if the file system supports it */
4251 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4252 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4254 #ifdef F_NOCACHE /* __APPLE__ */
4255 rc = fcntl(newfd, F_NOCACHE, 1);
4262 rc = mdb_env_copyfd(env, newfd);
4265 if (!(env->me_flags & MDB_NOSUBDIR))
4267 if (newfd != INVALID_HANDLE_VALUE)
4268 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4275 mdb_env_close(MDB_env *env)
4282 VGMEMP_DESTROY(env);
4283 while ((dp = env->me_dpages) != NULL) {
4284 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4285 env->me_dpages = dp->mp_next;
4289 mdb_env_close0(env, 0);
4293 /** Compare two items pointing at aligned size_t's */
4295 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4297 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4298 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4301 /** Compare two items pointing at aligned unsigned int's */
4303 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4305 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4306 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4309 /** Compare two items pointing at unsigned ints of unknown alignment.
4310 * Nodes and keys are guaranteed to be 2-byte aligned.
4313 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4315 #if BYTE_ORDER == LITTLE_ENDIAN
4316 unsigned short *u, *c;
4319 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4320 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4323 } while(!x && u > (unsigned short *)a->mv_data);
4326 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4330 /** Compare two items lexically */
4332 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4339 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4345 diff = memcmp(a->mv_data, b->mv_data, len);
4346 return diff ? diff : len_diff<0 ? -1 : len_diff;
4349 /** Compare two items in reverse byte order */
4351 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4353 const unsigned char *p1, *p2, *p1_lim;
4357 p1_lim = (const unsigned char *)a->mv_data;
4358 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4359 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4361 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4367 while (p1 > p1_lim) {
4368 diff = *--p1 - *--p2;
4372 return len_diff<0 ? -1 : len_diff;
4375 /** Search for key within a page, using binary search.
4376 * Returns the smallest entry larger or equal to the key.
4377 * If exactp is non-null, stores whether the found entry was an exact match
4378 * in *exactp (1 or 0).
4379 * Updates the cursor index with the index of the found entry.
4380 * If no entry larger or equal to the key is found, returns NULL.
4383 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4385 unsigned int i = 0, nkeys;
4388 MDB_page *mp = mc->mc_pg[mc->mc_top];
4389 MDB_node *node = NULL;
4394 nkeys = NUMKEYS(mp);
4399 COPY_PGNO(pgno, mp->mp_pgno);
4400 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4401 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4408 low = IS_LEAF(mp) ? 0 : 1;
4410 cmp = mc->mc_dbx->md_cmp;
4412 /* Branch pages have no data, so if using integer keys,
4413 * alignment is guaranteed. Use faster mdb_cmp_int.
4415 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4416 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4423 nodekey.mv_size = mc->mc_db->md_pad;
4424 node = NODEPTR(mp, 0); /* fake */
4425 while (low <= high) {
4426 i = (low + high) >> 1;
4427 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4428 rc = cmp(key, &nodekey);
4429 DPRINTF(("found leaf index %u [%s], rc = %i",
4430 i, DKEY(&nodekey), rc));
4439 while (low <= high) {
4440 i = (low + high) >> 1;
4442 node = NODEPTR(mp, i);
4443 nodekey.mv_size = NODEKSZ(node);
4444 nodekey.mv_data = NODEKEY(node);
4446 rc = cmp(key, &nodekey);
4449 DPRINTF(("found leaf index %u [%s], rc = %i",
4450 i, DKEY(&nodekey), rc));
4452 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4453 i, DKEY(&nodekey), NODEPGNO(node), rc));
4464 if (rc > 0) { /* Found entry is less than the key. */
4465 i++; /* Skip to get the smallest entry larger than key. */
4467 node = NODEPTR(mp, i);
4470 *exactp = (rc == 0);
4471 /* store the key index */
4472 mc->mc_ki[mc->mc_top] = i;
4474 /* There is no entry larger or equal to the key. */
4477 /* nodeptr is fake for LEAF2 */
4483 mdb_cursor_adjust(MDB_cursor *mc, func)
4487 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4488 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4495 /** Pop a page off the top of the cursor's stack. */
4497 mdb_cursor_pop(MDB_cursor *mc)
4501 MDB_page *top = mc->mc_pg[mc->mc_top];
4507 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4508 DDBI(mc), (void *) mc));
4512 /** Push a page onto the top of the cursor's stack. */
4514 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4516 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4517 DDBI(mc), (void *) mc));
4519 if (mc->mc_snum >= CURSOR_STACK) {
4520 assert(mc->mc_snum < CURSOR_STACK);
4521 return MDB_CURSOR_FULL;
4524 mc->mc_top = mc->mc_snum++;
4525 mc->mc_pg[mc->mc_top] = mp;
4526 mc->mc_ki[mc->mc_top] = 0;
4531 /** Find the address of the page corresponding to a given page number.
4532 * @param[in] txn the transaction for this access.
4533 * @param[in] pgno the page number for the page to retrieve.
4534 * @param[out] ret address of a pointer where the page's address will be stored.
4535 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4536 * @return 0 on success, non-zero on failure.
4539 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4541 MDB_env *env = txn->mt_env;
4545 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4549 MDB_ID2L dl = tx2->mt_u.dirty_list;
4551 /* Spilled pages were dirtied in this txn and flushed
4552 * because the dirty list got full. Bring this page
4553 * back in from the map (but don't unspill it here,
4554 * leave that unless page_touch happens again).
4556 if (tx2->mt_spill_pgs) {
4557 MDB_ID pn = pgno << 1;
4558 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4559 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4560 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4565 unsigned x = mdb_mid2l_search(dl, pgno);
4566 if (x <= dl[0].mid && dl[x].mid == pgno) {
4572 } while ((tx2 = tx2->mt_parent) != NULL);
4575 if (pgno < txn->mt_next_pgno) {
4577 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4579 DPRINTF(("page %"Z"u not found", pgno));
4581 return MDB_PAGE_NOTFOUND;
4591 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4592 * The cursor is at the root page, set up the rest of it.
4595 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4597 MDB_page *mp = mc->mc_pg[mc->mc_top];
4601 while (IS_BRANCH(mp)) {
4605 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4606 assert(NUMKEYS(mp) > 1);
4607 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4609 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4611 if (flags & MDB_PS_LAST)
4612 i = NUMKEYS(mp) - 1;
4615 node = mdb_node_search(mc, key, &exact);
4617 i = NUMKEYS(mp) - 1;
4619 i = mc->mc_ki[mc->mc_top];
4625 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4628 assert(i < NUMKEYS(mp));
4629 node = NODEPTR(mp, i);
4631 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4634 mc->mc_ki[mc->mc_top] = i;
4635 if ((rc = mdb_cursor_push(mc, mp)))
4638 if (flags & MDB_PS_MODIFY) {
4639 if ((rc = mdb_page_touch(mc)) != 0)
4641 mp = mc->mc_pg[mc->mc_top];
4646 DPRINTF(("internal error, index points to a %02X page!?",
4648 return MDB_CORRUPTED;
4651 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4652 key ? DKEY(key) : "null"));
4653 mc->mc_flags |= C_INITIALIZED;
4654 mc->mc_flags &= ~C_EOF;
4659 /** Search for the lowest key under the current branch page.
4660 * This just bypasses a NUMKEYS check in the current page
4661 * before calling mdb_page_search_root(), because the callers
4662 * are all in situations where the current page is known to
4666 mdb_page_search_lowest(MDB_cursor *mc)
4668 MDB_page *mp = mc->mc_pg[mc->mc_top];
4669 MDB_node *node = NODEPTR(mp, 0);
4672 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4675 mc->mc_ki[mc->mc_top] = 0;
4676 if ((rc = mdb_cursor_push(mc, mp)))
4678 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4681 /** Search for the page a given key should be in.
4682 * Push it and its parent pages on the cursor stack.
4683 * @param[in,out] mc the cursor for this operation.
4684 * @param[in] key the key to search for, or NULL for first/last page.
4685 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4686 * are touched (updated with new page numbers).
4687 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4688 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4689 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4690 * @return 0 on success, non-zero on failure.
4693 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4698 /* Make sure the txn is still viable, then find the root from
4699 * the txn's db table and set it as the root of the cursor's stack.
4701 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4702 DPUTS("transaction has failed, must abort");
4705 /* Make sure we're using an up-to-date root */
4706 if (*mc->mc_dbflag & DB_STALE) {
4708 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4709 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4716 MDB_node *leaf = mdb_node_search(&mc2,
4717 &mc->mc_dbx->md_name, &exact);
4719 return MDB_NOTFOUND;
4720 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4723 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4725 /* The txn may not know this DBI, or another process may
4726 * have dropped and recreated the DB with other flags.
4728 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4729 return MDB_INCOMPATIBLE;
4730 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4732 *mc->mc_dbflag &= ~DB_STALE;
4734 root = mc->mc_db->md_root;
4736 if (root == P_INVALID) { /* Tree is empty. */
4737 DPUTS("tree is empty");
4738 return MDB_NOTFOUND;
4743 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4744 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4750 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4751 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4753 if (flags & MDB_PS_MODIFY) {
4754 if ((rc = mdb_page_touch(mc)))
4758 if (flags & MDB_PS_ROOTONLY)
4761 return mdb_page_search_root(mc, key, flags);
4765 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4767 MDB_txn *txn = mc->mc_txn;
4768 pgno_t pg = mp->mp_pgno;
4769 unsigned x = 0, ovpages = mp->mp_pages;
4770 MDB_env *env = txn->mt_env;
4771 MDB_IDL sl = txn->mt_spill_pgs;
4772 MDB_ID pn = pg << 1;
4775 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4776 /* If the page is dirty or on the spill list we just acquired it,
4777 * so we should give it back to our current free list, if any.
4778 * Otherwise put it onto the list of pages we freed in this txn.
4780 * Won't create me_pghead: me_pglast must be inited along with it.
4781 * Unsupported in nested txns: They would need to hide the page
4782 * range in ancestor txns' dirty and spilled lists.
4784 if (env->me_pghead &&
4786 ((mp->mp_flags & P_DIRTY) ||
4787 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4791 MDB_ID2 *dl, ix, iy;
4792 rc = mdb_midl_need(&env->me_pghead, ovpages);
4795 if (!(mp->mp_flags & P_DIRTY)) {
4796 /* This page is no longer spilled */
4803 /* Remove from dirty list */
4804 dl = txn->mt_u.dirty_list;
4806 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4814 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4815 txn->mt_flags |= MDB_TXN_ERROR;
4816 return MDB_CORRUPTED;
4819 if (!(env->me_flags & MDB_WRITEMAP))
4820 mdb_dpage_free(env, mp);
4822 /* Insert in me_pghead */
4823 mop = env->me_pghead;
4824 j = mop[0] + ovpages;
4825 for (i = mop[0]; i && mop[i] < pg; i--)
4831 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4835 mc->mc_db->md_overflow_pages -= ovpages;
4839 /** Return the data associated with a given node.
4840 * @param[in] txn The transaction for this operation.
4841 * @param[in] leaf The node being read.
4842 * @param[out] data Updated to point to the node's data.
4843 * @return 0 on success, non-zero on failure.
4846 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4848 MDB_page *omp; /* overflow page */
4852 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4853 data->mv_size = NODEDSZ(leaf);
4854 data->mv_data = NODEDATA(leaf);
4858 /* Read overflow data.
4860 data->mv_size = NODEDSZ(leaf);
4861 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4862 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4863 DPRINTF(("read overflow page %"Z"u failed", pgno));
4866 data->mv_data = METADATA(omp);
4872 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4873 MDB_val *key, MDB_val *data)
4882 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4884 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4887 if (txn->mt_flags & MDB_TXN_ERROR)
4890 if (key->mv_size > MDB_MAXKEYSIZE) {
4891 return MDB_BAD_VALSIZE;
4894 mdb_cursor_init(&mc, txn, dbi, &mx);
4895 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4898 /** Find a sibling for a page.
4899 * Replaces the page at the top of the cursor's stack with the
4900 * specified sibling, if one exists.
4901 * @param[in] mc The cursor for this operation.
4902 * @param[in] move_right Non-zero if the right sibling is requested,
4903 * otherwise the left sibling.
4904 * @return 0 on success, non-zero on failure.
4907 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4913 if (mc->mc_snum < 2) {
4914 return MDB_NOTFOUND; /* root has no siblings */
4918 DPRINTF(("parent page is page %"Z"u, index %u",
4919 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4921 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4922 : (mc->mc_ki[mc->mc_top] == 0)) {
4923 DPRINTF(("no more keys left, moving to %s sibling",
4924 move_right ? "right" : "left"));
4925 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4926 /* undo cursor_pop before returning */
4933 mc->mc_ki[mc->mc_top]++;
4935 mc->mc_ki[mc->mc_top]--;
4936 DPRINTF(("just moving to %s index key %u",
4937 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4939 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4941 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4942 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4945 mdb_cursor_push(mc, mp);
4947 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4952 /** Move the cursor to the next data item. */
4954 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4960 if (mc->mc_flags & C_EOF) {
4961 return MDB_NOTFOUND;
4964 assert(mc->mc_flags & C_INITIALIZED);
4966 mp = mc->mc_pg[mc->mc_top];
4968 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4969 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4970 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4971 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4972 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4973 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
4974 if (rc == MDB_SUCCESS)
4975 MDB_GET_KEY(leaf, key);
4980 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4981 if (op == MDB_NEXT_DUP)
4982 return MDB_NOTFOUND;
4986 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4987 if (mc->mc_flags & C_DEL)
4990 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4991 DPUTS("=====> move to next sibling page");
4992 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4993 mc->mc_flags |= C_EOF;
4996 mp = mc->mc_pg[mc->mc_top];
4997 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
4999 mc->mc_ki[mc->mc_top]++;
5002 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5003 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5006 key->mv_size = mc->mc_db->md_pad;
5007 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5011 assert(IS_LEAF(mp));
5012 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5014 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5015 mdb_xcursor_init1(mc, leaf);
5018 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5021 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5022 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5023 if (rc != MDB_SUCCESS)
5028 MDB_GET_KEY(leaf, key);
5032 /** Move the cursor to the previous data item. */
5034 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5040 assert(mc->mc_flags & C_INITIALIZED);
5042 mp = mc->mc_pg[mc->mc_top];
5044 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5045 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5046 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5047 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5048 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5049 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5050 if (rc == MDB_SUCCESS)
5051 MDB_GET_KEY(leaf, key);
5055 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5056 if (op == MDB_PREV_DUP)
5057 return MDB_NOTFOUND;
5062 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5064 if (mc->mc_ki[mc->mc_top] == 0) {
5065 DPUTS("=====> move to prev sibling page");
5066 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5069 mp = mc->mc_pg[mc->mc_top];
5070 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5071 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5073 mc->mc_ki[mc->mc_top]--;
5075 mc->mc_flags &= ~C_EOF;
5077 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5078 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5081 key->mv_size = mc->mc_db->md_pad;
5082 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5086 assert(IS_LEAF(mp));
5087 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5089 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5090 mdb_xcursor_init1(mc, leaf);
5093 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5096 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5097 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5098 if (rc != MDB_SUCCESS)
5103 MDB_GET_KEY(leaf, key);
5107 /** Set the cursor on a specific data item. */
5109 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5110 MDB_cursor_op op, int *exactp)
5114 MDB_node *leaf = NULL;
5119 if (key->mv_size == 0)
5120 return MDB_BAD_VALSIZE;
5123 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5125 /* See if we're already on the right page */
5126 if (mc->mc_flags & C_INITIALIZED) {
5129 mp = mc->mc_pg[mc->mc_top];
5131 mc->mc_ki[mc->mc_top] = 0;
5132 return MDB_NOTFOUND;
5134 if (mp->mp_flags & P_LEAF2) {
5135 nodekey.mv_size = mc->mc_db->md_pad;
5136 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5138 leaf = NODEPTR(mp, 0);
5139 MDB_GET_KEY2(leaf, nodekey);
5141 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5143 /* Probably happens rarely, but first node on the page
5144 * was the one we wanted.
5146 mc->mc_ki[mc->mc_top] = 0;
5153 unsigned int nkeys = NUMKEYS(mp);
5155 if (mp->mp_flags & P_LEAF2) {
5156 nodekey.mv_data = LEAF2KEY(mp,
5157 nkeys-1, nodekey.mv_size);
5159 leaf = NODEPTR(mp, nkeys-1);
5160 MDB_GET_KEY2(leaf, nodekey);
5162 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5164 /* last node was the one we wanted */
5165 mc->mc_ki[mc->mc_top] = nkeys-1;
5171 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5172 /* This is definitely the right page, skip search_page */
5173 if (mp->mp_flags & P_LEAF2) {
5174 nodekey.mv_data = LEAF2KEY(mp,
5175 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5177 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5178 MDB_GET_KEY2(leaf, nodekey);
5180 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5182 /* current node was the one we wanted */
5192 /* If any parents have right-sibs, search.
5193 * Otherwise, there's nothing further.
5195 for (i=0; i<mc->mc_top; i++)
5197 NUMKEYS(mc->mc_pg[i])-1)
5199 if (i == mc->mc_top) {
5200 /* There are no other pages */
5201 mc->mc_ki[mc->mc_top] = nkeys;
5202 return MDB_NOTFOUND;
5206 /* There are no other pages */
5207 mc->mc_ki[mc->mc_top] = 0;
5208 if (op == MDB_SET_RANGE) {
5212 return MDB_NOTFOUND;
5216 rc = mdb_page_search(mc, key, 0);
5217 if (rc != MDB_SUCCESS)
5220 mp = mc->mc_pg[mc->mc_top];
5221 assert(IS_LEAF(mp));
5224 leaf = mdb_node_search(mc, key, exactp);
5225 if (exactp != NULL && !*exactp) {
5226 /* MDB_SET specified and not an exact match. */
5227 return MDB_NOTFOUND;
5231 DPUTS("===> inexact leaf not found, goto sibling");
5232 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5233 return rc; /* no entries matched */
5234 mp = mc->mc_pg[mc->mc_top];
5235 assert(IS_LEAF(mp));
5236 leaf = NODEPTR(mp, 0);
5240 mc->mc_flags |= C_INITIALIZED;
5241 mc->mc_flags &= ~C_EOF;
5244 key->mv_size = mc->mc_db->md_pad;
5245 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5249 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5250 mdb_xcursor_init1(mc, leaf);
5253 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5254 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5255 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5258 if (op == MDB_GET_BOTH) {
5264 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5265 if (rc != MDB_SUCCESS)
5268 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5270 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5272 rc = mc->mc_dbx->md_dcmp(data, &d2);
5274 if (op == MDB_GET_BOTH || rc > 0)
5275 return MDB_NOTFOUND;
5281 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5282 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5287 /* The key already matches in all other cases */
5288 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5289 MDB_GET_KEY(leaf, key);
5290 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5295 /** Move the cursor to the first item in the database. */
5297 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5303 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5305 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5306 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5307 if (rc != MDB_SUCCESS)
5310 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5312 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5313 mc->mc_flags |= C_INITIALIZED;
5314 mc->mc_flags &= ~C_EOF;
5316 mc->mc_ki[mc->mc_top] = 0;
5318 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5319 key->mv_size = mc->mc_db->md_pad;
5320 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5325 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5326 mdb_xcursor_init1(mc, leaf);
5327 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5331 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5335 MDB_GET_KEY(leaf, key);
5339 /** Move the cursor to the last item in the database. */
5341 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5347 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5349 if (!(mc->mc_flags & C_EOF)) {
5351 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5352 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5353 if (rc != MDB_SUCCESS)
5356 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5359 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5360 mc->mc_flags |= C_INITIALIZED|C_EOF;
5361 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5363 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5364 key->mv_size = mc->mc_db->md_pad;
5365 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5370 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5371 mdb_xcursor_init1(mc, leaf);
5372 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5376 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5381 MDB_GET_KEY(leaf, key);
5386 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5391 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5395 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5399 case MDB_GET_CURRENT:
5400 if (!(mc->mc_flags & C_INITIALIZED)) {
5403 MDB_page *mp = mc->mc_pg[mc->mc_top];
5405 mc->mc_ki[mc->mc_top] = 0;
5411 key->mv_size = mc->mc_db->md_pad;
5412 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5414 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5415 MDB_GET_KEY(leaf, key);
5417 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5418 if (mc->mc_flags & C_DEL)
5419 mdb_xcursor_init1(mc, leaf);
5420 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5422 rc = mdb_node_read(mc->mc_txn, leaf, data);
5429 case MDB_GET_BOTH_RANGE:
5434 if (mc->mc_xcursor == NULL) {
5435 rc = MDB_INCOMPATIBLE;
5444 } else if (key->mv_size > MDB_MAXKEYSIZE) {
5445 rc = MDB_BAD_VALSIZE;
5446 } else if (op == MDB_SET_RANGE)
5447 rc = mdb_cursor_set(mc, key, data, op, NULL);
5449 rc = mdb_cursor_set(mc, key, data, op, &exact);
5451 case MDB_GET_MULTIPLE:
5452 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5456 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5457 rc = MDB_INCOMPATIBLE;
5461 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5462 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5465 case MDB_NEXT_MULTIPLE:
5470 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5471 rc = MDB_INCOMPATIBLE;
5474 if (!(mc->mc_flags & C_INITIALIZED))
5475 rc = mdb_cursor_first(mc, key, data);
5477 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5478 if (rc == MDB_SUCCESS) {
5479 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5482 mx = &mc->mc_xcursor->mx_cursor;
5483 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5485 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5486 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5494 case MDB_NEXT_NODUP:
5495 if (!(mc->mc_flags & C_INITIALIZED))
5496 rc = mdb_cursor_first(mc, key, data);
5498 rc = mdb_cursor_next(mc, key, data, op);
5502 case MDB_PREV_NODUP:
5503 if (!(mc->mc_flags & C_INITIALIZED)) {
5504 rc = mdb_cursor_last(mc, key, data);
5507 mc->mc_flags |= C_INITIALIZED;
5508 mc->mc_ki[mc->mc_top]++;
5510 rc = mdb_cursor_prev(mc, key, data, op);
5513 rc = mdb_cursor_first(mc, key, data);
5516 mfunc = mdb_cursor_first;
5518 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5522 if (mc->mc_xcursor == NULL) {
5523 rc = MDB_INCOMPATIBLE;
5526 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5530 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5533 rc = mdb_cursor_last(mc, key, data);
5536 mfunc = mdb_cursor_last;
5539 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5544 if (mc->mc_flags & C_DEL)
5545 mc->mc_flags ^= C_DEL;
5550 /** Touch all the pages in the cursor stack.
5551 * Makes sure all the pages are writable, before attempting a write operation.
5552 * @param[in] mc The cursor to operate on.
5555 mdb_cursor_touch(MDB_cursor *mc)
5559 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5562 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5563 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5566 *mc->mc_dbflag |= DB_DIRTY;
5568 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5569 rc = mdb_page_touch(mc);
5573 mc->mc_top = mc->mc_snum-1;
5577 /** Do not spill pages to disk if txn is getting full, may fail instead */
5578 #define MDB_NOSPILL 0x8000
5581 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5584 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5585 MDB_node *leaf = NULL;
5586 MDB_val xdata, *rdata, dkey;
5589 int do_sub = 0, insert = 0;
5590 unsigned int mcount = 0, dcount = 0, nospill;
5594 char dbuf[MDB_MAXKEYSIZE+1];
5595 unsigned int nflags;
5598 /* Check this first so counter will always be zero on any
5601 if (flags & MDB_MULTIPLE) {
5602 dcount = data[1].mv_size;
5603 data[1].mv_size = 0;
5604 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5605 return MDB_INCOMPATIBLE;
5608 nospill = flags & MDB_NOSPILL;
5609 flags &= ~MDB_NOSPILL;
5611 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5612 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5614 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5615 return MDB_BAD_VALSIZE;
5617 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5618 return MDB_BAD_VALSIZE;
5620 #if SIZE_MAX > MAXDATASIZE
5621 if (data->mv_size > MAXDATASIZE)
5622 return MDB_BAD_VALSIZE;
5625 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5626 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5630 if (flags == MDB_CURRENT) {
5631 if (!(mc->mc_flags & C_INITIALIZED))
5634 } else if (mc->mc_db->md_root == P_INVALID) {
5635 /* new database, cursor has nothing to point to */
5637 mc->mc_flags &= ~C_INITIALIZED;
5642 if (flags & MDB_APPEND) {
5644 rc = mdb_cursor_last(mc, &k2, &d2);
5646 rc = mc->mc_dbx->md_cmp(key, &k2);
5649 mc->mc_ki[mc->mc_top]++;
5651 /* new key is <= last key */
5656 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5658 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5659 DPRINTF(("duplicate key [%s]", DKEY(key)));
5661 return MDB_KEYEXIST;
5663 if (rc && rc != MDB_NOTFOUND)
5667 if (mc->mc_flags & C_DEL)
5668 mc->mc_flags ^= C_DEL;
5670 /* Cursor is positioned, check for room in the dirty list */
5672 if (flags & MDB_MULTIPLE) {
5674 xdata.mv_size = data->mv_size * dcount;
5678 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5682 if (rc == MDB_NO_ROOT) {
5684 /* new database, write a root leaf page */
5685 DPUTS("allocating new root leaf page");
5686 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5689 mdb_cursor_push(mc, np);
5690 mc->mc_db->md_root = np->mp_pgno;
5691 mc->mc_db->md_depth++;
5692 *mc->mc_dbflag |= DB_DIRTY;
5693 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5695 np->mp_flags |= P_LEAF2;
5696 mc->mc_flags |= C_INITIALIZED;
5698 /* make sure all cursor pages are writable */
5699 rc2 = mdb_cursor_touch(mc);
5704 /* The key already exists */
5705 if (rc == MDB_SUCCESS) {
5706 /* there's only a key anyway, so this is a no-op */
5707 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5708 unsigned int ksize = mc->mc_db->md_pad;
5709 if (key->mv_size != ksize)
5710 return MDB_BAD_VALSIZE;
5711 if (flags == MDB_CURRENT) {
5712 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5713 memcpy(ptr, key->mv_data, ksize);
5718 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5721 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5722 /* Was a single item before, must convert now */
5724 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5725 /* Just overwrite the current item */
5726 if (flags == MDB_CURRENT)
5729 dkey.mv_size = NODEDSZ(leaf);
5730 dkey.mv_data = NODEDATA(leaf);
5731 #if UINT_MAX < SIZE_MAX
5732 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5733 #ifdef MISALIGNED_OK
5734 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5736 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5739 /* if data matches, skip it */
5740 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5741 if (flags & MDB_NODUPDATA)
5743 else if (flags & MDB_MULTIPLE)
5750 /* create a fake page for the dup items */
5751 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5752 dkey.mv_data = dbuf;
5753 fp = (MDB_page *)&pbuf;
5754 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5755 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5756 fp->mp_lower = PAGEHDRSZ;
5757 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5758 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5759 fp->mp_flags |= P_LEAF2;
5760 fp->mp_pad = data->mv_size;
5761 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5763 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5764 (dkey.mv_size & 1) + (data->mv_size & 1);
5766 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5769 xdata.mv_size = fp->mp_upper;
5774 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5775 /* See if we need to convert from fake page to subDB */
5777 unsigned int offset;
5781 fp = NODEDATA(leaf);
5782 if (flags == MDB_CURRENT) {
5784 fp->mp_flags |= P_DIRTY;
5785 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5786 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5790 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5791 offset = fp->mp_pad;
5792 if (SIZELEFT(fp) >= offset)
5794 offset *= 4; /* space for 4 more */
5796 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5798 offset += offset & 1;
5799 fp_flags = fp->mp_flags;
5800 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5801 offset >= mc->mc_txn->mt_env->me_nodemax) {
5802 /* yes, convert it */
5804 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5805 dummy.md_pad = fp->mp_pad;
5806 dummy.md_flags = MDB_DUPFIXED;
5807 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5808 dummy.md_flags |= MDB_INTEGERKEY;
5811 dummy.md_branch_pages = 0;
5812 dummy.md_leaf_pages = 1;
5813 dummy.md_overflow_pages = 0;
5814 dummy.md_entries = NUMKEYS(fp);
5816 xdata.mv_size = sizeof(MDB_db);
5817 xdata.mv_data = &dummy;
5818 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5820 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5821 flags |= F_DUPDATA|F_SUBDATA;
5822 dummy.md_root = mp->mp_pgno;
5823 fp_flags &= ~P_SUBP;
5825 /* no, just grow it */
5827 xdata.mv_size = NODEDSZ(leaf) + offset;
5828 xdata.mv_data = &pbuf;
5829 mp = (MDB_page *)&pbuf;
5830 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5833 mp->mp_flags = fp_flags | P_DIRTY;
5834 mp->mp_pad = fp->mp_pad;
5835 mp->mp_lower = fp->mp_lower;
5836 mp->mp_upper = fp->mp_upper + offset;
5838 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5840 nsize = NODEDSZ(leaf) - fp->mp_upper;
5841 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5842 for (i=0; i<NUMKEYS(fp); i++)
5843 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5845 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5849 /* data is on sub-DB, just store it */
5850 flags |= F_DUPDATA|F_SUBDATA;
5854 /* overflow page overwrites need special handling */
5855 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5858 unsigned psize = mc->mc_txn->mt_env->me_psize;
5859 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5861 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5862 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5864 ovpages = omp->mp_pages;
5866 /* Is the ov page large enough? */
5867 if (ovpages >= dpages) {
5868 if (!(omp->mp_flags & P_DIRTY) &&
5869 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5871 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5874 level = 0; /* dirty in this txn or clean */
5877 if (omp->mp_flags & P_DIRTY) {
5878 /* yes, overwrite it. Note in this case we don't
5879 * bother to try shrinking the page if the new data
5880 * is smaller than the overflow threshold.
5883 /* It is writable only in a parent txn */
5884 size_t sz = (size_t) psize * ovpages, off;
5885 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5891 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5892 if (!(flags & MDB_RESERVE)) {
5893 /* Copy end of page, adjusting alignment so
5894 * compiler may copy words instead of bytes.
5896 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5897 memcpy((size_t *)((char *)np + off),
5898 (size_t *)((char *)omp + off), sz - off);
5901 memcpy(np, omp, sz); /* Copy beginning of page */
5904 SETDSZ(leaf, data->mv_size);
5905 if (F_ISSET(flags, MDB_RESERVE))
5906 data->mv_data = METADATA(omp);
5908 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5912 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5914 } else if (NODEDSZ(leaf) == data->mv_size) {
5915 /* same size, just replace it. Note that we could
5916 * also reuse this node if the new data is smaller,
5917 * but instead we opt to shrink the node in that case.
5919 if (F_ISSET(flags, MDB_RESERVE))
5920 data->mv_data = NODEDATA(leaf);
5921 else if (data->mv_size)
5922 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5924 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5927 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5928 mc->mc_db->md_entries--;
5930 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5937 nflags = flags & NODE_ADD_FLAGS;
5938 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5939 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5940 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5941 nflags &= ~MDB_APPEND;
5943 nflags |= MDB_SPLIT_REPLACE;
5944 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5946 /* There is room already in this leaf page. */
5947 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5948 if (rc == 0 && !do_sub && insert) {
5949 /* Adjust other cursors pointing to mp */
5950 MDB_cursor *m2, *m3;
5951 MDB_dbi dbi = mc->mc_dbi;
5952 unsigned i = mc->mc_top;
5953 MDB_page *mp = mc->mc_pg[i];
5955 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5956 if (mc->mc_flags & C_SUB)
5957 m3 = &m2->mc_xcursor->mx_cursor;
5960 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5961 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5968 if (rc != MDB_SUCCESS)
5969 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5971 /* Now store the actual data in the child DB. Note that we're
5972 * storing the user data in the keys field, so there are strict
5973 * size limits on dupdata. The actual data fields of the child
5974 * DB are all zero size.
5981 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5982 if (flags & MDB_CURRENT) {
5983 xflags = MDB_CURRENT|MDB_NOSPILL;
5985 mdb_xcursor_init1(mc, leaf);
5986 xflags = (flags & MDB_NODUPDATA) ?
5987 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5989 /* converted, write the original data first */
5991 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5995 /* Adjust other cursors pointing to mp */
5997 unsigned i = mc->mc_top;
5998 MDB_page *mp = mc->mc_pg[i];
6000 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6001 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6002 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6003 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6004 mdb_xcursor_init1(m2, leaf);
6008 /* we've done our job */
6011 if (flags & MDB_APPENDDUP)
6012 xflags |= MDB_APPEND;
6013 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6014 if (flags & F_SUBDATA) {
6015 void *db = NODEDATA(leaf);
6016 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6019 /* sub-writes might have failed so check rc again.
6020 * Don't increment count if we just replaced an existing item.
6022 if (!rc && !(flags & MDB_CURRENT))
6023 mc->mc_db->md_entries++;
6024 if (flags & MDB_MULTIPLE) {
6028 /* let caller know how many succeeded, if any */
6029 data[1].mv_size = mcount;
6030 if (mcount < dcount) {
6031 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6032 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6039 /* If we succeeded and the key didn't exist before, make sure
6040 * the cursor is marked valid.
6043 mc->mc_flags |= C_INITIALIZED;
6048 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6053 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6054 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6056 if (!(mc->mc_flags & C_INITIALIZED))
6059 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6062 rc = mdb_cursor_touch(mc);
6066 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6068 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6069 if (!(flags & MDB_NODUPDATA)) {
6070 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6071 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6073 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6074 /* If sub-DB still has entries, we're done */
6075 if (mc->mc_xcursor->mx_db.md_entries) {
6076 if (leaf->mn_flags & F_SUBDATA) {
6077 /* update subDB info */
6078 void *db = NODEDATA(leaf);
6079 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6082 /* shrink fake page */
6083 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6084 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6085 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6086 /* fix other sub-DB cursors pointed at this fake page */
6087 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6088 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6089 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6090 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6091 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6094 mc->mc_db->md_entries--;
6095 mc->mc_flags |= C_DEL;
6098 /* otherwise fall thru and delete the sub-DB */
6101 if (leaf->mn_flags & F_SUBDATA) {
6102 /* add all the child DB's pages to the free list */
6103 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6104 if (rc == MDB_SUCCESS) {
6105 mc->mc_db->md_entries -=
6106 mc->mc_xcursor->mx_db.md_entries;
6111 return mdb_cursor_del0(mc, leaf);
6114 /** Allocate and initialize new pages for a database.
6115 * @param[in] mc a cursor on the database being added to.
6116 * @param[in] flags flags defining what type of page is being allocated.
6117 * @param[in] num the number of pages to allocate. This is usually 1,
6118 * unless allocating overflow pages for a large record.
6119 * @param[out] mp Address of a page, or NULL on failure.
6120 * @return 0 on success, non-zero on failure.
6123 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6128 if ((rc = mdb_page_alloc(mc, num, &np)))
6130 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6131 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6132 np->mp_flags = flags | P_DIRTY;
6133 np->mp_lower = PAGEHDRSZ;
6134 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6137 mc->mc_db->md_branch_pages++;
6138 else if (IS_LEAF(np))
6139 mc->mc_db->md_leaf_pages++;
6140 else if (IS_OVERFLOW(np)) {
6141 mc->mc_db->md_overflow_pages += num;
6149 /** Calculate the size of a leaf node.
6150 * The size depends on the environment's page size; if a data item
6151 * is too large it will be put onto an overflow page and the node
6152 * size will only include the key and not the data. Sizes are always
6153 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6154 * of the #MDB_node headers.
6155 * @param[in] env The environment handle.
6156 * @param[in] key The key for the node.
6157 * @param[in] data The data for the node.
6158 * @return The number of bytes needed to store the node.
6161 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6165 sz = LEAFSIZE(key, data);
6166 if (sz >= env->me_nodemax) {
6167 /* put on overflow page */
6168 sz -= data->mv_size - sizeof(pgno_t);
6172 return sz + sizeof(indx_t);
6175 /** Calculate the size of a branch node.
6176 * The size should depend on the environment's page size but since
6177 * we currently don't support spilling large keys onto overflow
6178 * pages, it's simply the size of the #MDB_node header plus the
6179 * size of the key. Sizes are always rounded up to an even number
6180 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6181 * @param[in] env The environment handle.
6182 * @param[in] key The key for the node.
6183 * @return The number of bytes needed to store the node.
6186 mdb_branch_size(MDB_env *env, MDB_val *key)
6191 if (sz >= env->me_nodemax) {
6192 /* put on overflow page */
6193 /* not implemented */
6194 /* sz -= key->size - sizeof(pgno_t); */
6197 return sz + sizeof(indx_t);
6200 /** Add a node to the page pointed to by the cursor.
6201 * @param[in] mc The cursor for this operation.
6202 * @param[in] indx The index on the page where the new node should be added.
6203 * @param[in] key The key for the new node.
6204 * @param[in] data The data for the new node, if any.
6205 * @param[in] pgno The page number, if adding a branch node.
6206 * @param[in] flags Flags for the node.
6207 * @return 0 on success, non-zero on failure. Possible errors are:
6209 * <li>ENOMEM - failed to allocate overflow pages for the node.
6210 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6211 * should never happen since all callers already calculate the
6212 * page's free space before calling this function.
6216 mdb_node_add(MDB_cursor *mc, indx_t indx,
6217 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6220 size_t node_size = NODESIZE;
6223 MDB_page *mp = mc->mc_pg[mc->mc_top];
6224 MDB_page *ofp = NULL; /* overflow page */
6227 assert(mp->mp_upper >= mp->mp_lower);
6229 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6230 IS_LEAF(mp) ? "leaf" : "branch",
6231 IS_SUBP(mp) ? "sub-" : "",
6232 mp->mp_pgno, indx, data ? data->mv_size : 0,
6233 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6236 /* Move higher keys up one slot. */
6237 int ksize = mc->mc_db->md_pad, dif;
6238 char *ptr = LEAF2KEY(mp, indx, ksize);
6239 dif = NUMKEYS(mp) - indx;
6241 memmove(ptr+ksize, ptr, dif*ksize);
6242 /* insert new key */
6243 memcpy(ptr, key->mv_data, ksize);
6245 /* Just using these for counting */
6246 mp->mp_lower += sizeof(indx_t);
6247 mp->mp_upper -= ksize - sizeof(indx_t);
6252 node_size += key->mv_size;
6256 if (F_ISSET(flags, F_BIGDATA)) {
6257 /* Data already on overflow page. */
6258 node_size += sizeof(pgno_t);
6259 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6260 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6262 /* Put data on overflow page. */
6263 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6264 data->mv_size, node_size+data->mv_size));
6265 node_size += sizeof(pgno_t);
6266 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6268 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6271 node_size += data->mv_size;
6274 node_size += node_size & 1;
6276 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6277 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6278 mp->mp_pgno, NUMKEYS(mp)));
6279 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6280 mp->mp_upper - mp->mp_lower));
6281 DPRINTF(("node size = %"Z"u", node_size));
6282 return MDB_PAGE_FULL;
6285 /* Move higher pointers up one slot. */
6286 for (i = NUMKEYS(mp); i > indx; i--)
6287 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6289 /* Adjust free space offsets. */
6290 ofs = mp->mp_upper - node_size;
6291 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6292 mp->mp_ptrs[indx] = ofs;
6294 mp->mp_lower += sizeof(indx_t);
6296 /* Write the node data. */
6297 node = NODEPTR(mp, indx);
6298 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6299 node->mn_flags = flags;
6301 SETDSZ(node,data->mv_size);
6306 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6311 if (F_ISSET(flags, F_BIGDATA))
6312 memcpy(node->mn_data + key->mv_size, data->mv_data,
6314 else if (F_ISSET(flags, MDB_RESERVE))
6315 data->mv_data = node->mn_data + key->mv_size;
6317 memcpy(node->mn_data + key->mv_size, data->mv_data,
6320 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6322 if (F_ISSET(flags, MDB_RESERVE))
6323 data->mv_data = METADATA(ofp);
6325 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6332 /** Delete the specified node from a page.
6333 * @param[in] mp The page to operate on.
6334 * @param[in] indx The index of the node to delete.
6335 * @param[in] ksize The size of a node. Only used if the page is
6336 * part of a #MDB_DUPFIXED database.
6339 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6342 indx_t i, j, numkeys, ptr;
6349 COPY_PGNO(pgno, mp->mp_pgno);
6350 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6351 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6354 assert(indx < NUMKEYS(mp));
6357 int x = NUMKEYS(mp) - 1 - indx;
6358 base = LEAF2KEY(mp, indx, ksize);
6360 memmove(base, base + ksize, x * ksize);
6361 mp->mp_lower -= sizeof(indx_t);
6362 mp->mp_upper += ksize - sizeof(indx_t);
6366 node = NODEPTR(mp, indx);
6367 sz = NODESIZE + node->mn_ksize;
6369 if (F_ISSET(node->mn_flags, F_BIGDATA))
6370 sz += sizeof(pgno_t);
6372 sz += NODEDSZ(node);
6376 ptr = mp->mp_ptrs[indx];
6377 numkeys = NUMKEYS(mp);
6378 for (i = j = 0; i < numkeys; i++) {
6380 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6381 if (mp->mp_ptrs[i] < ptr)
6382 mp->mp_ptrs[j] += sz;
6387 base = (char *)mp + mp->mp_upper;
6388 memmove(base + sz, base, ptr - mp->mp_upper);
6390 mp->mp_lower -= sizeof(indx_t);
6394 /** Compact the main page after deleting a node on a subpage.
6395 * @param[in] mp The main page to operate on.
6396 * @param[in] indx The index of the subpage on the main page.
6399 mdb_node_shrink(MDB_page *mp, indx_t indx)
6406 indx_t i, numkeys, ptr;
6408 node = NODEPTR(mp, indx);
6409 sp = (MDB_page *)NODEDATA(node);
6410 osize = NODEDSZ(node);
6412 delta = sp->mp_upper - sp->mp_lower;
6413 SETDSZ(node, osize - delta);
6414 xp = (MDB_page *)((char *)sp + delta);
6416 /* shift subpage upward */
6418 nsize = NUMKEYS(sp) * sp->mp_pad;
6419 memmove(METADATA(xp), METADATA(sp), nsize);
6422 nsize = osize - sp->mp_upper;
6423 numkeys = NUMKEYS(sp);
6424 for (i=numkeys-1; i>=0; i--)
6425 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6427 xp->mp_upper = sp->mp_lower;
6428 xp->mp_lower = sp->mp_lower;
6429 xp->mp_flags = sp->mp_flags;
6430 xp->mp_pad = sp->mp_pad;
6431 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6433 /* shift lower nodes upward */
6434 ptr = mp->mp_ptrs[indx];
6435 numkeys = NUMKEYS(mp);
6436 for (i = 0; i < numkeys; i++) {
6437 if (mp->mp_ptrs[i] <= ptr)
6438 mp->mp_ptrs[i] += delta;
6441 base = (char *)mp + mp->mp_upper;
6442 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6443 mp->mp_upper += delta;
6446 /** Initial setup of a sorted-dups cursor.
6447 * Sorted duplicates are implemented as a sub-database for the given key.
6448 * The duplicate data items are actually keys of the sub-database.
6449 * Operations on the duplicate data items are performed using a sub-cursor
6450 * initialized when the sub-database is first accessed. This function does
6451 * the preliminary setup of the sub-cursor, filling in the fields that
6452 * depend only on the parent DB.
6453 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6456 mdb_xcursor_init0(MDB_cursor *mc)
6458 MDB_xcursor *mx = mc->mc_xcursor;
6460 mx->mx_cursor.mc_xcursor = NULL;
6461 mx->mx_cursor.mc_txn = mc->mc_txn;
6462 mx->mx_cursor.mc_db = &mx->mx_db;
6463 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6464 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6465 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6466 mx->mx_cursor.mc_snum = 0;
6467 mx->mx_cursor.mc_top = 0;
6468 mx->mx_cursor.mc_flags = C_SUB;
6469 mx->mx_dbx.md_name.mv_size = 0;
6470 mx->mx_dbx.md_name.mv_data = NULL;
6471 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6472 mx->mx_dbx.md_dcmp = NULL;
6473 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6476 /** Final setup of a sorted-dups cursor.
6477 * Sets up the fields that depend on the data from the main cursor.
6478 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6479 * @param[in] node The data containing the #MDB_db record for the
6480 * sorted-dup database.
6483 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6485 MDB_xcursor *mx = mc->mc_xcursor;
6487 if (node->mn_flags & F_SUBDATA) {
6488 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6489 mx->mx_cursor.mc_pg[0] = 0;
6490 mx->mx_cursor.mc_snum = 0;
6491 mx->mx_cursor.mc_flags = C_SUB;
6493 MDB_page *fp = NODEDATA(node);
6494 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6495 mx->mx_db.md_flags = 0;
6496 mx->mx_db.md_depth = 1;
6497 mx->mx_db.md_branch_pages = 0;
6498 mx->mx_db.md_leaf_pages = 1;
6499 mx->mx_db.md_overflow_pages = 0;
6500 mx->mx_db.md_entries = NUMKEYS(fp);
6501 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6502 mx->mx_cursor.mc_snum = 1;
6503 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6504 mx->mx_cursor.mc_top = 0;
6505 mx->mx_cursor.mc_pg[0] = fp;
6506 mx->mx_cursor.mc_ki[0] = 0;
6507 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6508 mx->mx_db.md_flags = MDB_DUPFIXED;
6509 mx->mx_db.md_pad = fp->mp_pad;
6510 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6511 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6514 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6515 mx->mx_db.md_root));
6516 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6517 #if UINT_MAX < SIZE_MAX
6518 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6519 #ifdef MISALIGNED_OK
6520 mx->mx_dbx.md_cmp = mdb_cmp_long;
6522 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6527 /** Initialize a cursor for a given transaction and database. */
6529 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6532 mc->mc_backup = NULL;
6535 mc->mc_db = &txn->mt_dbs[dbi];
6536 mc->mc_dbx = &txn->mt_dbxs[dbi];
6537 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6542 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6544 mc->mc_xcursor = mx;
6545 mdb_xcursor_init0(mc);
6547 mc->mc_xcursor = NULL;
6549 if (*mc->mc_dbflag & DB_STALE) {
6550 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6555 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6558 size_t size = sizeof(MDB_cursor);
6560 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6563 if (txn->mt_flags & MDB_TXN_ERROR)
6566 /* Allow read access to the freelist */
6567 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6570 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6571 size += sizeof(MDB_xcursor);
6573 if ((mc = malloc(size)) != NULL) {
6574 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6575 if (txn->mt_cursors) {
6576 mc->mc_next = txn->mt_cursors[dbi];
6577 txn->mt_cursors[dbi] = mc;
6578 mc->mc_flags |= C_UNTRACK;
6590 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6592 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6595 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6598 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6602 /* Return the count of duplicate data items for the current key */
6604 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6608 if (mc == NULL || countp == NULL)
6611 if (mc->mc_xcursor == NULL)
6612 return MDB_INCOMPATIBLE;
6614 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6615 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6618 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6621 *countp = mc->mc_xcursor->mx_db.md_entries;
6627 mdb_cursor_close(MDB_cursor *mc)
6629 if (mc && !mc->mc_backup) {
6630 /* remove from txn, if tracked */
6631 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6632 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6633 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6635 *prev = mc->mc_next;
6642 mdb_cursor_txn(MDB_cursor *mc)
6644 if (!mc) return NULL;
6649 mdb_cursor_dbi(MDB_cursor *mc)
6655 /** Replace the key for a node with a new key.
6656 * @param[in] mc Cursor pointing to the node to operate on.
6657 * @param[in] key The new key to use.
6658 * @return 0 on success, non-zero on failure.
6661 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6668 indx_t ptr, i, numkeys, indx;
6671 indx = mc->mc_ki[mc->mc_top];
6672 mp = mc->mc_pg[mc->mc_top];
6673 node = NODEPTR(mp, indx);
6674 ptr = mp->mp_ptrs[indx];
6678 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6679 k2.mv_data = NODEKEY(node);
6680 k2.mv_size = node->mn_ksize;
6681 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6683 mdb_dkey(&k2, kbuf2),
6689 delta0 = delta = key->mv_size - node->mn_ksize;
6691 /* Must be 2-byte aligned. If new key is
6692 * shorter by 1, the shift will be skipped.
6694 delta += (delta & 1);
6696 if (delta > 0 && SIZELEFT(mp) < delta) {
6698 /* not enough space left, do a delete and split */
6699 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6700 pgno = NODEPGNO(node);
6701 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6702 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6705 numkeys = NUMKEYS(mp);
6706 for (i = 0; i < numkeys; i++) {
6707 if (mp->mp_ptrs[i] <= ptr)
6708 mp->mp_ptrs[i] -= delta;
6711 base = (char *)mp + mp->mp_upper;
6712 len = ptr - mp->mp_upper + NODESIZE;
6713 memmove(base - delta, base, len);
6714 mp->mp_upper -= delta;
6716 node = NODEPTR(mp, indx);
6719 /* But even if no shift was needed, update ksize */
6721 node->mn_ksize = key->mv_size;
6724 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6730 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6732 /** Move a node from csrc to cdst.
6735 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6742 unsigned short flags;
6746 /* Mark src and dst as dirty. */
6747 if ((rc = mdb_page_touch(csrc)) ||
6748 (rc = mdb_page_touch(cdst)))
6751 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6752 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6753 key.mv_size = csrc->mc_db->md_pad;
6754 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6756 data.mv_data = NULL;
6760 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6761 assert(!((long)srcnode&1));
6762 srcpg = NODEPGNO(srcnode);
6763 flags = srcnode->mn_flags;
6764 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6765 unsigned int snum = csrc->mc_snum;
6767 /* must find the lowest key below src */
6768 mdb_page_search_lowest(csrc);
6769 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6770 key.mv_size = csrc->mc_db->md_pad;
6771 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6773 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6774 key.mv_size = NODEKSZ(s2);
6775 key.mv_data = NODEKEY(s2);
6777 csrc->mc_snum = snum--;
6778 csrc->mc_top = snum;
6780 key.mv_size = NODEKSZ(srcnode);
6781 key.mv_data = NODEKEY(srcnode);
6783 data.mv_size = NODEDSZ(srcnode);
6784 data.mv_data = NODEDATA(srcnode);
6786 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6787 unsigned int snum = cdst->mc_snum;
6790 /* must find the lowest key below dst */
6791 mdb_page_search_lowest(cdst);
6792 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6793 bkey.mv_size = cdst->mc_db->md_pad;
6794 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6796 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6797 bkey.mv_size = NODEKSZ(s2);
6798 bkey.mv_data = NODEKEY(s2);
6800 cdst->mc_snum = snum--;
6801 cdst->mc_top = snum;
6802 mdb_cursor_copy(cdst, &mn);
6804 rc = mdb_update_key(&mn, &bkey);
6809 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6810 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6811 csrc->mc_ki[csrc->mc_top],
6813 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6814 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6816 /* Add the node to the destination page.
6818 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6819 if (rc != MDB_SUCCESS)
6822 /* Delete the node from the source page.
6824 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6827 /* Adjust other cursors pointing to mp */
6828 MDB_cursor *m2, *m3;
6829 MDB_dbi dbi = csrc->mc_dbi;
6830 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6832 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6833 if (csrc->mc_flags & C_SUB)
6834 m3 = &m2->mc_xcursor->mx_cursor;
6837 if (m3 == csrc) continue;
6838 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6839 csrc->mc_ki[csrc->mc_top]) {
6840 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6841 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6846 /* Update the parent separators.
6848 if (csrc->mc_ki[csrc->mc_top] == 0) {
6849 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6850 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6851 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6853 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6854 key.mv_size = NODEKSZ(srcnode);
6855 key.mv_data = NODEKEY(srcnode);
6857 DPRINTF(("update separator for source page %"Z"u to [%s]",
6858 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6859 mdb_cursor_copy(csrc, &mn);
6862 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6865 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6867 indx_t ix = csrc->mc_ki[csrc->mc_top];
6868 nullkey.mv_size = 0;
6869 csrc->mc_ki[csrc->mc_top] = 0;
6870 rc = mdb_update_key(csrc, &nullkey);
6871 csrc->mc_ki[csrc->mc_top] = ix;
6872 assert(rc == MDB_SUCCESS);
6876 if (cdst->mc_ki[cdst->mc_top] == 0) {
6877 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6878 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6879 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6881 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6882 key.mv_size = NODEKSZ(srcnode);
6883 key.mv_data = NODEKEY(srcnode);
6885 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6886 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6887 mdb_cursor_copy(cdst, &mn);
6890 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6893 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6895 indx_t ix = cdst->mc_ki[cdst->mc_top];
6896 nullkey.mv_size = 0;
6897 cdst->mc_ki[cdst->mc_top] = 0;
6898 rc = mdb_update_key(cdst, &nullkey);
6899 cdst->mc_ki[cdst->mc_top] = ix;
6900 assert(rc == MDB_SUCCESS);
6907 /** Merge one page into another.
6908 * The nodes from the page pointed to by \b csrc will
6909 * be copied to the page pointed to by \b cdst and then
6910 * the \b csrc page will be freed.
6911 * @param[in] csrc Cursor pointing to the source page.
6912 * @param[in] cdst Cursor pointing to the destination page.
6915 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6923 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6924 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6926 assert(csrc->mc_snum > 1); /* can't merge root page */
6927 assert(cdst->mc_snum > 1);
6929 /* Mark dst as dirty. */
6930 if ((rc = mdb_page_touch(cdst)))
6933 /* Move all nodes from src to dst.
6935 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6936 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6937 key.mv_size = csrc->mc_db->md_pad;
6938 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6939 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6940 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6941 if (rc != MDB_SUCCESS)
6943 key.mv_data = (char *)key.mv_data + key.mv_size;
6946 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6947 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6948 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6949 unsigned int snum = csrc->mc_snum;
6951 /* must find the lowest key below src */
6952 mdb_page_search_lowest(csrc);
6953 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6954 key.mv_size = csrc->mc_db->md_pad;
6955 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6957 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6958 key.mv_size = NODEKSZ(s2);
6959 key.mv_data = NODEKEY(s2);
6961 csrc->mc_snum = snum--;
6962 csrc->mc_top = snum;
6964 key.mv_size = srcnode->mn_ksize;
6965 key.mv_data = NODEKEY(srcnode);
6968 data.mv_size = NODEDSZ(srcnode);
6969 data.mv_data = NODEDATA(srcnode);
6970 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6971 if (rc != MDB_SUCCESS)
6976 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6977 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
6978 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
6980 /* Unlink the src page from parent and add to free list.
6982 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6983 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6986 rc = mdb_update_key(csrc, &key);
6992 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6993 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6996 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6997 csrc->mc_db->md_leaf_pages--;
6999 csrc->mc_db->md_branch_pages--;
7001 /* Adjust other cursors pointing to mp */
7002 MDB_cursor *m2, *m3;
7003 MDB_dbi dbi = csrc->mc_dbi;
7004 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7006 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7007 if (csrc->mc_flags & C_SUB)
7008 m3 = &m2->mc_xcursor->mx_cursor;
7011 if (m3 == csrc) continue;
7012 if (m3->mc_snum < csrc->mc_snum) continue;
7013 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7014 m3->mc_pg[csrc->mc_top] = mp;
7015 m3->mc_ki[csrc->mc_top] += nkeys;
7019 mdb_cursor_pop(csrc);
7021 return mdb_rebalance(csrc);
7024 /** Copy the contents of a cursor.
7025 * @param[in] csrc The cursor to copy from.
7026 * @param[out] cdst The cursor to copy to.
7029 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7033 cdst->mc_txn = csrc->mc_txn;
7034 cdst->mc_dbi = csrc->mc_dbi;
7035 cdst->mc_db = csrc->mc_db;
7036 cdst->mc_dbx = csrc->mc_dbx;
7037 cdst->mc_snum = csrc->mc_snum;
7038 cdst->mc_top = csrc->mc_top;
7039 cdst->mc_flags = csrc->mc_flags;
7041 for (i=0; i<csrc->mc_snum; i++) {
7042 cdst->mc_pg[i] = csrc->mc_pg[i];
7043 cdst->mc_ki[i] = csrc->mc_ki[i];
7047 /** Rebalance the tree after a delete operation.
7048 * @param[in] mc Cursor pointing to the page where rebalancing
7050 * @return 0 on success, non-zero on failure.
7053 mdb_rebalance(MDB_cursor *mc)
7057 unsigned int ptop, minkeys;
7060 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7064 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7065 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7066 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7067 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7068 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7072 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7073 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7076 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7077 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7083 if (mc->mc_snum < 2) {
7084 MDB_page *mp = mc->mc_pg[0];
7086 DPUTS("Can't rebalance a subpage, ignoring");
7089 if (NUMKEYS(mp) == 0) {
7090 DPUTS("tree is completely empty");
7091 mc->mc_db->md_root = P_INVALID;
7092 mc->mc_db->md_depth = 0;
7093 mc->mc_db->md_leaf_pages = 0;
7094 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7097 /* Adjust cursors pointing to mp */
7101 MDB_cursor *m2, *m3;
7102 MDB_dbi dbi = mc->mc_dbi;
7104 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7105 if (mc->mc_flags & C_SUB)
7106 m3 = &m2->mc_xcursor->mx_cursor;
7109 if (m3->mc_snum < mc->mc_snum) continue;
7110 if (m3->mc_pg[0] == mp) {
7116 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7117 DPUTS("collapsing root page!");
7118 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7121 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7122 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7125 mc->mc_db->md_depth--;
7126 mc->mc_db->md_branch_pages--;
7127 mc->mc_ki[0] = mc->mc_ki[1];
7129 /* Adjust other cursors pointing to mp */
7130 MDB_cursor *m2, *m3;
7131 MDB_dbi dbi = mc->mc_dbi;
7133 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7134 if (mc->mc_flags & C_SUB)
7135 m3 = &m2->mc_xcursor->mx_cursor;
7138 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7139 if (m3->mc_pg[0] == mp) {
7140 m3->mc_pg[0] = mc->mc_pg[0];
7143 m3->mc_ki[0] = m3->mc_ki[1];
7148 DPUTS("root page doesn't need rebalancing");
7152 /* The parent (branch page) must have at least 2 pointers,
7153 * otherwise the tree is invalid.
7155 ptop = mc->mc_top-1;
7156 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7158 /* Leaf page fill factor is below the threshold.
7159 * Try to move keys from left or right neighbor, or
7160 * merge with a neighbor page.
7165 mdb_cursor_copy(mc, &mn);
7166 mn.mc_xcursor = NULL;
7168 if (mc->mc_ki[ptop] == 0) {
7169 /* We're the leftmost leaf in our parent.
7171 DPUTS("reading right neighbor");
7173 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7174 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7177 mn.mc_ki[mn.mc_top] = 0;
7178 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7180 /* There is at least one neighbor to the left.
7182 DPUTS("reading left neighbor");
7184 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7185 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7188 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7189 mc->mc_ki[mc->mc_top] = 0;
7192 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7193 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7194 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7196 /* If the neighbor page is above threshold and has enough keys,
7197 * move one key from it. Otherwise we should try to merge them.
7198 * (A branch page must never have less than 2 keys.)
7200 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7201 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7202 return mdb_node_move(&mn, mc);
7204 if (mc->mc_ki[ptop] == 0)
7205 rc = mdb_page_merge(&mn, mc);
7207 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7208 rc = mdb_page_merge(mc, &mn);
7209 mdb_cursor_copy(&mn, mc);
7211 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7216 /** Complete a delete operation started by #mdb_cursor_del(). */
7218 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7225 mp = mc->mc_pg[mc->mc_top];
7226 ki = mc->mc_ki[mc->mc_top];
7228 /* add overflow pages to free list */
7229 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7233 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7234 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7235 (rc = mdb_ovpage_free(mc, omp)))
7238 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7239 mc->mc_db->md_entries--;
7240 rc = mdb_rebalance(mc);
7241 if (rc != MDB_SUCCESS)
7242 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7245 MDB_dbi dbi = mc->mc_dbi;
7247 mp = mc->mc_pg[mc->mc_top];
7248 nkeys = NUMKEYS(mp);
7250 /* if mc points past last node in page, find next sibling */
7251 if (mc->mc_ki[mc->mc_top] >= nkeys)
7252 mdb_cursor_sibling(mc, 1);
7254 /* Adjust other cursors pointing to mp */
7255 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7258 if (!(m2->mc_flags & C_INITIALIZED))
7260 if (m2->mc_pg[mc->mc_top] == mp) {
7261 if (m2->mc_ki[mc->mc_top] >= ki) {
7262 m2->mc_flags |= C_DEL;
7263 if (m2->mc_ki[mc->mc_top] > ki)
7264 m2->mc_ki[mc->mc_top]--;
7266 if (m2->mc_ki[mc->mc_top] >= nkeys)
7267 mdb_cursor_sibling(m2, 1);
7270 mc->mc_flags |= C_DEL;
7277 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7278 MDB_val *key, MDB_val *data)
7283 MDB_val rdata, *xdata;
7287 assert(key != NULL);
7289 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7291 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7294 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7295 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7297 if (key->mv_size > MDB_MAXKEYSIZE) {
7298 return MDB_BAD_VALSIZE;
7301 mdb_cursor_init(&mc, txn, dbi, &mx);
7304 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7305 /* must ignore any data */
7316 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7318 /* let mdb_page_split know about this cursor if needed:
7319 * delete will trigger a rebalance; if it needs to move
7320 * a node from one page to another, it will have to
7321 * update the parent's separator key(s). If the new sepkey
7322 * is larger than the current one, the parent page may
7323 * run out of space, triggering a split. We need this
7324 * cursor to be consistent until the end of the rebalance.
7326 mc.mc_flags |= C_UNTRACK;
7327 mc.mc_next = txn->mt_cursors[dbi];
7328 txn->mt_cursors[dbi] = &mc;
7329 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7330 txn->mt_cursors[dbi] = mc.mc_next;
7335 /** Split a page and insert a new node.
7336 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7337 * The cursor will be updated to point to the actual page and index where
7338 * the node got inserted after the split.
7339 * @param[in] newkey The key for the newly inserted node.
7340 * @param[in] newdata The data for the newly inserted node.
7341 * @param[in] newpgno The page number, if the new node is a branch node.
7342 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7343 * @return 0 on success, non-zero on failure.
7346 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7347 unsigned int nflags)
7350 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7353 unsigned int i, j, split_indx, nkeys, pmax;
7355 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7357 MDB_page *mp, *rp, *pp;
7362 mp = mc->mc_pg[mc->mc_top];
7363 newindx = mc->mc_ki[mc->mc_top];
7365 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7366 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7367 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7369 /* Create a right sibling. */
7370 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7372 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7374 if (mc->mc_snum < 2) {
7375 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7377 /* shift current top to make room for new parent */
7378 mc->mc_pg[1] = mc->mc_pg[0];
7379 mc->mc_ki[1] = mc->mc_ki[0];
7382 mc->mc_db->md_root = pp->mp_pgno;
7383 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7384 mc->mc_db->md_depth++;
7387 /* Add left (implicit) pointer. */
7388 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7389 /* undo the pre-push */
7390 mc->mc_pg[0] = mc->mc_pg[1];
7391 mc->mc_ki[0] = mc->mc_ki[1];
7392 mc->mc_db->md_root = mp->mp_pgno;
7393 mc->mc_db->md_depth--;
7400 ptop = mc->mc_top-1;
7401 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7404 mc->mc_flags |= C_SPLITTING;
7405 mdb_cursor_copy(mc, &mn);
7406 mn.mc_pg[mn.mc_top] = rp;
7407 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7409 if (nflags & MDB_APPEND) {
7410 mn.mc_ki[mn.mc_top] = 0;
7412 split_indx = newindx;
7417 nkeys = NUMKEYS(mp);
7418 split_indx = nkeys / 2;
7419 if (newindx < split_indx)
7425 unsigned int lsize, rsize, ksize;
7426 /* Move half of the keys to the right sibling */
7428 x = mc->mc_ki[mc->mc_top] - split_indx;
7429 ksize = mc->mc_db->md_pad;
7430 split = LEAF2KEY(mp, split_indx, ksize);
7431 rsize = (nkeys - split_indx) * ksize;
7432 lsize = (nkeys - split_indx) * sizeof(indx_t);
7433 mp->mp_lower -= lsize;
7434 rp->mp_lower += lsize;
7435 mp->mp_upper += rsize - lsize;
7436 rp->mp_upper -= rsize - lsize;
7437 sepkey.mv_size = ksize;
7438 if (newindx == split_indx) {
7439 sepkey.mv_data = newkey->mv_data;
7441 sepkey.mv_data = split;
7444 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7445 memcpy(rp->mp_ptrs, split, rsize);
7446 sepkey.mv_data = rp->mp_ptrs;
7447 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7448 memcpy(ins, newkey->mv_data, ksize);
7449 mp->mp_lower += sizeof(indx_t);
7450 mp->mp_upper -= ksize - sizeof(indx_t);
7453 memcpy(rp->mp_ptrs, split, x * ksize);
7454 ins = LEAF2KEY(rp, x, ksize);
7455 memcpy(ins, newkey->mv_data, ksize);
7456 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7457 rp->mp_lower += sizeof(indx_t);
7458 rp->mp_upper -= ksize - sizeof(indx_t);
7459 mc->mc_ki[mc->mc_top] = x;
7460 mc->mc_pg[mc->mc_top] = rp;
7465 /* For leaf pages, check the split point based on what
7466 * fits where, since otherwise mdb_node_add can fail.
7468 * This check is only needed when the data items are
7469 * relatively large, such that being off by one will
7470 * make the difference between success or failure.
7472 * It's also relevant if a page happens to be laid out
7473 * such that one half of its nodes are all "small" and
7474 * the other half of its nodes are "large." If the new
7475 * item is also "large" and falls on the half with
7476 * "large" nodes, it also may not fit.
7479 unsigned int psize, nsize;
7480 /* Maximum free space in an empty page */
7481 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7482 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7483 if ((nkeys < 20) || (nsize > pmax/16)) {
7484 if (newindx <= split_indx) {
7487 for (i=0; i<split_indx; i++) {
7488 node = NODEPTR(mp, i);
7489 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7490 if (F_ISSET(node->mn_flags, F_BIGDATA))
7491 psize += sizeof(pgno_t);
7493 psize += NODEDSZ(node);
7497 split_indx = newindx;
7508 for (i=nkeys-1; i>=split_indx; i--) {
7509 node = NODEPTR(mp, i);
7510 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7511 if (F_ISSET(node->mn_flags, F_BIGDATA))
7512 psize += sizeof(pgno_t);
7514 psize += NODEDSZ(node);
7518 split_indx = newindx;
7529 /* First find the separating key between the split pages.
7530 * The case where newindx == split_indx is ambiguous; the
7531 * new item could go to the new page or stay on the original
7532 * page. If newpos == 1 it goes to the new page.
7534 if (newindx == split_indx && newpos) {
7535 sepkey.mv_size = newkey->mv_size;
7536 sepkey.mv_data = newkey->mv_data;
7538 node = NODEPTR(mp, split_indx);
7539 sepkey.mv_size = node->mn_ksize;
7540 sepkey.mv_data = NODEKEY(node);
7544 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7546 /* Copy separator key to the parent.
7548 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7552 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7555 if (mn.mc_snum == mc->mc_snum) {
7556 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7557 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7558 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7559 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7564 /* Right page might now have changed parent.
7565 * Check if left page also changed parent.
7567 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7568 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7569 for (i=0; i<ptop; i++) {
7570 mc->mc_pg[i] = mn.mc_pg[i];
7571 mc->mc_ki[i] = mn.mc_ki[i];
7573 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7574 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7578 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7581 mc->mc_flags ^= C_SPLITTING;
7582 if (rc != MDB_SUCCESS) {
7585 if (nflags & MDB_APPEND) {
7586 mc->mc_pg[mc->mc_top] = rp;
7587 mc->mc_ki[mc->mc_top] = 0;
7588 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7591 for (i=0; i<mc->mc_top; i++)
7592 mc->mc_ki[i] = mn.mc_ki[i];
7599 /* Move half of the keys to the right sibling. */
7601 /* grab a page to hold a temporary copy */
7602 copy = mdb_page_malloc(mc->mc_txn, 1);
7606 copy->mp_pgno = mp->mp_pgno;
7607 copy->mp_flags = mp->mp_flags;
7608 copy->mp_lower = PAGEHDRSZ;
7609 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7610 mc->mc_pg[mc->mc_top] = copy;
7611 for (i = j = 0; i <= nkeys; j++) {
7612 if (i == split_indx) {
7613 /* Insert in right sibling. */
7614 /* Reset insert index for right sibling. */
7615 if (i != newindx || (newpos ^ ins_new)) {
7617 mc->mc_pg[mc->mc_top] = rp;
7621 if (i == newindx && !ins_new) {
7622 /* Insert the original entry that caused the split. */
7623 rkey.mv_data = newkey->mv_data;
7624 rkey.mv_size = newkey->mv_size;
7633 /* Update index for the new key. */
7634 mc->mc_ki[mc->mc_top] = j;
7635 } else if (i == nkeys) {
7638 node = NODEPTR(mp, i);
7639 rkey.mv_data = NODEKEY(node);
7640 rkey.mv_size = node->mn_ksize;
7642 xdata.mv_data = NODEDATA(node);
7643 xdata.mv_size = NODEDSZ(node);
7646 pgno = NODEPGNO(node);
7647 flags = node->mn_flags;
7652 if (!IS_LEAF(mp) && j == 0) {
7653 /* First branch index doesn't need key data. */
7657 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7661 nkeys = NUMKEYS(copy);
7662 for (i=0; i<nkeys; i++)
7663 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7664 mp->mp_lower = copy->mp_lower;
7665 mp->mp_upper = copy->mp_upper;
7666 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7667 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7669 /* reset back to original page */
7670 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7671 mc->mc_pg[mc->mc_top] = mp;
7672 if (nflags & MDB_RESERVE) {
7673 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7674 if (!(node->mn_flags & F_BIGDATA))
7675 newdata->mv_data = NODEDATA(node);
7679 /* Make sure mc_ki is still valid.
7681 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7682 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7683 for (i=0; i<ptop; i++) {
7684 mc->mc_pg[i] = mn.mc_pg[i];
7685 mc->mc_ki[i] = mn.mc_ki[i];
7687 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7688 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7692 /* return tmp page to freelist */
7693 mdb_page_free(mc->mc_txn->mt_env, copy);
7696 /* Adjust other cursors pointing to mp */
7697 MDB_cursor *m2, *m3;
7698 MDB_dbi dbi = mc->mc_dbi;
7699 int fixup = NUMKEYS(mp);
7701 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7702 if (mc->mc_flags & C_SUB)
7703 m3 = &m2->mc_xcursor->mx_cursor;
7708 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7710 if (m3->mc_flags & C_SPLITTING)
7715 for (k=m3->mc_top; k>=0; k--) {
7716 m3->mc_ki[k+1] = m3->mc_ki[k];
7717 m3->mc_pg[k+1] = m3->mc_pg[k];
7719 if (m3->mc_ki[0] >= split_indx) {
7724 m3->mc_pg[0] = mc->mc_pg[0];
7728 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7729 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7730 m3->mc_ki[mc->mc_top]++;
7731 if (m3->mc_ki[mc->mc_top] >= fixup) {
7732 m3->mc_pg[mc->mc_top] = rp;
7733 m3->mc_ki[mc->mc_top] -= fixup;
7734 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7736 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7737 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7746 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7747 MDB_val *key, MDB_val *data, unsigned int flags)
7752 assert(key != NULL);
7753 assert(data != NULL);
7755 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
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 mdb_default_cmp(txn, slot);
7976 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7978 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7981 if (txn->mt_dbflags[dbi] & DB_STALE) {
7984 /* Stale, must read the DB's root. cursor_init does it for us. */
7985 mdb_cursor_init(&mc, txn, dbi, &mx);
7987 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7990 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7993 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7995 ptr = env->me_dbxs[dbi].md_name.mv_data;
7996 env->me_dbxs[dbi].md_name.mv_data = NULL;
7997 env->me_dbxs[dbi].md_name.mv_size = 0;
7998 env->me_dbflags[dbi] = 0;
8002 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8004 /* We could return the flags for the FREE_DBI too but what's the point? */
8005 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8007 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8011 /** Add all the DB's pages to the free list.
8012 * @param[in] mc Cursor on the DB to free.
8013 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8014 * @return 0 on success, non-zero on failure.
8017 mdb_drop0(MDB_cursor *mc, int subs)
8021 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8022 if (rc == MDB_SUCCESS) {
8023 MDB_txn *txn = mc->mc_txn;
8028 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8029 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8032 mdb_cursor_copy(mc, &mx);
8033 while (mc->mc_snum > 0) {
8034 MDB_page *mp = mc->mc_pg[mc->mc_top];
8035 unsigned n = NUMKEYS(mp);
8037 for (i=0; i<n; i++) {
8038 ni = NODEPTR(mp, i);
8039 if (ni->mn_flags & F_BIGDATA) {
8042 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8043 rc = mdb_page_get(txn, pg, &omp, NULL);
8046 assert(IS_OVERFLOW(omp));
8047 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8051 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8052 mdb_xcursor_init1(mc, ni);
8053 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8059 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8061 for (i=0; i<n; i++) {
8063 ni = NODEPTR(mp, i);
8066 mdb_midl_xappend(txn->mt_free_pgs, pg);
8071 mc->mc_ki[mc->mc_top] = i;
8072 rc = mdb_cursor_sibling(mc, 1);
8074 /* no more siblings, go back to beginning
8075 * of previous level.
8079 for (i=1; i<mc->mc_snum; i++) {
8081 mc->mc_pg[i] = mx.mc_pg[i];
8086 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8087 } else if (rc == MDB_NOTFOUND) {
8093 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8095 MDB_cursor *mc, *m2;
8098 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8101 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8104 rc = mdb_cursor_open(txn, dbi, &mc);
8108 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8109 /* Invalidate the dropped DB's cursors */
8110 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8111 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8115 /* Can't delete the main DB */
8116 if (del && dbi > MAIN_DBI) {
8117 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8119 txn->mt_dbflags[dbi] = DB_STALE;
8120 mdb_dbi_close(txn->mt_env, dbi);
8123 /* reset the DB record, mark it dirty */
8124 txn->mt_dbflags[dbi] |= DB_DIRTY;
8125 txn->mt_dbs[dbi].md_depth = 0;
8126 txn->mt_dbs[dbi].md_branch_pages = 0;
8127 txn->mt_dbs[dbi].md_leaf_pages = 0;
8128 txn->mt_dbs[dbi].md_overflow_pages = 0;
8129 txn->mt_dbs[dbi].md_entries = 0;
8130 txn->mt_dbs[dbi].md_root = P_INVALID;
8132 txn->mt_flags |= MDB_TXN_DIRTY;
8135 mdb_cursor_close(mc);
8139 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8141 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8144 txn->mt_dbxs[dbi].md_cmp = cmp;
8148 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8150 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8153 txn->mt_dbxs[dbi].md_dcmp = cmp;
8157 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8159 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8162 txn->mt_dbxs[dbi].md_rel = rel;
8166 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8168 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8171 txn->mt_dbxs[dbi].md_relctx = ctx;
8175 int mdb_env_get_maxkeysize(MDB_env *env)
8177 return MDB_MAXKEYSIZE;
8180 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8182 unsigned int i, rdrs;
8189 if (!env->me_txns) {
8190 return func("(no reader locks)\n", ctx);
8192 rdrs = env->me_txns->mti_numreaders;
8193 mr = env->me_txns->mti_readers;
8194 for (i=0; i<rdrs; i++) {
8199 if (mr[i].mr_txnid == (txnid_t)-1) {
8200 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8202 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8206 func(" pid thread txnid\n", ctx);
8208 rc = func(buf, ctx);
8214 func("(no active readers)\n", ctx);
8219 /** Insert pid into list if not already present.
8220 * return -1 if already present.
8222 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8224 /* binary search of pid in list */
8226 unsigned cursor = 1;
8228 unsigned n = ids[0];
8231 unsigned pivot = n >> 1;
8232 cursor = base + pivot + 1;
8233 val = pid - ids[cursor];
8238 } else if ( val > 0 ) {
8243 /* found, so it's a duplicate */
8252 for (n = ids[0]; n > cursor; n--)
8258 int mdb_reader_check(MDB_env *env, int *dead)
8260 unsigned int i, j, rdrs;
8271 rdrs = env->me_txns->mti_numreaders;
8272 pids = malloc((rdrs+1) * sizeof(pid_t));
8276 mr = env->me_txns->mti_readers;
8278 for (i=0; i<rdrs; i++) {
8279 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8281 if (mdb_pid_insert(pids, pid) == 0) {
8282 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8284 /* Recheck, a new process may have reused pid */
8285 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8286 for (j=i; j<rdrs; j++)
8287 if (mr[j].mr_pid == pid) {
8288 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8289 (unsigned) pid, mr[j].mr_txnid));
8294 UNLOCK_MUTEX_R(env);