2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define MDB_PIDLOCK 0
144 #define pthread_t DWORD
145 #define pthread_mutex_t HANDLE
146 #define pthread_key_t DWORD
147 #define pthread_self() GetCurrentThreadId()
148 #define pthread_key_create(x,y) \
149 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
150 #define pthread_key_delete(x) TlsFree(x)
151 #define pthread_getspecific(x) TlsGetValue(x)
152 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
153 #define pthread_mutex_unlock(x) ReleaseMutex(x)
154 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
155 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
156 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
157 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
158 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
159 #define getpid() GetCurrentProcessId()
160 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
161 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
162 #define ErrCode() GetLastError()
163 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
164 #define close(fd) (CloseHandle(fd) ? 0 : -1)
165 #define munmap(ptr,len) UnmapViewOfFile(ptr)
166 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
167 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
169 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
176 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
177 #define MDB_PIDLOCK 1
179 #ifdef MDB_USE_POSIX_SEM
181 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
182 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
183 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
184 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
187 mdb_sem_wait(sem_t *sem)
190 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
195 /** Lock the reader mutex.
197 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
198 /** Unlock the reader mutex.
200 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
202 /** Lock the writer mutex.
203 * Only a single write transaction is allowed at a time. Other writers
204 * will block waiting for this mutex.
206 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
207 /** Unlock the writer mutex.
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
210 #endif /* MDB_USE_POSIX_SEM */
212 /** Get the error code for the last failed system function.
214 #define ErrCode() errno
216 /** An abstraction for a file handle.
217 * On POSIX systems file handles are small integers. On Windows
218 * they're opaque pointers.
222 /** A value for an invalid file handle.
223 * Mainly used to initialize file variables and signify that they are
226 #define INVALID_HANDLE_VALUE (-1)
228 /** Get the size of a memory page for the system.
229 * This is the basic size that the platform's memory manager uses, and is
230 * fundamental to the use of memory-mapped files.
232 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
235 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
238 #define MNAME_LEN (sizeof(pthread_mutex_t))
244 /** A flag for opening a file and requesting synchronous data writes.
245 * This is only used when writing a meta page. It's not strictly needed;
246 * we could just do a normal write and then immediately perform a flush.
247 * But if this flag is available it saves us an extra system call.
249 * @note If O_DSYNC is undefined but exists in /usr/include,
250 * preferably set some compiler flag to get the definition.
251 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
254 # define MDB_DSYNC O_DSYNC
258 /** Function for flushing the data of a file. Define this to fsync
259 * if fdatasync() is not supported.
261 #ifndef MDB_FDATASYNC
262 # define MDB_FDATASYNC fdatasync
266 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
277 /** A page number in the database.
278 * Note that 64 bit page numbers are overkill, since pages themselves
279 * already represent 12-13 bits of addressable memory, and the OS will
280 * always limit applications to a maximum of 63 bits of address space.
282 * @note In the #MDB_node structure, we only store 48 bits of this value,
283 * which thus limits us to only 60 bits of addressable data.
285 typedef MDB_ID pgno_t;
287 /** A transaction ID.
288 * See struct MDB_txn.mt_txnid for details.
290 typedef MDB_ID txnid_t;
292 /** @defgroup debug Debug Macros
296 /** Enable debug output. Needs variable argument macros (a C99 feature).
297 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
298 * read from and written to the database (used for free space management).
304 static int mdb_debug;
305 static txnid_t mdb_debug_start;
307 /** Print a debug message with printf formatting.
308 * Requires double parenthesis around 2 or more args.
310 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
311 # define DPRINTF0(fmt, ...) \
312 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
314 # define DPRINTF(args) ((void) 0)
316 /** Print a debug string.
317 * The string is printed literally, with no format processing.
319 #define DPUTS(arg) DPRINTF(("%s", arg))
322 /** A default memory page size.
323 * The actual size is platform-dependent, but we use this for
324 * boot-strapping. We probably should not be using this any more.
325 * The #GET_PAGESIZE() macro is used to get the actual size.
327 * Note that we don't currently support Huge pages. On Linux,
328 * regular data files cannot use Huge pages, and in general
329 * Huge pages aren't actually pageable. We rely on the OS
330 * demand-pager to read our data and page it out when memory
331 * pressure from other processes is high. So until OSs have
332 * actual paging support for Huge pages, they're not viable.
334 #define MDB_PAGESIZE 4096
336 /** The minimum number of keys required in a database page.
337 * Setting this to a larger value will place a smaller bound on the
338 * maximum size of a data item. Data items larger than this size will
339 * be pushed into overflow pages instead of being stored directly in
340 * the B-tree node. This value used to default to 4. With a page size
341 * of 4096 bytes that meant that any item larger than 1024 bytes would
342 * go into an overflow page. That also meant that on average 2-3KB of
343 * each overflow page was wasted space. The value cannot be lower than
344 * 2 because then there would no longer be a tree structure. With this
345 * value, items larger than 2KB will go into overflow pages, and on
346 * average only 1KB will be wasted.
348 #define MDB_MINKEYS 2
350 /** A stamp that identifies a file as an MDB file.
351 * There's nothing special about this value other than that it is easily
352 * recognizable, and it will reflect any byte order mismatches.
354 #define MDB_MAGIC 0xBEEFC0DE
356 /** The version number for a database's datafile format. */
357 #define MDB_DATA_VERSION 1
358 /** The version number for a database's lockfile format. */
359 #define MDB_LOCK_VERSION 1
361 /** @brief The maximum size of a key in the database.
363 * The library rejects bigger keys, and cannot deal with records
364 * with bigger keys stored by a library with bigger max keysize.
366 * We require that keys all fit onto a regular page. This limit
367 * could be raised a bit further if needed; to something just
368 * under #MDB_PAGESIZE / #MDB_MINKEYS.
370 * Note that data items in an #MDB_DUPSORT database are actually keys
371 * of a subDB, so they're also limited to this size.
373 #ifndef MDB_MAXKEYSIZE
374 #define MDB_MAXKEYSIZE 511
377 /** @brief The maximum size of a data item.
379 * We only store a 32 bit value for node sizes.
381 #define MAXDATASIZE 0xffffffffUL
386 * This is used for printing a hex dump of a key's contents.
388 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
389 /** Display a key in hex.
391 * Invoke a function to display a key in hex.
393 #define DKEY(x) mdb_dkey(x, kbuf)
395 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
399 /** An invalid page number.
400 * Mainly used to denote an empty tree.
402 #define P_INVALID (~(pgno_t)0)
404 /** Test if the flags \b f are set in a flag word \b w. */
405 #define F_ISSET(w, f) (((w) & (f)) == (f))
407 /** Used for offsets within a single page.
408 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
411 typedef uint16_t indx_t;
413 /** Default size of memory map.
414 * This is certainly too small for any actual applications. Apps should always set
415 * the size explicitly using #mdb_env_set_mapsize().
417 #define DEFAULT_MAPSIZE 1048576
419 /** @defgroup readers Reader Lock Table
420 * Readers don't acquire any locks for their data access. Instead, they
421 * simply record their transaction ID in the reader table. The reader
422 * mutex is needed just to find an empty slot in the reader table. The
423 * slot's address is saved in thread-specific data so that subsequent read
424 * transactions started by the same thread need no further locking to proceed.
426 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
428 * No reader table is used if the database is on a read-only filesystem.
430 * Since the database uses multi-version concurrency control, readers don't
431 * actually need any locking. This table is used to keep track of which
432 * readers are using data from which old transactions, so that we'll know
433 * when a particular old transaction is no longer in use. Old transactions
434 * that have discarded any data pages can then have those pages reclaimed
435 * for use by a later write transaction.
437 * The lock table is constructed such that reader slots are aligned with the
438 * processor's cache line size. Any slot is only ever used by one thread.
439 * This alignment guarantees that there will be no contention or cache
440 * thrashing as threads update their own slot info, and also eliminates
441 * any need for locking when accessing a slot.
443 * A writer thread will scan every slot in the table to determine the oldest
444 * outstanding reader transaction. Any freed pages older than this will be
445 * reclaimed by the writer. The writer doesn't use any locks when scanning
446 * this table. This means that there's no guarantee that the writer will
447 * see the most up-to-date reader info, but that's not required for correct
448 * operation - all we need is to know the upper bound on the oldest reader,
449 * we don't care at all about the newest reader. So the only consequence of
450 * reading stale information here is that old pages might hang around a
451 * while longer before being reclaimed. That's actually good anyway, because
452 * the longer we delay reclaiming old pages, the more likely it is that a
453 * string of contiguous pages can be found after coalescing old pages from
454 * many old transactions together.
457 /** Number of slots in the reader table.
458 * This value was chosen somewhat arbitrarily. 126 readers plus a
459 * couple mutexes fit exactly into 8KB on my development machine.
460 * Applications should set the table size using #mdb_env_set_maxreaders().
462 #define DEFAULT_READERS 126
464 /** The size of a CPU cache line in bytes. We want our lock structures
465 * aligned to this size to avoid false cache line sharing in the
467 * This value works for most CPUs. For Itanium this should be 128.
473 /** The information we store in a single slot of the reader table.
474 * In addition to a transaction ID, we also record the process and
475 * thread ID that owns a slot, so that we can detect stale information,
476 * e.g. threads or processes that went away without cleaning up.
477 * @note We currently don't check for stale records. We simply re-init
478 * the table when we know that we're the only process opening the
481 typedef struct MDB_rxbody {
482 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
483 * Multiple readers that start at the same time will probably have the
484 * same ID here. Again, it's not important to exclude them from
485 * anything; all we need to know is which version of the DB they
486 * started from so we can avoid overwriting any data used in that
487 * particular version.
490 /** The process ID of the process owning this reader txn. */
492 /** The thread ID of the thread owning this txn. */
496 /** The actual reader record, with cacheline padding. */
497 typedef struct MDB_reader {
500 /** shorthand for mrb_txnid */
501 #define mr_txnid mru.mrx.mrb_txnid
502 #define mr_pid mru.mrx.mrb_pid
503 #define mr_tid mru.mrx.mrb_tid
504 /** cache line alignment */
505 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
509 /** The header for the reader table.
510 * The table resides in a memory-mapped file. (This is a different file
511 * than is used for the main database.)
513 * For POSIX the actual mutexes reside in the shared memory of this
514 * mapped file. On Windows, mutexes are named objects allocated by the
515 * kernel; we store the mutex names in this mapped file so that other
516 * processes can grab them. This same approach is also used on
517 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
518 * process-shared POSIX mutexes. For these cases where a named object
519 * is used, the object name is derived from a 64 bit FNV hash of the
520 * environment pathname. As such, naming collisions are extremely
521 * unlikely. If a collision occurs, the results are unpredictable.
523 typedef struct MDB_txbody {
524 /** Stamp identifying this as an MDB file. It must be set
527 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
529 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
530 char mtb_rmname[MNAME_LEN];
532 /** Mutex protecting access to this table.
533 * This is the reader lock that #LOCK_MUTEX_R acquires.
535 pthread_mutex_t mtb_mutex;
537 /** The ID of the last transaction committed to the database.
538 * This is recorded here only for convenience; the value can always
539 * be determined by reading the main database meta pages.
542 /** The number of slots that have been used in the reader table.
543 * This always records the maximum count, it is not decremented
544 * when readers release their slots.
546 unsigned mtb_numreaders;
549 /** The actual reader table definition. */
550 typedef struct MDB_txninfo {
553 #define mti_magic mt1.mtb.mtb_magic
554 #define mti_format mt1.mtb.mtb_format
555 #define mti_mutex mt1.mtb.mtb_mutex
556 #define mti_rmname mt1.mtb.mtb_rmname
557 #define mti_txnid mt1.mtb.mtb_txnid
558 #define mti_numreaders mt1.mtb.mtb_numreaders
559 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
562 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
563 char mt2_wmname[MNAME_LEN];
564 #define mti_wmname mt2.mt2_wmname
566 pthread_mutex_t mt2_wmutex;
567 #define mti_wmutex mt2.mt2_wmutex
569 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
571 MDB_reader mti_readers[1];
574 /** Lockfile format signature: version, features and field layout */
575 #define MDB_LOCK_FORMAT \
577 ((MDB_LOCK_VERSION) \
578 /* Flags which describe functionality */ \
579 + (((MDB_PIDLOCK) != 0) << 16)))
582 /** Common header for all page types.
583 * Overflow records occupy a number of contiguous pages with no
584 * headers on any page after the first.
586 typedef struct MDB_page {
587 #define mp_pgno mp_p.p_pgno
588 #define mp_next mp_p.p_next
590 pgno_t p_pgno; /**< page number */
591 void * p_next; /**< for in-memory list of freed structs */
594 /** @defgroup mdb_page Page Flags
596 * Flags for the page headers.
599 #define P_BRANCH 0x01 /**< branch page */
600 #define P_LEAF 0x02 /**< leaf page */
601 #define P_OVERFLOW 0x04 /**< overflow page */
602 #define P_META 0x08 /**< meta page */
603 #define P_DIRTY 0x10 /**< dirty page */
604 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
605 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
606 #define P_KEEP 0x8000 /**< leave this page alone during spill */
608 uint16_t mp_flags; /**< @ref mdb_page */
609 #define mp_lower mp_pb.pb.pb_lower
610 #define mp_upper mp_pb.pb.pb_upper
611 #define mp_pages mp_pb.pb_pages
614 indx_t pb_lower; /**< lower bound of free space */
615 indx_t pb_upper; /**< upper bound of free space */
617 uint32_t pb_pages; /**< number of overflow pages */
619 indx_t mp_ptrs[1]; /**< dynamic size */
622 /** Size of the page header, excluding dynamic data at the end */
623 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
625 /** Address of first usable data byte in a page, after the header */
626 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
628 /** Number of nodes on a page */
629 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
631 /** The amount of space remaining in the page */
632 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
634 /** The percentage of space used in the page, in tenths of a percent. */
635 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
636 ((env)->me_psize - PAGEHDRSZ))
637 /** The minimum page fill factor, in tenths of a percent.
638 * Pages emptier than this are candidates for merging.
640 #define FILL_THRESHOLD 250
642 /** Test if a page is a leaf page */
643 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
644 /** Test if a page is a LEAF2 page */
645 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
646 /** Test if a page is a branch page */
647 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
648 /** Test if a page is an overflow page */
649 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
650 /** Test if a page is a sub page */
651 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
653 /** The number of overflow pages needed to store the given size. */
654 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
656 /** Header for a single key/data pair within a page.
657 * We guarantee 2-byte alignment for nodes.
659 typedef struct MDB_node {
660 /** lo and hi are used for data size on leaf nodes and for
661 * child pgno on branch nodes. On 64 bit platforms, flags
662 * is also used for pgno. (Branch nodes have no flags).
663 * They are in host byte order in case that lets some
664 * accesses be optimized into a 32-bit word access.
666 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
667 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
668 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
669 /** @defgroup mdb_node Node Flags
671 * Flags for node headers.
674 #define F_BIGDATA 0x01 /**< data put on overflow page */
675 #define F_SUBDATA 0x02 /**< data is a sub-database */
676 #define F_DUPDATA 0x04 /**< data has duplicates */
678 /** valid flags for #mdb_node_add() */
679 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
682 unsigned short mn_flags; /**< @ref mdb_node */
683 unsigned short mn_ksize; /**< key size */
684 char mn_data[1]; /**< key and data are appended here */
687 /** Size of the node header, excluding dynamic data at the end */
688 #define NODESIZE offsetof(MDB_node, mn_data)
690 /** Bit position of top word in page number, for shifting mn_flags */
691 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
693 /** Size of a node in a branch page with a given key.
694 * This is just the node header plus the key, there is no data.
696 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
698 /** Size of a node in a leaf page with a given key and data.
699 * This is node header plus key plus data size.
701 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
703 /** Address of node \b i in page \b p */
704 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
706 /** Address of the key for the node */
707 #define NODEKEY(node) (void *)((node)->mn_data)
709 /** Address of the data for a node */
710 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
712 /** Get the page number pointed to by a branch node */
713 #define NODEPGNO(node) \
714 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
715 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
716 /** Set the page number in a branch node */
717 #define SETPGNO(node,pgno) do { \
718 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
719 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
721 /** Get the size of the data in a leaf node */
722 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
723 /** Set the size of the data for a leaf node */
724 #define SETDSZ(node,size) do { \
725 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
726 /** The size of a key in a node */
727 #define NODEKSZ(node) ((node)->mn_ksize)
729 /** Copy a page number from src to dst */
731 #define COPY_PGNO(dst,src) dst = src
733 #if SIZE_MAX > 4294967295UL
734 #define COPY_PGNO(dst,src) do { \
735 unsigned short *s, *d; \
736 s = (unsigned short *)&(src); \
737 d = (unsigned short *)&(dst); \
744 #define COPY_PGNO(dst,src) do { \
745 unsigned short *s, *d; \
746 s = (unsigned short *)&(src); \
747 d = (unsigned short *)&(dst); \
753 /** The address of a key in a LEAF2 page.
754 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
755 * There are no node headers, keys are stored contiguously.
757 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
759 /** Set the \b node's key into \b key, if requested. */
760 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
761 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
763 /** Information about a single database in the environment. */
764 typedef struct MDB_db {
765 uint32_t md_pad; /**< also ksize for LEAF2 pages */
766 uint16_t md_flags; /**< @ref mdb_dbi_open */
767 uint16_t md_depth; /**< depth of this tree */
768 pgno_t md_branch_pages; /**< number of internal pages */
769 pgno_t md_leaf_pages; /**< number of leaf pages */
770 pgno_t md_overflow_pages; /**< number of overflow pages */
771 size_t md_entries; /**< number of data items */
772 pgno_t md_root; /**< the root page of this tree */
775 /** mdb_dbi_open flags */
776 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
777 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
778 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
779 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
781 /** Handle for the DB used to track free pages. */
783 /** Handle for the default DB. */
786 /** Meta page content. */
787 typedef struct MDB_meta {
788 /** Stamp identifying this as an MDB file. It must be set
791 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
793 void *mm_address; /**< address for fixed mapping */
794 size_t mm_mapsize; /**< size of mmap region */
795 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
796 /** The size of pages used in this DB */
797 #define mm_psize mm_dbs[0].md_pad
798 /** Any persistent environment flags. @ref mdb_env */
799 #define mm_flags mm_dbs[0].md_flags
800 pgno_t mm_last_pg; /**< last used page in file */
801 txnid_t mm_txnid; /**< txnid that committed this page */
804 /** Buffer for a stack-allocated dirty page.
805 * The members define size and alignment, and silence type
806 * aliasing warnings. They are not used directly; that could
807 * mean incorrectly using several union members in parallel.
809 typedef union MDB_pagebuf {
810 char mb_raw[MDB_PAGESIZE];
813 char mm_pad[PAGEHDRSZ];
818 /** Auxiliary DB info.
819 * The information here is mostly static/read-only. There is
820 * only a single copy of this record in the environment.
822 typedef struct MDB_dbx {
823 MDB_val md_name; /**< name of the database */
824 MDB_cmp_func *md_cmp; /**< function for comparing keys */
825 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
826 MDB_rel_func *md_rel; /**< user relocate function */
827 void *md_relctx; /**< user-provided context for md_rel */
830 /** A database transaction.
831 * Every operation requires a transaction handle.
834 MDB_txn *mt_parent; /**< parent of a nested txn */
835 MDB_txn *mt_child; /**< nested txn under this txn */
836 pgno_t mt_next_pgno; /**< next unallocated page */
837 /** The ID of this transaction. IDs are integers incrementing from 1.
838 * Only committed write transactions increment the ID. If a transaction
839 * aborts, the ID may be re-used by the next writer.
842 MDB_env *mt_env; /**< the DB environment */
843 /** The list of pages that became unused during this transaction.
846 /** The sorted list of dirty pages we temporarily wrote to disk
847 * because the dirty list was full.
849 MDB_IDL mt_spill_pgs;
851 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
853 /** For read txns: This thread/txn's reader table slot, or NULL. */
856 /** Array of records for each DB known in the environment. */
858 /** Array of MDB_db records for each known DB */
860 /** @defgroup mt_dbflag Transaction DB Flags
864 #define DB_DIRTY 0x01 /**< DB was written in this txn */
865 #define DB_STALE 0x02 /**< DB record is older than txnID */
866 #define DB_NEW 0x04 /**< DB handle opened in this txn */
867 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
869 /** In write txns, array of cursors for each DB */
870 MDB_cursor **mt_cursors;
871 /** Array of flags for each DB */
872 unsigned char *mt_dbflags;
873 /** Number of DB records in use. This number only ever increments;
874 * we don't decrement it when individual DB handles are closed.
878 /** @defgroup mdb_txn Transaction Flags
882 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
883 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
884 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
885 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
887 unsigned int mt_flags; /**< @ref mdb_txn */
888 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
889 unsigned int mt_dirty_room;
890 /** Tracks which of the two meta pages was used at the start
891 * of this transaction.
893 unsigned int mt_toggle;
896 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
897 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
898 * raise this on a 64 bit machine.
900 #define CURSOR_STACK 32
904 /** Cursors are used for all DB operations */
906 /** Next cursor on this DB in this txn */
908 /** Backup of the original cursor if this cursor is a shadow */
909 MDB_cursor *mc_backup;
910 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
911 struct MDB_xcursor *mc_xcursor;
912 /** The transaction that owns this cursor */
914 /** The database handle this cursor operates on */
916 /** The database record for this cursor */
918 /** The database auxiliary record for this cursor */
920 /** The @ref mt_dbflag for this database */
921 unsigned char *mc_dbflag;
922 unsigned short mc_snum; /**< number of pushed pages */
923 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
924 /** @defgroup mdb_cursor Cursor Flags
926 * Cursor state flags.
929 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
930 #define C_EOF 0x02 /**< No more data */
931 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
932 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
933 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
935 unsigned int mc_flags; /**< @ref mdb_cursor */
936 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
937 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
940 /** Context for sorted-dup records.
941 * We could have gone to a fully recursive design, with arbitrarily
942 * deep nesting of sub-databases. But for now we only handle these
943 * levels - main DB, optional sub-DB, sorted-duplicate DB.
945 typedef struct MDB_xcursor {
946 /** A sub-cursor for traversing the Dup DB */
947 MDB_cursor mx_cursor;
948 /** The database record for this Dup DB */
950 /** The auxiliary DB record for this Dup DB */
952 /** The @ref mt_dbflag for this Dup DB */
953 unsigned char mx_dbflag;
956 /** State of FreeDB old pages, stored in the MDB_env */
957 typedef struct MDB_pgstate {
958 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
959 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
962 /** The database environment. */
964 HANDLE me_fd; /**< The main data file */
965 HANDLE me_lfd; /**< The lock file */
966 HANDLE me_mfd; /**< just for writing the meta pages */
967 /** Failed to update the meta page. Probably an I/O error. */
968 #define MDB_FATAL_ERROR 0x80000000U
969 /** Some fields are initialized. */
970 #define MDB_ENV_ACTIVE 0x20000000U
971 /** me_txkey is set */
972 #define MDB_ENV_TXKEY 0x10000000U
973 /** Have liveness lock in reader table */
974 #define MDB_LIVE_READER 0x08000000U
975 uint32_t me_flags; /**< @ref mdb_env */
976 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
977 unsigned int me_maxreaders; /**< size of the reader table */
978 unsigned int me_numreaders; /**< max numreaders set by this env */
979 MDB_dbi me_numdbs; /**< number of DBs opened */
980 MDB_dbi me_maxdbs; /**< size of the DB table */
981 pid_t me_pid; /**< process ID of this env */
982 char *me_path; /**< path to the DB files */
983 char *me_map; /**< the memory map of the data file */
984 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
985 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
986 MDB_txn *me_txn; /**< current write transaction */
987 size_t me_mapsize; /**< size of the data memory map */
988 off_t me_size; /**< current file size */
989 pgno_t me_maxpg; /**< me_mapsize / me_psize */
990 MDB_dbx *me_dbxs; /**< array of static DB info */
991 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
992 pthread_key_t me_txkey; /**< thread-key for readers */
993 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
994 # define me_pglast me_pgstate.mf_pglast
995 # define me_pghead me_pgstate.mf_pghead
996 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
997 /** IDL of pages that became unused in a write txn */
999 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1000 MDB_ID2L me_dirty_list;
1001 /** Max number of freelist items that can fit in a single overflow page */
1003 /** Max size of a node on a page */
1004 unsigned int me_nodemax;
1006 int me_pidquery; /**< Used in OpenProcess */
1007 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1009 #elif defined(MDB_USE_POSIX_SEM)
1010 sem_t *me_rmutex; /* Shared mutexes are not supported */
1015 /** Nested transaction */
1016 typedef struct MDB_ntxn {
1017 MDB_txn mnt_txn; /* the transaction */
1018 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
1021 /** max number of pages to commit in one writev() call */
1022 #define MDB_COMMIT_PAGES 64
1023 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1024 #undef MDB_COMMIT_PAGES
1025 #define MDB_COMMIT_PAGES IOV_MAX
1028 /* max bytes to write in one call */
1029 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1031 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1032 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1033 static int mdb_page_touch(MDB_cursor *mc);
1035 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1036 static int mdb_page_search_root(MDB_cursor *mc,
1037 MDB_val *key, int modify);
1038 #define MDB_PS_MODIFY 1
1039 #define MDB_PS_ROOTONLY 2
1040 static int mdb_page_search(MDB_cursor *mc,
1041 MDB_val *key, int flags);
1042 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1044 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1045 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1046 pgno_t newpgno, unsigned int nflags);
1048 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1049 static int mdb_env_pick_meta(const MDB_env *env);
1050 static int mdb_env_write_meta(MDB_txn *txn);
1051 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1052 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1054 static void mdb_env_close0(MDB_env *env, int excl);
1056 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1057 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1058 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1059 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1060 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1061 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1062 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1063 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1064 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1066 static int mdb_rebalance(MDB_cursor *mc);
1067 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1069 static void mdb_cursor_pop(MDB_cursor *mc);
1070 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1072 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1073 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1074 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1075 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1076 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1078 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1079 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1081 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1082 static void mdb_xcursor_init0(MDB_cursor *mc);
1083 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1085 static int mdb_drop0(MDB_cursor *mc, int subs);
1086 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1089 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1093 static SECURITY_DESCRIPTOR mdb_null_sd;
1094 static SECURITY_ATTRIBUTES mdb_all_sa;
1095 static int mdb_sec_inited;
1098 /** Return the library version info. */
1100 mdb_version(int *major, int *minor, int *patch)
1102 if (major) *major = MDB_VERSION_MAJOR;
1103 if (minor) *minor = MDB_VERSION_MINOR;
1104 if (patch) *patch = MDB_VERSION_PATCH;
1105 return MDB_VERSION_STRING;
1108 /** Table of descriptions for MDB @ref errors */
1109 static char *const mdb_errstr[] = {
1110 "MDB_KEYEXIST: Key/data pair already exists",
1111 "MDB_NOTFOUND: No matching key/data pair found",
1112 "MDB_PAGE_NOTFOUND: Requested page not found",
1113 "MDB_CORRUPTED: Located page was wrong type",
1114 "MDB_PANIC: Update of meta page failed",
1115 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1116 "MDB_INVALID: File is not an MDB file",
1117 "MDB_MAP_FULL: Environment mapsize limit reached",
1118 "MDB_DBS_FULL: Environment maxdbs limit reached",
1119 "MDB_READERS_FULL: Environment maxreaders limit reached",
1120 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1121 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1122 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1123 "MDB_PAGE_FULL: Internal error - page has no more space",
1124 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1125 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1126 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1127 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1128 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1132 mdb_strerror(int err)
1136 return ("Successful return: 0");
1138 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1139 i = err - MDB_KEYEXIST;
1140 return mdb_errstr[i];
1143 return strerror(err);
1147 /** Display a key in hexadecimal and return the address of the result.
1148 * @param[in] key the key to display
1149 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1150 * @return The key in hexadecimal form.
1153 mdb_dkey(MDB_val *key, char *buf)
1156 unsigned char *c = key->mv_data;
1162 if (key->mv_size > MDB_MAXKEYSIZE)
1163 return "MDB_MAXKEYSIZE";
1164 /* may want to make this a dynamic check: if the key is mostly
1165 * printable characters, print it as-is instead of converting to hex.
1169 for (i=0; i<key->mv_size; i++)
1170 ptr += sprintf(ptr, "%02x", *c++);
1172 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1177 /** Display all the keys in the page. */
1179 mdb_page_list(MDB_page *mp)
1182 unsigned int i, nkeys, nsize;
1186 nkeys = NUMKEYS(mp);
1187 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1188 for (i=0; i<nkeys; i++) {
1189 node = NODEPTR(mp, i);
1190 key.mv_size = node->mn_ksize;
1191 key.mv_data = node->mn_data;
1192 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1193 if (IS_BRANCH(mp)) {
1194 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1197 if (F_ISSET(node->mn_flags, F_BIGDATA))
1198 nsize += sizeof(pgno_t);
1200 nsize += NODEDSZ(node);
1201 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1207 mdb_cursor_chk(MDB_cursor *mc)
1213 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1214 for (i=0; i<mc->mc_top; i++) {
1216 node = NODEPTR(mp, mc->mc_ki[i]);
1217 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1220 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1226 /** Count all the pages in each DB and in the freelist
1227 * and make sure it matches the actual number of pages
1230 static void mdb_audit(MDB_txn *txn)
1234 MDB_ID freecount, count;
1239 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1240 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1241 freecount += *(MDB_ID *)data.mv_data;
1244 for (i = 0; i<txn->mt_numdbs; i++) {
1246 mdb_cursor_init(&mc, txn, i, &mx);
1247 if (txn->mt_dbs[i].md_root == P_INVALID)
1249 count += txn->mt_dbs[i].md_branch_pages +
1250 txn->mt_dbs[i].md_leaf_pages +
1251 txn->mt_dbs[i].md_overflow_pages;
1252 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1253 mdb_page_search(&mc, NULL, 0);
1257 mp = mc.mc_pg[mc.mc_top];
1258 for (j=0; j<NUMKEYS(mp); j++) {
1259 MDB_node *leaf = NODEPTR(mp, j);
1260 if (leaf->mn_flags & F_SUBDATA) {
1262 memcpy(&db, NODEDATA(leaf), sizeof(db));
1263 count += db.md_branch_pages + db.md_leaf_pages +
1264 db.md_overflow_pages;
1268 while (mdb_cursor_sibling(&mc, 1) == 0);
1271 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1272 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1273 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1279 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1281 return txn->mt_dbxs[dbi].md_cmp(a, b);
1285 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1287 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1290 /** Allocate memory for a page.
1291 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1294 mdb_page_malloc(MDB_txn *txn, unsigned num)
1296 MDB_env *env = txn->mt_env;
1297 MDB_page *ret = env->me_dpages;
1298 size_t sz = env->me_psize;
1301 VGMEMP_ALLOC(env, ret, sz);
1302 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1303 env->me_dpages = ret->mp_next;
1309 if ((ret = malloc(sz)) != NULL) {
1310 VGMEMP_ALLOC(env, ret, sz);
1315 /** Free a single page.
1316 * Saves single pages to a list, for future reuse.
1317 * (This is not used for multi-page overflow pages.)
1320 mdb_page_free(MDB_env *env, MDB_page *mp)
1322 mp->mp_next = env->me_dpages;
1323 VGMEMP_FREE(env, mp);
1324 env->me_dpages = mp;
1327 /* Free a dirty page */
1329 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1331 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1332 mdb_page_free(env, dp);
1334 /* large pages just get freed directly */
1335 VGMEMP_FREE(env, dp);
1340 /** Return all dirty pages to dpage list */
1342 mdb_dlist_free(MDB_txn *txn)
1344 MDB_env *env = txn->mt_env;
1345 MDB_ID2L dl = txn->mt_u.dirty_list;
1346 unsigned i, n = dl[0].mid;
1348 for (i = 1; i <= n; i++) {
1349 mdb_dpage_free(env, dl[i].mptr);
1354 /* Set or clear P_KEEP in non-overflow, non-sub pages in this txn's cursors.
1355 * @param[in] mc A cursor handle for the current operation.
1356 * @param[in] pflags Flags of the pages to update:
1357 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1360 mdb_cursorpages_mark(MDB_cursor *mc, unsigned pflags)
1362 MDB_txn *txn = mc->mc_txn;
1367 if (mc->mc_flags & C_UNTRACK)
1368 mc = NULL; /* will find mc in mt_cursors */
1369 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1370 for (; mc; mc=mc->mc_next) {
1371 for (m3 = mc; m3->mc_flags & C_INITIALIZED; m3 = &mx->mx_cursor) {
1372 for (j=0; j<m3->mc_snum; j++)
1373 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1375 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1376 mx = m3->mc_xcursor;
1386 static int mdb_page_flush(MDB_txn *txn);
1388 /** Spill pages from the dirty list back to disk.
1389 * This is intended to prevent running into #MDB_TXN_FULL situations,
1390 * but note that they may still occur in a few cases:
1391 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1392 * are too many of these dirtied in one txn, the txn may still get
1394 * 2) child txns may run out of space if their parents dirtied a
1395 * lot of pages and never spilled them. TODO: we probably should do
1396 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1397 * the parent's dirty_room is below a given threshold.
1398 * 3) our estimate of the txn size could be too small. At the
1399 * moment this seems unlikely.
1401 * Otherwise, if not using nested txns, it is expected that apps will
1402 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1403 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1404 * If the txn never references them again, they can be left alone.
1405 * If the txn only reads them, they can be used without any fuss.
1406 * If the txn writes them again, they can be dirtied immediately without
1407 * going thru all of the work of #mdb_page_touch(). Such references are
1408 * handled by #mdb_page_unspill().
1410 * Also note, we never spill DB root pages, nor pages of active cursors,
1411 * because we'll need these back again soon anyway. And in nested txns,
1412 * we can't spill a page in a child txn if it was already spilled in a
1413 * parent txn. That would alter the parent txns' data even though
1414 * the child hasn't committed yet, and we'd have no way to undo it if
1415 * the child aborted.
1417 * @param[in] m0 cursor A cursor handle identifying the transaction and
1418 * database for which we are checking space.
1419 * @param[in] key For a put operation, the key being stored.
1420 * @param[in] data For a put operation, the data being stored.
1421 * @return 0 on success, non-zero on failure.
1424 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1426 MDB_txn *txn = m0->mc_txn;
1428 MDB_ID2L dl = txn->mt_u.dirty_list;
1432 if (m0->mc_flags & C_SUB)
1435 /* Estimate how much space this op will take */
1436 i = m0->mc_db->md_depth;
1437 /* Named DBs also dirty the main DB */
1438 if (m0->mc_dbi > MAIN_DBI)
1439 i += txn->mt_dbs[MAIN_DBI].md_depth;
1440 /* For puts, roughly factor in the key+data size */
1442 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1443 i += i; /* double it for good measure */
1445 if (txn->mt_dirty_room > i)
1448 if (!txn->mt_spill_pgs) {
1449 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1450 if (!txn->mt_spill_pgs)
1454 /* Mark all the dirty root pages we want to preserve */
1455 for (i=0; i<txn->mt_numdbs; i++) {
1456 if (txn->mt_dbflags[i] & DB_DIRTY) {
1457 pgno_t pgno = txn->mt_dbs[i].md_root;
1458 if (pgno == P_INVALID)
1460 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1462 if ((dp->mp_flags & P_DIRTY) && level <= 1)
1463 dp->mp_flags |= P_KEEP;
1467 /* Preserve pages used by cursors */
1468 mdb_cursorpages_mark(m0, P_DIRTY);
1470 /* Save the page IDs of all the pages we're flushing */
1471 for (i=1; i<=dl[0].mid; i++) {
1473 if (dp->mp_flags & P_KEEP)
1475 /* Can't spill twice, make sure it's not already in a parent's
1478 if (txn->mt_parent) {
1480 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1481 if (tx2->mt_spill_pgs) {
1482 j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
1483 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
1484 dp->mp_flags |= P_KEEP;
1492 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
1495 mdb_midl_sort(txn->mt_spill_pgs);
1497 rc = mdb_page_flush(txn);
1499 mdb_cursorpages_mark(m0, P_DIRTY|P_KEEP);
1503 if (txn->mt_parent) {
1505 pgno_t pgno = dl[i].mid;
1506 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1507 /* dirty pages that are dirty in an ancestor don't
1508 * count against this txn's dirty_room.
1510 for (i=1; i<=dl[0].mid; i++) {
1511 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1512 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1513 if (j <= tx2->mt_u.dirty_list[0].mid &&
1514 tx2->mt_u.dirty_list[j].mid == pgno) {
1515 txn->mt_dirty_room++;
1521 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1523 txn->mt_flags |= MDB_TXN_SPILLS;
1525 txn->mt_flags |= MDB_TXN_ERROR;
1530 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1532 mdb_find_oldest(MDB_txn *txn)
1535 txnid_t mr, oldest = txn->mt_txnid - 1;
1536 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1537 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1547 /** Add a page to the txn's dirty list */
1549 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1552 int (*insert)(MDB_ID2L, MDB_ID2 *);
1554 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1555 insert = mdb_mid2l_append;
1557 insert = mdb_mid2l_insert;
1559 mid.mid = mp->mp_pgno;
1561 insert(txn->mt_u.dirty_list, &mid);
1562 txn->mt_dirty_room--;
1565 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1566 * me_pghead and mt_next_pgno.
1568 * If there are free pages available from older transactions, they
1569 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1570 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1571 * and move me_pglast to say which records were consumed. Only this
1572 * function can create me_pghead and move me_pglast/mt_next_pgno.
1573 * @param[in] mc cursor A cursor handle identifying the transaction and
1574 * database for which we are allocating.
1575 * @param[in] num the number of pages to allocate.
1576 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1577 * will always be satisfied by a single contiguous chunk of memory.
1578 * @return 0 on success, non-zero on failure.
1581 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1583 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1584 /* Get at most <Max_retries> more freeDB records once me_pghead
1585 * has enough pages. If not enough, use new pages from the map.
1586 * If <Paranoid> and mc is updating the freeDB, only get new
1587 * records if me_pghead is empty. Then the freelist cannot play
1588 * catch-up with itself by growing while trying to save it.
1590 enum { Paranoid = 1, Max_retries = 500 };
1592 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1594 int rc, n2 = num-1, retry = Max_retries;
1595 MDB_txn *txn = mc->mc_txn;
1596 MDB_env *env = txn->mt_env;
1597 pgno_t pgno, *mop = env->me_pghead;
1598 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1600 txnid_t oldest = 0, last;
1606 /* If our dirty list is already full, we can't do anything */
1607 if (txn->mt_dirty_room == 0)
1608 return MDB_TXN_FULL;
1610 for (op = MDB_FIRST;; op = MDB_NEXT) {
1613 pgno_t *idl, old_id, new_id;
1615 /* Seek a big enough contiguous page range. Prefer
1616 * pages at the tail, just truncating the list.
1618 if (mop_len >= (unsigned)num) {
1622 if (mop[i-n2] == pgno+n2)
1624 } while (--i >= (unsigned)num);
1625 if (Max_retries < INT_MAX && --retry < 0)
1629 if (op == MDB_FIRST) { /* 1st iteration */
1630 /* Prepare to fetch more and coalesce */
1631 oldest = mdb_find_oldest(txn);
1632 last = env->me_pglast;
1633 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1636 key.mv_data = &last; /* will look up last+1 */
1637 key.mv_size = sizeof(last);
1639 if (Paranoid && mc->mc_dbi == FREE_DBI)
1642 if (Paranoid && retry < 0 && mop_len)
1646 /* Do not fetch more if the record will be too recent */
1649 rc = mdb_cursor_get(&m2, &key, NULL, op);
1651 if (rc == MDB_NOTFOUND)
1655 last = *(txnid_t*)key.mv_data;
1658 np = m2.mc_pg[m2.mc_top];
1659 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1660 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1663 idl = (MDB_ID *) data.mv_data;
1666 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1669 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1671 mop = env->me_pghead;
1673 env->me_pglast = last;
1675 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1676 last, txn->mt_dbs[FREE_DBI].md_root, i));
1678 DPRINTF(("IDL %"Z"u", idl[k]));
1680 /* Merge in descending sorted order */
1683 mop[0] = (pgno_t)-1;
1687 for (; old_id < new_id; old_id = mop[--j])
1694 /* Use new pages from the map when nothing suitable in the freeDB */
1696 pgno = txn->mt_next_pgno;
1697 if (pgno + num >= env->me_maxpg) {
1698 DPUTS("DB size maxed out");
1699 return MDB_MAP_FULL;
1703 if (env->me_flags & MDB_WRITEMAP) {
1704 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1706 if (!(np = mdb_page_malloc(txn, num)))
1710 mop[0] = mop_len -= num;
1711 /* Move any stragglers down */
1712 for (j = i-num; j < mop_len; )
1713 mop[++j] = mop[++i];
1715 txn->mt_next_pgno = pgno + num;
1718 mdb_page_dirty(txn, np);
1724 /** Copy the used portions of a non-overflow page.
1725 * @param[in] dst page to copy into
1726 * @param[in] src page to copy from
1727 * @param[in] psize size of a page
1730 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1732 enum { Align = sizeof(pgno_t) };
1733 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1735 /* If page isn't full, just copy the used portion. Adjust
1736 * alignment so memcpy may copy words instead of bytes.
1738 if ((unused &= -Align) && !IS_LEAF2(src)) {
1740 memcpy(dst, src, (lower + (Align-1)) & -Align);
1741 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1744 memcpy(dst, src, psize - unused);
1748 /** Pull a page off the txn's spill list, if present.
1749 * If a page being referenced was spilled to disk in this txn, bring
1750 * it back and make it dirty/writable again.
1751 * @param[in] tx0 the transaction handle.
1752 * @param[in] mp the page being referenced.
1753 * @param[out] ret the writable page, if any. ret is unchanged if
1754 * mp wasn't spilled.
1757 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1759 MDB_env *env = tx0->mt_env;
1762 pgno_t pgno = mp->mp_pgno;
1764 for (txn = tx0; txn; txn=txn->mt_parent) {
1765 if (!txn->mt_spill_pgs)
1767 x = mdb_midl_search(txn->mt_spill_pgs, pgno);
1768 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
1771 if (IS_OVERFLOW(mp))
1775 if (env->me_flags & MDB_WRITEMAP) {
1778 np = mdb_page_malloc(txn, num);
1782 memcpy(np, mp, num * env->me_psize);
1784 mdb_page_copy(np, mp, env->me_psize);
1787 /* If in current txn, this page is no longer spilled */
1788 for (; x < txn->mt_spill_pgs[0]; x++)
1789 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
1790 txn->mt_spill_pgs[0]--;
1791 } /* otherwise, if belonging to a parent txn, the
1792 * page remains spilled until child commits
1795 if (txn->mt_parent) {
1797 /* If this page is also in a parent's dirty list, then
1798 * it's already accounted in dirty_room, and we need to
1799 * cancel out the decrement that mdb_page_dirty does.
1801 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1802 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1803 if (x <= tx2->mt_u.dirty_list[0].mid &&
1804 tx2->mt_u.dirty_list[x].mid == pgno) {
1805 txn->mt_dirty_room++;
1810 mdb_page_dirty(tx0, np);
1811 np->mp_flags |= P_DIRTY;
1819 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1820 * @param[in] mc cursor pointing to the page to be touched
1821 * @return 0 on success, non-zero on failure.
1824 mdb_page_touch(MDB_cursor *mc)
1826 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1827 MDB_txn *txn = mc->mc_txn;
1828 MDB_cursor *m2, *m3;
1833 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1834 if (txn->mt_flags & MDB_TXN_SPILLS) {
1836 rc = mdb_page_unspill(txn, mp, &np);
1842 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1843 (rc = mdb_page_alloc(mc, 1, &np)))
1846 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1847 assert(mp->mp_pgno != pgno);
1848 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1849 /* Update the parent page, if any, to point to the new page */
1851 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1852 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1853 SETPGNO(node, pgno);
1855 mc->mc_db->md_root = pgno;
1857 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1858 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1860 /* If txn has a parent, make sure the page is in our
1864 unsigned x = mdb_mid2l_search(dl, pgno);
1865 if (x <= dl[0].mid && dl[x].mid == pgno) {
1866 if (mp != dl[x].mptr) { /* bad cursor? */
1867 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1868 return MDB_CORRUPTED;
1873 assert(dl[0].mid < MDB_IDL_UM_MAX);
1875 np = mdb_page_malloc(txn, 1);
1880 mdb_mid2l_insert(dl, &mid);
1885 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1887 np->mp_flags |= P_DIRTY;
1890 /* Adjust cursors pointing to mp */
1891 mc->mc_pg[mc->mc_top] = np;
1893 if (mc->mc_flags & C_SUB) {
1895 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1896 m3 = &m2->mc_xcursor->mx_cursor;
1897 if (m3->mc_snum < mc->mc_snum) continue;
1898 if (m3->mc_pg[mc->mc_top] == mp)
1899 m3->mc_pg[mc->mc_top] = np;
1902 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1903 if (m2->mc_snum < mc->mc_snum) continue;
1904 if (m2->mc_pg[mc->mc_top] == mp) {
1905 m2->mc_pg[mc->mc_top] = np;
1906 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1907 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1909 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1910 if (!(leaf->mn_flags & F_SUBDATA))
1911 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1920 mdb_env_sync(MDB_env *env, int force)
1923 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1924 if (env->me_flags & MDB_WRITEMAP) {
1925 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1926 ? MS_ASYNC : MS_SYNC;
1927 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1930 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1934 if (MDB_FDATASYNC(env->me_fd))
1941 /** Back up parent txn's cursors, then grab the originals for tracking */
1943 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1945 MDB_cursor *mc, *bk;
1950 for (i = src->mt_numdbs; --i >= 0; ) {
1951 if ((mc = src->mt_cursors[i]) != NULL) {
1952 size = sizeof(MDB_cursor);
1954 size += sizeof(MDB_xcursor);
1955 for (; mc; mc = bk->mc_next) {
1961 mc->mc_db = &dst->mt_dbs[i];
1962 /* Kill pointers into src - and dst to reduce abuse: The
1963 * user may not use mc until dst ends. Otherwise we'd...
1965 mc->mc_txn = NULL; /* ...set this to dst */
1966 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1967 if ((mx = mc->mc_xcursor) != NULL) {
1968 *(MDB_xcursor *)(bk+1) = *mx;
1969 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
1971 mc->mc_next = dst->mt_cursors[i];
1972 dst->mt_cursors[i] = mc;
1979 /** Close this write txn's cursors, give parent txn's cursors back to parent.
1980 * @param[in] txn the transaction handle.
1981 * @param[in] merge true to keep changes to parent cursors, false to revert.
1982 * @return 0 on success, non-zero on failure.
1985 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1987 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
1991 for (i = txn->mt_numdbs; --i >= 0; ) {
1992 for (mc = cursors[i]; mc; mc = next) {
1994 if ((bk = mc->mc_backup) != NULL) {
1996 /* Commit changes to parent txn */
1997 mc->mc_next = bk->mc_next;
1998 mc->mc_backup = bk->mc_backup;
1999 mc->mc_txn = bk->mc_txn;
2000 mc->mc_db = bk->mc_db;
2001 mc->mc_dbflag = bk->mc_dbflag;
2002 if ((mx = mc->mc_xcursor) != NULL)
2003 mx->mx_cursor.mc_txn = bk->mc_txn;
2005 /* Abort nested txn */
2007 if ((mx = mc->mc_xcursor) != NULL)
2008 *mx = *(MDB_xcursor *)(bk+1);
2012 /* Only malloced cursors are permanently tracked. */
2020 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2023 mdb_txn_reset0(MDB_txn *txn, const char *act);
2025 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2031 Pidset = F_SETLK, Pidcheck = F_GETLK
2035 /** Set or check a pid lock. Set returns 0 on success.
2036 * Check returns 0 if the process is certainly dead, nonzero if it may
2037 * be alive (the lock exists or an error happened so we do not know).
2039 * On Windows Pidset is a no-op, we merely check for the existence
2040 * of the process with the given pid. On POSIX we use a single byte
2041 * lock on the lockfile, set at an offset equal to the pid.
2044 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2046 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2049 if (op == Pidcheck) {
2050 h = OpenProcess(env->me_pidquery, FALSE, pid);
2051 /* No documented "no such process" code, but other program use this: */
2053 return ErrCode() != ERROR_INVALID_PARAMETER;
2054 /* A process exists until all handles to it close. Has it exited? */
2055 ret = WaitForSingleObject(h, 0) != 0;
2062 struct flock lock_info;
2063 memset(&lock_info, 0, sizeof(lock_info));
2064 lock_info.l_type = F_WRLCK;
2065 lock_info.l_whence = SEEK_SET;
2066 lock_info.l_start = pid;
2067 lock_info.l_len = 1;
2068 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2069 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2071 } else if ((rc = ErrCode()) == EINTR) {
2079 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2080 * @param[in] txn the transaction handle to initialize
2081 * @return 0 on success, non-zero on failure.
2084 mdb_txn_renew0(MDB_txn *txn)
2086 MDB_env *env = txn->mt_env;
2089 int rc, new_notls = 0;
2092 txn->mt_numdbs = env->me_numdbs;
2093 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2095 if (txn->mt_flags & MDB_TXN_RDONLY) {
2096 if (!env->me_txns) {
2097 i = mdb_env_pick_meta(env);
2098 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2099 txn->mt_u.reader = NULL;
2101 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2102 pthread_getspecific(env->me_txkey);
2104 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2105 return MDB_BAD_RSLOT;
2107 pid_t pid = env->me_pid;
2108 pthread_t tid = pthread_self();
2110 if (!(env->me_flags & MDB_LIVE_READER)) {
2111 rc = mdb_reader_pid(env, Pidset, pid);
2113 UNLOCK_MUTEX_R(env);
2116 env->me_flags |= MDB_LIVE_READER;
2120 for (i=0; i<env->me_txns->mti_numreaders; i++)
2121 if (env->me_txns->mti_readers[i].mr_pid == 0)
2123 if (i == env->me_maxreaders) {
2124 UNLOCK_MUTEX_R(env);
2125 return MDB_READERS_FULL;
2127 env->me_txns->mti_readers[i].mr_pid = pid;
2128 env->me_txns->mti_readers[i].mr_tid = tid;
2129 if (i >= env->me_txns->mti_numreaders)
2130 env->me_txns->mti_numreaders = i+1;
2131 /* Save numreaders for un-mutexed mdb_env_close() */
2132 env->me_numreaders = env->me_txns->mti_numreaders;
2133 UNLOCK_MUTEX_R(env);
2134 r = &env->me_txns->mti_readers[i];
2135 new_notls = (env->me_flags & MDB_NOTLS);
2136 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2141 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2142 txn->mt_u.reader = r;
2144 txn->mt_toggle = txn->mt_txnid & 1;
2148 txn->mt_txnid = env->me_txns->mti_txnid;
2149 txn->mt_toggle = txn->mt_txnid & 1;
2152 if (txn->mt_txnid == mdb_debug_start)
2155 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2156 txn->mt_u.dirty_list = env->me_dirty_list;
2157 txn->mt_u.dirty_list[0].mid = 0;
2158 txn->mt_free_pgs = env->me_free_pgs;
2159 txn->mt_free_pgs[0] = 0;
2160 txn->mt_spill_pgs = NULL;
2164 /* Copy the DB info and flags */
2165 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2167 /* Moved to here to avoid a data race in read TXNs */
2168 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2170 for (i=2; i<txn->mt_numdbs; i++) {
2171 x = env->me_dbflags[i];
2172 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2173 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2175 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2177 if (env->me_maxpg < txn->mt_next_pgno) {
2178 mdb_txn_reset0(txn, "renew0-mapfail");
2180 txn->mt_u.reader->mr_pid = 0;
2181 txn->mt_u.reader = NULL;
2183 return MDB_MAP_RESIZED;
2190 mdb_txn_renew(MDB_txn *txn)
2194 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2197 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2198 DPUTS("environment had fatal error, must shutdown!");
2202 rc = mdb_txn_renew0(txn);
2203 if (rc == MDB_SUCCESS) {
2204 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2205 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2206 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2212 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2216 int rc, size, tsize = sizeof(MDB_txn);
2218 if (env->me_flags & MDB_FATAL_ERROR) {
2219 DPUTS("environment had fatal error, must shutdown!");
2222 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2225 /* Nested transactions: Max 1 child, write txns only, no writemap */
2226 if (parent->mt_child ||
2227 (flags & MDB_RDONLY) ||
2228 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2229 (env->me_flags & MDB_WRITEMAP))
2231 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2233 tsize = sizeof(MDB_ntxn);
2235 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2236 if (!(flags & MDB_RDONLY))
2237 size += env->me_maxdbs * sizeof(MDB_cursor *);
2239 if ((txn = calloc(1, size)) == NULL) {
2240 DPRINTF(("calloc: %s", strerror(ErrCode())));
2243 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2244 if (flags & MDB_RDONLY) {
2245 txn->mt_flags |= MDB_TXN_RDONLY;
2246 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2248 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2249 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2255 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2256 if (!txn->mt_u.dirty_list ||
2257 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2259 free(txn->mt_u.dirty_list);
2263 txn->mt_txnid = parent->mt_txnid;
2264 txn->mt_toggle = parent->mt_toggle;
2265 txn->mt_dirty_room = parent->mt_dirty_room;
2266 txn->mt_u.dirty_list[0].mid = 0;
2267 txn->mt_spill_pgs = NULL;
2268 txn->mt_next_pgno = parent->mt_next_pgno;
2269 parent->mt_child = txn;
2270 txn->mt_parent = parent;
2271 txn->mt_numdbs = parent->mt_numdbs;
2272 txn->mt_flags = parent->mt_flags;
2273 txn->mt_dbxs = parent->mt_dbxs;
2274 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2275 /* Copy parent's mt_dbflags, but clear DB_NEW */
2276 for (i=0; i<txn->mt_numdbs; i++)
2277 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2279 ntxn = (MDB_ntxn *)txn;
2280 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2281 if (env->me_pghead) {
2282 size = MDB_IDL_SIZEOF(env->me_pghead);
2283 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2285 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2290 rc = mdb_cursor_shadow(parent, txn);
2292 mdb_txn_reset0(txn, "beginchild-fail");
2294 rc = mdb_txn_renew0(txn);
2300 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2301 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2302 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2309 mdb_txn_env(MDB_txn *txn)
2311 if(!txn) return NULL;
2315 /** Export or close DBI handles opened in this txn. */
2317 mdb_dbis_update(MDB_txn *txn, int keep)
2320 MDB_dbi n = txn->mt_numdbs;
2321 MDB_env *env = txn->mt_env;
2322 unsigned char *tdbflags = txn->mt_dbflags;
2324 for (i = n; --i >= 2;) {
2325 if (tdbflags[i] & DB_NEW) {
2327 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2329 char *ptr = env->me_dbxs[i].md_name.mv_data;
2330 env->me_dbxs[i].md_name.mv_data = NULL;
2331 env->me_dbxs[i].md_name.mv_size = 0;
2332 env->me_dbflags[i] = 0;
2337 if (keep && env->me_numdbs < n)
2341 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2342 * May be called twice for readonly txns: First reset it, then abort.
2343 * @param[in] txn the transaction handle to reset
2344 * @param[in] act why the transaction is being reset
2347 mdb_txn_reset0(MDB_txn *txn, const char *act)
2349 MDB_env *env = txn->mt_env;
2351 /* Close any DBI handles opened in this txn */
2352 mdb_dbis_update(txn, 0);
2354 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2355 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2356 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2358 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2359 if (txn->mt_u.reader) {
2360 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2361 if (!(env->me_flags & MDB_NOTLS))
2362 txn->mt_u.reader = NULL; /* txn does not own reader */
2364 txn->mt_numdbs = 0; /* close nothing if called again */
2365 txn->mt_dbxs = NULL; /* mark txn as reset */
2367 mdb_cursors_close(txn, 0);
2369 if (!(env->me_flags & MDB_WRITEMAP)) {
2370 mdb_dlist_free(txn);
2372 mdb_midl_free(env->me_pghead);
2374 if (txn->mt_parent) {
2375 txn->mt_parent->mt_child = NULL;
2376 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2377 mdb_midl_free(txn->mt_free_pgs);
2378 mdb_midl_free(txn->mt_spill_pgs);
2379 free(txn->mt_u.dirty_list);
2383 if (mdb_midl_shrink(&txn->mt_free_pgs))
2384 env->me_free_pgs = txn->mt_free_pgs;
2385 env->me_pghead = NULL;
2389 /* The writer mutex was locked in mdb_txn_begin. */
2390 UNLOCK_MUTEX_W(env);
2395 mdb_txn_reset(MDB_txn *txn)
2400 /* This call is only valid for read-only txns */
2401 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2404 mdb_txn_reset0(txn, "reset");
2408 mdb_txn_abort(MDB_txn *txn)
2414 mdb_txn_abort(txn->mt_child);
2416 mdb_txn_reset0(txn, "abort");
2417 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2418 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2419 txn->mt_u.reader->mr_pid = 0;
2424 /** Save the freelist as of this transaction to the freeDB.
2425 * This changes the freelist. Keep trying until it stabilizes.
2428 mdb_freelist_save(MDB_txn *txn)
2430 /* env->me_pghead[] can grow and shrink during this call.
2431 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2432 * Page numbers cannot disappear from txn->mt_free_pgs[].
2435 MDB_env *env = txn->mt_env;
2436 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2437 txnid_t pglast = 0, head_id = 0;
2438 pgno_t freecnt = 0, *free_pgs, *mop;
2439 ssize_t head_room = 0, total_room = 0, mop_len;
2441 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2443 if (env->me_pghead) {
2444 /* Make sure first page of freeDB is touched and on freelist */
2445 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2446 if (rc && rc != MDB_NOTFOUND)
2451 /* Come back here after each Put() in case freelist changed */
2454 /* If using records from freeDB which we have not yet
2455 * deleted, delete them and any we reserved for me_pghead.
2457 while (pglast < env->me_pglast) {
2458 rc = mdb_cursor_first(&mc, &key, NULL);
2461 pglast = head_id = *(txnid_t *)key.mv_data;
2462 total_room = head_room = 0;
2463 assert(pglast <= env->me_pglast);
2464 rc = mdb_cursor_del(&mc, 0);
2469 /* Save the IDL of pages freed by this txn, to a single record */
2470 if (freecnt < txn->mt_free_pgs[0]) {
2472 /* Make sure last page of freeDB is touched and on freelist */
2473 key.mv_size = MDB_MAXKEYSIZE+1;
2475 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2476 if (rc && rc != MDB_NOTFOUND)
2479 free_pgs = txn->mt_free_pgs;
2480 /* Write to last page of freeDB */
2481 key.mv_size = sizeof(txn->mt_txnid);
2482 key.mv_data = &txn->mt_txnid;
2484 freecnt = free_pgs[0];
2485 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2486 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2489 /* Retry if mt_free_pgs[] grew during the Put() */
2490 free_pgs = txn->mt_free_pgs;
2491 } while (freecnt < free_pgs[0]);
2492 mdb_midl_sort(free_pgs);
2493 memcpy(data.mv_data, free_pgs, data.mv_size);
2496 unsigned int i = free_pgs[0];
2497 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2498 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2500 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2506 mop = env->me_pghead;
2507 mop_len = mop ? mop[0] : 0;
2509 /* Reserve records for me_pghead[]. Split it if multi-page,
2510 * to avoid searching freeDB for a page range. Use keys in
2511 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2513 if (total_room >= mop_len) {
2514 if (total_room == mop_len || --more < 0)
2516 } else if (head_room >= maxfree_1pg && head_id > 1) {
2517 /* Keep current record (overflow page), add a new one */
2521 /* (Re)write {key = head_id, IDL length = head_room} */
2522 total_room -= head_room;
2523 head_room = mop_len - total_room;
2524 if (head_room > maxfree_1pg && head_id > 1) {
2525 /* Overflow multi-page for part of me_pghead */
2526 head_room /= head_id; /* amortize page sizes */
2527 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2528 } else if (head_room < 0) {
2529 /* Rare case, not bothering to delete this record */
2532 key.mv_size = sizeof(head_id);
2533 key.mv_data = &head_id;
2534 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2535 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2538 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2539 total_room += head_room;
2542 /* Fill in the reserved, touched me_pghead records */
2548 rc = mdb_cursor_first(&mc, &key, &data);
2549 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2550 unsigned flags = MDB_CURRENT;
2551 txnid_t id = *(txnid_t *)key.mv_data;
2552 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2555 assert(len >= 0 && id <= env->me_pglast);
2557 if (len > mop_len) {
2559 data.mv_size = (len + 1) * sizeof(MDB_ID);
2562 data.mv_data = mop -= len;
2565 rc = mdb_cursor_put(&mc, &key, &data, flags);
2567 if (rc || !(mop_len -= len))
2574 /** Flush dirty pages to the map, after clearing their dirty flag.
2577 mdb_page_flush(MDB_txn *txn)
2579 MDB_env *env = txn->mt_env;
2580 MDB_ID2L dl = txn->mt_u.dirty_list;
2581 unsigned psize = env->me_psize, j;
2582 int i, pagecount = dl[0].mid, rc;
2583 size_t size = 0, pos = 0;
2585 MDB_page *dp = NULL;
2589 struct iovec iov[MDB_COMMIT_PAGES];
2590 ssize_t wpos = 0, wsize = 0, wres;
2591 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2596 if (env->me_flags & MDB_WRITEMAP) {
2597 /* Clear dirty flags */
2598 for (i=1; i<=pagecount; i++) {
2600 /* Don't flush this page yet */
2601 if (dp->mp_flags & P_KEEP) {
2602 dp->mp_flags ^= P_KEEP;
2606 dp->mp_flags &= ~P_DIRTY;
2612 /* Write the pages */
2614 if (i <= pagecount) {
2616 /* Don't flush this page yet */
2617 if (dp->mp_flags & P_KEEP) {
2618 dp->mp_flags ^= P_KEEP;
2623 /* clear dirty flag */
2624 dp->mp_flags &= ~P_DIRTY;
2627 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2632 /* Windows actually supports scatter/gather I/O, but only on
2633 * unbuffered file handles. Since we're relying on the OS page
2634 * cache for all our data, that's self-defeating. So we just
2635 * write pages one at a time. We use the ov structure to set
2636 * the write offset, to at least save the overhead of a Seek
2639 DPRINTF(("committing page %"Z"u", pgno));
2640 memset(&ov, 0, sizeof(ov));
2641 ov.Offset = pos & 0xffffffff;
2642 ov.OffsetHigh = pos >> 16 >> 16;
2643 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2645 DPRINTF(("WriteFile: %d", rc));
2649 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2650 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2652 /* Write previous page(s) */
2653 #ifdef MDB_USE_PWRITEV
2654 wres = pwritev(env->me_fd, iov, n, wpos);
2657 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2659 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2661 DPRINTF(("lseek: %s", strerror(rc)));
2664 wres = writev(env->me_fd, iov, n);
2667 if (wres != wsize) {
2670 DPRINTF(("Write error: %s", strerror(rc)));
2672 rc = EIO; /* TODO: Use which error code? */
2673 DPUTS("short write, filesystem full?");
2684 DPRINTF(("committing page %"Z"u", pgno));
2685 next_pos = pos + size;
2686 iov[n].iov_len = size;
2687 iov[n].iov_base = (char *)dp;
2694 for (i=1; i<=pagecount; i++) {
2696 /* This is a page we skipped above */
2699 dl[j].mid = dp->mp_pgno;
2702 mdb_dpage_free(env, dp);
2710 mdb_txn_commit(MDB_txn *txn)
2716 assert(txn != NULL);
2717 assert(txn->mt_env != NULL);
2719 if (txn->mt_child) {
2720 rc = mdb_txn_commit(txn->mt_child);
2721 txn->mt_child = NULL;
2728 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2729 mdb_dbis_update(txn, 1);
2730 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2735 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2736 DPUTS("error flag is set, can't commit");
2738 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2743 if (txn->mt_parent) {
2744 MDB_txn *parent = txn->mt_parent;
2748 /* Append our free list to parent's */
2749 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2752 mdb_midl_free(txn->mt_free_pgs);
2754 parent->mt_next_pgno = txn->mt_next_pgno;
2755 parent->mt_flags = txn->mt_flags;
2757 /* Merge our cursors into parent's and close them */
2758 mdb_cursors_close(txn, 1);
2760 /* Update parent's DB table. */
2761 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2762 parent->mt_numdbs = txn->mt_numdbs;
2763 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2764 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2765 for (i=2; i<txn->mt_numdbs; i++) {
2766 /* preserve parent's DB_NEW status */
2767 x = parent->mt_dbflags[i] & DB_NEW;
2768 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2771 dst = parent->mt_u.dirty_list;
2772 src = txn->mt_u.dirty_list;
2773 /* Remove anything in our dirty list from parent's spill list */
2774 if (parent->mt_spill_pgs) {
2775 x = parent->mt_spill_pgs[0];
2777 /* zero out our dirty pages in parent spill list */
2778 for (i=1; i<=src[0].mid; i++) {
2779 if (src[i].mid < parent->mt_spill_pgs[x])
2781 if (src[i].mid > parent->mt_spill_pgs[x]) {
2787 parent->mt_spill_pgs[x] = 0;
2790 /* OK, we had a few hits, squash zeros from the spill list */
2791 if (len < parent->mt_spill_pgs[0]) {
2793 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2794 if (parent->mt_spill_pgs[y]) {
2796 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2801 parent->mt_spill_pgs[0] = len;
2804 /* Find len = length of merging our dirty list with parent's */
2806 dst[0].mid = 0; /* simplify loops */
2807 if (parent->mt_parent) {
2808 len = x + src[0].mid;
2809 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2810 for (i = x; y && i; y--) {
2811 pgno_t yp = src[y].mid;
2812 while (yp < dst[i].mid)
2814 if (yp == dst[i].mid) {
2819 } else { /* Simplify the above for single-ancestor case */
2820 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2822 /* Merge our dirty list with parent's */
2824 for (i = len; y; dst[i--] = src[y--]) {
2825 pgno_t yp = src[y].mid;
2826 while (yp < dst[x].mid)
2827 dst[i--] = dst[x--];
2828 if (yp == dst[x].mid)
2829 free(dst[x--].mptr);
2833 free(txn->mt_u.dirty_list);
2834 parent->mt_dirty_room = txn->mt_dirty_room;
2835 if (txn->mt_spill_pgs) {
2836 if (parent->mt_spill_pgs) {
2837 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2838 mdb_midl_free(txn->mt_spill_pgs);
2839 mdb_midl_sort(parent->mt_spill_pgs);
2841 parent->mt_spill_pgs = txn->mt_spill_pgs;
2845 parent->mt_child = NULL;
2846 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2851 if (txn != env->me_txn) {
2852 DPUTS("attempt to commit unknown transaction");
2857 mdb_cursors_close(txn, 0);
2859 if (!txn->mt_u.dirty_list[0].mid &&
2860 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2863 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2864 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2866 /* Update DB root pointers */
2867 if (txn->mt_numdbs > 2) {
2871 data.mv_size = sizeof(MDB_db);
2873 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2874 for (i = 2; i < txn->mt_numdbs; i++) {
2875 if (txn->mt_dbflags[i] & DB_DIRTY) {
2876 data.mv_data = &txn->mt_dbs[i];
2877 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2884 rc = mdb_freelist_save(txn);
2888 mdb_midl_free(env->me_pghead);
2889 env->me_pghead = NULL;
2890 if (mdb_midl_shrink(&txn->mt_free_pgs))
2891 env->me_free_pgs = txn->mt_free_pgs;
2897 if ((rc = mdb_page_flush(txn)) ||
2898 (rc = mdb_env_sync(env, 0)) ||
2899 (rc = mdb_env_write_meta(txn)))
2905 mdb_dbis_update(txn, 1);
2907 UNLOCK_MUTEX_W(env);
2917 /** Read the environment parameters of a DB environment before
2918 * mapping it into memory.
2919 * @param[in] env the environment handle
2920 * @param[out] meta address of where to store the meta information
2921 * @return 0 on success, non-zero on failure.
2924 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2931 /* We don't know the page size yet, so use a minimum value.
2932 * Read both meta pages so we can use the latest one.
2935 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2939 memset(&ov, 0, sizeof(ov));
2941 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2942 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2945 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2947 if (rc != MDB_PAGESIZE) {
2948 if (rc == 0 && off == 0)
2950 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2951 DPRINTF(("read: %s", mdb_strerror(rc)));
2955 p = (MDB_page *)&pbuf;
2957 if (!F_ISSET(p->mp_flags, P_META)) {
2958 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
2963 if (m->mm_magic != MDB_MAGIC) {
2964 DPUTS("meta has invalid magic");
2968 if (m->mm_version != MDB_DATA_VERSION) {
2969 DPRINTF(("database is version %u, expected version %u",
2970 m->mm_version, MDB_DATA_VERSION));
2971 return MDB_VERSION_MISMATCH;
2974 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2980 /** Write the environment parameters of a freshly created DB environment.
2981 * @param[in] env the environment handle
2982 * @param[out] meta address of where to store the meta information
2983 * @return 0 on success, non-zero on failure.
2986 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2994 memset(&ov, 0, sizeof(ov));
2995 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
2997 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3000 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3001 len = pwrite(fd, ptr, size, pos); \
3002 rc = (len >= 0); } while(0)
3005 DPUTS("writing new meta page");
3007 GET_PAGESIZE(psize);
3009 meta->mm_magic = MDB_MAGIC;
3010 meta->mm_version = MDB_DATA_VERSION;
3011 meta->mm_mapsize = env->me_mapsize;
3012 meta->mm_psize = psize;
3013 meta->mm_last_pg = 1;
3014 meta->mm_flags = env->me_flags & 0xffff;
3015 meta->mm_flags |= MDB_INTEGERKEY;
3016 meta->mm_dbs[0].md_root = P_INVALID;
3017 meta->mm_dbs[1].md_root = P_INVALID;
3019 p = calloc(2, psize);
3021 p->mp_flags = P_META;
3022 *(MDB_meta *)METADATA(p) = *meta;
3024 q = (MDB_page *)((char *)p + psize);
3026 q->mp_flags = P_META;
3027 *(MDB_meta *)METADATA(q) = *meta;
3029 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3032 else if ((unsigned) len == psize * 2)
3040 /** Update the environment info to commit a transaction.
3041 * @param[in] txn the transaction that's being committed
3042 * @return 0 on success, non-zero on failure.
3045 mdb_env_write_meta(MDB_txn *txn)
3048 MDB_meta meta, metab, *mp;
3050 int rc, len, toggle;
3059 assert(txn != NULL);
3060 assert(txn->mt_env != NULL);
3062 toggle = !txn->mt_toggle;
3063 DPRINTF(("writing meta page %d for root page %"Z"u",
3064 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3067 mp = env->me_metas[toggle];
3069 if (env->me_flags & MDB_WRITEMAP) {
3070 /* Persist any increases of mapsize config */
3071 if (env->me_mapsize > mp->mm_mapsize)
3072 mp->mm_mapsize = env->me_mapsize;
3073 mp->mm_dbs[0] = txn->mt_dbs[0];
3074 mp->mm_dbs[1] = txn->mt_dbs[1];
3075 mp->mm_last_pg = txn->mt_next_pgno - 1;
3076 mp->mm_txnid = txn->mt_txnid;
3077 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3078 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3081 ptr += env->me_psize;
3082 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3089 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3090 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3092 ptr = (char *)&meta;
3093 if (env->me_mapsize > mp->mm_mapsize) {
3094 /* Persist any increases of mapsize config */
3095 meta.mm_mapsize = env->me_mapsize;
3096 off = offsetof(MDB_meta, mm_mapsize);
3098 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3100 len = sizeof(MDB_meta) - off;
3103 meta.mm_dbs[0] = txn->mt_dbs[0];
3104 meta.mm_dbs[1] = txn->mt_dbs[1];
3105 meta.mm_last_pg = txn->mt_next_pgno - 1;
3106 meta.mm_txnid = txn->mt_txnid;
3109 off += env->me_psize;
3112 /* Write to the SYNC fd */
3113 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3114 env->me_fd : env->me_mfd;
3117 memset(&ov, 0, sizeof(ov));
3119 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3123 rc = pwrite(mfd, ptr, len, off);
3126 rc = rc < 0 ? ErrCode() : EIO;
3127 DPUTS("write failed, disk error?");
3128 /* On a failure, the pagecache still contains the new data.
3129 * Write some old data back, to prevent it from being used.
3130 * Use the non-SYNC fd; we know it will fail anyway.
3132 meta.mm_last_pg = metab.mm_last_pg;
3133 meta.mm_txnid = metab.mm_txnid;
3135 memset(&ov, 0, sizeof(ov));
3137 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3139 r2 = pwrite(env->me_fd, ptr, len, off);
3142 env->me_flags |= MDB_FATAL_ERROR;
3146 /* Memory ordering issues are irrelevant; since the entire writer
3147 * is wrapped by wmutex, all of these changes will become visible
3148 * after the wmutex is unlocked. Since the DB is multi-version,
3149 * readers will get consistent data regardless of how fresh or
3150 * how stale their view of these values is.
3152 env->me_txns->mti_txnid = txn->mt_txnid;
3157 /** Check both meta pages to see which one is newer.
3158 * @param[in] env the environment handle
3159 * @return meta toggle (0 or 1).
3162 mdb_env_pick_meta(const MDB_env *env)
3164 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3168 mdb_env_create(MDB_env **env)
3172 e = calloc(1, sizeof(MDB_env));
3176 e->me_maxreaders = DEFAULT_READERS;
3177 e->me_maxdbs = e->me_numdbs = 2;
3178 e->me_fd = INVALID_HANDLE_VALUE;
3179 e->me_lfd = INVALID_HANDLE_VALUE;
3180 e->me_mfd = INVALID_HANDLE_VALUE;
3181 #ifdef MDB_USE_POSIX_SEM
3182 e->me_rmutex = SEM_FAILED;
3183 e->me_wmutex = SEM_FAILED;
3185 e->me_pid = getpid();
3186 VGMEMP_CREATE(e,0,0);
3192 mdb_env_set_mapsize(MDB_env *env, size_t size)
3196 env->me_mapsize = size;
3198 env->me_maxpg = env->me_mapsize / env->me_psize;
3203 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3207 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3212 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3214 if (env->me_map || readers < 1)
3216 env->me_maxreaders = readers;
3221 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3223 if (!env || !readers)
3225 *readers = env->me_maxreaders;
3229 /** Further setup required for opening an MDB environment
3232 mdb_env_open2(MDB_env *env)
3234 unsigned int flags = env->me_flags;
3242 memset(&meta, 0, sizeof(meta));
3244 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3247 DPUTS("new mdbenv");
3251 /* Was a mapsize configured? */
3252 if (!env->me_mapsize) {
3253 /* If this is a new environment, take the default,
3254 * else use the size recorded in the existing env.
3256 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3257 } else if (env->me_mapsize < meta.mm_mapsize) {
3258 /* If the configured size is smaller, make sure it's
3259 * still big enough. Silently round up to minimum if not.
3261 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3262 if (env->me_mapsize < minsize)
3263 env->me_mapsize = minsize;
3270 LONG sizelo, sizehi;
3271 sizelo = env->me_mapsize & 0xffffffff;
3272 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3274 /* See if we should use QueryLimited */
3276 if ((rc & 0xff) > 5)
3277 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3279 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3281 /* Windows won't create mappings for zero length files.
3282 * Just allocate the maxsize right now.
3285 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3286 || !SetEndOfFile(env->me_fd)
3287 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3290 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3291 PAGE_READWRITE : PAGE_READONLY,
3292 sizehi, sizelo, NULL);
3295 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3296 FILE_MAP_WRITE : FILE_MAP_READ,
3297 0, 0, env->me_mapsize, meta.mm_address);
3298 rc = env->me_map ? 0 : ErrCode();
3306 if (flags & MDB_WRITEMAP) {
3308 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3311 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3313 if (env->me_map == MAP_FAILED) {
3317 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3319 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3321 #ifdef POSIX_MADV_RANDOM
3322 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3323 #endif /* POSIX_MADV_RANDOM */
3324 #endif /* MADV_RANDOM */
3328 if (flags & MDB_FIXEDMAP)
3329 meta.mm_address = env->me_map;
3330 i = mdb_env_init_meta(env, &meta);
3331 if (i != MDB_SUCCESS) {
3334 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3335 /* Can happen because the address argument to mmap() is just a
3336 * hint. mmap() can pick another, e.g. if the range is in use.
3337 * The MAP_FIXED flag would prevent that, but then mmap could
3338 * instead unmap existing pages to make room for the new map.
3340 return EBUSY; /* TODO: Make a new MDB_* error code? */
3342 env->me_psize = meta.mm_psize;
3343 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3344 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3346 env->me_maxpg = env->me_mapsize / env->me_psize;
3348 p = (MDB_page *)env->me_map;
3349 env->me_metas[0] = METADATA(p);
3350 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3354 int toggle = mdb_env_pick_meta(env);
3355 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3357 DPRINTF(("opened database version %u, pagesize %u",
3358 env->me_metas[0]->mm_version, env->me_psize));
3359 DPRINTF(("using meta page %d", toggle));
3360 DPRINTF(("depth: %u", db->md_depth));
3361 DPRINTF(("entries: %"Z"u", db->md_entries));
3362 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3363 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3364 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3365 DPRINTF(("root: %"Z"u", db->md_root));
3373 /** Release a reader thread's slot in the reader lock table.
3374 * This function is called automatically when a thread exits.
3375 * @param[in] ptr This points to the slot in the reader lock table.
3378 mdb_env_reader_dest(void *ptr)
3380 MDB_reader *reader = ptr;
3386 /** Junk for arranging thread-specific callbacks on Windows. This is
3387 * necessarily platform and compiler-specific. Windows supports up
3388 * to 1088 keys. Let's assume nobody opens more than 64 environments
3389 * in a single process, for now. They can override this if needed.
3391 #ifndef MAX_TLS_KEYS
3392 #define MAX_TLS_KEYS 64
3394 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3395 static int mdb_tls_nkeys;
3397 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3401 case DLL_PROCESS_ATTACH: break;
3402 case DLL_THREAD_ATTACH: break;
3403 case DLL_THREAD_DETACH:
3404 for (i=0; i<mdb_tls_nkeys; i++) {
3405 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3406 mdb_env_reader_dest(r);
3409 case DLL_PROCESS_DETACH: break;
3414 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3416 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3420 /* Force some symbol references.
3421 * _tls_used forces the linker to create the TLS directory if not already done
3422 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3424 #pragma comment(linker, "/INCLUDE:_tls_used")
3425 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3426 #pragma const_seg(".CRT$XLB")
3427 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3428 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3431 #pragma comment(linker, "/INCLUDE:__tls_used")
3432 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3433 #pragma data_seg(".CRT$XLB")
3434 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3436 #endif /* WIN 32/64 */
3437 #endif /* !__GNUC__ */
3440 /** Downgrade the exclusive lock on the region back to shared */
3442 mdb_env_share_locks(MDB_env *env, int *excl)
3444 int rc = 0, toggle = mdb_env_pick_meta(env);
3446 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3451 /* First acquire a shared lock. The Unlock will
3452 * then release the existing exclusive lock.
3454 memset(&ov, 0, sizeof(ov));
3455 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3458 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3464 struct flock lock_info;
3465 /* The shared lock replaces the existing lock */
3466 memset((void *)&lock_info, 0, sizeof(lock_info));
3467 lock_info.l_type = F_RDLCK;
3468 lock_info.l_whence = SEEK_SET;
3469 lock_info.l_start = 0;
3470 lock_info.l_len = 1;
3471 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3472 (rc = ErrCode()) == EINTR) ;
3473 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3480 /** Try to get exlusive lock, otherwise shared.
3481 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3484 mdb_env_excl_lock(MDB_env *env, int *excl)
3488 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3492 memset(&ov, 0, sizeof(ov));
3493 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3500 struct flock lock_info;
3501 memset((void *)&lock_info, 0, sizeof(lock_info));
3502 lock_info.l_type = F_WRLCK;
3503 lock_info.l_whence = SEEK_SET;
3504 lock_info.l_start = 0;
3505 lock_info.l_len = 1;
3506 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3507 (rc = ErrCode()) == EINTR) ;
3511 # ifdef MDB_USE_POSIX_SEM
3512 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3515 lock_info.l_type = F_RDLCK;
3516 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3517 (rc = ErrCode()) == EINTR) ;
3525 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3527 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3529 * @(#) $Revision: 5.1 $
3530 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3531 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3533 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3537 * Please do not copyright this code. This code is in the public domain.
3539 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3540 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3541 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3542 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3543 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3544 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3545 * PERFORMANCE OF THIS SOFTWARE.
3548 * chongo <Landon Curt Noll> /\oo/\
3549 * http://www.isthe.com/chongo/
3551 * Share and Enjoy! :-)
3554 typedef unsigned long long mdb_hash_t;
3555 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3557 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3558 * @param[in] val value to hash
3559 * @param[in] hval initial value for hash
3560 * @return 64 bit hash
3562 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3563 * hval arg on the first call.
3566 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3568 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3569 unsigned char *end = s + val->mv_size;
3571 * FNV-1a hash each octet of the string
3574 /* xor the bottom with the current octet */
3575 hval ^= (mdb_hash_t)*s++;
3577 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3578 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3579 (hval << 7) + (hval << 8) + (hval << 40);
3581 /* return our new hash value */
3585 /** Hash the string and output the encoded hash.
3586 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3587 * very short name limits. We don't care about the encoding being reversible,
3588 * we just want to preserve as many bits of the input as possible in a
3589 * small printable string.
3590 * @param[in] str string to hash
3591 * @param[out] encbuf an array of 11 chars to hold the hash
3593 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3596 mdb_pack85(unsigned long l, char *out)
3600 for (i=0; i<5; i++) {
3601 *out++ = mdb_a85[l % 85];
3607 mdb_hash_enc(MDB_val *val, char *encbuf)
3609 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3610 unsigned long *l = (unsigned long *)&h;
3612 mdb_pack85(l[0], encbuf);
3613 mdb_pack85(l[1], encbuf+5);
3618 /** Open and/or initialize the lock region for the environment.
3619 * @param[in] env The MDB environment.
3620 * @param[in] lpath The pathname of the file used for the lock region.
3621 * @param[in] mode The Unix permissions for the file, if we create it.
3622 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3623 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3624 * @return 0 on success, non-zero on failure.
3627 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3630 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3632 # define MDB_ERRCODE_ROFS EROFS
3633 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3634 # define MDB_CLOEXEC O_CLOEXEC
3637 # define MDB_CLOEXEC 0
3644 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3645 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3646 FILE_ATTRIBUTE_NORMAL, NULL);
3648 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3650 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3652 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3657 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3658 /* Lose record locks when exec*() */
3659 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3660 fcntl(env->me_lfd, F_SETFD, fdflags);
3663 if (!(env->me_flags & MDB_NOTLS)) {
3664 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3667 env->me_flags |= MDB_ENV_TXKEY;
3669 /* Windows TLS callbacks need help finding their TLS info. */
3670 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3674 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3678 /* Try to get exclusive lock. If we succeed, then
3679 * nobody is using the lock region and we should initialize it.
3681 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3684 size = GetFileSize(env->me_lfd, NULL);
3686 size = lseek(env->me_lfd, 0, SEEK_END);
3687 if (size == -1) goto fail_errno;
3689 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3690 if (size < rsize && *excl > 0) {
3692 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3693 || !SetEndOfFile(env->me_lfd))
3696 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3700 size = rsize - sizeof(MDB_txninfo);
3701 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3706 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3708 if (!mh) goto fail_errno;
3709 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3711 if (!env->me_txns) goto fail_errno;
3713 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3715 if (m == MAP_FAILED) goto fail_errno;
3721 BY_HANDLE_FILE_INFORMATION stbuf;
3730 if (!mdb_sec_inited) {
3731 InitializeSecurityDescriptor(&mdb_null_sd,
3732 SECURITY_DESCRIPTOR_REVISION);
3733 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3734 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3735 mdb_all_sa.bInheritHandle = FALSE;
3736 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3739 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3740 idbuf.volume = stbuf.dwVolumeSerialNumber;
3741 idbuf.nhigh = stbuf.nFileIndexHigh;
3742 idbuf.nlow = stbuf.nFileIndexLow;
3743 val.mv_data = &idbuf;
3744 val.mv_size = sizeof(idbuf);
3745 mdb_hash_enc(&val, encbuf);
3746 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3747 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3748 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3749 if (!env->me_rmutex) goto fail_errno;
3750 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3751 if (!env->me_wmutex) goto fail_errno;
3752 #elif defined(MDB_USE_POSIX_SEM)
3761 #if defined(__NetBSD__)
3762 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3764 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3765 idbuf.dev = stbuf.st_dev;
3766 idbuf.ino = stbuf.st_ino;
3767 val.mv_data = &idbuf;
3768 val.mv_size = sizeof(idbuf);
3769 mdb_hash_enc(&val, encbuf);
3770 #ifdef MDB_SHORT_SEMNAMES
3771 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3773 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3774 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3775 /* Clean up after a previous run, if needed: Try to
3776 * remove both semaphores before doing anything else.
3778 sem_unlink(env->me_txns->mti_rmname);
3779 sem_unlink(env->me_txns->mti_wmname);
3780 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3781 O_CREAT|O_EXCL, mode, 1);
3782 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3783 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3784 O_CREAT|O_EXCL, mode, 1);
3785 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3786 #else /* MDB_USE_POSIX_SEM */
3787 pthread_mutexattr_t mattr;
3789 if ((rc = pthread_mutexattr_init(&mattr))
3790 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3791 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3792 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3794 pthread_mutexattr_destroy(&mattr);
3795 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3797 env->me_txns->mti_magic = MDB_MAGIC;
3798 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3799 env->me_txns->mti_txnid = 0;
3800 env->me_txns->mti_numreaders = 0;
3803 if (env->me_txns->mti_magic != MDB_MAGIC) {
3804 DPUTS("lock region has invalid magic");
3808 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3809 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3810 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3811 rc = MDB_VERSION_MISMATCH;
3815 if (rc && rc != EACCES && rc != EAGAIN) {
3819 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3820 if (!env->me_rmutex) goto fail_errno;
3821 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3822 if (!env->me_wmutex) goto fail_errno;
3823 #elif defined(MDB_USE_POSIX_SEM)
3824 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3825 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3826 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3827 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3838 /** The name of the lock file in the DB environment */
3839 #define LOCKNAME "/lock.mdb"
3840 /** The name of the data file in the DB environment */
3841 #define DATANAME "/data.mdb"
3842 /** The suffix of the lock file when no subdir is used */
3843 #define LOCKSUFF "-lock"
3844 /** Only a subset of the @ref mdb_env flags can be changed
3845 * at runtime. Changing other flags requires closing the
3846 * environment and re-opening it with the new flags.
3848 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3849 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3852 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3854 int oflags, rc, len, excl = -1;
3855 char *lpath, *dpath;
3857 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3861 if (flags & MDB_NOSUBDIR) {
3862 rc = len + sizeof(LOCKSUFF) + len + 1;
3864 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3869 if (flags & MDB_NOSUBDIR) {
3870 dpath = lpath + len + sizeof(LOCKSUFF);
3871 sprintf(lpath, "%s" LOCKSUFF, path);
3872 strcpy(dpath, path);
3874 dpath = lpath + len + sizeof(LOCKNAME);
3875 sprintf(lpath, "%s" LOCKNAME, path);
3876 sprintf(dpath, "%s" DATANAME, path);
3880 flags |= env->me_flags;
3881 if (flags & MDB_RDONLY) {
3882 /* silently ignore WRITEMAP when we're only getting read access */
3883 flags &= ~MDB_WRITEMAP;
3885 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3886 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3889 env->me_flags = flags |= MDB_ENV_ACTIVE;
3893 env->me_path = strdup(path);
3894 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3895 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3896 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3901 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3906 if (F_ISSET(flags, MDB_RDONLY)) {
3907 oflags = GENERIC_READ;
3908 len = OPEN_EXISTING;
3910 oflags = GENERIC_READ|GENERIC_WRITE;
3913 mode = FILE_ATTRIBUTE_NORMAL;
3914 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3915 NULL, len, mode, NULL);
3917 if (F_ISSET(flags, MDB_RDONLY))
3920 oflags = O_RDWR | O_CREAT;
3922 env->me_fd = open(dpath, oflags, mode);
3924 if (env->me_fd == INVALID_HANDLE_VALUE) {
3929 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3930 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3931 env->me_mfd = env->me_fd;
3933 /* Synchronous fd for meta writes. Needed even with
3934 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3937 env->me_mfd = CreateFile(dpath, oflags,
3938 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3939 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3941 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3943 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3948 DPRINTF(("opened dbenv %p", (void *) env));
3950 rc = mdb_env_share_locks(env, &excl);
3956 mdb_env_close0(env, excl);
3962 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3964 mdb_env_close0(MDB_env *env, int excl)
3968 if (!(env->me_flags & MDB_ENV_ACTIVE))
3971 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3972 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3973 free(env->me_dbxs[i].md_name.mv_data);
3975 free(env->me_dbflags);
3978 free(env->me_dirty_list);
3979 mdb_midl_free(env->me_free_pgs);
3981 if (env->me_flags & MDB_ENV_TXKEY) {
3982 pthread_key_delete(env->me_txkey);
3984 /* Delete our key from the global list */
3985 for (i=0; i<mdb_tls_nkeys; i++)
3986 if (mdb_tls_keys[i] == env->me_txkey) {
3987 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3995 munmap(env->me_map, env->me_mapsize);
3997 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3998 (void) close(env->me_mfd);
3999 if (env->me_fd != INVALID_HANDLE_VALUE)
4000 (void) close(env->me_fd);
4002 pid_t pid = env->me_pid;
4003 /* Clearing readers is done in this function because
4004 * me_txkey with its destructor must be disabled first.
4006 for (i = env->me_numreaders; --i >= 0; )
4007 if (env->me_txns->mti_readers[i].mr_pid == pid)
4008 env->me_txns->mti_readers[i].mr_pid = 0;
4010 if (env->me_rmutex) {
4011 CloseHandle(env->me_rmutex);
4012 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4014 /* Windows automatically destroys the mutexes when
4015 * the last handle closes.
4017 #elif defined(MDB_USE_POSIX_SEM)
4018 if (env->me_rmutex != SEM_FAILED) {
4019 sem_close(env->me_rmutex);
4020 if (env->me_wmutex != SEM_FAILED)
4021 sem_close(env->me_wmutex);
4022 /* If we have the filelock: If we are the
4023 * only remaining user, clean up semaphores.
4026 mdb_env_excl_lock(env, &excl);
4028 sem_unlink(env->me_txns->mti_rmname);
4029 sem_unlink(env->me_txns->mti_wmname);
4033 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4035 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4038 /* Unlock the lockfile. Windows would have unlocked it
4039 * after closing anyway, but not necessarily at once.
4041 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4044 (void) close(env->me_lfd);
4047 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4051 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4053 MDB_txn *txn = NULL;
4059 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4063 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4066 /* Do the lock/unlock of the reader mutex before starting the
4067 * write txn. Otherwise other read txns could block writers.
4069 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4074 /* We must start the actual read txn after blocking writers */
4075 mdb_txn_reset0(txn, "reset-stage1");
4077 /* Temporarily block writers until we snapshot the meta pages */
4080 rc = mdb_txn_renew0(txn);
4082 UNLOCK_MUTEX_W(env);
4087 wsize = env->me_psize * 2;
4091 DO_WRITE(rc, fd, ptr, w2, len);
4095 } else if (len > 0) {
4101 /* Non-blocking or async handles are not supported */
4107 UNLOCK_MUTEX_W(env);
4112 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4114 if (wsize > MAX_WRITE)
4118 DO_WRITE(rc, fd, ptr, w2, len);
4122 } else if (len > 0) {
4139 mdb_env_copy(MDB_env *env, const char *path)
4143 HANDLE newfd = INVALID_HANDLE_VALUE;
4145 if (env->me_flags & MDB_NOSUBDIR) {
4146 lpath = (char *)path;
4149 len += sizeof(DATANAME);
4150 lpath = malloc(len);
4153 sprintf(lpath, "%s" DATANAME, path);
4156 /* The destination path must exist, but the destination file must not.
4157 * We don't want the OS to cache the writes, since the source data is
4158 * already in the OS cache.
4161 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4162 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4164 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4170 if (newfd == INVALID_HANDLE_VALUE) {
4175 #ifdef F_NOCACHE /* __APPLE__ */
4176 rc = fcntl(newfd, F_NOCACHE, 1);
4183 rc = mdb_env_copyfd(env, newfd);
4186 if (!(env->me_flags & MDB_NOSUBDIR))
4188 if (newfd != INVALID_HANDLE_VALUE)
4189 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4196 mdb_env_close(MDB_env *env)
4203 VGMEMP_DESTROY(env);
4204 while ((dp = env->me_dpages) != NULL) {
4205 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4206 env->me_dpages = dp->mp_next;
4210 mdb_env_close0(env, 0);
4214 /** Compare two items pointing at aligned size_t's */
4216 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4218 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4219 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4222 /** Compare two items pointing at aligned int's */
4224 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4226 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4227 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4230 /** Compare two items pointing at ints of unknown alignment.
4231 * Nodes and keys are guaranteed to be 2-byte aligned.
4234 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4236 #if BYTE_ORDER == LITTLE_ENDIAN
4237 unsigned short *u, *c;
4240 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4241 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4244 } while(!x && u > (unsigned short *)a->mv_data);
4247 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4251 /** Compare two items lexically */
4253 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4260 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4266 diff = memcmp(a->mv_data, b->mv_data, len);
4267 return diff ? diff : len_diff<0 ? -1 : len_diff;
4270 /** Compare two items in reverse byte order */
4272 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4274 const unsigned char *p1, *p2, *p1_lim;
4278 p1_lim = (const unsigned char *)a->mv_data;
4279 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4280 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4282 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4288 while (p1 > p1_lim) {
4289 diff = *--p1 - *--p2;
4293 return len_diff<0 ? -1 : len_diff;
4296 /** Search for key within a page, using binary search.
4297 * Returns the smallest entry larger or equal to the key.
4298 * If exactp is non-null, stores whether the found entry was an exact match
4299 * in *exactp (1 or 0).
4300 * Updates the cursor index with the index of the found entry.
4301 * If no entry larger or equal to the key is found, returns NULL.
4304 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4306 unsigned int i = 0, nkeys;
4309 MDB_page *mp = mc->mc_pg[mc->mc_top];
4310 MDB_node *node = NULL;
4315 nkeys = NUMKEYS(mp);
4320 COPY_PGNO(pgno, mp->mp_pgno);
4321 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4322 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4329 low = IS_LEAF(mp) ? 0 : 1;
4331 cmp = mc->mc_dbx->md_cmp;
4333 /* Branch pages have no data, so if using integer keys,
4334 * alignment is guaranteed. Use faster mdb_cmp_int.
4336 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4337 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4344 nodekey.mv_size = mc->mc_db->md_pad;
4345 node = NODEPTR(mp, 0); /* fake */
4346 while (low <= high) {
4347 i = (low + high) >> 1;
4348 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4349 rc = cmp(key, &nodekey);
4350 DPRINTF(("found leaf index %u [%s], rc = %i",
4351 i, DKEY(&nodekey), rc));
4360 while (low <= high) {
4361 i = (low + high) >> 1;
4363 node = NODEPTR(mp, i);
4364 nodekey.mv_size = NODEKSZ(node);
4365 nodekey.mv_data = NODEKEY(node);
4367 rc = cmp(key, &nodekey);
4370 DPRINTF(("found leaf index %u [%s], rc = %i",
4371 i, DKEY(&nodekey), rc));
4373 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4374 i, DKEY(&nodekey), NODEPGNO(node), rc));
4385 if (rc > 0) { /* Found entry is less than the key. */
4386 i++; /* Skip to get the smallest entry larger than key. */
4388 node = NODEPTR(mp, i);
4391 *exactp = (rc == 0);
4392 /* store the key index */
4393 mc->mc_ki[mc->mc_top] = i;
4395 /* There is no entry larger or equal to the key. */
4398 /* nodeptr is fake for LEAF2 */
4404 mdb_cursor_adjust(MDB_cursor *mc, func)
4408 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4409 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4416 /** Pop a page off the top of the cursor's stack. */
4418 mdb_cursor_pop(MDB_cursor *mc)
4422 MDB_page *top = mc->mc_pg[mc->mc_top];
4428 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4429 mc->mc_dbi, (void *) mc));
4433 /** Push a page onto the top of the cursor's stack. */
4435 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4437 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4438 mc->mc_dbi, (void *) mc));
4440 if (mc->mc_snum >= CURSOR_STACK) {
4441 assert(mc->mc_snum < CURSOR_STACK);
4442 return MDB_CURSOR_FULL;
4445 mc->mc_top = mc->mc_snum++;
4446 mc->mc_pg[mc->mc_top] = mp;
4447 mc->mc_ki[mc->mc_top] = 0;
4452 /** Find the address of the page corresponding to a given page number.
4453 * @param[in] txn the transaction for this access.
4454 * @param[in] pgno the page number for the page to retrieve.
4455 * @param[out] ret address of a pointer where the page's address will be stored.
4456 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4457 * @return 0 on success, non-zero on failure.
4460 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4462 MDB_env *env = txn->mt_env;
4466 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4470 MDB_ID2L dl = tx2->mt_u.dirty_list;
4472 /* Spilled pages were dirtied in this txn and flushed
4473 * because the dirty list got full. Bring this page
4474 * back in from the map (but don't unspill it here,
4475 * leave that unless page_touch happens again).
4477 if (tx2->mt_spill_pgs) {
4478 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4479 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4480 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4485 unsigned x = mdb_mid2l_search(dl, pgno);
4486 if (x <= dl[0].mid && dl[x].mid == pgno) {
4492 } while ((tx2 = tx2->mt_parent) != NULL);
4495 if (pgno < txn->mt_next_pgno) {
4497 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4499 DPRINTF(("page %"Z"u not found", pgno));
4501 return MDB_PAGE_NOTFOUND;
4511 /** Search for the page a given key should be in.
4512 * Pushes parent pages on the cursor stack. This function continues a
4513 * search on a cursor that has already been initialized. (Usually by
4514 * #mdb_page_search() but also by #mdb_node_move().)
4515 * @param[in,out] mc the cursor for this operation.
4516 * @param[in] key the key to search for. If NULL, search for the lowest
4517 * page. (This is used by #mdb_cursor_first().)
4518 * @param[in] modify If true, visited pages are updated with new page numbers.
4519 * @return 0 on success, non-zero on failure.
4522 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4524 MDB_page *mp = mc->mc_pg[mc->mc_top];
4529 while (IS_BRANCH(mp)) {
4533 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4534 assert(NUMKEYS(mp) > 1);
4535 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4537 if (key == NULL) /* Initialize cursor to first page. */
4539 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4540 /* cursor to last page */
4544 node = mdb_node_search(mc, key, &exact);
4546 i = NUMKEYS(mp) - 1;
4548 i = mc->mc_ki[mc->mc_top];
4557 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4558 assert(i < NUMKEYS(mp));
4559 node = NODEPTR(mp, i);
4561 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4564 mc->mc_ki[mc->mc_top] = i;
4565 if ((rc = mdb_cursor_push(mc, mp)))
4569 if ((rc = mdb_page_touch(mc)) != 0)
4571 mp = mc->mc_pg[mc->mc_top];
4576 DPRINTF(("internal error, index points to a %02X page!?",
4578 return MDB_CORRUPTED;
4581 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4582 key ? DKEY(key) : NULL));
4583 mc->mc_flags |= C_INITIALIZED;
4584 mc->mc_flags &= ~C_EOF;
4589 /** Search for the lowest key under the current branch page.
4590 * This just bypasses a NUMKEYS check in the current page
4591 * before calling mdb_page_search_root(), because the callers
4592 * are all in situations where the current page is known to
4596 mdb_page_search_lowest(MDB_cursor *mc)
4598 MDB_page *mp = mc->mc_pg[mc->mc_top];
4599 MDB_node *node = NODEPTR(mp, 0);
4602 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4605 mc->mc_ki[mc->mc_top] = 0;
4606 if ((rc = mdb_cursor_push(mc, mp)))
4608 return mdb_page_search_root(mc, NULL, 0);
4611 /** Search for the page a given key should be in.
4612 * Pushes parent pages on the cursor stack. This function just sets up
4613 * the search; it finds the root page for \b mc's database and sets this
4614 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4615 * called to complete the search.
4616 * @param[in,out] mc the cursor for this operation.
4617 * @param[in] key the key to search for. If NULL, search for the lowest
4618 * page. (This is used by #mdb_cursor_first().)
4619 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4620 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4621 * @return 0 on success, non-zero on failure.
4624 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4629 /* Make sure the txn is still viable, then find the root from
4630 * the txn's db table.
4632 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4633 DPUTS("transaction has failed, must abort");
4636 /* Make sure we're using an up-to-date root */
4637 if (mc->mc_dbi > MAIN_DBI) {
4638 if ((*mc->mc_dbflag & DB_STALE) ||
4639 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4641 unsigned char dbflag = 0;
4642 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4643 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4646 if (*mc->mc_dbflag & DB_STALE) {
4650 MDB_node *leaf = mdb_node_search(&mc2,
4651 &mc->mc_dbx->md_name, &exact);
4653 return MDB_NOTFOUND;
4654 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4657 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4659 /* The txn may not know this DBI, or another process may
4660 * have dropped and recreated the DB with other flags.
4662 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4663 return MDB_INCOMPATIBLE;
4664 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4666 if (flags & MDB_PS_MODIFY)
4668 *mc->mc_dbflag &= ~DB_STALE;
4669 *mc->mc_dbflag |= dbflag;
4672 root = mc->mc_db->md_root;
4674 if (root == P_INVALID) { /* Tree is empty. */
4675 DPUTS("tree is empty");
4676 return MDB_NOTFOUND;
4681 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4682 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4688 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4689 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4691 if (flags & MDB_PS_MODIFY) {
4692 if ((rc = mdb_page_touch(mc)))
4696 if (flags & MDB_PS_ROOTONLY)
4699 return mdb_page_search_root(mc, key, flags);
4703 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4705 MDB_txn *txn = mc->mc_txn;
4706 pgno_t pg = mp->mp_pgno;
4707 unsigned x = 0, ovpages = mp->mp_pages;
4708 MDB_env *env = txn->mt_env;
4709 MDB_IDL sl = txn->mt_spill_pgs;
4712 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4713 /* If the page is dirty or on the spill list we just acquired it,
4714 * so we should give it back to our current free list, if any.
4715 * Otherwise put it onto the list of pages we freed in this txn.
4717 * Won't create me_pghead: me_pglast must be inited along with it.
4718 * Unsupported in nested txns: They would need to hide the page
4719 * range in ancestor txns' dirty and spilled lists.
4721 if (env->me_pghead &&
4723 ((mp->mp_flags & P_DIRTY) ||
4724 (sl && (x = mdb_midl_search(sl, pg)) <= sl[0] && sl[x] == pg)))
4728 MDB_ID2 *dl, ix, iy;
4729 rc = mdb_midl_need(&env->me_pghead, ovpages);
4732 if (!(mp->mp_flags & P_DIRTY)) {
4733 /* This page is no longer spilled */
4734 for (; x < sl[0]; x++)
4739 /* Remove from dirty list */
4740 dl = txn->mt_u.dirty_list;
4742 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4750 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4751 txn->mt_flags |= MDB_TXN_ERROR;
4752 return MDB_CORRUPTED;
4755 if (!(env->me_flags & MDB_WRITEMAP))
4756 mdb_dpage_free(env, mp);
4758 /* Insert in me_pghead */
4759 mop = env->me_pghead;
4760 j = mop[0] + ovpages;
4761 for (i = mop[0]; i && mop[i] < pg; i--)
4767 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4771 mc->mc_db->md_overflow_pages -= ovpages;
4775 /** Return the data associated with a given node.
4776 * @param[in] txn The transaction for this operation.
4777 * @param[in] leaf The node being read.
4778 * @param[out] data Updated to point to the node's data.
4779 * @return 0 on success, non-zero on failure.
4782 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4784 MDB_page *omp; /* overflow page */
4788 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4789 data->mv_size = NODEDSZ(leaf);
4790 data->mv_data = NODEDATA(leaf);
4794 /* Read overflow data.
4796 data->mv_size = NODEDSZ(leaf);
4797 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4798 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4799 DPRINTF(("read overflow page %"Z"u failed", pgno));
4802 data->mv_data = METADATA(omp);
4808 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4809 MDB_val *key, MDB_val *data)
4818 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4820 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4823 if (txn->mt_flags & MDB_TXN_ERROR)
4826 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4827 return MDB_BAD_VALSIZE;
4830 mdb_cursor_init(&mc, txn, dbi, &mx);
4831 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4834 /** Find a sibling for a page.
4835 * Replaces the page at the top of the cursor's stack with the
4836 * specified sibling, if one exists.
4837 * @param[in] mc The cursor for this operation.
4838 * @param[in] move_right Non-zero if the right sibling is requested,
4839 * otherwise the left sibling.
4840 * @return 0 on success, non-zero on failure.
4843 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4849 if (mc->mc_snum < 2) {
4850 return MDB_NOTFOUND; /* root has no siblings */
4854 DPRINTF(("parent page is page %"Z"u, index %u",
4855 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4857 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4858 : (mc->mc_ki[mc->mc_top] == 0)) {
4859 DPRINTF(("no more keys left, moving to %s sibling",
4860 move_right ? "right" : "left"));
4861 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4862 /* undo cursor_pop before returning */
4869 mc->mc_ki[mc->mc_top]++;
4871 mc->mc_ki[mc->mc_top]--;
4872 DPRINTF(("just moving to %s index key %u",
4873 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4875 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4877 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4878 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4881 mdb_cursor_push(mc, mp);
4883 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4888 /** Move the cursor to the next data item. */
4890 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4896 if (mc->mc_flags & C_EOF) {
4897 return MDB_NOTFOUND;
4900 assert(mc->mc_flags & C_INITIALIZED);
4902 mp = mc->mc_pg[mc->mc_top];
4904 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4905 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4906 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4907 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4908 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4909 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4913 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4914 if (op == MDB_NEXT_DUP)
4915 return MDB_NOTFOUND;
4919 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4921 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4922 DPUTS("=====> move to next sibling page");
4923 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4924 mc->mc_flags |= C_EOF;
4927 mp = mc->mc_pg[mc->mc_top];
4928 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
4930 mc->mc_ki[mc->mc_top]++;
4932 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
4933 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
4936 key->mv_size = mc->mc_db->md_pad;
4937 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4941 assert(IS_LEAF(mp));
4942 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4944 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4945 mdb_xcursor_init1(mc, leaf);
4948 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4951 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4952 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4953 if (rc != MDB_SUCCESS)
4958 MDB_GET_KEY(leaf, key);
4962 /** Move the cursor to the previous data item. */
4964 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4970 assert(mc->mc_flags & C_INITIALIZED);
4972 mp = mc->mc_pg[mc->mc_top];
4974 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4975 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4976 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4977 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4978 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4979 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4982 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4983 if (op == MDB_PREV_DUP)
4984 return MDB_NOTFOUND;
4989 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4991 if (mc->mc_ki[mc->mc_top] == 0) {
4992 DPUTS("=====> move to prev sibling page");
4993 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4996 mp = mc->mc_pg[mc->mc_top];
4997 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4998 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5000 mc->mc_ki[mc->mc_top]--;
5002 mc->mc_flags &= ~C_EOF;
5004 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5005 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5008 key->mv_size = mc->mc_db->md_pad;
5009 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5013 assert(IS_LEAF(mp));
5014 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5016 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5017 mdb_xcursor_init1(mc, leaf);
5020 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5023 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5024 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5025 if (rc != MDB_SUCCESS)
5030 MDB_GET_KEY(leaf, key);
5034 /** Set the cursor on a specific data item. */
5036 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5037 MDB_cursor_op op, int *exactp)
5041 MDB_node *leaf = NULL;
5046 assert(key->mv_size > 0);
5049 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5051 /* See if we're already on the right page */
5052 if (mc->mc_flags & C_INITIALIZED) {
5055 mp = mc->mc_pg[mc->mc_top];
5057 mc->mc_ki[mc->mc_top] = 0;
5058 return MDB_NOTFOUND;
5060 if (mp->mp_flags & P_LEAF2) {
5061 nodekey.mv_size = mc->mc_db->md_pad;
5062 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5064 leaf = NODEPTR(mp, 0);
5065 MDB_GET_KEY(leaf, &nodekey);
5067 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5069 /* Probably happens rarely, but first node on the page
5070 * was the one we wanted.
5072 mc->mc_ki[mc->mc_top] = 0;
5079 unsigned int nkeys = NUMKEYS(mp);
5081 if (mp->mp_flags & P_LEAF2) {
5082 nodekey.mv_data = LEAF2KEY(mp,
5083 nkeys-1, nodekey.mv_size);
5085 leaf = NODEPTR(mp, nkeys-1);
5086 MDB_GET_KEY(leaf, &nodekey);
5088 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5090 /* last node was the one we wanted */
5091 mc->mc_ki[mc->mc_top] = nkeys-1;
5097 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5098 /* This is definitely the right page, skip search_page */
5099 if (mp->mp_flags & P_LEAF2) {
5100 nodekey.mv_data = LEAF2KEY(mp,
5101 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5103 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5104 MDB_GET_KEY(leaf, &nodekey);
5106 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5108 /* current node was the one we wanted */
5118 /* If any parents have right-sibs, search.
5119 * Otherwise, there's nothing further.
5121 for (i=0; i<mc->mc_top; i++)
5123 NUMKEYS(mc->mc_pg[i])-1)
5125 if (i == mc->mc_top) {
5126 /* There are no other pages */
5127 mc->mc_ki[mc->mc_top] = nkeys;
5128 return MDB_NOTFOUND;
5132 /* There are no other pages */
5133 mc->mc_ki[mc->mc_top] = 0;
5134 return MDB_NOTFOUND;
5138 rc = mdb_page_search(mc, key, 0);
5139 if (rc != MDB_SUCCESS)
5142 mp = mc->mc_pg[mc->mc_top];
5143 assert(IS_LEAF(mp));
5146 leaf = mdb_node_search(mc, key, exactp);
5147 if (exactp != NULL && !*exactp) {
5148 /* MDB_SET specified and not an exact match. */
5149 return MDB_NOTFOUND;
5153 DPUTS("===> inexact leaf not found, goto sibling");
5154 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5155 return rc; /* no entries matched */
5156 mp = mc->mc_pg[mc->mc_top];
5157 assert(IS_LEAF(mp));
5158 leaf = NODEPTR(mp, 0);
5162 mc->mc_flags |= C_INITIALIZED;
5163 mc->mc_flags &= ~C_EOF;
5166 key->mv_size = mc->mc_db->md_pad;
5167 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5171 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5172 mdb_xcursor_init1(mc, leaf);
5175 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5176 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5177 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5180 if (op == MDB_GET_BOTH) {
5186 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5187 if (rc != MDB_SUCCESS)
5190 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5192 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5194 rc = mc->mc_dbx->md_dcmp(data, &d2);
5196 if (op == MDB_GET_BOTH || rc > 0)
5197 return MDB_NOTFOUND;
5202 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5203 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5208 /* The key already matches in all other cases */
5209 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5210 MDB_GET_KEY(leaf, key);
5211 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5216 /** Move the cursor to the first item in the database. */
5218 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5224 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5226 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5227 rc = mdb_page_search(mc, NULL, 0);
5228 if (rc != MDB_SUCCESS)
5231 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5233 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5234 mc->mc_flags |= C_INITIALIZED;
5235 mc->mc_flags &= ~C_EOF;
5237 mc->mc_ki[mc->mc_top] = 0;
5239 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5240 key->mv_size = mc->mc_db->md_pad;
5241 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5246 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5247 mdb_xcursor_init1(mc, leaf);
5248 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5252 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5256 MDB_GET_KEY(leaf, key);
5260 /** Move the cursor to the last item in the database. */
5262 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5268 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5270 if (!(mc->mc_flags & C_EOF)) {
5272 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5275 lkey.mv_size = MDB_MAXKEYSIZE+1;
5276 lkey.mv_data = NULL;
5277 rc = mdb_page_search(mc, &lkey, 0);
5278 if (rc != MDB_SUCCESS)
5281 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5284 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5285 mc->mc_flags |= C_INITIALIZED|C_EOF;
5286 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5288 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5289 key->mv_size = mc->mc_db->md_pad;
5290 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5295 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5296 mdb_xcursor_init1(mc, leaf);
5297 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5301 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5306 MDB_GET_KEY(leaf, key);
5311 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5316 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5320 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5324 case MDB_GET_CURRENT:
5325 if (!(mc->mc_flags & C_INITIALIZED)) {
5328 MDB_page *mp = mc->mc_pg[mc->mc_top];
5330 mc->mc_ki[mc->mc_top] = 0;
5336 key->mv_size = mc->mc_db->md_pad;
5337 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5339 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5340 MDB_GET_KEY(leaf, key);
5342 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5343 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5345 rc = mdb_node_read(mc->mc_txn, leaf, data);
5352 case MDB_GET_BOTH_RANGE:
5357 if (mc->mc_xcursor == NULL) {
5358 rc = MDB_INCOMPATIBLE;
5367 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5368 rc = MDB_BAD_VALSIZE;
5369 } else if (op == MDB_SET_RANGE)
5370 rc = mdb_cursor_set(mc, key, data, op, NULL);
5372 rc = mdb_cursor_set(mc, key, data, op, &exact);
5374 case MDB_GET_MULTIPLE:
5375 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5379 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5380 rc = MDB_INCOMPATIBLE;
5384 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5385 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5388 case MDB_NEXT_MULTIPLE:
5393 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5394 rc = MDB_INCOMPATIBLE;
5397 if (!(mc->mc_flags & C_INITIALIZED))
5398 rc = mdb_cursor_first(mc, key, data);
5400 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5401 if (rc == MDB_SUCCESS) {
5402 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5405 mx = &mc->mc_xcursor->mx_cursor;
5406 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5408 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5409 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5417 case MDB_NEXT_NODUP:
5418 if (!(mc->mc_flags & C_INITIALIZED))
5419 rc = mdb_cursor_first(mc, key, data);
5421 rc = mdb_cursor_next(mc, key, data, op);
5425 case MDB_PREV_NODUP:
5426 if (!(mc->mc_flags & C_INITIALIZED)) {
5427 rc = mdb_cursor_last(mc, key, data);
5430 mc->mc_flags |= C_INITIALIZED;
5431 mc->mc_ki[mc->mc_top]++;
5433 rc = mdb_cursor_prev(mc, key, data, op);
5436 rc = mdb_cursor_first(mc, key, data);
5439 mfunc = mdb_cursor_first;
5441 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5445 if (mc->mc_xcursor == NULL) {
5446 rc = MDB_INCOMPATIBLE;
5449 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5453 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5456 rc = mdb_cursor_last(mc, key, data);
5459 mfunc = mdb_cursor_last;
5462 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5470 /** Touch all the pages in the cursor stack.
5471 * Makes sure all the pages are writable, before attempting a write operation.
5472 * @param[in] mc The cursor to operate on.
5475 mdb_cursor_touch(MDB_cursor *mc)
5479 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5482 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5483 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5486 *mc->mc_dbflag |= DB_DIRTY;
5488 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5489 rc = mdb_page_touch(mc);
5493 mc->mc_top = mc->mc_snum-1;
5497 /** Do not spill pages to disk if txn is getting full, may fail instead */
5498 #define MDB_NOSPILL 0x8000
5501 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5504 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5505 MDB_node *leaf = NULL;
5506 MDB_val xdata, *rdata, dkey;
5509 int do_sub = 0, insert = 0;
5510 unsigned int mcount = 0, dcount = 0, nospill;
5514 char dbuf[MDB_MAXKEYSIZE+1];
5515 unsigned int nflags;
5518 /* Check this first so counter will always be zero on any
5521 if (flags & MDB_MULTIPLE) {
5522 dcount = data[1].mv_size;
5523 data[1].mv_size = 0;
5524 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5525 return MDB_INCOMPATIBLE;
5528 nospill = flags & MDB_NOSPILL;
5529 flags &= ~MDB_NOSPILL;
5531 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5532 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5534 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5535 return MDB_BAD_VALSIZE;
5537 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5538 return MDB_BAD_VALSIZE;
5540 #if SIZE_MAX > MAXDATASIZE
5541 if (data->mv_size > MAXDATASIZE)
5542 return MDB_BAD_VALSIZE;
5545 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5546 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5550 if (flags == MDB_CURRENT) {
5551 if (!(mc->mc_flags & C_INITIALIZED))
5554 } else if (mc->mc_db->md_root == P_INVALID) {
5555 /* new database, cursor has nothing to point to */
5557 mc->mc_flags &= ~C_INITIALIZED;
5562 if (flags & MDB_APPEND) {
5564 rc = mdb_cursor_last(mc, &k2, &d2);
5566 rc = mc->mc_dbx->md_cmp(key, &k2);
5569 mc->mc_ki[mc->mc_top]++;
5571 /* new key is <= last key */
5576 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5578 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5579 DPRINTF(("duplicate key [%s]", DKEY(key)));
5581 return MDB_KEYEXIST;
5583 if (rc && rc != MDB_NOTFOUND)
5587 /* Cursor is positioned, check for room in the dirty list */
5589 if (flags & MDB_MULTIPLE) {
5591 xdata.mv_size = data->mv_size * dcount;
5595 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5599 if (rc == MDB_NO_ROOT) {
5601 /* new database, write a root leaf page */
5602 DPUTS("allocating new root leaf page");
5603 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5606 mdb_cursor_push(mc, np);
5607 mc->mc_db->md_root = np->mp_pgno;
5608 mc->mc_db->md_depth++;
5609 *mc->mc_dbflag |= DB_DIRTY;
5610 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5612 np->mp_flags |= P_LEAF2;
5613 mc->mc_flags |= C_INITIALIZED;
5615 /* make sure all cursor pages are writable */
5616 rc2 = mdb_cursor_touch(mc);
5621 /* The key already exists */
5622 if (rc == MDB_SUCCESS) {
5623 /* there's only a key anyway, so this is a no-op */
5624 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5625 unsigned int ksize = mc->mc_db->md_pad;
5626 if (key->mv_size != ksize)
5627 return MDB_BAD_VALSIZE;
5628 if (flags == MDB_CURRENT) {
5629 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5630 memcpy(ptr, key->mv_data, ksize);
5635 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5638 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5639 /* Was a single item before, must convert now */
5641 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5642 /* Just overwrite the current item */
5643 if (flags == MDB_CURRENT)
5646 dkey.mv_size = NODEDSZ(leaf);
5647 dkey.mv_data = NODEDATA(leaf);
5648 #if UINT_MAX < SIZE_MAX
5649 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5650 #ifdef MISALIGNED_OK
5651 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5653 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5656 /* if data matches, skip it */
5657 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5658 if (flags == MDB_NODUPDATA)
5660 else if (flags & MDB_MULTIPLE)
5667 /* create a fake page for the dup items */
5668 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5669 dkey.mv_data = dbuf;
5670 fp = (MDB_page *)&pbuf;
5671 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5672 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5673 fp->mp_lower = PAGEHDRSZ;
5674 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5675 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5676 fp->mp_flags |= P_LEAF2;
5677 fp->mp_pad = data->mv_size;
5678 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5680 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5681 (dkey.mv_size & 1) + (data->mv_size & 1);
5683 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5686 xdata.mv_size = fp->mp_upper;
5691 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5692 /* See if we need to convert from fake page to subDB */
5694 unsigned int offset;
5698 fp = NODEDATA(leaf);
5699 if (flags == MDB_CURRENT) {
5701 fp->mp_flags |= P_DIRTY;
5702 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5703 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5707 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5708 offset = fp->mp_pad;
5709 if (SIZELEFT(fp) >= offset)
5711 offset *= 4; /* space for 4 more */
5713 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5715 offset += offset & 1;
5716 fp_flags = fp->mp_flags;
5717 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5718 offset >= mc->mc_txn->mt_env->me_nodemax) {
5719 /* yes, convert it */
5721 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5722 dummy.md_pad = fp->mp_pad;
5723 dummy.md_flags = MDB_DUPFIXED;
5724 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5725 dummy.md_flags |= MDB_INTEGERKEY;
5728 dummy.md_branch_pages = 0;
5729 dummy.md_leaf_pages = 1;
5730 dummy.md_overflow_pages = 0;
5731 dummy.md_entries = NUMKEYS(fp);
5733 xdata.mv_size = sizeof(MDB_db);
5734 xdata.mv_data = &dummy;
5735 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5737 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5738 flags |= F_DUPDATA|F_SUBDATA;
5739 dummy.md_root = mp->mp_pgno;
5740 fp_flags &= ~P_SUBP;
5742 /* no, just grow it */
5744 xdata.mv_size = NODEDSZ(leaf) + offset;
5745 xdata.mv_data = &pbuf;
5746 mp = (MDB_page *)&pbuf;
5747 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5750 mp->mp_flags = fp_flags | P_DIRTY;
5751 mp->mp_pad = fp->mp_pad;
5752 mp->mp_lower = fp->mp_lower;
5753 mp->mp_upper = fp->mp_upper + offset;
5755 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5757 nsize = NODEDSZ(leaf) - fp->mp_upper;
5758 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5759 for (i=0; i<NUMKEYS(fp); i++)
5760 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5762 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5766 /* data is on sub-DB, just store it */
5767 flags |= F_DUPDATA|F_SUBDATA;
5771 /* overflow page overwrites need special handling */
5772 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5775 unsigned psize = mc->mc_txn->mt_env->me_psize;
5776 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5778 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5779 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5781 ovpages = omp->mp_pages;
5783 /* Is the ov page large enough? */
5784 if (ovpages >= dpages) {
5785 if (!(omp->mp_flags & P_DIRTY) &&
5786 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5788 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5791 level = 0; /* dirty in this txn or clean */
5794 if (omp->mp_flags & P_DIRTY) {
5795 /* yes, overwrite it. Note in this case we don't
5796 * bother to try shrinking the page if the new data
5797 * is smaller than the overflow threshold.
5800 /* It is writable only in a parent txn */
5801 size_t sz = (size_t) psize * ovpages, off;
5802 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5808 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5809 if (!(flags & MDB_RESERVE)) {
5810 /* Copy end of page, adjusting alignment so
5811 * compiler may copy words instead of bytes.
5813 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5814 memcpy((size_t *)((char *)np + off),
5815 (size_t *)((char *)omp + off), sz - off);
5818 memcpy(np, omp, sz); /* Copy beginning of page */
5821 SETDSZ(leaf, data->mv_size);
5822 if (F_ISSET(flags, MDB_RESERVE))
5823 data->mv_data = METADATA(omp);
5825 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5829 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5831 } else if (NODEDSZ(leaf) == data->mv_size) {
5832 /* same size, just replace it. Note that we could
5833 * also reuse this node if the new data is smaller,
5834 * but instead we opt to shrink the node in that case.
5836 if (F_ISSET(flags, MDB_RESERVE))
5837 data->mv_data = NODEDATA(leaf);
5838 else if (data->mv_size)
5839 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5841 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5844 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5845 mc->mc_db->md_entries--;
5847 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5854 nflags = flags & NODE_ADD_FLAGS;
5855 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5856 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5857 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5858 nflags &= ~MDB_APPEND;
5860 nflags |= MDB_SPLIT_REPLACE;
5861 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5863 /* There is room already in this leaf page. */
5864 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5865 if (rc == 0 && !do_sub && insert) {
5866 /* Adjust other cursors pointing to mp */
5867 MDB_cursor *m2, *m3;
5868 MDB_dbi dbi = mc->mc_dbi;
5869 unsigned i = mc->mc_top;
5870 MDB_page *mp = mc->mc_pg[i];
5872 if (mc->mc_flags & C_SUB)
5875 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5876 if (mc->mc_flags & C_SUB)
5877 m3 = &m2->mc_xcursor->mx_cursor;
5880 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5881 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5888 if (rc != MDB_SUCCESS)
5889 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5891 /* Now store the actual data in the child DB. Note that we're
5892 * storing the user data in the keys field, so there are strict
5893 * size limits on dupdata. The actual data fields of the child
5894 * DB are all zero size.
5901 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5902 if (flags & MDB_CURRENT) {
5903 xflags = MDB_CURRENT|MDB_NOSPILL;
5905 mdb_xcursor_init1(mc, leaf);
5906 xflags = (flags & MDB_NODUPDATA) ?
5907 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5909 /* converted, write the original data first */
5911 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5915 /* Adjust other cursors pointing to mp */
5917 unsigned i = mc->mc_top;
5918 MDB_page *mp = mc->mc_pg[i];
5920 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5921 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5922 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5923 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5924 mdb_xcursor_init1(m2, leaf);
5928 /* we've done our job */
5931 if (flags & MDB_APPENDDUP)
5932 xflags |= MDB_APPEND;
5933 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5934 if (flags & F_SUBDATA) {
5935 void *db = NODEDATA(leaf);
5936 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5939 /* sub-writes might have failed so check rc again.
5940 * Don't increment count if we just replaced an existing item.
5942 if (!rc && !(flags & MDB_CURRENT))
5943 mc->mc_db->md_entries++;
5944 if (flags & MDB_MULTIPLE) {
5948 /* let caller know how many succeeded, if any */
5949 data[1].mv_size = mcount;
5950 if (mcount < dcount) {
5951 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5952 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5959 /* If we succeeded and the key didn't exist before, make sure
5960 * the cursor is marked valid.
5963 mc->mc_flags |= C_INITIALIZED;
5968 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5973 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5974 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5976 if (!(mc->mc_flags & C_INITIALIZED))
5979 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
5981 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
5983 rc = mdb_cursor_touch(mc);
5987 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5989 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5990 if (flags != MDB_NODUPDATA) {
5991 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5992 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5994 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
5995 /* If sub-DB still has entries, we're done */
5996 if (mc->mc_xcursor->mx_db.md_entries) {
5997 if (leaf->mn_flags & F_SUBDATA) {
5998 /* update subDB info */
5999 void *db = NODEDATA(leaf);
6000 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6003 /* shrink fake page */
6004 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6005 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6006 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6007 /* fix other sub-DB cursors pointed at this fake page */
6008 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6009 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6010 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6011 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6012 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6015 mc->mc_db->md_entries--;
6018 /* otherwise fall thru and delete the sub-DB */
6021 if (leaf->mn_flags & F_SUBDATA) {
6022 /* add all the child DB's pages to the free list */
6023 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6024 if (rc == MDB_SUCCESS) {
6025 mc->mc_db->md_entries -=
6026 mc->mc_xcursor->mx_db.md_entries;
6031 return mdb_cursor_del0(mc, leaf);
6034 /** Allocate and initialize new pages for a database.
6035 * @param[in] mc a cursor on the database being added to.
6036 * @param[in] flags flags defining what type of page is being allocated.
6037 * @param[in] num the number of pages to allocate. This is usually 1,
6038 * unless allocating overflow pages for a large record.
6039 * @param[out] mp Address of a page, or NULL on failure.
6040 * @return 0 on success, non-zero on failure.
6043 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6048 if ((rc = mdb_page_alloc(mc, num, &np)))
6050 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6051 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6052 np->mp_flags = flags | P_DIRTY;
6053 np->mp_lower = PAGEHDRSZ;
6054 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6057 mc->mc_db->md_branch_pages++;
6058 else if (IS_LEAF(np))
6059 mc->mc_db->md_leaf_pages++;
6060 else if (IS_OVERFLOW(np)) {
6061 mc->mc_db->md_overflow_pages += num;
6069 /** Calculate the size of a leaf node.
6070 * The size depends on the environment's page size; if a data item
6071 * is too large it will be put onto an overflow page and the node
6072 * size will only include the key and not the data. Sizes are always
6073 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6074 * of the #MDB_node headers.
6075 * @param[in] env The environment handle.
6076 * @param[in] key The key for the node.
6077 * @param[in] data The data for the node.
6078 * @return The number of bytes needed to store the node.
6081 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6085 sz = LEAFSIZE(key, data);
6086 if (sz >= env->me_nodemax) {
6087 /* put on overflow page */
6088 sz -= data->mv_size - sizeof(pgno_t);
6092 return sz + sizeof(indx_t);
6095 /** Calculate the size of a branch node.
6096 * The size should depend on the environment's page size but since
6097 * we currently don't support spilling large keys onto overflow
6098 * pages, it's simply the size of the #MDB_node header plus the
6099 * size of the key. Sizes are always rounded up to an even number
6100 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6101 * @param[in] env The environment handle.
6102 * @param[in] key The key for the node.
6103 * @return The number of bytes needed to store the node.
6106 mdb_branch_size(MDB_env *env, MDB_val *key)
6111 if (sz >= env->me_nodemax) {
6112 /* put on overflow page */
6113 /* not implemented */
6114 /* sz -= key->size - sizeof(pgno_t); */
6117 return sz + sizeof(indx_t);
6120 /** Add a node to the page pointed to by the cursor.
6121 * @param[in] mc The cursor for this operation.
6122 * @param[in] indx The index on the page where the new node should be added.
6123 * @param[in] key The key for the new node.
6124 * @param[in] data The data for the new node, if any.
6125 * @param[in] pgno The page number, if adding a branch node.
6126 * @param[in] flags Flags for the node.
6127 * @return 0 on success, non-zero on failure. Possible errors are:
6129 * <li>ENOMEM - failed to allocate overflow pages for the node.
6130 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6131 * should never happen since all callers already calculate the
6132 * page's free space before calling this function.
6136 mdb_node_add(MDB_cursor *mc, indx_t indx,
6137 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6140 size_t node_size = NODESIZE;
6143 MDB_page *mp = mc->mc_pg[mc->mc_top];
6144 MDB_page *ofp = NULL; /* overflow page */
6147 assert(mp->mp_upper >= mp->mp_lower);
6149 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6150 IS_LEAF(mp) ? "leaf" : "branch",
6151 IS_SUBP(mp) ? "sub-" : "",
6152 mp->mp_pgno, indx, data ? data->mv_size : 0,
6153 key ? key->mv_size : 0, key ? DKEY(key) : NULL));
6156 /* Move higher keys up one slot. */
6157 int ksize = mc->mc_db->md_pad, dif;
6158 char *ptr = LEAF2KEY(mp, indx, ksize);
6159 dif = NUMKEYS(mp) - indx;
6161 memmove(ptr+ksize, ptr, dif*ksize);
6162 /* insert new key */
6163 memcpy(ptr, key->mv_data, ksize);
6165 /* Just using these for counting */
6166 mp->mp_lower += sizeof(indx_t);
6167 mp->mp_upper -= ksize - sizeof(indx_t);
6172 node_size += key->mv_size;
6176 if (F_ISSET(flags, F_BIGDATA)) {
6177 /* Data already on overflow page. */
6178 node_size += sizeof(pgno_t);
6179 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6180 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6182 /* Put data on overflow page. */
6183 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6184 data->mv_size, node_size+data->mv_size));
6185 node_size += sizeof(pgno_t);
6186 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6188 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6191 node_size += data->mv_size;
6194 node_size += node_size & 1;
6196 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6197 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6198 mp->mp_pgno, NUMKEYS(mp)));
6199 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6200 mp->mp_upper - mp->mp_lower));
6201 DPRINTF(("node size = %"Z"u", node_size));
6202 return MDB_PAGE_FULL;
6205 /* Move higher pointers up one slot. */
6206 for (i = NUMKEYS(mp); i > indx; i--)
6207 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6209 /* Adjust free space offsets. */
6210 ofs = mp->mp_upper - node_size;
6211 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6212 mp->mp_ptrs[indx] = ofs;
6214 mp->mp_lower += sizeof(indx_t);
6216 /* Write the node data. */
6217 node = NODEPTR(mp, indx);
6218 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6219 node->mn_flags = flags;
6221 SETDSZ(node,data->mv_size);
6226 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6231 if (F_ISSET(flags, F_BIGDATA))
6232 memcpy(node->mn_data + key->mv_size, data->mv_data,
6234 else if (F_ISSET(flags, MDB_RESERVE))
6235 data->mv_data = node->mn_data + key->mv_size;
6237 memcpy(node->mn_data + key->mv_size, data->mv_data,
6240 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6242 if (F_ISSET(flags, MDB_RESERVE))
6243 data->mv_data = METADATA(ofp);
6245 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6252 /** Delete the specified node from a page.
6253 * @param[in] mp The page to operate on.
6254 * @param[in] indx The index of the node to delete.
6255 * @param[in] ksize The size of a node. Only used if the page is
6256 * part of a #MDB_DUPFIXED database.
6259 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6262 indx_t i, j, numkeys, ptr;
6269 COPY_PGNO(pgno, mp->mp_pgno);
6270 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6271 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6274 assert(indx < NUMKEYS(mp));
6277 int x = NUMKEYS(mp) - 1 - indx;
6278 base = LEAF2KEY(mp, indx, ksize);
6280 memmove(base, base + ksize, x * ksize);
6281 mp->mp_lower -= sizeof(indx_t);
6282 mp->mp_upper += ksize - sizeof(indx_t);
6286 node = NODEPTR(mp, indx);
6287 sz = NODESIZE + node->mn_ksize;
6289 if (F_ISSET(node->mn_flags, F_BIGDATA))
6290 sz += sizeof(pgno_t);
6292 sz += NODEDSZ(node);
6296 ptr = mp->mp_ptrs[indx];
6297 numkeys = NUMKEYS(mp);
6298 for (i = j = 0; i < numkeys; i++) {
6300 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6301 if (mp->mp_ptrs[i] < ptr)
6302 mp->mp_ptrs[j] += sz;
6307 base = (char *)mp + mp->mp_upper;
6308 memmove(base + sz, base, ptr - mp->mp_upper);
6310 mp->mp_lower -= sizeof(indx_t);
6314 /** Compact the main page after deleting a node on a subpage.
6315 * @param[in] mp The main page to operate on.
6316 * @param[in] indx The index of the subpage on the main page.
6319 mdb_node_shrink(MDB_page *mp, indx_t indx)
6326 indx_t i, numkeys, ptr;
6328 node = NODEPTR(mp, indx);
6329 sp = (MDB_page *)NODEDATA(node);
6330 osize = NODEDSZ(node);
6332 delta = sp->mp_upper - sp->mp_lower;
6333 SETDSZ(node, osize - delta);
6334 xp = (MDB_page *)((char *)sp + delta);
6336 /* shift subpage upward */
6338 nsize = NUMKEYS(sp) * sp->mp_pad;
6339 memmove(METADATA(xp), METADATA(sp), nsize);
6342 nsize = osize - sp->mp_upper;
6343 numkeys = NUMKEYS(sp);
6344 for (i=numkeys-1; i>=0; i--)
6345 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6347 xp->mp_upper = sp->mp_lower;
6348 xp->mp_lower = sp->mp_lower;
6349 xp->mp_flags = sp->mp_flags;
6350 xp->mp_pad = sp->mp_pad;
6351 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6353 /* shift lower nodes upward */
6354 ptr = mp->mp_ptrs[indx];
6355 numkeys = NUMKEYS(mp);
6356 for (i = 0; i < numkeys; i++) {
6357 if (mp->mp_ptrs[i] <= ptr)
6358 mp->mp_ptrs[i] += delta;
6361 base = (char *)mp + mp->mp_upper;
6362 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6363 mp->mp_upper += delta;
6366 /** Initial setup of a sorted-dups cursor.
6367 * Sorted duplicates are implemented as a sub-database for the given key.
6368 * The duplicate data items are actually keys of the sub-database.
6369 * Operations on the duplicate data items are performed using a sub-cursor
6370 * initialized when the sub-database is first accessed. This function does
6371 * the preliminary setup of the sub-cursor, filling in the fields that
6372 * depend only on the parent DB.
6373 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6376 mdb_xcursor_init0(MDB_cursor *mc)
6378 MDB_xcursor *mx = mc->mc_xcursor;
6380 mx->mx_cursor.mc_xcursor = NULL;
6381 mx->mx_cursor.mc_txn = mc->mc_txn;
6382 mx->mx_cursor.mc_db = &mx->mx_db;
6383 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6384 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6385 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6386 mx->mx_cursor.mc_snum = 0;
6387 mx->mx_cursor.mc_top = 0;
6388 mx->mx_cursor.mc_flags = C_SUB;
6389 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6390 mx->mx_dbx.md_dcmp = NULL;
6391 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6394 /** Final setup of a sorted-dups cursor.
6395 * Sets up the fields that depend on the data from the main cursor.
6396 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6397 * @param[in] node The data containing the #MDB_db record for the
6398 * sorted-dup database.
6401 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6403 MDB_xcursor *mx = mc->mc_xcursor;
6405 if (node->mn_flags & F_SUBDATA) {
6406 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6407 mx->mx_cursor.mc_pg[0] = 0;
6408 mx->mx_cursor.mc_snum = 0;
6409 mx->mx_cursor.mc_flags = C_SUB;
6411 MDB_page *fp = NODEDATA(node);
6412 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6413 mx->mx_db.md_flags = 0;
6414 mx->mx_db.md_depth = 1;
6415 mx->mx_db.md_branch_pages = 0;
6416 mx->mx_db.md_leaf_pages = 1;
6417 mx->mx_db.md_overflow_pages = 0;
6418 mx->mx_db.md_entries = NUMKEYS(fp);
6419 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6420 mx->mx_cursor.mc_snum = 1;
6421 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6422 mx->mx_cursor.mc_top = 0;
6423 mx->mx_cursor.mc_pg[0] = fp;
6424 mx->mx_cursor.mc_ki[0] = 0;
6425 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6426 mx->mx_db.md_flags = MDB_DUPFIXED;
6427 mx->mx_db.md_pad = fp->mp_pad;
6428 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6429 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6432 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6433 mx->mx_db.md_root));
6434 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6436 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6437 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6438 #if UINT_MAX < SIZE_MAX
6439 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6440 #ifdef MISALIGNED_OK
6441 mx->mx_dbx.md_cmp = mdb_cmp_long;
6443 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6448 /** Initialize a cursor for a given transaction and database. */
6450 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6453 mc->mc_backup = NULL;
6456 mc->mc_db = &txn->mt_dbs[dbi];
6457 mc->mc_dbx = &txn->mt_dbxs[dbi];
6458 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6463 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6465 mc->mc_xcursor = mx;
6466 mdb_xcursor_init0(mc);
6468 mc->mc_xcursor = NULL;
6470 if (*mc->mc_dbflag & DB_STALE) {
6471 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6476 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6479 size_t size = sizeof(MDB_cursor);
6481 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6484 if (txn->mt_flags & MDB_TXN_ERROR)
6487 /* Allow read access to the freelist */
6488 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6491 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6492 size += sizeof(MDB_xcursor);
6494 if ((mc = malloc(size)) != NULL) {
6495 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6496 if (txn->mt_cursors) {
6497 mc->mc_next = txn->mt_cursors[dbi];
6498 txn->mt_cursors[dbi] = mc;
6499 mc->mc_flags |= C_UNTRACK;
6511 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6513 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6516 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6519 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6523 /* Return the count of duplicate data items for the current key */
6525 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6529 if (mc == NULL || countp == NULL)
6532 if (mc->mc_xcursor == NULL)
6533 return MDB_INCOMPATIBLE;
6535 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6536 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6539 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6542 *countp = mc->mc_xcursor->mx_db.md_entries;
6548 mdb_cursor_close(MDB_cursor *mc)
6550 if (mc && !mc->mc_backup) {
6551 /* remove from txn, if tracked */
6552 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6553 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6554 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6556 *prev = mc->mc_next;
6563 mdb_cursor_txn(MDB_cursor *mc)
6565 if (!mc) return NULL;
6570 mdb_cursor_dbi(MDB_cursor *mc)
6576 /** Replace the key for a node with a new key.
6577 * @param[in] mc Cursor pointing to the node to operate on.
6578 * @param[in] key The new key to use.
6579 * @return 0 on success, non-zero on failure.
6582 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6589 indx_t ptr, i, numkeys, indx;
6592 indx = mc->mc_ki[mc->mc_top];
6593 mp = mc->mc_pg[mc->mc_top];
6594 node = NODEPTR(mp, indx);
6595 ptr = mp->mp_ptrs[indx];
6599 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6600 k2.mv_data = NODEKEY(node);
6601 k2.mv_size = node->mn_ksize;
6602 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6604 mdb_dkey(&k2, kbuf2),
6610 delta0 = delta = key->mv_size - node->mn_ksize;
6612 /* Must be 2-byte aligned. If new key is
6613 * shorter by 1, the shift will be skipped.
6615 delta += (delta & 1);
6617 if (delta > 0 && SIZELEFT(mp) < delta) {
6619 /* not enough space left, do a delete and split */
6620 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6621 pgno = NODEPGNO(node);
6622 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6623 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6626 numkeys = NUMKEYS(mp);
6627 for (i = 0; i < numkeys; i++) {
6628 if (mp->mp_ptrs[i] <= ptr)
6629 mp->mp_ptrs[i] -= delta;
6632 base = (char *)mp + mp->mp_upper;
6633 len = ptr - mp->mp_upper + NODESIZE;
6634 memmove(base - delta, base, len);
6635 mp->mp_upper -= delta;
6637 node = NODEPTR(mp, indx);
6640 /* But even if no shift was needed, update ksize */
6642 node->mn_ksize = key->mv_size;
6645 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6651 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6653 /** Move a node from csrc to cdst.
6656 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6663 unsigned short flags;
6667 /* Mark src and dst as dirty. */
6668 if ((rc = mdb_page_touch(csrc)) ||
6669 (rc = mdb_page_touch(cdst)))
6672 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6673 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6674 key.mv_size = csrc->mc_db->md_pad;
6675 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6677 data.mv_data = NULL;
6681 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6682 assert(!((long)srcnode&1));
6683 srcpg = NODEPGNO(srcnode);
6684 flags = srcnode->mn_flags;
6685 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6686 unsigned int snum = csrc->mc_snum;
6688 /* must find the lowest key below src */
6689 mdb_page_search_lowest(csrc);
6690 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6691 key.mv_size = csrc->mc_db->md_pad;
6692 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6694 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6695 key.mv_size = NODEKSZ(s2);
6696 key.mv_data = NODEKEY(s2);
6698 csrc->mc_snum = snum--;
6699 csrc->mc_top = snum;
6701 key.mv_size = NODEKSZ(srcnode);
6702 key.mv_data = NODEKEY(srcnode);
6704 data.mv_size = NODEDSZ(srcnode);
6705 data.mv_data = NODEDATA(srcnode);
6707 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6708 unsigned int snum = cdst->mc_snum;
6711 /* must find the lowest key below dst */
6712 mdb_page_search_lowest(cdst);
6713 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6714 bkey.mv_size = cdst->mc_db->md_pad;
6715 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6717 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6718 bkey.mv_size = NODEKSZ(s2);
6719 bkey.mv_data = NODEKEY(s2);
6721 cdst->mc_snum = snum--;
6722 cdst->mc_top = snum;
6723 mdb_cursor_copy(cdst, &mn);
6725 rc = mdb_update_key(&mn, &bkey);
6730 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6731 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6732 csrc->mc_ki[csrc->mc_top],
6734 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6735 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6737 /* Add the node to the destination page.
6739 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6740 if (rc != MDB_SUCCESS)
6743 /* Delete the node from the source page.
6745 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6748 /* Adjust other cursors pointing to mp */
6749 MDB_cursor *m2, *m3;
6750 MDB_dbi dbi = csrc->mc_dbi;
6751 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6753 if (csrc->mc_flags & C_SUB)
6756 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6757 if (csrc->mc_flags & C_SUB)
6758 m3 = &m2->mc_xcursor->mx_cursor;
6761 if (m3 == csrc) continue;
6762 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6763 csrc->mc_ki[csrc->mc_top]) {
6764 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6765 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6770 /* Update the parent separators.
6772 if (csrc->mc_ki[csrc->mc_top] == 0) {
6773 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6774 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6775 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6777 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6778 key.mv_size = NODEKSZ(srcnode);
6779 key.mv_data = NODEKEY(srcnode);
6781 DPRINTF(("update separator for source page %"Z"u to [%s]",
6782 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6783 mdb_cursor_copy(csrc, &mn);
6786 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6789 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6791 indx_t ix = csrc->mc_ki[csrc->mc_top];
6792 nullkey.mv_size = 0;
6793 csrc->mc_ki[csrc->mc_top] = 0;
6794 rc = mdb_update_key(csrc, &nullkey);
6795 csrc->mc_ki[csrc->mc_top] = ix;
6796 assert(rc == MDB_SUCCESS);
6800 if (cdst->mc_ki[cdst->mc_top] == 0) {
6801 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6802 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6803 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6805 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6806 key.mv_size = NODEKSZ(srcnode);
6807 key.mv_data = NODEKEY(srcnode);
6809 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6810 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6811 mdb_cursor_copy(cdst, &mn);
6814 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6817 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6819 indx_t ix = cdst->mc_ki[cdst->mc_top];
6820 nullkey.mv_size = 0;
6821 cdst->mc_ki[cdst->mc_top] = 0;
6822 rc = mdb_update_key(cdst, &nullkey);
6823 cdst->mc_ki[cdst->mc_top] = ix;
6824 assert(rc == MDB_SUCCESS);
6831 /** Merge one page into another.
6832 * The nodes from the page pointed to by \b csrc will
6833 * be copied to the page pointed to by \b cdst and then
6834 * the \b csrc page will be freed.
6835 * @param[in] csrc Cursor pointing to the source page.
6836 * @param[in] cdst Cursor pointing to the destination page.
6839 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6847 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6848 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6850 assert(csrc->mc_snum > 1); /* can't merge root page */
6851 assert(cdst->mc_snum > 1);
6853 /* Mark dst as dirty. */
6854 if ((rc = mdb_page_touch(cdst)))
6857 /* Move all nodes from src to dst.
6859 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6860 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6861 key.mv_size = csrc->mc_db->md_pad;
6862 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6863 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6864 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6865 if (rc != MDB_SUCCESS)
6867 key.mv_data = (char *)key.mv_data + key.mv_size;
6870 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6871 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6872 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6873 unsigned int snum = csrc->mc_snum;
6875 /* must find the lowest key below src */
6876 mdb_page_search_lowest(csrc);
6877 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6878 key.mv_size = csrc->mc_db->md_pad;
6879 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6881 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6882 key.mv_size = NODEKSZ(s2);
6883 key.mv_data = NODEKEY(s2);
6885 csrc->mc_snum = snum--;
6886 csrc->mc_top = snum;
6888 key.mv_size = srcnode->mn_ksize;
6889 key.mv_data = NODEKEY(srcnode);
6892 data.mv_size = NODEDSZ(srcnode);
6893 data.mv_data = NODEDATA(srcnode);
6894 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6895 if (rc != MDB_SUCCESS)
6900 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6901 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
6902 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
6904 /* Unlink the src page from parent and add to free list.
6906 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6907 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6910 rc = mdb_update_key(csrc, &key);
6916 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6917 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6920 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6921 csrc->mc_db->md_leaf_pages--;
6923 csrc->mc_db->md_branch_pages--;
6925 /* Adjust other cursors pointing to mp */
6926 MDB_cursor *m2, *m3;
6927 MDB_dbi dbi = csrc->mc_dbi;
6928 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6930 if (csrc->mc_flags & C_SUB)
6933 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6934 if (csrc->mc_flags & C_SUB)
6935 m3 = &m2->mc_xcursor->mx_cursor;
6938 if (m3 == csrc) continue;
6939 if (m3->mc_snum < csrc->mc_snum) continue;
6940 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6941 m3->mc_pg[csrc->mc_top] = mp;
6942 m3->mc_ki[csrc->mc_top] += nkeys;
6946 mdb_cursor_pop(csrc);
6948 return mdb_rebalance(csrc);
6951 /** Copy the contents of a cursor.
6952 * @param[in] csrc The cursor to copy from.
6953 * @param[out] cdst The cursor to copy to.
6956 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6960 cdst->mc_txn = csrc->mc_txn;
6961 cdst->mc_dbi = csrc->mc_dbi;
6962 cdst->mc_db = csrc->mc_db;
6963 cdst->mc_dbx = csrc->mc_dbx;
6964 cdst->mc_snum = csrc->mc_snum;
6965 cdst->mc_top = csrc->mc_top;
6966 cdst->mc_flags = csrc->mc_flags;
6968 for (i=0; i<csrc->mc_snum; i++) {
6969 cdst->mc_pg[i] = csrc->mc_pg[i];
6970 cdst->mc_ki[i] = csrc->mc_ki[i];
6974 /** Rebalance the tree after a delete operation.
6975 * @param[in] mc Cursor pointing to the page where rebalancing
6977 * @return 0 on success, non-zero on failure.
6980 mdb_rebalance(MDB_cursor *mc)
6984 unsigned int ptop, minkeys;
6987 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6991 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6992 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
6993 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6994 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
6995 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
6999 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7000 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7003 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7004 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7010 if (mc->mc_snum < 2) {
7011 MDB_page *mp = mc->mc_pg[0];
7013 DPUTS("Can't rebalance a subpage, ignoring");
7016 if (NUMKEYS(mp) == 0) {
7017 DPUTS("tree is completely empty");
7018 mc->mc_db->md_root = P_INVALID;
7019 mc->mc_db->md_depth = 0;
7020 mc->mc_db->md_leaf_pages = 0;
7021 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7024 /* Adjust cursors pointing to mp */
7028 MDB_cursor *m2, *m3;
7029 MDB_dbi dbi = mc->mc_dbi;
7031 if (mc->mc_flags & C_SUB)
7034 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7035 if (mc->mc_flags & C_SUB)
7036 m3 = &m2->mc_xcursor->mx_cursor;
7039 if (m3->mc_snum < mc->mc_snum) continue;
7040 if (m3->mc_pg[0] == mp) {
7046 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7047 DPUTS("collapsing root page!");
7048 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7051 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7052 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7055 mc->mc_db->md_depth--;
7056 mc->mc_db->md_branch_pages--;
7057 mc->mc_ki[0] = mc->mc_ki[1];
7059 /* Adjust other cursors pointing to mp */
7060 MDB_cursor *m2, *m3;
7061 MDB_dbi dbi = mc->mc_dbi;
7063 if (mc->mc_flags & C_SUB)
7066 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7067 if (mc->mc_flags & C_SUB)
7068 m3 = &m2->mc_xcursor->mx_cursor;
7071 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7072 if (m3->mc_pg[0] == mp) {
7073 m3->mc_pg[0] = mc->mc_pg[0];
7076 m3->mc_ki[0] = m3->mc_ki[1];
7081 DPUTS("root page doesn't need rebalancing");
7085 /* The parent (branch page) must have at least 2 pointers,
7086 * otherwise the tree is invalid.
7088 ptop = mc->mc_top-1;
7089 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7091 /* Leaf page fill factor is below the threshold.
7092 * Try to move keys from left or right neighbor, or
7093 * merge with a neighbor page.
7098 mdb_cursor_copy(mc, &mn);
7099 mn.mc_xcursor = NULL;
7101 if (mc->mc_ki[ptop] == 0) {
7102 /* We're the leftmost leaf in our parent.
7104 DPUTS("reading right neighbor");
7106 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7107 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7110 mn.mc_ki[mn.mc_top] = 0;
7111 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7113 /* There is at least one neighbor to the left.
7115 DPUTS("reading left neighbor");
7117 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7118 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7121 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7122 mc->mc_ki[mc->mc_top] = 0;
7125 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7126 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7127 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7129 /* If the neighbor page is above threshold and has enough keys,
7130 * move one key from it. Otherwise we should try to merge them.
7131 * (A branch page must never have less than 2 keys.)
7133 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7134 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7135 return mdb_node_move(&mn, mc);
7137 if (mc->mc_ki[ptop] == 0)
7138 rc = mdb_page_merge(&mn, mc);
7140 rc = mdb_page_merge(mc, &mn);
7141 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7146 /** Complete a delete operation started by #mdb_cursor_del(). */
7148 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7154 mp = mc->mc_pg[mc->mc_top];
7155 ki = mc->mc_ki[mc->mc_top];
7157 /* add overflow pages to free list */
7158 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7162 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7163 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7164 (rc = mdb_ovpage_free(mc, omp)))
7167 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7168 mc->mc_db->md_entries--;
7169 rc = mdb_rebalance(mc);
7170 if (rc != MDB_SUCCESS)
7171 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7172 /* if mc points past last node in page, invalidate */
7173 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7174 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7177 /* Adjust other cursors pointing to mp */
7180 MDB_dbi dbi = mc->mc_dbi;
7182 mp = mc->mc_pg[mc->mc_top];
7183 nkeys = NUMKEYS(mp);
7184 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7187 if (!(m2->mc_flags & C_INITIALIZED))
7189 if (m2->mc_pg[mc->mc_top] == mp) {
7190 if (m2->mc_ki[mc->mc_top] > ki)
7191 m2->mc_ki[mc->mc_top]--;
7192 if (m2->mc_ki[mc->mc_top] >= nkeys)
7193 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7202 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7203 MDB_val *key, MDB_val *data)
7208 MDB_val rdata, *xdata;
7212 assert(key != NULL);
7214 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7216 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7219 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7220 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7222 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7223 return MDB_BAD_VALSIZE;
7226 mdb_cursor_init(&mc, txn, dbi, &mx);
7229 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7230 /* must ignore any data */
7241 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7243 /* let mdb_page_split know about this cursor if needed:
7244 * delete will trigger a rebalance; if it needs to move
7245 * a node from one page to another, it will have to
7246 * update the parent's separator key(s). If the new sepkey
7247 * is larger than the current one, the parent page may
7248 * run out of space, triggering a split. We need this
7249 * cursor to be consistent until the end of the rebalance.
7251 mc.mc_flags |= C_UNTRACK;
7252 mc.mc_next = txn->mt_cursors[dbi];
7253 txn->mt_cursors[dbi] = &mc;
7254 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7255 txn->mt_cursors[dbi] = mc.mc_next;
7260 /** Split a page and insert a new node.
7261 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7262 * The cursor will be updated to point to the actual page and index where
7263 * the node got inserted after the split.
7264 * @param[in] newkey The key for the newly inserted node.
7265 * @param[in] newdata The data for the newly inserted node.
7266 * @param[in] newpgno The page number, if the new node is a branch node.
7267 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7268 * @return 0 on success, non-zero on failure.
7271 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7272 unsigned int nflags)
7275 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7278 unsigned int i, j, split_indx, nkeys, pmax;
7280 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7282 MDB_page *mp, *rp, *pp;
7287 mp = mc->mc_pg[mc->mc_top];
7288 newindx = mc->mc_ki[mc->mc_top];
7290 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7291 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7292 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7294 /* Create a right sibling. */
7295 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7297 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7299 if (mc->mc_snum < 2) {
7300 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7302 /* shift current top to make room for new parent */
7303 mc->mc_pg[1] = mc->mc_pg[0];
7304 mc->mc_ki[1] = mc->mc_ki[0];
7307 mc->mc_db->md_root = pp->mp_pgno;
7308 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7309 mc->mc_db->md_depth++;
7312 /* Add left (implicit) pointer. */
7313 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7314 /* undo the pre-push */
7315 mc->mc_pg[0] = mc->mc_pg[1];
7316 mc->mc_ki[0] = mc->mc_ki[1];
7317 mc->mc_db->md_root = mp->mp_pgno;
7318 mc->mc_db->md_depth--;
7325 ptop = mc->mc_top-1;
7326 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7329 mc->mc_flags |= C_SPLITTING;
7330 mdb_cursor_copy(mc, &mn);
7331 mn.mc_pg[mn.mc_top] = rp;
7332 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7334 if (nflags & MDB_APPEND) {
7335 mn.mc_ki[mn.mc_top] = 0;
7337 split_indx = newindx;
7342 nkeys = NUMKEYS(mp);
7343 split_indx = nkeys / 2;
7344 if (newindx < split_indx)
7350 unsigned int lsize, rsize, ksize;
7351 /* Move half of the keys to the right sibling */
7353 x = mc->mc_ki[mc->mc_top] - split_indx;
7354 ksize = mc->mc_db->md_pad;
7355 split = LEAF2KEY(mp, split_indx, ksize);
7356 rsize = (nkeys - split_indx) * ksize;
7357 lsize = (nkeys - split_indx) * sizeof(indx_t);
7358 mp->mp_lower -= lsize;
7359 rp->mp_lower += lsize;
7360 mp->mp_upper += rsize - lsize;
7361 rp->mp_upper -= rsize - lsize;
7362 sepkey.mv_size = ksize;
7363 if (newindx == split_indx) {
7364 sepkey.mv_data = newkey->mv_data;
7366 sepkey.mv_data = split;
7369 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7370 memcpy(rp->mp_ptrs, split, rsize);
7371 sepkey.mv_data = rp->mp_ptrs;
7372 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7373 memcpy(ins, newkey->mv_data, ksize);
7374 mp->mp_lower += sizeof(indx_t);
7375 mp->mp_upper -= ksize - sizeof(indx_t);
7378 memcpy(rp->mp_ptrs, split, x * ksize);
7379 ins = LEAF2KEY(rp, x, ksize);
7380 memcpy(ins, newkey->mv_data, ksize);
7381 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7382 rp->mp_lower += sizeof(indx_t);
7383 rp->mp_upper -= ksize - sizeof(indx_t);
7384 mc->mc_ki[mc->mc_top] = x;
7385 mc->mc_pg[mc->mc_top] = rp;
7390 /* For leaf pages, check the split point based on what
7391 * fits where, since otherwise mdb_node_add can fail.
7393 * This check is only needed when the data items are
7394 * relatively large, such that being off by one will
7395 * make the difference between success or failure.
7397 * It's also relevant if a page happens to be laid out
7398 * such that one half of its nodes are all "small" and
7399 * the other half of its nodes are "large." If the new
7400 * item is also "large" and falls on the half with
7401 * "large" nodes, it also may not fit.
7404 unsigned int psize, nsize;
7405 /* Maximum free space in an empty page */
7406 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7407 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7408 if ((nkeys < 20) || (nsize > pmax/16)) {
7409 if (newindx <= split_indx) {
7412 for (i=0; i<split_indx; i++) {
7413 node = NODEPTR(mp, i);
7414 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7415 if (F_ISSET(node->mn_flags, F_BIGDATA))
7416 psize += sizeof(pgno_t);
7418 psize += NODEDSZ(node);
7422 split_indx = newindx;
7433 for (i=nkeys-1; i>=split_indx; i--) {
7434 node = NODEPTR(mp, i);
7435 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7436 if (F_ISSET(node->mn_flags, F_BIGDATA))
7437 psize += sizeof(pgno_t);
7439 psize += NODEDSZ(node);
7443 split_indx = newindx;
7454 /* First find the separating key between the split pages.
7455 * The case where newindx == split_indx is ambiguous; the
7456 * new item could go to the new page or stay on the original
7457 * page. If newpos == 1 it goes to the new page.
7459 if (newindx == split_indx && newpos) {
7460 sepkey.mv_size = newkey->mv_size;
7461 sepkey.mv_data = newkey->mv_data;
7463 node = NODEPTR(mp, split_indx);
7464 sepkey.mv_size = node->mn_ksize;
7465 sepkey.mv_data = NODEKEY(node);
7469 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7471 /* Copy separator key to the parent.
7473 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7477 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7480 if (mn.mc_snum == mc->mc_snum) {
7481 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7482 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7483 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7484 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7489 /* Right page might now have changed parent.
7490 * Check if left page also changed parent.
7492 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7493 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7494 for (i=0; i<ptop; i++) {
7495 mc->mc_pg[i] = mn.mc_pg[i];
7496 mc->mc_ki[i] = mn.mc_ki[i];
7498 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7499 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7503 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7506 mc->mc_flags ^= C_SPLITTING;
7507 if (rc != MDB_SUCCESS) {
7510 if (nflags & MDB_APPEND) {
7511 mc->mc_pg[mc->mc_top] = rp;
7512 mc->mc_ki[mc->mc_top] = 0;
7513 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7516 for (i=0; i<mc->mc_top; i++)
7517 mc->mc_ki[i] = mn.mc_ki[i];
7524 /* Move half of the keys to the right sibling. */
7526 /* grab a page to hold a temporary copy */
7527 copy = mdb_page_malloc(mc->mc_txn, 1);
7531 copy->mp_pgno = mp->mp_pgno;
7532 copy->mp_flags = mp->mp_flags;
7533 copy->mp_lower = PAGEHDRSZ;
7534 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7535 mc->mc_pg[mc->mc_top] = copy;
7536 for (i = j = 0; i <= nkeys; j++) {
7537 if (i == split_indx) {
7538 /* Insert in right sibling. */
7539 /* Reset insert index for right sibling. */
7540 if (i != newindx || (newpos ^ ins_new)) {
7542 mc->mc_pg[mc->mc_top] = rp;
7546 if (i == newindx && !ins_new) {
7547 /* Insert the original entry that caused the split. */
7548 rkey.mv_data = newkey->mv_data;
7549 rkey.mv_size = newkey->mv_size;
7558 /* Update index for the new key. */
7559 mc->mc_ki[mc->mc_top] = j;
7560 } else if (i == nkeys) {
7563 node = NODEPTR(mp, i);
7564 rkey.mv_data = NODEKEY(node);
7565 rkey.mv_size = node->mn_ksize;
7567 xdata.mv_data = NODEDATA(node);
7568 xdata.mv_size = NODEDSZ(node);
7571 pgno = NODEPGNO(node);
7572 flags = node->mn_flags;
7577 if (!IS_LEAF(mp) && j == 0) {
7578 /* First branch index doesn't need key data. */
7582 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7586 nkeys = NUMKEYS(copy);
7587 for (i=0; i<nkeys; i++)
7588 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7589 mp->mp_lower = copy->mp_lower;
7590 mp->mp_upper = copy->mp_upper;
7591 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7592 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7594 /* reset back to original page */
7595 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7596 mc->mc_pg[mc->mc_top] = mp;
7597 if (nflags & MDB_RESERVE) {
7598 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7599 if (!(node->mn_flags & F_BIGDATA))
7600 newdata->mv_data = NODEDATA(node);
7604 /* Make sure mc_ki is still valid.
7606 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7607 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7608 for (i=0; i<ptop; i++) {
7609 mc->mc_pg[i] = mn.mc_pg[i];
7610 mc->mc_ki[i] = mn.mc_ki[i];
7612 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7613 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7617 /* return tmp page to freelist */
7618 mdb_page_free(mc->mc_txn->mt_env, copy);
7621 /* Adjust other cursors pointing to mp */
7622 MDB_cursor *m2, *m3;
7623 MDB_dbi dbi = mc->mc_dbi;
7624 int fixup = NUMKEYS(mp);
7626 if (mc->mc_flags & C_SUB)
7629 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7630 if (mc->mc_flags & C_SUB)
7631 m3 = &m2->mc_xcursor->mx_cursor;
7636 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7638 if (m3->mc_flags & C_SPLITTING)
7643 for (k=m3->mc_top; k>=0; k--) {
7644 m3->mc_ki[k+1] = m3->mc_ki[k];
7645 m3->mc_pg[k+1] = m3->mc_pg[k];
7647 if (m3->mc_ki[0] >= split_indx) {
7652 m3->mc_pg[0] = mc->mc_pg[0];
7656 if (m3->mc_pg[mc->mc_top] == mp) {
7657 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7658 m3->mc_ki[mc->mc_top]++;
7659 if (m3->mc_ki[mc->mc_top] >= fixup) {
7660 m3->mc_pg[mc->mc_top] = rp;
7661 m3->mc_ki[mc->mc_top] -= fixup;
7662 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7664 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7665 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7674 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7675 MDB_val *key, MDB_val *data, unsigned int flags)
7680 assert(key != NULL);
7681 assert(data != NULL);
7683 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7686 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7687 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7689 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7690 return MDB_BAD_VALSIZE;
7693 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7696 mdb_cursor_init(&mc, txn, dbi, &mx);
7697 return mdb_cursor_put(&mc, key, data, flags);
7701 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7703 if ((flag & CHANGEABLE) != flag)
7706 env->me_flags |= flag;
7708 env->me_flags &= ~flag;
7713 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7718 *arg = env->me_flags;
7723 mdb_env_get_path(MDB_env *env, const char **arg)
7728 *arg = env->me_path;
7732 /** Common code for #mdb_stat() and #mdb_env_stat().
7733 * @param[in] env the environment to operate in.
7734 * @param[in] db the #MDB_db record containing the stats to return.
7735 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7736 * @return 0, this function always succeeds.
7739 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7741 arg->ms_psize = env->me_psize;
7742 arg->ms_depth = db->md_depth;
7743 arg->ms_branch_pages = db->md_branch_pages;
7744 arg->ms_leaf_pages = db->md_leaf_pages;
7745 arg->ms_overflow_pages = db->md_overflow_pages;
7746 arg->ms_entries = db->md_entries;
7751 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7755 if (env == NULL || arg == NULL)
7758 toggle = mdb_env_pick_meta(env);
7760 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7764 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7768 if (env == NULL || arg == NULL)
7771 toggle = mdb_env_pick_meta(env);
7772 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7773 arg->me_mapsize = env->me_mapsize;
7774 arg->me_maxreaders = env->me_maxreaders;
7776 /* me_numreaders may be zero if this process never used any readers. Use
7777 * the shared numreader count if it exists.
7779 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7781 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7782 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7786 /** Set the default comparison functions for a database.
7787 * Called immediately after a database is opened to set the defaults.
7788 * The user can then override them with #mdb_set_compare() or
7789 * #mdb_set_dupsort().
7790 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7791 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7794 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7796 uint16_t f = txn->mt_dbs[dbi].md_flags;
7798 txn->mt_dbxs[dbi].md_cmp =
7799 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7800 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7802 txn->mt_dbxs[dbi].md_dcmp =
7803 !(f & MDB_DUPSORT) ? 0 :
7804 ((f & MDB_INTEGERDUP)
7805 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7806 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7809 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7814 int rc, dbflag, exact;
7815 unsigned int unused = 0;
7818 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7819 mdb_default_cmp(txn, FREE_DBI);
7822 if ((flags & VALID_FLAGS) != flags)
7824 if (txn->mt_flags & MDB_TXN_ERROR)
7830 if (flags & PERSISTENT_FLAGS) {
7831 uint16_t f2 = flags & PERSISTENT_FLAGS;
7832 /* make sure flag changes get committed */
7833 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7834 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7835 txn->mt_flags |= MDB_TXN_DIRTY;
7838 mdb_default_cmp(txn, MAIN_DBI);
7842 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7843 mdb_default_cmp(txn, MAIN_DBI);
7846 /* Is the DB already open? */
7848 for (i=2; i<txn->mt_numdbs; i++) {
7849 if (!txn->mt_dbxs[i].md_name.mv_size) {
7850 /* Remember this free slot */
7851 if (!unused) unused = i;
7854 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7855 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7861 /* If no free slot and max hit, fail */
7862 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7863 return MDB_DBS_FULL;
7865 /* Cannot mix named databases with some mainDB flags */
7866 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7867 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7869 /* Find the DB info */
7870 dbflag = DB_NEW|DB_VALID;
7873 key.mv_data = (void *)name;
7874 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7875 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7876 if (rc == MDB_SUCCESS) {
7877 /* make sure this is actually a DB */
7878 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7879 if (!(node->mn_flags & F_SUBDATA))
7880 return MDB_INCOMPATIBLE;
7881 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7882 /* Create if requested */
7884 data.mv_size = sizeof(MDB_db);
7885 data.mv_data = &dummy;
7886 memset(&dummy, 0, sizeof(dummy));
7887 dummy.md_root = P_INVALID;
7888 dummy.md_flags = flags & PERSISTENT_FLAGS;
7889 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7893 /* OK, got info, add to table */
7894 if (rc == MDB_SUCCESS) {
7895 unsigned int slot = unused ? unused : txn->mt_numdbs;
7896 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7897 txn->mt_dbxs[slot].md_name.mv_size = len;
7898 txn->mt_dbxs[slot].md_rel = NULL;
7899 txn->mt_dbflags[slot] = dbflag;
7900 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7902 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7903 mdb_default_cmp(txn, slot);
7912 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7914 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7917 if (txn->mt_dbflags[dbi] & DB_STALE) {
7920 /* Stale, must read the DB's root. cursor_init does it for us. */
7921 mdb_cursor_init(&mc, txn, dbi, &mx);
7923 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7926 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7929 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7931 ptr = env->me_dbxs[dbi].md_name.mv_data;
7932 env->me_dbxs[dbi].md_name.mv_data = NULL;
7933 env->me_dbxs[dbi].md_name.mv_size = 0;
7934 env->me_dbflags[dbi] = 0;
7938 int mdb_dbi_flags(MDB_env *env, MDB_dbi dbi, unsigned int *flags)
7940 /* We could return the flags for the FREE_DBI too but what's the point? */
7941 if (dbi < MAIN_DBI || dbi >= env->me_numdbs)
7943 *flags = env->me_dbflags[dbi];
7947 /** Add all the DB's pages to the free list.
7948 * @param[in] mc Cursor on the DB to free.
7949 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7950 * @return 0 on success, non-zero on failure.
7953 mdb_drop0(MDB_cursor *mc, int subs)
7957 rc = mdb_page_search(mc, NULL, 0);
7958 if (rc == MDB_SUCCESS) {
7959 MDB_txn *txn = mc->mc_txn;
7964 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7965 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7968 mdb_cursor_copy(mc, &mx);
7969 while (mc->mc_snum > 0) {
7970 MDB_page *mp = mc->mc_pg[mc->mc_top];
7971 unsigned n = NUMKEYS(mp);
7973 for (i=0; i<n; i++) {
7974 ni = NODEPTR(mp, i);
7975 if (ni->mn_flags & F_BIGDATA) {
7978 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7979 rc = mdb_page_get(txn, pg, &omp, NULL);
7982 assert(IS_OVERFLOW(omp));
7983 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7987 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7988 mdb_xcursor_init1(mc, ni);
7989 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7995 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7997 for (i=0; i<n; i++) {
7999 ni = NODEPTR(mp, i);
8002 mdb_midl_xappend(txn->mt_free_pgs, pg);
8007 mc->mc_ki[mc->mc_top] = i;
8008 rc = mdb_cursor_sibling(mc, 1);
8010 /* no more siblings, go back to beginning
8011 * of previous level.
8015 for (i=1; i<mc->mc_snum; i++) {
8017 mc->mc_pg[i] = mx.mc_pg[i];
8022 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8023 } else if (rc == MDB_NOTFOUND) {
8029 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8031 MDB_cursor *mc, *m2;
8034 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8037 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8040 rc = mdb_cursor_open(txn, dbi, &mc);
8044 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8045 /* Invalidate the dropped DB's cursors */
8046 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8047 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8051 /* Can't delete the main DB */
8052 if (del && dbi > MAIN_DBI) {
8053 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8055 txn->mt_dbflags[dbi] = DB_STALE;
8056 mdb_dbi_close(txn->mt_env, dbi);
8059 /* reset the DB record, mark it dirty */
8060 txn->mt_dbflags[dbi] |= DB_DIRTY;
8061 txn->mt_dbs[dbi].md_depth = 0;
8062 txn->mt_dbs[dbi].md_branch_pages = 0;
8063 txn->mt_dbs[dbi].md_leaf_pages = 0;
8064 txn->mt_dbs[dbi].md_overflow_pages = 0;
8065 txn->mt_dbs[dbi].md_entries = 0;
8066 txn->mt_dbs[dbi].md_root = P_INVALID;
8068 txn->mt_flags |= MDB_TXN_DIRTY;
8071 mdb_cursor_close(mc);
8075 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8077 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8080 txn->mt_dbxs[dbi].md_cmp = cmp;
8084 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8086 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8089 txn->mt_dbxs[dbi].md_dcmp = cmp;
8093 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8095 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8098 txn->mt_dbxs[dbi].md_rel = rel;
8102 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8104 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8107 txn->mt_dbxs[dbi].md_relctx = ctx;
8111 int mdb_env_get_maxkeysize(MDB_env *env)
8113 return MDB_MAXKEYSIZE;
8116 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8118 unsigned int i, rdrs;
8125 if (!env->me_txns) {
8126 return func("(no reader locks)\n", ctx);
8128 rdrs = env->me_txns->mti_numreaders;
8129 mr = env->me_txns->mti_readers;
8130 for (i=0; i<rdrs; i++) {
8135 if (mr[i].mr_txnid == (txnid_t)-1) {
8136 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8138 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8142 func(" pid thread txnid\n", ctx);
8144 rc = func(buf, ctx);
8150 func("(no active readers)\n", ctx);
8155 /* insert pid into list if not already present.
8156 * return -1 if already present.
8158 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8160 /* binary search of pid in list */
8162 unsigned cursor = 1;
8164 unsigned n = ids[0];
8167 unsigned pivot = n >> 1;
8168 cursor = base + pivot + 1;
8169 val = pid - ids[cursor];
8174 } else if ( val > 0 ) {
8179 /* found, so it's a duplicate */
8188 for (n = ids[0]; n > cursor; n--)
8194 int mdb_reader_check(MDB_env *env, int *dead)
8196 unsigned int i, j, rdrs;
8207 rdrs = env->me_txns->mti_numreaders;
8208 pids = malloc((rdrs+1) * sizeof(pid_t));
8212 mr = env->me_txns->mti_readers;
8214 for (i=0; i<rdrs; i++) {
8215 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8217 if (mdb_pid_insert(pids, pid) == 0) {
8218 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8220 /* Recheck, a new process may have reused pid */
8221 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8222 for (j=i; j<rdrs; j++)
8223 if (mr[j].mr_pid == pid) {
8228 UNLOCK_MUTEX_R(env);