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>
42 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
43 * as int64 which is wrong. MSVC doesn't define it at all, so just
48 # include <sys/param.h>
50 # define LITTLE_ENDIAN 1234
51 # define BIG_ENDIAN 4321
52 # define BYTE_ORDER LITTLE_ENDIAN
54 # define SSIZE_MAX INT_MAX
58 #define MDB_PID_T pid_t
59 #include <sys/param.h>
62 #ifdef HAVE_SYS_FILE_H
79 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
80 #include <netinet/in.h>
81 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
84 #if defined(__APPLE__) || defined (BSD)
85 # define MDB_USE_POSIX_SEM 1
86 # define MDB_FDATASYNC fsync
87 #elif defined(ANDROID)
88 # define MDB_FDATASYNC fsync
93 #ifdef MDB_USE_POSIX_SEM
94 # define MDB_USE_HASH 1
95 #include <semaphore.h>
100 #include <valgrind/memcheck.h>
101 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
102 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
103 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
104 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
105 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
107 #define VGMEMP_CREATE(h,r,z)
108 #define VGMEMP_ALLOC(h,a,s)
109 #define VGMEMP_FREE(h,a)
110 #define VGMEMP_DESTROY(h)
111 #define VGMEMP_DEFINED(a,s)
115 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
116 /* Solaris just defines one or the other */
117 # define LITTLE_ENDIAN 1234
118 # define BIG_ENDIAN 4321
119 # ifdef _LITTLE_ENDIAN
120 # define BYTE_ORDER LITTLE_ENDIAN
122 # define BYTE_ORDER BIG_ENDIAN
125 # define BYTE_ORDER __BYTE_ORDER
129 #ifndef LITTLE_ENDIAN
130 #define LITTLE_ENDIAN __LITTLE_ENDIAN
133 #define BIG_ENDIAN __BIG_ENDIAN
136 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
137 #define MISALIGNED_OK 1
143 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
144 # error "Unknown or unsupported endianness (BYTE_ORDER)"
145 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
146 # error "Two's complement, reasonably sized integer types, please"
149 /** @defgroup internal MDB Internals
152 /** @defgroup compat Compatibility Macros
153 * A bunch of macros to minimize the amount of platform-specific ifdefs
154 * needed throughout the rest of the code. When the features this library
155 * needs are similar enough to POSIX to be hidden in a one-or-two line
156 * replacement, this macro approach is used.
160 /** Wrapper around __func__, which is a C99 feature */
161 #if __STDC_VERSION__ >= 199901L
162 # define mdb_func_ __func__
163 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
164 # define mdb_func_ __FUNCTION__
166 /* If a debug message says <mdb_unknown>(), update the #if statements above */
167 # define mdb_func_ "<mdb_unknown>"
171 #define MDB_USE_HASH 1
172 #define MDB_PIDLOCK 0
173 #define pthread_t DWORD
174 #define pthread_mutex_t HANDLE
175 #define pthread_key_t DWORD
176 #define pthread_self() GetCurrentThreadId()
177 #define pthread_key_create(x,y) \
178 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
179 #define pthread_key_delete(x) TlsFree(x)
180 #define pthread_getspecific(x) TlsGetValue(x)
181 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
182 #define pthread_mutex_unlock(x) ReleaseMutex(x)
183 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
184 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
185 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
186 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
187 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
188 #define getpid() GetCurrentProcessId()
189 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
190 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
191 #define ErrCode() GetLastError()
192 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
193 #define close(fd) (CloseHandle(fd) ? 0 : -1)
194 #define munmap(ptr,len) UnmapViewOfFile(ptr)
195 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
196 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
198 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
203 #define Z "z" /**< printf format modifier for size_t */
205 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
206 #define MDB_PIDLOCK 1
208 #ifdef MDB_USE_POSIX_SEM
210 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
211 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
212 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
213 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
216 mdb_sem_wait(sem_t *sem)
219 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
224 /** Lock the reader mutex.
226 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
227 /** Unlock the reader mutex.
229 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
231 /** Lock the writer mutex.
232 * Only a single write transaction is allowed at a time. Other writers
233 * will block waiting for this mutex.
235 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
236 /** Unlock the writer mutex.
238 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
239 #endif /* MDB_USE_POSIX_SEM */
241 /** Get the error code for the last failed system function.
243 #define ErrCode() errno
245 /** An abstraction for a file handle.
246 * On POSIX systems file handles are small integers. On Windows
247 * they're opaque pointers.
251 /** A value for an invalid file handle.
252 * Mainly used to initialize file variables and signify that they are
255 #define INVALID_HANDLE_VALUE (-1)
257 /** Get the size of a memory page for the system.
258 * This is the basic size that the platform's memory manager uses, and is
259 * fundamental to the use of memory-mapped files.
261 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
264 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
267 #define MNAME_LEN (sizeof(pthread_mutex_t))
273 /** A flag for opening a file and requesting synchronous data writes.
274 * This is only used when writing a meta page. It's not strictly needed;
275 * we could just do a normal write and then immediately perform a flush.
276 * But if this flag is available it saves us an extra system call.
278 * @note If O_DSYNC is undefined but exists in /usr/include,
279 * preferably set some compiler flag to get the definition.
280 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
283 # define MDB_DSYNC O_DSYNC
287 /** Function for flushing the data of a file. Define this to fsync
288 * if fdatasync() is not supported.
290 #ifndef MDB_FDATASYNC
291 # define MDB_FDATASYNC fdatasync
295 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
306 /** A page number in the database.
307 * Note that 64 bit page numbers are overkill, since pages themselves
308 * already represent 12-13 bits of addressable memory, and the OS will
309 * always limit applications to a maximum of 63 bits of address space.
311 * @note In the #MDB_node structure, we only store 48 bits of this value,
312 * which thus limits us to only 60 bits of addressable data.
314 typedef MDB_ID pgno_t;
316 /** A transaction ID.
317 * See struct MDB_txn.mt_txnid for details.
319 typedef MDB_ID txnid_t;
321 /** @defgroup debug Debug Macros
325 /** Enable debug output. Needs variable argument macros (a C99 feature).
326 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
327 * read from and written to the database (used for free space management).
333 static int mdb_debug;
334 static txnid_t mdb_debug_start;
336 /** Print a debug message with printf formatting.
337 * Requires double parenthesis around 2 or more args.
339 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
340 # define DPRINTF0(fmt, ...) \
341 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
343 # define DPRINTF(args) ((void) 0)
345 /** Print a debug string.
346 * The string is printed literally, with no format processing.
348 #define DPUTS(arg) DPRINTF(("%s", arg))
349 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
351 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
354 /** @brief The maximum size of a database page.
356 * This is 32k, since it must fit in #MDB_page.#mp_upper.
358 * LMDB will use database pages < OS pages if needed.
359 * That causes more I/O in write transactions: The OS must
360 * know (read) the whole page before writing a partial page.
362 * Note that we don't currently support Huge pages. On Linux,
363 * regular data files cannot use Huge pages, and in general
364 * Huge pages aren't actually pageable. We rely on the OS
365 * demand-pager to read our data and page it out when memory
366 * pressure from other processes is high. So until OSs have
367 * actual paging support for Huge pages, they're not viable.
369 #define MAX_PAGESIZE 0x8000
371 /** The minimum number of keys required in a database page.
372 * Setting this to a larger value will place a smaller bound on the
373 * maximum size of a data item. Data items larger than this size will
374 * be pushed into overflow pages instead of being stored directly in
375 * the B-tree node. This value used to default to 4. With a page size
376 * of 4096 bytes that meant that any item larger than 1024 bytes would
377 * go into an overflow page. That also meant that on average 2-3KB of
378 * each overflow page was wasted space. The value cannot be lower than
379 * 2 because then there would no longer be a tree structure. With this
380 * value, items larger than 2KB will go into overflow pages, and on
381 * average only 1KB will be wasted.
383 #define MDB_MINKEYS 2
385 /** A stamp that identifies a file as an MDB file.
386 * There's nothing special about this value other than that it is easily
387 * recognizable, and it will reflect any byte order mismatches.
389 #define MDB_MAGIC 0xBEEFC0DE
391 /** The version number for a database's datafile format. */
392 #define MDB_DATA_VERSION 1
393 /** The version number for a database's lockfile format. */
394 #define MDB_LOCK_VERSION 1
396 /** @brief The max size of a key we can write, or 0 for dynamic max.
398 * Define this as 0 to compute the max from the page size. 511
399 * is default for backwards compat: liblmdb <= 0.9.10 can break
400 * when modifying a DB with keys/dupsort data bigger than its max.
402 * Data items in an #MDB_DUPSORT database are also limited to
403 * this size, since they're actually keys of a sub-DB. Keys and
404 * #MDB_DUPSORT data items must fit on a node in a regular page.
406 #ifndef MDB_MAXKEYSIZE
407 #define MDB_MAXKEYSIZE 511
410 /** The maximum size of a key we can write to the environment. */
412 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
414 #define ENV_MAXKEY(env) ((env)->me_maxkey)
417 /** @brief The maximum size of a data item.
419 * We only store a 32 bit value for node sizes.
421 #define MAXDATASIZE 0xffffffffUL
424 /** Key size which fits in a #DKBUF.
427 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
430 * This is used for printing a hex dump of a key's contents.
432 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
433 /** Display a key in hex.
435 * Invoke a function to display a key in hex.
437 #define DKEY(x) mdb_dkey(x, kbuf)
443 /** An invalid page number.
444 * Mainly used to denote an empty tree.
446 #define P_INVALID (~(pgno_t)0)
448 /** Test if the flags \b f are set in a flag word \b w. */
449 #define F_ISSET(w, f) (((w) & (f)) == (f))
451 /** Round \b n up to an even number. */
452 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
454 /** Used for offsets within a single page.
455 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
458 typedef uint16_t indx_t;
460 /** Default size of memory map.
461 * This is certainly too small for any actual applications. Apps should always set
462 * the size explicitly using #mdb_env_set_mapsize().
464 #define DEFAULT_MAPSIZE 1048576
466 /** @defgroup readers Reader Lock Table
467 * Readers don't acquire any locks for their data access. Instead, they
468 * simply record their transaction ID in the reader table. The reader
469 * mutex is needed just to find an empty slot in the reader table. The
470 * slot's address is saved in thread-specific data so that subsequent read
471 * transactions started by the same thread need no further locking to proceed.
473 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
475 * No reader table is used if the database is on a read-only filesystem, or
476 * if #MDB_NOLOCK is set.
478 * Since the database uses multi-version concurrency control, readers don't
479 * actually need any locking. This table is used to keep track of which
480 * readers are using data from which old transactions, so that we'll know
481 * when a particular old transaction is no longer in use. Old transactions
482 * that have discarded any data pages can then have those pages reclaimed
483 * for use by a later write transaction.
485 * The lock table is constructed such that reader slots are aligned with the
486 * processor's cache line size. Any slot is only ever used by one thread.
487 * This alignment guarantees that there will be no contention or cache
488 * thrashing as threads update their own slot info, and also eliminates
489 * any need for locking when accessing a slot.
491 * A writer thread will scan every slot in the table to determine the oldest
492 * outstanding reader transaction. Any freed pages older than this will be
493 * reclaimed by the writer. The writer doesn't use any locks when scanning
494 * this table. This means that there's no guarantee that the writer will
495 * see the most up-to-date reader info, but that's not required for correct
496 * operation - all we need is to know the upper bound on the oldest reader,
497 * we don't care at all about the newest reader. So the only consequence of
498 * reading stale information here is that old pages might hang around a
499 * while longer before being reclaimed. That's actually good anyway, because
500 * the longer we delay reclaiming old pages, the more likely it is that a
501 * string of contiguous pages can be found after coalescing old pages from
502 * many old transactions together.
505 /** Number of slots in the reader table.
506 * This value was chosen somewhat arbitrarily. 126 readers plus a
507 * couple mutexes fit exactly into 8KB on my development machine.
508 * Applications should set the table size using #mdb_env_set_maxreaders().
510 #define DEFAULT_READERS 126
512 /** The size of a CPU cache line in bytes. We want our lock structures
513 * aligned to this size to avoid false cache line sharing in the
515 * This value works for most CPUs. For Itanium this should be 128.
521 /** The information we store in a single slot of the reader table.
522 * In addition to a transaction ID, we also record the process and
523 * thread ID that owns a slot, so that we can detect stale information,
524 * e.g. threads or processes that went away without cleaning up.
525 * @note We currently don't check for stale records. We simply re-init
526 * the table when we know that we're the only process opening the
529 typedef struct MDB_rxbody {
530 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
531 * Multiple readers that start at the same time will probably have the
532 * same ID here. Again, it's not important to exclude them from
533 * anything; all we need to know is which version of the DB they
534 * started from so we can avoid overwriting any data used in that
535 * particular version.
538 /** The process ID of the process owning this reader txn. */
540 /** The thread ID of the thread owning this txn. */
544 /** The actual reader record, with cacheline padding. */
545 typedef struct MDB_reader {
548 /** shorthand for mrb_txnid */
549 #define mr_txnid mru.mrx.mrb_txnid
550 #define mr_pid mru.mrx.mrb_pid
551 #define mr_tid mru.mrx.mrb_tid
552 /** cache line alignment */
553 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
557 /** The header for the reader table.
558 * The table resides in a memory-mapped file. (This is a different file
559 * than is used for the main database.)
561 * For POSIX the actual mutexes reside in the shared memory of this
562 * mapped file. On Windows, mutexes are named objects allocated by the
563 * kernel; we store the mutex names in this mapped file so that other
564 * processes can grab them. This same approach is also used on
565 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
566 * process-shared POSIX mutexes. For these cases where a named object
567 * is used, the object name is derived from a 64 bit FNV hash of the
568 * environment pathname. As such, naming collisions are extremely
569 * unlikely. If a collision occurs, the results are unpredictable.
571 typedef struct MDB_txbody {
572 /** Stamp identifying this as an MDB file. It must be set
575 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
577 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
578 char mtb_rmname[MNAME_LEN];
580 /** Mutex protecting access to this table.
581 * This is the reader lock that #LOCK_MUTEX_R acquires.
583 pthread_mutex_t mtb_mutex;
585 /** The ID of the last transaction committed to the database.
586 * This is recorded here only for convenience; the value can always
587 * be determined by reading the main database meta pages.
590 /** The number of slots that have been used in the reader table.
591 * This always records the maximum count, it is not decremented
592 * when readers release their slots.
594 unsigned mtb_numreaders;
597 /** The actual reader table definition. */
598 typedef struct MDB_txninfo {
601 #define mti_magic mt1.mtb.mtb_magic
602 #define mti_format mt1.mtb.mtb_format
603 #define mti_mutex mt1.mtb.mtb_mutex
604 #define mti_rmname mt1.mtb.mtb_rmname
605 #define mti_txnid mt1.mtb.mtb_txnid
606 #define mti_numreaders mt1.mtb.mtb_numreaders
607 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
610 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
611 char mt2_wmname[MNAME_LEN];
612 #define mti_wmname mt2.mt2_wmname
614 pthread_mutex_t mt2_wmutex;
615 #define mti_wmutex mt2.mt2_wmutex
617 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
619 MDB_reader mti_readers[1];
622 /** Lockfile format signature: version, features and field layout */
623 #define MDB_LOCK_FORMAT \
625 ((MDB_LOCK_VERSION) \
626 /* Flags which describe functionality */ \
627 + (((MDB_PIDLOCK) != 0) << 16)))
630 /** Common header for all page types.
631 * Overflow records occupy a number of contiguous pages with no
632 * headers on any page after the first.
634 typedef struct MDB_page {
635 #define mp_pgno mp_p.p_pgno
636 #define mp_next mp_p.p_next
638 pgno_t p_pgno; /**< page number */
639 void * p_next; /**< for in-memory list of freed structs */
642 /** @defgroup mdb_page Page Flags
644 * Flags for the page headers.
647 #define P_BRANCH 0x01 /**< branch page */
648 #define P_LEAF 0x02 /**< leaf page */
649 #define P_OVERFLOW 0x04 /**< overflow page */
650 #define P_META 0x08 /**< meta page */
651 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
652 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
653 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
654 #define P_KEEP 0x8000 /**< leave this page alone during spill */
656 uint16_t mp_flags; /**< @ref mdb_page */
657 #define mp_lower mp_pb.pb.pb_lower
658 #define mp_upper mp_pb.pb.pb_upper
659 #define mp_pages mp_pb.pb_pages
662 indx_t pb_lower; /**< lower bound of free space */
663 indx_t pb_upper; /**< upper bound of free space */
665 uint32_t pb_pages; /**< number of overflow pages */
667 indx_t mp_ptrs[1]; /**< dynamic size */
670 /** Size of the page header, excluding dynamic data at the end */
671 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
673 /** Address of first usable data byte in a page, after the header */
674 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
676 /** Number of nodes on a page */
677 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
679 /** The amount of space remaining in the page */
680 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
682 /** The percentage of space used in the page, in tenths of a percent. */
683 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
684 ((env)->me_psize - PAGEHDRSZ))
685 /** The minimum page fill factor, in tenths of a percent.
686 * Pages emptier than this are candidates for merging.
688 #define FILL_THRESHOLD 250
690 /** Test if a page is a leaf page */
691 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
692 /** Test if a page is a LEAF2 page */
693 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
694 /** Test if a page is a branch page */
695 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
696 /** Test if a page is an overflow page */
697 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
698 /** Test if a page is a sub page */
699 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
701 /** The number of overflow pages needed to store the given size. */
702 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
704 /** Header for a single key/data pair within a page.
705 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
706 * We guarantee 2-byte alignment for 'MDB_node's.
708 typedef struct MDB_node {
709 /** lo and hi are used for data size on leaf nodes and for
710 * child pgno on branch nodes. On 64 bit platforms, flags
711 * is also used for pgno. (Branch nodes have no flags).
712 * They are in host byte order in case that lets some
713 * accesses be optimized into a 32-bit word access.
715 #if BYTE_ORDER == LITTLE_ENDIAN
716 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
718 unsigned short mn_hi, mn_lo;
720 /** @defgroup mdb_node Node Flags
722 * Flags for node headers.
725 #define F_BIGDATA 0x01 /**< data put on overflow page */
726 #define F_SUBDATA 0x02 /**< data is a sub-database */
727 #define F_DUPDATA 0x04 /**< data has duplicates */
729 /** valid flags for #mdb_node_add() */
730 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
733 unsigned short mn_flags; /**< @ref mdb_node */
734 unsigned short mn_ksize; /**< key size */
735 char mn_data[1]; /**< key and data are appended here */
738 /** Size of the node header, excluding dynamic data at the end */
739 #define NODESIZE offsetof(MDB_node, mn_data)
741 /** Bit position of top word in page number, for shifting mn_flags */
742 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
744 /** Size of a node in a branch page with a given key.
745 * This is just the node header plus the key, there is no data.
747 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
749 /** Size of a node in a leaf page with a given key and data.
750 * This is node header plus key plus data size.
752 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
754 /** Address of node \b i in page \b p */
755 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
757 /** Address of the key for the node */
758 #define NODEKEY(node) (void *)((node)->mn_data)
760 /** Address of the data for a node */
761 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
763 /** Get the page number pointed to by a branch node */
764 #define NODEPGNO(node) \
765 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
766 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
767 /** Set the page number in a branch node */
768 #define SETPGNO(node,pgno) do { \
769 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
770 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
772 /** Get the size of the data in a leaf node */
773 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
774 /** Set the size of the data for a leaf node */
775 #define SETDSZ(node,size) do { \
776 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
777 /** The size of a key in a node */
778 #define NODEKSZ(node) ((node)->mn_ksize)
780 /** Copy a page number from src to dst */
782 #define COPY_PGNO(dst,src) dst = src
784 #if SIZE_MAX > 4294967295UL
785 #define COPY_PGNO(dst,src) do { \
786 unsigned short *s, *d; \
787 s = (unsigned short *)&(src); \
788 d = (unsigned short *)&(dst); \
795 #define COPY_PGNO(dst,src) do { \
796 unsigned short *s, *d; \
797 s = (unsigned short *)&(src); \
798 d = (unsigned short *)&(dst); \
804 /** The address of a key in a LEAF2 page.
805 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
806 * There are no node headers, keys are stored contiguously.
808 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
810 /** Set the \b node's key into \b keyptr, if requested. */
811 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
812 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
814 /** Set the \b node's key into \b key. */
815 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
817 /** Information about a single database in the environment. */
818 typedef struct MDB_db {
819 uint32_t md_pad; /**< also ksize for LEAF2 pages */
820 uint16_t md_flags; /**< @ref mdb_dbi_open */
821 uint16_t md_depth; /**< depth of this tree */
822 pgno_t md_branch_pages; /**< number of internal pages */
823 pgno_t md_leaf_pages; /**< number of leaf pages */
824 pgno_t md_overflow_pages; /**< number of overflow pages */
825 size_t md_entries; /**< number of data items */
826 pgno_t md_root; /**< the root page of this tree */
829 /** mdb_dbi_open flags */
830 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
831 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
832 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
833 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
835 /** Handle for the DB used to track free pages. */
837 /** Handle for the default DB. */
840 /** Meta page content.
841 * A meta page is the start point for accessing a database snapshot.
842 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
844 typedef struct MDB_meta {
845 /** Stamp identifying this as an MDB file. It must be set
848 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
850 void *mm_address; /**< address for fixed mapping */
851 size_t mm_mapsize; /**< size of mmap region */
852 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
853 /** The size of pages used in this DB */
854 #define mm_psize mm_dbs[0].md_pad
855 /** Any persistent environment flags. @ref mdb_env */
856 #define mm_flags mm_dbs[0].md_flags
857 pgno_t mm_last_pg; /**< last used page in file */
858 txnid_t mm_txnid; /**< txnid that committed this page */
861 /** Buffer for a stack-allocated meta page.
862 * The members define size and alignment, and silence type
863 * aliasing warnings. They are not used directly; that could
864 * mean incorrectly using several union members in parallel.
866 typedef union MDB_metabuf {
869 char mm_pad[PAGEHDRSZ];
874 /** Auxiliary DB info.
875 * The information here is mostly static/read-only. There is
876 * only a single copy of this record in the environment.
878 typedef struct MDB_dbx {
879 MDB_val md_name; /**< name of the database */
880 MDB_cmp_func *md_cmp; /**< function for comparing keys */
881 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
882 MDB_rel_func *md_rel; /**< user relocate function */
883 void *md_relctx; /**< user-provided context for md_rel */
886 /** A database transaction.
887 * Every operation requires a transaction handle.
890 MDB_txn *mt_parent; /**< parent of a nested txn */
891 MDB_txn *mt_child; /**< nested txn under this txn */
892 pgno_t mt_next_pgno; /**< next unallocated page */
893 /** The ID of this transaction. IDs are integers incrementing from 1.
894 * Only committed write transactions increment the ID. If a transaction
895 * aborts, the ID may be re-used by the next writer.
898 MDB_env *mt_env; /**< the DB environment */
899 /** The list of pages that became unused during this transaction.
902 /** The sorted list of dirty pages we temporarily wrote to disk
903 * because the dirty list was full. page numbers in here are
904 * shifted left by 1, deleted slots have the LSB set.
906 MDB_IDL mt_spill_pgs;
908 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
910 /** For read txns: This thread/txn's reader table slot, or NULL. */
913 /** Array of records for each DB known in the environment. */
915 /** Array of MDB_db records for each known DB */
917 /** @defgroup mt_dbflag Transaction DB Flags
921 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
922 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
923 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
924 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
926 /** In write txns, array of cursors for each DB */
927 MDB_cursor **mt_cursors;
928 /** Array of flags for each DB */
929 unsigned char *mt_dbflags;
930 /** Number of DB records in use. This number only ever increments;
931 * we don't decrement it when individual DB handles are closed.
935 /** @defgroup mdb_txn Transaction Flags
939 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
940 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
941 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
942 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
944 unsigned int mt_flags; /**< @ref mdb_txn */
945 /** dirty_list room: Array size - #dirty pages visible to this txn.
946 * Includes ancestor txns' dirty pages not hidden by other txns'
947 * dirty/spilled pages. Thus commit(nested txn) has room to merge
948 * dirty_list into mt_parent after freeing hidden mt_parent pages.
950 unsigned int mt_dirty_room;
953 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
954 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
955 * raise this on a 64 bit machine.
957 #define CURSOR_STACK 32
961 /** Cursors are used for all DB operations.
962 * A cursor holds a path of (page pointer, key index) from the DB
963 * root to a position in the DB, plus other state. #MDB_DUPSORT
964 * cursors include an xcursor to the current data item. Write txns
965 * track their cursors and keep them up to date when data moves.
966 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
967 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
970 /** Next cursor on this DB in this txn */
972 /** Backup of the original cursor if this cursor is a shadow */
973 MDB_cursor *mc_backup;
974 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
975 struct MDB_xcursor *mc_xcursor;
976 /** The transaction that owns this cursor */
978 /** The database handle this cursor operates on */
980 /** The database record for this cursor */
982 /** The database auxiliary record for this cursor */
984 /** The @ref mt_dbflag for this database */
985 unsigned char *mc_dbflag;
986 unsigned short mc_snum; /**< number of pushed pages */
987 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
988 /** @defgroup mdb_cursor Cursor Flags
990 * Cursor state flags.
993 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
994 #define C_EOF 0x02 /**< No more data */
995 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
996 #define C_DEL 0x08 /**< last op was a cursor_del */
997 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
998 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1000 unsigned int mc_flags; /**< @ref mdb_cursor */
1001 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1002 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1005 /** Context for sorted-dup records.
1006 * We could have gone to a fully recursive design, with arbitrarily
1007 * deep nesting of sub-databases. But for now we only handle these
1008 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1010 typedef struct MDB_xcursor {
1011 /** A sub-cursor for traversing the Dup DB */
1012 MDB_cursor mx_cursor;
1013 /** The database record for this Dup DB */
1015 /** The auxiliary DB record for this Dup DB */
1017 /** The @ref mt_dbflag for this Dup DB */
1018 unsigned char mx_dbflag;
1021 /** State of FreeDB old pages, stored in the MDB_env */
1022 typedef struct MDB_pgstate {
1023 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1024 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1027 /** The database environment. */
1029 HANDLE me_fd; /**< The main data file */
1030 HANDLE me_lfd; /**< The lock file */
1031 HANDLE me_mfd; /**< just for writing the meta pages */
1032 /** Failed to update the meta page. Probably an I/O error. */
1033 #define MDB_FATAL_ERROR 0x80000000U
1034 /** Some fields are initialized. */
1035 #define MDB_ENV_ACTIVE 0x20000000U
1036 /** me_txkey is set */
1037 #define MDB_ENV_TXKEY 0x10000000U
1038 /** Have liveness lock in reader table */
1039 #define MDB_LIVE_READER 0x08000000U
1040 uint32_t me_flags; /**< @ref mdb_env */
1041 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1042 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1043 unsigned int me_maxreaders; /**< size of the reader table */
1044 unsigned int me_numreaders; /**< max numreaders set by this env */
1045 MDB_dbi me_numdbs; /**< number of DBs opened */
1046 MDB_dbi me_maxdbs; /**< size of the DB table */
1047 MDB_PID_T me_pid; /**< process ID of this env */
1048 char *me_path; /**< path to the DB files */
1049 char *me_map; /**< the memory map of the data file */
1050 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1051 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1052 void *me_pbuf; /**< scratch area for DUPSORT put() */
1053 MDB_txn *me_txn; /**< current write transaction */
1054 size_t me_mapsize; /**< size of the data memory map */
1055 off_t me_size; /**< current file size */
1056 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1057 MDB_dbx *me_dbxs; /**< array of static DB info */
1058 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1059 pthread_key_t me_txkey; /**< thread-key for readers */
1060 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1061 # define me_pglast me_pgstate.mf_pglast
1062 # define me_pghead me_pgstate.mf_pghead
1063 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1064 /** IDL of pages that became unused in a write txn */
1065 MDB_IDL me_free_pgs;
1066 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1067 MDB_ID2L me_dirty_list;
1068 /** Max number of freelist items that can fit in a single overflow page */
1070 /** Max size of a node on a page */
1071 unsigned int me_nodemax;
1072 #if !(MDB_MAXKEYSIZE)
1073 unsigned int me_maxkey; /**< max size of a key */
1076 int me_pidquery; /**< Used in OpenProcess */
1077 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1079 #elif defined(MDB_USE_POSIX_SEM)
1080 sem_t *me_rmutex; /* Shared mutexes are not supported */
1083 void *me_userctx; /**< User-settable context */
1084 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1087 /** Nested transaction */
1088 typedef struct MDB_ntxn {
1089 MDB_txn mnt_txn; /**< the transaction */
1090 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1093 /** max number of pages to commit in one writev() call */
1094 #define MDB_COMMIT_PAGES 64
1095 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1096 #undef MDB_COMMIT_PAGES
1097 #define MDB_COMMIT_PAGES IOV_MAX
1100 /* max bytes to write in one call */
1101 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1103 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1104 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1105 static int mdb_page_touch(MDB_cursor *mc);
1107 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1108 static int mdb_page_search_root(MDB_cursor *mc,
1109 MDB_val *key, int modify);
1110 #define MDB_PS_MODIFY 1
1111 #define MDB_PS_ROOTONLY 2
1112 #define MDB_PS_FIRST 4
1113 #define MDB_PS_LAST 8
1114 static int mdb_page_search(MDB_cursor *mc,
1115 MDB_val *key, int flags);
1116 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1118 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1119 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1120 pgno_t newpgno, unsigned int nflags);
1122 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1123 static int mdb_env_pick_meta(const MDB_env *env);
1124 static int mdb_env_write_meta(MDB_txn *txn);
1125 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1126 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1128 static void mdb_env_close0(MDB_env *env, int excl);
1130 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1131 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1132 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1133 static void mdb_node_del(MDB_cursor *mc, int ksize);
1134 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1135 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1136 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1137 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1138 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1140 static int mdb_rebalance(MDB_cursor *mc);
1141 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1143 static void mdb_cursor_pop(MDB_cursor *mc);
1144 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1146 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1147 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1148 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1149 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1150 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1152 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1153 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1155 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1156 static void mdb_xcursor_init0(MDB_cursor *mc);
1157 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1159 static int mdb_drop0(MDB_cursor *mc, int subs);
1160 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1163 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1167 static SECURITY_DESCRIPTOR mdb_null_sd;
1168 static SECURITY_ATTRIBUTES mdb_all_sa;
1169 static int mdb_sec_inited;
1172 /** Return the library version info. */
1174 mdb_version(int *major, int *minor, int *patch)
1176 if (major) *major = MDB_VERSION_MAJOR;
1177 if (minor) *minor = MDB_VERSION_MINOR;
1178 if (patch) *patch = MDB_VERSION_PATCH;
1179 return MDB_VERSION_STRING;
1182 /** Table of descriptions for MDB @ref errors */
1183 static char *const mdb_errstr[] = {
1184 "MDB_KEYEXIST: Key/data pair already exists",
1185 "MDB_NOTFOUND: No matching key/data pair found",
1186 "MDB_PAGE_NOTFOUND: Requested page not found",
1187 "MDB_CORRUPTED: Located page was wrong type",
1188 "MDB_PANIC: Update of meta page failed",
1189 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1190 "MDB_INVALID: File is not an MDB file",
1191 "MDB_MAP_FULL: Environment mapsize limit reached",
1192 "MDB_DBS_FULL: Environment maxdbs limit reached",
1193 "MDB_READERS_FULL: Environment maxreaders limit reached",
1194 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1195 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1196 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1197 "MDB_PAGE_FULL: Internal error - page has no more space",
1198 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1199 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1200 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1201 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1202 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1206 mdb_strerror(int err)
1210 return ("Successful return: 0");
1212 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1213 i = err - MDB_KEYEXIST;
1214 return mdb_errstr[i];
1217 return strerror(err);
1220 /** assert(3) variant in cursor context */
1221 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1222 /** assert(3) variant in transaction context */
1223 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1224 /** assert(3) variant in environment context */
1225 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1228 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1229 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1232 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1233 const char *func, const char *file, int line)
1236 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1237 file, line, expr_txt, func);
1238 if (env->me_assert_func)
1239 env->me_assert_func(env, buf);
1240 fprintf(stderr, "%s\n", buf);
1244 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1248 /** Return the page number of \b mp which may be sub-page, for debug output */
1250 mdb_dbg_pgno(MDB_page *mp)
1253 COPY_PGNO(ret, mp->mp_pgno);
1257 /** Display a key in hexadecimal and return the address of the result.
1258 * @param[in] key the key to display
1259 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1260 * @return The key in hexadecimal form.
1263 mdb_dkey(MDB_val *key, char *buf)
1266 unsigned char *c = key->mv_data;
1272 if (key->mv_size > DKBUF_MAXKEYSIZE)
1273 return "MDB_MAXKEYSIZE";
1274 /* may want to make this a dynamic check: if the key is mostly
1275 * printable characters, print it as-is instead of converting to hex.
1279 for (i=0; i<key->mv_size; i++)
1280 ptr += sprintf(ptr, "%02x", *c++);
1282 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1287 /** Display all the keys in the page. */
1289 mdb_page_list(MDB_page *mp)
1292 unsigned int i, nkeys, nsize, total = 0;
1296 nkeys = NUMKEYS(mp);
1297 fprintf(stderr, "Page %"Z"u numkeys %d\n", mdb_dbg_pgno(mp), nkeys);
1298 for (i=0; i<nkeys; i++) {
1299 node = NODEPTR(mp, i);
1300 key.mv_size = node->mn_ksize;
1301 key.mv_data = node->mn_data;
1302 nsize = NODESIZE + key.mv_size;
1303 if (IS_BRANCH(mp)) {
1304 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1308 if (F_ISSET(node->mn_flags, F_BIGDATA))
1309 nsize += sizeof(pgno_t);
1311 nsize += NODEDSZ(node);
1313 nsize += sizeof(indx_t);
1314 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1316 total = EVEN(total);
1318 fprintf(stderr, "Total: %d\n", total);
1322 mdb_cursor_chk(MDB_cursor *mc)
1328 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1329 for (i=0; i<mc->mc_top; i++) {
1331 node = NODEPTR(mp, mc->mc_ki[i]);
1332 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1335 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1341 /** Count all the pages in each DB and in the freelist
1342 * and make sure it matches the actual number of pages
1345 static void mdb_audit(MDB_txn *txn)
1349 MDB_ID freecount, count;
1354 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1355 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1356 freecount += *(MDB_ID *)data.mv_data;
1359 for (i = 0; i<txn->mt_numdbs; i++) {
1361 mdb_cursor_init(&mc, txn, i, &mx);
1362 if (txn->mt_dbs[i].md_root == P_INVALID)
1364 count += txn->mt_dbs[i].md_branch_pages +
1365 txn->mt_dbs[i].md_leaf_pages +
1366 txn->mt_dbs[i].md_overflow_pages;
1367 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1368 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1372 mp = mc.mc_pg[mc.mc_top];
1373 for (j=0; j<NUMKEYS(mp); j++) {
1374 MDB_node *leaf = NODEPTR(mp, j);
1375 if (leaf->mn_flags & F_SUBDATA) {
1377 memcpy(&db, NODEDATA(leaf), sizeof(db));
1378 count += db.md_branch_pages + db.md_leaf_pages +
1379 db.md_overflow_pages;
1383 while (mdb_cursor_sibling(&mc, 1) == 0);
1386 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1387 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1388 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1394 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1396 return txn->mt_dbxs[dbi].md_cmp(a, b);
1400 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1402 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1405 /** Allocate memory for a page.
1406 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1409 mdb_page_malloc(MDB_txn *txn, unsigned num)
1411 MDB_env *env = txn->mt_env;
1412 MDB_page *ret = env->me_dpages;
1413 size_t psize = env->me_psize, sz = psize, off;
1414 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1415 * For a single page alloc, we init everything after the page header.
1416 * For multi-page, we init the final page; if the caller needed that
1417 * many pages they will be filling in at least up to the last page.
1421 VGMEMP_ALLOC(env, ret, sz);
1422 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1423 env->me_dpages = ret->mp_next;
1426 psize -= off = PAGEHDRSZ;
1431 if ((ret = malloc(sz)) != NULL) {
1432 VGMEMP_ALLOC(env, ret, sz);
1433 if (!(env->me_flags & MDB_NOMEMINIT)) {
1434 memset((char *)ret + off, 0, psize);
1438 txn->mt_flags |= MDB_TXN_ERROR;
1443 /** Free a single page.
1444 * Saves single pages to a list, for future reuse.
1445 * (This is not used for multi-page overflow pages.)
1448 mdb_page_free(MDB_env *env, MDB_page *mp)
1450 mp->mp_next = env->me_dpages;
1451 VGMEMP_FREE(env, mp);
1452 env->me_dpages = mp;
1455 /** Free a dirty page */
1457 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1459 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1460 mdb_page_free(env, dp);
1462 /* large pages just get freed directly */
1463 VGMEMP_FREE(env, dp);
1468 /** Return all dirty pages to dpage list */
1470 mdb_dlist_free(MDB_txn *txn)
1472 MDB_env *env = txn->mt_env;
1473 MDB_ID2L dl = txn->mt_u.dirty_list;
1474 unsigned i, n = dl[0].mid;
1476 for (i = 1; i <= n; i++) {
1477 mdb_dpage_free(env, dl[i].mptr);
1482 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1483 * @param[in] mc A cursor handle for the current operation.
1484 * @param[in] pflags Flags of the pages to update:
1485 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1486 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1487 * @return 0 on success, non-zero on failure.
1490 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1492 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1493 MDB_txn *txn = mc->mc_txn;
1499 int rc = MDB_SUCCESS, level;
1501 /* Mark pages seen by cursors */
1502 if (mc->mc_flags & C_UNTRACK)
1503 mc = NULL; /* will find mc in mt_cursors */
1504 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1505 for (; mc; mc=mc->mc_next) {
1506 if (!(mc->mc_flags & C_INITIALIZED))
1508 for (m3 = mc;; m3 = &mx->mx_cursor) {
1510 for (j=0; j<m3->mc_snum; j++) {
1512 if ((mp->mp_flags & Mask) == pflags)
1513 mp->mp_flags ^= P_KEEP;
1515 mx = m3->mc_xcursor;
1516 /* Proceed to mx if it is at a sub-database */
1517 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1519 if (! (mp && (mp->mp_flags & P_LEAF)))
1521 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1522 if (!(leaf->mn_flags & F_SUBDATA))
1531 /* Mark dirty root pages */
1532 for (i=0; i<txn->mt_numdbs; i++) {
1533 if (txn->mt_dbflags[i] & DB_DIRTY) {
1534 pgno_t pgno = txn->mt_dbs[i].md_root;
1535 if (pgno == P_INVALID)
1537 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1539 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1540 dp->mp_flags ^= P_KEEP;
1548 static int mdb_page_flush(MDB_txn *txn, int keep);
1550 /** Spill pages from the dirty list back to disk.
1551 * This is intended to prevent running into #MDB_TXN_FULL situations,
1552 * but note that they may still occur in a few cases:
1553 * 1) our estimate of the txn size could be too small. Currently this
1554 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1555 * 2) child txns may run out of space if their parents dirtied a
1556 * lot of pages and never spilled them. TODO: we probably should do
1557 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1558 * the parent's dirty_room is below a given threshold.
1560 * Otherwise, if not using nested txns, it is expected that apps will
1561 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1562 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1563 * If the txn never references them again, they can be left alone.
1564 * If the txn only reads them, they can be used without any fuss.
1565 * If the txn writes them again, they can be dirtied immediately without
1566 * going thru all of the work of #mdb_page_touch(). Such references are
1567 * handled by #mdb_page_unspill().
1569 * Also note, we never spill DB root pages, nor pages of active cursors,
1570 * because we'll need these back again soon anyway. And in nested txns,
1571 * we can't spill a page in a child txn if it was already spilled in a
1572 * parent txn. That would alter the parent txns' data even though
1573 * the child hasn't committed yet, and we'd have no way to undo it if
1574 * the child aborted.
1576 * @param[in] m0 cursor A cursor handle identifying the transaction and
1577 * database for which we are checking space.
1578 * @param[in] key For a put operation, the key being stored.
1579 * @param[in] data For a put operation, the data being stored.
1580 * @return 0 on success, non-zero on failure.
1583 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1585 MDB_txn *txn = m0->mc_txn;
1587 MDB_ID2L dl = txn->mt_u.dirty_list;
1588 unsigned int i, j, need;
1591 if (m0->mc_flags & C_SUB)
1594 /* Estimate how much space this op will take */
1595 i = m0->mc_db->md_depth;
1596 /* Named DBs also dirty the main DB */
1597 if (m0->mc_dbi > MAIN_DBI)
1598 i += txn->mt_dbs[MAIN_DBI].md_depth;
1599 /* For puts, roughly factor in the key+data size */
1601 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1602 i += i; /* double it for good measure */
1605 if (txn->mt_dirty_room > i)
1608 if (!txn->mt_spill_pgs) {
1609 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1610 if (!txn->mt_spill_pgs)
1613 /* purge deleted slots */
1614 MDB_IDL sl = txn->mt_spill_pgs;
1615 unsigned int num = sl[0];
1617 for (i=1; i<=num; i++) {
1624 /* Preserve pages which may soon be dirtied again */
1625 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1628 /* Less aggressive spill - we originally spilled the entire dirty list,
1629 * with a few exceptions for cursor pages and DB root pages. But this
1630 * turns out to be a lot of wasted effort because in a large txn many
1631 * of those pages will need to be used again. So now we spill only 1/8th
1632 * of the dirty pages. Testing revealed this to be a good tradeoff,
1633 * better than 1/2, 1/4, or 1/10.
1635 if (need < MDB_IDL_UM_MAX / 8)
1636 need = MDB_IDL_UM_MAX / 8;
1638 /* Save the page IDs of all the pages we're flushing */
1639 /* flush from the tail forward, this saves a lot of shifting later on. */
1640 for (i=dl[0].mid; i && need; i--) {
1641 MDB_ID pn = dl[i].mid << 1;
1643 if (dp->mp_flags & P_KEEP)
1645 /* Can't spill twice, make sure it's not already in a parent's
1648 if (txn->mt_parent) {
1650 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1651 if (tx2->mt_spill_pgs) {
1652 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1653 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1654 dp->mp_flags |= P_KEEP;
1662 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1666 mdb_midl_sort(txn->mt_spill_pgs);
1668 /* Flush the spilled part of dirty list */
1669 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1672 /* Reset any dirty pages we kept that page_flush didn't see */
1673 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1676 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1680 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1682 mdb_find_oldest(MDB_txn *txn)
1685 txnid_t mr, oldest = txn->mt_txnid - 1;
1686 if (txn->mt_env->me_txns) {
1687 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1688 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1699 /** Add a page to the txn's dirty list */
1701 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1704 int (*insert)(MDB_ID2L, MDB_ID2 *);
1706 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1707 insert = mdb_mid2l_append;
1709 insert = mdb_mid2l_insert;
1711 mid.mid = mp->mp_pgno;
1713 insert(txn->mt_u.dirty_list, &mid);
1714 txn->mt_dirty_room--;
1717 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1718 * me_pghead and mt_next_pgno.
1720 * If there are free pages available from older transactions, they
1721 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1722 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1723 * and move me_pglast to say which records were consumed. Only this
1724 * function can create me_pghead and move me_pglast/mt_next_pgno.
1725 * @param[in] mc cursor A cursor handle identifying the transaction and
1726 * database for which we are allocating.
1727 * @param[in] num the number of pages to allocate.
1728 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1729 * will always be satisfied by a single contiguous chunk of memory.
1730 * @return 0 on success, non-zero on failure.
1733 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1735 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1736 /* Get at most <Max_retries> more freeDB records once me_pghead
1737 * has enough pages. If not enough, use new pages from the map.
1738 * If <Paranoid> and mc is updating the freeDB, only get new
1739 * records if me_pghead is empty. Then the freelist cannot play
1740 * catch-up with itself by growing while trying to save it.
1742 enum { Paranoid = 1, Max_retries = 500 };
1744 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1746 int rc, retry = Max_retries;
1747 MDB_txn *txn = mc->mc_txn;
1748 MDB_env *env = txn->mt_env;
1749 pgno_t pgno, *mop = env->me_pghead;
1750 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1752 txnid_t oldest = 0, last;
1758 /* If our dirty list is already full, we can't do anything */
1759 if (txn->mt_dirty_room == 0) {
1764 for (op = MDB_FIRST;; op = MDB_NEXT) {
1767 pgno_t *idl, old_id, new_id;
1769 /* Seek a big enough contiguous page range. Prefer
1770 * pages at the tail, just truncating the list.
1776 if (mop[i-n2] == pgno+n2)
1779 if (Max_retries < INT_MAX && --retry < 0)
1783 if (op == MDB_FIRST) { /* 1st iteration */
1784 /* Prepare to fetch more and coalesce */
1785 oldest = mdb_find_oldest(txn);
1786 last = env->me_pglast;
1787 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1790 key.mv_data = &last; /* will look up last+1 */
1791 key.mv_size = sizeof(last);
1793 if (Paranoid && mc->mc_dbi == FREE_DBI)
1796 if (Paranoid && retry < 0 && mop_len)
1800 /* Do not fetch more if the record will be too recent */
1803 rc = mdb_cursor_get(&m2, &key, NULL, op);
1805 if (rc == MDB_NOTFOUND)
1809 last = *(txnid_t*)key.mv_data;
1812 np = m2.mc_pg[m2.mc_top];
1813 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1814 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1817 idl = (MDB_ID *) data.mv_data;
1820 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1825 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1827 mop = env->me_pghead;
1829 env->me_pglast = last;
1831 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1832 last, txn->mt_dbs[FREE_DBI].md_root, i));
1834 DPRINTF(("IDL %"Z"u", idl[k]));
1836 /* Merge in descending sorted order */
1839 mop[0] = (pgno_t)-1;
1843 for (; old_id < new_id; old_id = mop[--j])
1850 /* Use new pages from the map when nothing suitable in the freeDB */
1852 pgno = txn->mt_next_pgno;
1853 if (pgno + num >= env->me_maxpg) {
1854 DPUTS("DB size maxed out");
1860 if (env->me_flags & MDB_WRITEMAP) {
1861 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1863 if (!(np = mdb_page_malloc(txn, num))) {
1869 mop[0] = mop_len -= num;
1870 /* Move any stragglers down */
1871 for (j = i-num; j < mop_len; )
1872 mop[++j] = mop[++i];
1874 txn->mt_next_pgno = pgno + num;
1877 mdb_page_dirty(txn, np);
1883 txn->mt_flags |= MDB_TXN_ERROR;
1887 /** Copy the used portions of a non-overflow page.
1888 * @param[in] dst page to copy into
1889 * @param[in] src page to copy from
1890 * @param[in] psize size of a page
1893 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1895 enum { Align = sizeof(pgno_t) };
1896 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1898 /* If page isn't full, just copy the used portion. Adjust
1899 * alignment so memcpy may copy words instead of bytes.
1901 if ((unused &= -Align) && !IS_LEAF2(src)) {
1903 memcpy(dst, src, (lower + (Align-1)) & -Align);
1904 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1907 memcpy(dst, src, psize - unused);
1911 /** Pull a page off the txn's spill list, if present.
1912 * If a page being referenced was spilled to disk in this txn, bring
1913 * it back and make it dirty/writable again.
1914 * @param[in] txn the transaction handle.
1915 * @param[in] mp the page being referenced. It must not be dirty.
1916 * @param[out] ret the writable page, if any. ret is unchanged if
1917 * mp wasn't spilled.
1920 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1922 MDB_env *env = txn->mt_env;
1925 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1927 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1928 if (!tx2->mt_spill_pgs)
1930 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1931 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1934 if (txn->mt_dirty_room == 0)
1935 return MDB_TXN_FULL;
1936 if (IS_OVERFLOW(mp))
1940 if (env->me_flags & MDB_WRITEMAP) {
1943 np = mdb_page_malloc(txn, num);
1947 memcpy(np, mp, num * env->me_psize);
1949 mdb_page_copy(np, mp, env->me_psize);
1952 /* If in current txn, this page is no longer spilled.
1953 * If it happens to be the last page, truncate the spill list.
1954 * Otherwise mark it as deleted by setting the LSB.
1956 if (x == txn->mt_spill_pgs[0])
1957 txn->mt_spill_pgs[0]--;
1959 txn->mt_spill_pgs[x] |= 1;
1960 } /* otherwise, if belonging to a parent txn, the
1961 * page remains spilled until child commits
1964 mdb_page_dirty(txn, np);
1965 np->mp_flags |= P_DIRTY;
1973 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1974 * @param[in] mc cursor pointing to the page to be touched
1975 * @return 0 on success, non-zero on failure.
1978 mdb_page_touch(MDB_cursor *mc)
1980 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1981 MDB_txn *txn = mc->mc_txn;
1982 MDB_cursor *m2, *m3;
1986 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1987 if (txn->mt_flags & MDB_TXN_SPILLS) {
1989 rc = mdb_page_unspill(txn, mp, &np);
1995 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1996 (rc = mdb_page_alloc(mc, 1, &np)))
1999 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2000 mp->mp_pgno, pgno));
2001 mdb_cassert(mc, mp->mp_pgno != pgno);
2002 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2003 /* Update the parent page, if any, to point to the new page */
2005 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2006 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2007 SETPGNO(node, pgno);
2009 mc->mc_db->md_root = pgno;
2011 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2012 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2014 /* If txn has a parent, make sure the page is in our
2018 unsigned x = mdb_mid2l_search(dl, pgno);
2019 if (x <= dl[0].mid && dl[x].mid == pgno) {
2020 if (mp != dl[x].mptr) { /* bad cursor? */
2021 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2022 txn->mt_flags |= MDB_TXN_ERROR;
2023 return MDB_CORRUPTED;
2028 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2030 np = mdb_page_malloc(txn, 1);
2035 mdb_mid2l_insert(dl, &mid);
2040 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2042 np->mp_flags |= P_DIRTY;
2045 /* Adjust cursors pointing to mp */
2046 mc->mc_pg[mc->mc_top] = np;
2047 m2 = txn->mt_cursors[mc->mc_dbi];
2048 if (mc->mc_flags & C_SUB) {
2049 for (; m2; m2=m2->mc_next) {
2050 m3 = &m2->mc_xcursor->mx_cursor;
2051 if (m3->mc_snum < mc->mc_snum) continue;
2052 if (m3->mc_pg[mc->mc_top] == mp)
2053 m3->mc_pg[mc->mc_top] = np;
2056 for (; m2; m2=m2->mc_next) {
2057 if (m2->mc_snum < mc->mc_snum) continue;
2058 if (m2->mc_pg[mc->mc_top] == mp) {
2059 m2->mc_pg[mc->mc_top] = np;
2060 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2061 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2063 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2064 if (!(leaf->mn_flags & F_SUBDATA))
2065 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2073 txn->mt_flags |= MDB_TXN_ERROR;
2078 mdb_env_sync(MDB_env *env, int force)
2081 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2082 if (env->me_flags & MDB_WRITEMAP) {
2083 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2084 ? MS_ASYNC : MS_SYNC;
2085 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2088 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2092 if (MDB_FDATASYNC(env->me_fd))
2099 /** Back up parent txn's cursors, then grab the originals for tracking */
2101 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2103 MDB_cursor *mc, *bk;
2108 for (i = src->mt_numdbs; --i >= 0; ) {
2109 if ((mc = src->mt_cursors[i]) != NULL) {
2110 size = sizeof(MDB_cursor);
2112 size += sizeof(MDB_xcursor);
2113 for (; mc; mc = bk->mc_next) {
2119 mc->mc_db = &dst->mt_dbs[i];
2120 /* Kill pointers into src - and dst to reduce abuse: The
2121 * user may not use mc until dst ends. Otherwise we'd...
2123 mc->mc_txn = NULL; /* ...set this to dst */
2124 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2125 if ((mx = mc->mc_xcursor) != NULL) {
2126 *(MDB_xcursor *)(bk+1) = *mx;
2127 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2129 mc->mc_next = dst->mt_cursors[i];
2130 dst->mt_cursors[i] = mc;
2137 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2138 * @param[in] txn the transaction handle.
2139 * @param[in] merge true to keep changes to parent cursors, false to revert.
2140 * @return 0 on success, non-zero on failure.
2143 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2145 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2149 for (i = txn->mt_numdbs; --i >= 0; ) {
2150 for (mc = cursors[i]; mc; mc = next) {
2152 if ((bk = mc->mc_backup) != NULL) {
2154 /* Commit changes to parent txn */
2155 mc->mc_next = bk->mc_next;
2156 mc->mc_backup = bk->mc_backup;
2157 mc->mc_txn = bk->mc_txn;
2158 mc->mc_db = bk->mc_db;
2159 mc->mc_dbflag = bk->mc_dbflag;
2160 if ((mx = mc->mc_xcursor) != NULL)
2161 mx->mx_cursor.mc_txn = bk->mc_txn;
2163 /* Abort nested txn */
2165 if ((mx = mc->mc_xcursor) != NULL)
2166 *mx = *(MDB_xcursor *)(bk+1);
2170 /* Only malloced cursors are permanently tracked. */
2178 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2181 mdb_txn_reset0(MDB_txn *txn, const char *act);
2183 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2189 Pidset = F_SETLK, Pidcheck = F_GETLK
2193 /** Set or check a pid lock. Set returns 0 on success.
2194 * Check returns 0 if the process is certainly dead, nonzero if it may
2195 * be alive (the lock exists or an error happened so we do not know).
2197 * On Windows Pidset is a no-op, we merely check for the existence
2198 * of the process with the given pid. On POSIX we use a single byte
2199 * lock on the lockfile, set at an offset equal to the pid.
2202 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2204 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2207 if (op == Pidcheck) {
2208 h = OpenProcess(env->me_pidquery, FALSE, pid);
2209 /* No documented "no such process" code, but other program use this: */
2211 return ErrCode() != ERROR_INVALID_PARAMETER;
2212 /* A process exists until all handles to it close. Has it exited? */
2213 ret = WaitForSingleObject(h, 0) != 0;
2220 struct flock lock_info;
2221 memset(&lock_info, 0, sizeof(lock_info));
2222 lock_info.l_type = F_WRLCK;
2223 lock_info.l_whence = SEEK_SET;
2224 lock_info.l_start = pid;
2225 lock_info.l_len = 1;
2226 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2227 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2229 } else if ((rc = ErrCode()) == EINTR) {
2237 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2238 * @param[in] txn the transaction handle to initialize
2239 * @return 0 on success, non-zero on failure.
2242 mdb_txn_renew0(MDB_txn *txn)
2244 MDB_env *env = txn->mt_env;
2245 MDB_txninfo *ti = env->me_txns;
2249 int rc, new_notls = 0;
2252 txn->mt_numdbs = env->me_numdbs;
2253 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2255 if (txn->mt_flags & MDB_TXN_RDONLY) {
2257 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2258 txn->mt_txnid = meta->mm_txnid;
2259 txn->mt_u.reader = NULL;
2261 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2262 pthread_getspecific(env->me_txkey);
2264 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2265 return MDB_BAD_RSLOT;
2267 MDB_PID_T pid = env->me_pid;
2268 pthread_t tid = pthread_self();
2270 if (!(env->me_flags & MDB_LIVE_READER)) {
2271 rc = mdb_reader_pid(env, Pidset, pid);
2274 env->me_flags |= MDB_LIVE_READER;
2278 nr = ti->mti_numreaders;
2279 for (i=0; i<nr; i++)
2280 if (ti->mti_readers[i].mr_pid == 0)
2282 if (i == env->me_maxreaders) {
2283 UNLOCK_MUTEX_R(env);
2284 return MDB_READERS_FULL;
2286 ti->mti_readers[i].mr_pid = pid;
2287 ti->mti_readers[i].mr_tid = tid;
2289 ti->mti_numreaders = ++nr;
2290 /* Save numreaders for un-mutexed mdb_env_close() */
2291 env->me_numreaders = nr;
2292 UNLOCK_MUTEX_R(env);
2294 r = &ti->mti_readers[i];
2295 new_notls = (env->me_flags & MDB_NOTLS);
2296 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2301 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2302 txn->mt_u.reader = r;
2303 meta = env->me_metas[txn->mt_txnid & 1];
2309 txn->mt_txnid = ti->mti_txnid;
2310 meta = env->me_metas[txn->mt_txnid & 1];
2312 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2313 txn->mt_txnid = meta->mm_txnid;
2317 if (txn->mt_txnid == mdb_debug_start)
2320 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2321 txn->mt_u.dirty_list = env->me_dirty_list;
2322 txn->mt_u.dirty_list[0].mid = 0;
2323 txn->mt_free_pgs = env->me_free_pgs;
2324 txn->mt_free_pgs[0] = 0;
2325 txn->mt_spill_pgs = NULL;
2329 /* Copy the DB info and flags */
2330 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2332 /* Moved to here to avoid a data race in read TXNs */
2333 txn->mt_next_pgno = meta->mm_last_pg+1;
2335 for (i=2; i<txn->mt_numdbs; i++) {
2336 x = env->me_dbflags[i];
2337 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2338 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2340 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2342 if (env->me_maxpg < txn->mt_next_pgno) {
2343 mdb_txn_reset0(txn, "renew0-mapfail");
2345 txn->mt_u.reader->mr_pid = 0;
2346 txn->mt_u.reader = NULL;
2348 return MDB_MAP_RESIZED;
2355 mdb_txn_renew(MDB_txn *txn)
2359 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2362 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2363 DPUTS("environment had fatal error, must shutdown!");
2367 rc = mdb_txn_renew0(txn);
2368 if (rc == MDB_SUCCESS) {
2369 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2370 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2371 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2377 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2381 int rc, size, tsize = sizeof(MDB_txn);
2383 if (env->me_flags & MDB_FATAL_ERROR) {
2384 DPUTS("environment had fatal error, must shutdown!");
2387 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2390 /* Nested transactions: Max 1 child, write txns only, no writemap */
2391 if (parent->mt_child ||
2392 (flags & MDB_RDONLY) ||
2393 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2394 (env->me_flags & MDB_WRITEMAP))
2396 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2398 tsize = sizeof(MDB_ntxn);
2400 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2401 if (!(flags & MDB_RDONLY))
2402 size += env->me_maxdbs * sizeof(MDB_cursor *);
2404 if ((txn = calloc(1, size)) == NULL) {
2405 DPRINTF(("calloc: %s", strerror(ErrCode())));
2408 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2409 if (flags & MDB_RDONLY) {
2410 txn->mt_flags |= MDB_TXN_RDONLY;
2411 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2413 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2414 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2420 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2421 if (!txn->mt_u.dirty_list ||
2422 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2424 free(txn->mt_u.dirty_list);
2428 txn->mt_txnid = parent->mt_txnid;
2429 txn->mt_dirty_room = parent->mt_dirty_room;
2430 txn->mt_u.dirty_list[0].mid = 0;
2431 txn->mt_spill_pgs = NULL;
2432 txn->mt_next_pgno = parent->mt_next_pgno;
2433 parent->mt_child = txn;
2434 txn->mt_parent = parent;
2435 txn->mt_numdbs = parent->mt_numdbs;
2436 txn->mt_flags = parent->mt_flags;
2437 txn->mt_dbxs = parent->mt_dbxs;
2438 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2439 /* Copy parent's mt_dbflags, but clear DB_NEW */
2440 for (i=0; i<txn->mt_numdbs; i++)
2441 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2443 ntxn = (MDB_ntxn *)txn;
2444 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2445 if (env->me_pghead) {
2446 size = MDB_IDL_SIZEOF(env->me_pghead);
2447 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2449 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2454 rc = mdb_cursor_shadow(parent, txn);
2456 mdb_txn_reset0(txn, "beginchild-fail");
2458 rc = mdb_txn_renew0(txn);
2464 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2465 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2466 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2473 mdb_txn_env(MDB_txn *txn)
2475 if(!txn) return NULL;
2479 /** Export or close DBI handles opened in this txn. */
2481 mdb_dbis_update(MDB_txn *txn, int keep)
2484 MDB_dbi n = txn->mt_numdbs;
2485 MDB_env *env = txn->mt_env;
2486 unsigned char *tdbflags = txn->mt_dbflags;
2488 for (i = n; --i >= 2;) {
2489 if (tdbflags[i] & DB_NEW) {
2491 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2493 char *ptr = env->me_dbxs[i].md_name.mv_data;
2494 env->me_dbxs[i].md_name.mv_data = NULL;
2495 env->me_dbxs[i].md_name.mv_size = 0;
2496 env->me_dbflags[i] = 0;
2501 if (keep && env->me_numdbs < n)
2505 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2506 * May be called twice for readonly txns: First reset it, then abort.
2507 * @param[in] txn the transaction handle to reset
2508 * @param[in] act why the transaction is being reset
2511 mdb_txn_reset0(MDB_txn *txn, const char *act)
2513 MDB_env *env = txn->mt_env;
2515 /* Close any DBI handles opened in this txn */
2516 mdb_dbis_update(txn, 0);
2518 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2519 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2520 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2522 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2523 if (txn->mt_u.reader) {
2524 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2525 if (!(env->me_flags & MDB_NOTLS))
2526 txn->mt_u.reader = NULL; /* txn does not own reader */
2528 txn->mt_numdbs = 0; /* close nothing if called again */
2529 txn->mt_dbxs = NULL; /* mark txn as reset */
2531 mdb_cursors_close(txn, 0);
2533 if (!(env->me_flags & MDB_WRITEMAP)) {
2534 mdb_dlist_free(txn);
2536 mdb_midl_free(env->me_pghead);
2538 if (txn->mt_parent) {
2539 txn->mt_parent->mt_child = NULL;
2540 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2541 mdb_midl_free(txn->mt_free_pgs);
2542 mdb_midl_free(txn->mt_spill_pgs);
2543 free(txn->mt_u.dirty_list);
2547 if (mdb_midl_shrink(&txn->mt_free_pgs))
2548 env->me_free_pgs = txn->mt_free_pgs;
2549 env->me_pghead = NULL;
2553 /* The writer mutex was locked in mdb_txn_begin. */
2555 UNLOCK_MUTEX_W(env);
2560 mdb_txn_reset(MDB_txn *txn)
2565 /* This call is only valid for read-only txns */
2566 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2569 mdb_txn_reset0(txn, "reset");
2573 mdb_txn_abort(MDB_txn *txn)
2579 mdb_txn_abort(txn->mt_child);
2581 mdb_txn_reset0(txn, "abort");
2582 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2583 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2584 txn->mt_u.reader->mr_pid = 0;
2589 /** Save the freelist as of this transaction to the freeDB.
2590 * This changes the freelist. Keep trying until it stabilizes.
2593 mdb_freelist_save(MDB_txn *txn)
2595 /* env->me_pghead[] can grow and shrink during this call.
2596 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2597 * Page numbers cannot disappear from txn->mt_free_pgs[].
2600 MDB_env *env = txn->mt_env;
2601 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2602 txnid_t pglast = 0, head_id = 0;
2603 pgno_t freecnt = 0, *free_pgs, *mop;
2604 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2606 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2608 if (env->me_pghead) {
2609 /* Make sure first page of freeDB is touched and on freelist */
2610 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2611 if (rc && rc != MDB_NOTFOUND)
2615 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2616 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2617 ? SSIZE_MAX : maxfree_1pg;
2620 /* Come back here after each Put() in case freelist changed */
2625 /* If using records from freeDB which we have not yet
2626 * deleted, delete them and any we reserved for me_pghead.
2628 while (pglast < env->me_pglast) {
2629 rc = mdb_cursor_first(&mc, &key, NULL);
2632 pglast = head_id = *(txnid_t *)key.mv_data;
2633 total_room = head_room = 0;
2634 mdb_tassert(txn, pglast <= env->me_pglast);
2635 rc = mdb_cursor_del(&mc, 0);
2640 /* Save the IDL of pages freed by this txn, to a single record */
2641 if (freecnt < txn->mt_free_pgs[0]) {
2643 /* Make sure last page of freeDB is touched and on freelist */
2644 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2645 if (rc && rc != MDB_NOTFOUND)
2648 free_pgs = txn->mt_free_pgs;
2649 /* Write to last page of freeDB */
2650 key.mv_size = sizeof(txn->mt_txnid);
2651 key.mv_data = &txn->mt_txnid;
2653 freecnt = free_pgs[0];
2654 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2655 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2658 /* Retry if mt_free_pgs[] grew during the Put() */
2659 free_pgs = txn->mt_free_pgs;
2660 } while (freecnt < free_pgs[0]);
2661 mdb_midl_sort(free_pgs);
2662 memcpy(data.mv_data, free_pgs, data.mv_size);
2665 unsigned int i = free_pgs[0];
2666 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2667 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2669 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2675 mop = env->me_pghead;
2676 mop_len = mop ? mop[0] : 0;
2678 /* Reserve records for me_pghead[]. Split it if multi-page,
2679 * to avoid searching freeDB for a page range. Use keys in
2680 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2682 if (total_room >= mop_len) {
2683 if (total_room == mop_len || --more < 0)
2685 } else if (head_room >= maxfree_1pg && head_id > 1) {
2686 /* Keep current record (overflow page), add a new one */
2690 /* (Re)write {key = head_id, IDL length = head_room} */
2691 total_room -= head_room;
2692 head_room = mop_len - total_room;
2693 if (head_room > maxfree_1pg && head_id > 1) {
2694 /* Overflow multi-page for part of me_pghead */
2695 head_room /= head_id; /* amortize page sizes */
2696 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2697 } else if (head_room < 0) {
2698 /* Rare case, not bothering to delete this record */
2701 key.mv_size = sizeof(head_id);
2702 key.mv_data = &head_id;
2703 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2704 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2707 /* IDL is initially empty, zero out at least the length */
2708 pgs = (pgno_t *)data.mv_data;
2709 j = head_room > clean_limit ? head_room : 0;
2713 total_room += head_room;
2716 /* Fill in the reserved me_pghead records */
2722 rc = mdb_cursor_first(&mc, &key, &data);
2723 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2724 unsigned flags = MDB_CURRENT;
2725 txnid_t id = *(txnid_t *)key.mv_data;
2726 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2729 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2731 if (len > mop_len) {
2733 data.mv_size = (len + 1) * sizeof(MDB_ID);
2736 data.mv_data = mop -= len;
2739 rc = mdb_cursor_put(&mc, &key, &data, flags);
2741 if (rc || !(mop_len -= len))
2748 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2749 * @param[in] txn the transaction that's being committed
2750 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2751 * @return 0 on success, non-zero on failure.
2754 mdb_page_flush(MDB_txn *txn, int keep)
2756 MDB_env *env = txn->mt_env;
2757 MDB_ID2L dl = txn->mt_u.dirty_list;
2758 unsigned psize = env->me_psize, j;
2759 int i, pagecount = dl[0].mid, rc;
2760 size_t size = 0, pos = 0;
2762 MDB_page *dp = NULL;
2766 struct iovec iov[MDB_COMMIT_PAGES];
2767 ssize_t wpos = 0, wsize = 0, wres;
2768 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2774 if (env->me_flags & MDB_WRITEMAP) {
2775 /* Clear dirty flags */
2776 while (++i <= pagecount) {
2778 /* Don't flush this page yet */
2779 if (dp->mp_flags & P_KEEP) {
2780 dp->mp_flags ^= P_KEEP;
2784 dp->mp_flags &= ~P_DIRTY;
2789 /* Write the pages */
2791 if (++i <= pagecount) {
2793 /* Don't flush this page yet */
2794 if (dp->mp_flags & P_KEEP) {
2795 dp->mp_flags ^= P_KEEP;
2800 /* clear dirty flag */
2801 dp->mp_flags &= ~P_DIRTY;
2804 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2809 /* Windows actually supports scatter/gather I/O, but only on
2810 * unbuffered file handles. Since we're relying on the OS page
2811 * cache for all our data, that's self-defeating. So we just
2812 * write pages one at a time. We use the ov structure to set
2813 * the write offset, to at least save the overhead of a Seek
2816 DPRINTF(("committing page %"Z"u", pgno));
2817 memset(&ov, 0, sizeof(ov));
2818 ov.Offset = pos & 0xffffffff;
2819 ov.OffsetHigh = pos >> 16 >> 16;
2820 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2822 DPRINTF(("WriteFile: %d", rc));
2826 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2827 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2829 /* Write previous page(s) */
2830 #ifdef MDB_USE_PWRITEV
2831 wres = pwritev(env->me_fd, iov, n, wpos);
2834 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2836 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2838 DPRINTF(("lseek: %s", strerror(rc)));
2841 wres = writev(env->me_fd, iov, n);
2844 if (wres != wsize) {
2847 DPRINTF(("Write error: %s", strerror(rc)));
2849 rc = EIO; /* TODO: Use which error code? */
2850 DPUTS("short write, filesystem full?");
2861 DPRINTF(("committing page %"Z"u", pgno));
2862 next_pos = pos + size;
2863 iov[n].iov_len = size;
2864 iov[n].iov_base = (char *)dp;
2870 for (i = keep; ++i <= pagecount; ) {
2872 /* This is a page we skipped above */
2875 dl[j].mid = dp->mp_pgno;
2878 mdb_dpage_free(env, dp);
2883 txn->mt_dirty_room += i - j;
2889 mdb_txn_commit(MDB_txn *txn)
2895 if (txn == NULL || txn->mt_env == NULL)
2898 if (txn->mt_child) {
2899 rc = mdb_txn_commit(txn->mt_child);
2900 txn->mt_child = NULL;
2907 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2908 mdb_dbis_update(txn, 1);
2909 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2914 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2915 DPUTS("error flag is set, can't commit");
2917 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2922 if (txn->mt_parent) {
2923 MDB_txn *parent = txn->mt_parent;
2926 unsigned x, y, len, ps_len;
2928 /* Append our free list to parent's */
2929 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2932 mdb_midl_free(txn->mt_free_pgs);
2933 /* Failures after this must either undo the changes
2934 * to the parent or set MDB_TXN_ERROR in the parent.
2937 parent->mt_next_pgno = txn->mt_next_pgno;
2938 parent->mt_flags = txn->mt_flags;
2940 /* Merge our cursors into parent's and close them */
2941 mdb_cursors_close(txn, 1);
2943 /* Update parent's DB table. */
2944 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2945 parent->mt_numdbs = txn->mt_numdbs;
2946 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2947 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2948 for (i=2; i<txn->mt_numdbs; i++) {
2949 /* preserve parent's DB_NEW status */
2950 x = parent->mt_dbflags[i] & DB_NEW;
2951 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2954 dst = parent->mt_u.dirty_list;
2955 src = txn->mt_u.dirty_list;
2956 /* Remove anything in our dirty list from parent's spill list */
2957 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2959 pspill[0] = (pgno_t)-1;
2960 /* Mark our dirty pages as deleted in parent spill list */
2961 for (i=0, len=src[0].mid; ++i <= len; ) {
2962 MDB_ID pn = src[i].mid << 1;
2963 while (pn > pspill[x])
2965 if (pn == pspill[x]) {
2970 /* Squash deleted pagenums if we deleted any */
2971 for (x=y; ++x <= ps_len; )
2972 if (!(pspill[x] & 1))
2973 pspill[++y] = pspill[x];
2977 /* Find len = length of merging our dirty list with parent's */
2979 dst[0].mid = 0; /* simplify loops */
2980 if (parent->mt_parent) {
2981 len = x + src[0].mid;
2982 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2983 for (i = x; y && i; y--) {
2984 pgno_t yp = src[y].mid;
2985 while (yp < dst[i].mid)
2987 if (yp == dst[i].mid) {
2992 } else { /* Simplify the above for single-ancestor case */
2993 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2995 /* Merge our dirty list with parent's */
2997 for (i = len; y; dst[i--] = src[y--]) {
2998 pgno_t yp = src[y].mid;
2999 while (yp < dst[x].mid)
3000 dst[i--] = dst[x--];
3001 if (yp == dst[x].mid)
3002 free(dst[x--].mptr);
3004 mdb_tassert(txn, i == x);
3006 free(txn->mt_u.dirty_list);
3007 parent->mt_dirty_room = txn->mt_dirty_room;
3008 if (txn->mt_spill_pgs) {
3009 if (parent->mt_spill_pgs) {
3010 /* TODO: Prevent failure here, so parent does not fail */
3011 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3013 parent->mt_flags |= MDB_TXN_ERROR;
3014 mdb_midl_free(txn->mt_spill_pgs);
3015 mdb_midl_sort(parent->mt_spill_pgs);
3017 parent->mt_spill_pgs = txn->mt_spill_pgs;
3021 parent->mt_child = NULL;
3022 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3027 if (txn != env->me_txn) {
3028 DPUTS("attempt to commit unknown transaction");
3033 mdb_cursors_close(txn, 0);
3035 if (!txn->mt_u.dirty_list[0].mid &&
3036 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3039 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3040 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3042 /* Update DB root pointers */
3043 if (txn->mt_numdbs > 2) {
3047 data.mv_size = sizeof(MDB_db);
3049 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3050 for (i = 2; i < txn->mt_numdbs; i++) {
3051 if (txn->mt_dbflags[i] & DB_DIRTY) {
3052 data.mv_data = &txn->mt_dbs[i];
3053 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3060 rc = mdb_freelist_save(txn);
3064 mdb_midl_free(env->me_pghead);
3065 env->me_pghead = NULL;
3066 if (mdb_midl_shrink(&txn->mt_free_pgs))
3067 env->me_free_pgs = txn->mt_free_pgs;
3073 if ((rc = mdb_page_flush(txn, 0)) ||
3074 (rc = mdb_env_sync(env, 0)) ||
3075 (rc = mdb_env_write_meta(txn)))
3081 mdb_dbis_update(txn, 1);
3084 UNLOCK_MUTEX_W(env);
3094 /** Read the environment parameters of a DB environment before
3095 * mapping it into memory.
3096 * @param[in] env the environment handle
3097 * @param[out] meta address of where to store the meta information
3098 * @return 0 on success, non-zero on failure.
3101 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3107 enum { Size = sizeof(pbuf) };
3109 /* We don't know the page size yet, so use a minimum value.
3110 * Read both meta pages so we can use the latest one.
3113 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3117 memset(&ov, 0, sizeof(ov));
3119 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3120 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3123 rc = pread(env->me_fd, &pbuf, Size, off);
3126 if (rc == 0 && off == 0)
3128 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3129 DPRINTF(("read: %s", mdb_strerror(rc)));
3133 p = (MDB_page *)&pbuf;
3135 if (!F_ISSET(p->mp_flags, P_META)) {
3136 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3141 if (m->mm_magic != MDB_MAGIC) {
3142 DPUTS("meta has invalid magic");
3146 if (m->mm_version != MDB_DATA_VERSION) {
3147 DPRINTF(("database is version %u, expected version %u",
3148 m->mm_version, MDB_DATA_VERSION));
3149 return MDB_VERSION_MISMATCH;
3152 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3158 /** Write the environment parameters of a freshly created DB environment.
3159 * @param[in] env the environment handle
3160 * @param[out] meta address of where to store the meta information
3161 * @return 0 on success, non-zero on failure.
3164 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3172 memset(&ov, 0, sizeof(ov));
3173 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3175 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3178 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3179 len = pwrite(fd, ptr, size, pos); \
3180 rc = (len >= 0); } while(0)
3183 DPUTS("writing new meta page");
3185 psize = env->me_psize;
3187 meta->mm_magic = MDB_MAGIC;
3188 meta->mm_version = MDB_DATA_VERSION;
3189 meta->mm_mapsize = env->me_mapsize;
3190 meta->mm_psize = psize;
3191 meta->mm_last_pg = 1;
3192 meta->mm_flags = env->me_flags & 0xffff;
3193 meta->mm_flags |= MDB_INTEGERKEY;
3194 meta->mm_dbs[0].md_root = P_INVALID;
3195 meta->mm_dbs[1].md_root = P_INVALID;
3197 p = calloc(2, psize);
3199 p->mp_flags = P_META;
3200 *(MDB_meta *)METADATA(p) = *meta;
3202 q = (MDB_page *)((char *)p + psize);
3204 q->mp_flags = P_META;
3205 *(MDB_meta *)METADATA(q) = *meta;
3207 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3210 else if ((unsigned) len == psize * 2)
3218 /** Update the environment info to commit a transaction.
3219 * @param[in] txn the transaction that's being committed
3220 * @return 0 on success, non-zero on failure.
3223 mdb_env_write_meta(MDB_txn *txn)
3226 MDB_meta meta, metab, *mp;
3228 int rc, len, toggle;
3237 toggle = txn->mt_txnid & 1;
3238 DPRINTF(("writing meta page %d for root page %"Z"u",
3239 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3242 mp = env->me_metas[toggle];
3244 if (env->me_flags & MDB_WRITEMAP) {
3245 /* Persist any increases of mapsize config */
3246 if (env->me_mapsize > mp->mm_mapsize)
3247 mp->mm_mapsize = env->me_mapsize;
3248 mp->mm_dbs[0] = txn->mt_dbs[0];
3249 mp->mm_dbs[1] = txn->mt_dbs[1];
3250 mp->mm_last_pg = txn->mt_next_pgno - 1;
3251 mp->mm_txnid = txn->mt_txnid;
3252 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3253 unsigned meta_size = env->me_psize;
3254 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3257 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3258 if (meta_size < env->me_os_psize)
3259 meta_size += meta_size;
3264 if (MDB_MSYNC(ptr, meta_size, rc)) {
3271 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3272 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3274 ptr = (char *)&meta;
3275 if (env->me_mapsize > mp->mm_mapsize) {
3276 /* Persist any increases of mapsize config */
3277 meta.mm_mapsize = env->me_mapsize;
3278 off = offsetof(MDB_meta, mm_mapsize);
3280 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3282 len = sizeof(MDB_meta) - off;
3285 meta.mm_dbs[0] = txn->mt_dbs[0];
3286 meta.mm_dbs[1] = txn->mt_dbs[1];
3287 meta.mm_last_pg = txn->mt_next_pgno - 1;
3288 meta.mm_txnid = txn->mt_txnid;
3291 off += env->me_psize;
3294 /* Write to the SYNC fd */
3295 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3296 env->me_fd : env->me_mfd;
3299 memset(&ov, 0, sizeof(ov));
3301 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3305 rc = pwrite(mfd, ptr, len, off);
3308 rc = rc < 0 ? ErrCode() : EIO;
3309 DPUTS("write failed, disk error?");
3310 /* On a failure, the pagecache still contains the new data.
3311 * Write some old data back, to prevent it from being used.
3312 * Use the non-SYNC fd; we know it will fail anyway.
3314 meta.mm_last_pg = metab.mm_last_pg;
3315 meta.mm_txnid = metab.mm_txnid;
3317 memset(&ov, 0, sizeof(ov));
3319 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3321 r2 = pwrite(env->me_fd, ptr, len, off);
3322 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3325 env->me_flags |= MDB_FATAL_ERROR;
3329 /* Memory ordering issues are irrelevant; since the entire writer
3330 * is wrapped by wmutex, all of these changes will become visible
3331 * after the wmutex is unlocked. Since the DB is multi-version,
3332 * readers will get consistent data regardless of how fresh or
3333 * how stale their view of these values is.
3336 env->me_txns->mti_txnid = txn->mt_txnid;
3341 /** Check both meta pages to see which one is newer.
3342 * @param[in] env the environment handle
3343 * @return meta toggle (0 or 1).
3346 mdb_env_pick_meta(const MDB_env *env)
3348 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3352 mdb_env_create(MDB_env **env)
3356 e = calloc(1, sizeof(MDB_env));
3360 e->me_maxreaders = DEFAULT_READERS;
3361 e->me_maxdbs = e->me_numdbs = 2;
3362 e->me_fd = INVALID_HANDLE_VALUE;
3363 e->me_lfd = INVALID_HANDLE_VALUE;
3364 e->me_mfd = INVALID_HANDLE_VALUE;
3365 #ifdef MDB_USE_POSIX_SEM
3366 e->me_rmutex = SEM_FAILED;
3367 e->me_wmutex = SEM_FAILED;
3369 e->me_pid = getpid();
3370 GET_PAGESIZE(e->me_os_psize);
3371 VGMEMP_CREATE(e,0,0);
3377 mdb_env_map(MDB_env *env, void *addr, int newsize)
3380 unsigned int flags = env->me_flags;
3384 LONG sizelo, sizehi;
3385 sizelo = env->me_mapsize & 0xffffffff;
3386 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3388 /* Windows won't create mappings for zero length files.
3389 * Just allocate the maxsize right now.
3392 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3393 || !SetEndOfFile(env->me_fd)
3394 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3397 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3398 PAGE_READWRITE : PAGE_READONLY,
3399 sizehi, sizelo, NULL);
3402 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3403 FILE_MAP_WRITE : FILE_MAP_READ,
3404 0, 0, env->me_mapsize, addr);
3405 rc = env->me_map ? 0 : ErrCode();
3410 int prot = PROT_READ;
3411 if (flags & MDB_WRITEMAP) {
3413 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3416 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3418 if (env->me_map == MAP_FAILED) {
3423 if (flags & MDB_NORDAHEAD) {
3424 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3426 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3428 #ifdef POSIX_MADV_RANDOM
3429 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3430 #endif /* POSIX_MADV_RANDOM */
3431 #endif /* MADV_RANDOM */
3435 /* Can happen because the address argument to mmap() is just a
3436 * hint. mmap() can pick another, e.g. if the range is in use.
3437 * The MAP_FIXED flag would prevent that, but then mmap could
3438 * instead unmap existing pages to make room for the new map.
3440 if (addr && env->me_map != addr)
3441 return EBUSY; /* TODO: Make a new MDB_* error code? */
3443 p = (MDB_page *)env->me_map;
3444 env->me_metas[0] = METADATA(p);
3445 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3451 mdb_env_set_mapsize(MDB_env *env, size_t size)
3453 /* If env is already open, caller is responsible for making
3454 * sure there are no active txns.
3462 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3463 else if (size < env->me_mapsize) {
3464 /* If the configured size is smaller, make sure it's
3465 * still big enough. Silently round up to minimum if not.
3467 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3471 munmap(env->me_map, env->me_mapsize);
3472 env->me_mapsize = size;
3473 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3474 rc = mdb_env_map(env, old, 1);
3478 env->me_mapsize = size;
3480 env->me_maxpg = env->me_mapsize / env->me_psize;
3485 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3489 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3494 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3496 if (env->me_map || readers < 1)
3498 env->me_maxreaders = readers;
3503 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3505 if (!env || !readers)
3507 *readers = env->me_maxreaders;
3511 /** Further setup required for opening an MDB environment
3514 mdb_env_open2(MDB_env *env)
3516 unsigned int flags = env->me_flags;
3517 int i, newenv = 0, rc;
3521 /* See if we should use QueryLimited */
3523 if ((rc & 0xff) > 5)
3524 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3526 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3529 memset(&meta, 0, sizeof(meta));
3531 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3534 DPUTS("new mdbenv");
3536 env->me_psize = env->me_os_psize;
3537 if (env->me_psize > MAX_PAGESIZE)
3538 env->me_psize = MAX_PAGESIZE;
3540 env->me_psize = meta.mm_psize;
3543 /* Was a mapsize configured? */
3544 if (!env->me_mapsize) {
3545 /* If this is a new environment, take the default,
3546 * else use the size recorded in the existing env.
3548 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3549 } else if (env->me_mapsize < meta.mm_mapsize) {
3550 /* If the configured size is smaller, make sure it's
3551 * still big enough. Silently round up to minimum if not.
3553 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3554 if (env->me_mapsize < minsize)
3555 env->me_mapsize = minsize;
3558 rc = mdb_env_map(env, meta.mm_address, newenv);
3563 if (flags & MDB_FIXEDMAP)
3564 meta.mm_address = env->me_map;
3565 i = mdb_env_init_meta(env, &meta);
3566 if (i != MDB_SUCCESS) {
3571 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3572 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3574 #if !(MDB_MAXKEYSIZE)
3575 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3577 env->me_maxpg = env->me_mapsize / env->me_psize;
3581 int toggle = mdb_env_pick_meta(env);
3582 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3584 DPRINTF(("opened database version %u, pagesize %u",
3585 env->me_metas[0]->mm_version, env->me_psize));
3586 DPRINTF(("using meta page %d", toggle));
3587 DPRINTF(("depth: %u", db->md_depth));
3588 DPRINTF(("entries: %"Z"u", db->md_entries));
3589 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3590 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3591 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3592 DPRINTF(("root: %"Z"u", db->md_root));
3600 /** Release a reader thread's slot in the reader lock table.
3601 * This function is called automatically when a thread exits.
3602 * @param[in] ptr This points to the slot in the reader lock table.
3605 mdb_env_reader_dest(void *ptr)
3607 MDB_reader *reader = ptr;
3613 /** Junk for arranging thread-specific callbacks on Windows. This is
3614 * necessarily platform and compiler-specific. Windows supports up
3615 * to 1088 keys. Let's assume nobody opens more than 64 environments
3616 * in a single process, for now. They can override this if needed.
3618 #ifndef MAX_TLS_KEYS
3619 #define MAX_TLS_KEYS 64
3621 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3622 static int mdb_tls_nkeys;
3624 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3628 case DLL_PROCESS_ATTACH: break;
3629 case DLL_THREAD_ATTACH: break;
3630 case DLL_THREAD_DETACH:
3631 for (i=0; i<mdb_tls_nkeys; i++) {
3632 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3633 mdb_env_reader_dest(r);
3636 case DLL_PROCESS_DETACH: break;
3641 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3643 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3647 /* Force some symbol references.
3648 * _tls_used forces the linker to create the TLS directory if not already done
3649 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3651 #pragma comment(linker, "/INCLUDE:_tls_used")
3652 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3653 #pragma const_seg(".CRT$XLB")
3654 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3655 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3658 #pragma comment(linker, "/INCLUDE:__tls_used")
3659 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3660 #pragma data_seg(".CRT$XLB")
3661 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3663 #endif /* WIN 32/64 */
3664 #endif /* !__GNUC__ */
3667 /** Downgrade the exclusive lock on the region back to shared */
3669 mdb_env_share_locks(MDB_env *env, int *excl)
3671 int rc = 0, toggle = mdb_env_pick_meta(env);
3673 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3678 /* First acquire a shared lock. The Unlock will
3679 * then release the existing exclusive lock.
3681 memset(&ov, 0, sizeof(ov));
3682 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3685 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3691 struct flock lock_info;
3692 /* The shared lock replaces the existing lock */
3693 memset((void *)&lock_info, 0, sizeof(lock_info));
3694 lock_info.l_type = F_RDLCK;
3695 lock_info.l_whence = SEEK_SET;
3696 lock_info.l_start = 0;
3697 lock_info.l_len = 1;
3698 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3699 (rc = ErrCode()) == EINTR) ;
3700 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3707 /** Try to get exlusive lock, otherwise shared.
3708 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3711 mdb_env_excl_lock(MDB_env *env, int *excl)
3715 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3719 memset(&ov, 0, sizeof(ov));
3720 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3727 struct flock lock_info;
3728 memset((void *)&lock_info, 0, sizeof(lock_info));
3729 lock_info.l_type = F_WRLCK;
3730 lock_info.l_whence = SEEK_SET;
3731 lock_info.l_start = 0;
3732 lock_info.l_len = 1;
3733 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3734 (rc = ErrCode()) == EINTR) ;
3738 # ifdef MDB_USE_POSIX_SEM
3739 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3742 lock_info.l_type = F_RDLCK;
3743 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3744 (rc = ErrCode()) == EINTR) ;
3754 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3756 * @(#) $Revision: 5.1 $
3757 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3758 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3760 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3764 * Please do not copyright this code. This code is in the public domain.
3766 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3767 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3768 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3769 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3770 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3771 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3772 * PERFORMANCE OF THIS SOFTWARE.
3775 * chongo <Landon Curt Noll> /\oo/\
3776 * http://www.isthe.com/chongo/
3778 * Share and Enjoy! :-)
3781 typedef unsigned long long mdb_hash_t;
3782 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3784 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3785 * @param[in] val value to hash
3786 * @param[in] hval initial value for hash
3787 * @return 64 bit hash
3789 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3790 * hval arg on the first call.
3793 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3795 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3796 unsigned char *end = s + val->mv_size;
3798 * FNV-1a hash each octet of the string
3801 /* xor the bottom with the current octet */
3802 hval ^= (mdb_hash_t)*s++;
3804 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3805 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3806 (hval << 7) + (hval << 8) + (hval << 40);
3808 /* return our new hash value */
3812 /** Hash the string and output the encoded hash.
3813 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3814 * very short name limits. We don't care about the encoding being reversible,
3815 * we just want to preserve as many bits of the input as possible in a
3816 * small printable string.
3817 * @param[in] str string to hash
3818 * @param[out] encbuf an array of 11 chars to hold the hash
3820 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3823 mdb_pack85(unsigned long l, char *out)
3827 for (i=0; i<5; i++) {
3828 *out++ = mdb_a85[l % 85];
3834 mdb_hash_enc(MDB_val *val, char *encbuf)
3836 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3838 mdb_pack85(h, encbuf);
3839 mdb_pack85(h>>32, encbuf+5);
3844 /** Open and/or initialize the lock region for the environment.
3845 * @param[in] env The MDB environment.
3846 * @param[in] lpath The pathname of the file used for the lock region.
3847 * @param[in] mode The Unix permissions for the file, if we create it.
3848 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3849 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3850 * @return 0 on success, non-zero on failure.
3853 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3856 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3858 # define MDB_ERRCODE_ROFS EROFS
3859 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3860 # define MDB_CLOEXEC O_CLOEXEC
3863 # define MDB_CLOEXEC 0
3870 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3871 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3872 FILE_ATTRIBUTE_NORMAL, NULL);
3874 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3876 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3878 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3883 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3884 /* Lose record locks when exec*() */
3885 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3886 fcntl(env->me_lfd, F_SETFD, fdflags);
3889 if (!(env->me_flags & MDB_NOTLS)) {
3890 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3893 env->me_flags |= MDB_ENV_TXKEY;
3895 /* Windows TLS callbacks need help finding their TLS info. */
3896 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3900 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3904 /* Try to get exclusive lock. If we succeed, then
3905 * nobody is using the lock region and we should initialize it.
3907 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3910 size = GetFileSize(env->me_lfd, NULL);
3912 size = lseek(env->me_lfd, 0, SEEK_END);
3913 if (size == -1) goto fail_errno;
3915 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3916 if (size < rsize && *excl > 0) {
3918 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3919 || !SetEndOfFile(env->me_lfd))
3922 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3926 size = rsize - sizeof(MDB_txninfo);
3927 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3932 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3934 if (!mh) goto fail_errno;
3935 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3937 if (!env->me_txns) goto fail_errno;
3939 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3941 if (m == MAP_FAILED) goto fail_errno;
3947 BY_HANDLE_FILE_INFORMATION stbuf;
3956 if (!mdb_sec_inited) {
3957 InitializeSecurityDescriptor(&mdb_null_sd,
3958 SECURITY_DESCRIPTOR_REVISION);
3959 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3960 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3961 mdb_all_sa.bInheritHandle = FALSE;
3962 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3965 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3966 idbuf.volume = stbuf.dwVolumeSerialNumber;
3967 idbuf.nhigh = stbuf.nFileIndexHigh;
3968 idbuf.nlow = stbuf.nFileIndexLow;
3969 val.mv_data = &idbuf;
3970 val.mv_size = sizeof(idbuf);
3971 mdb_hash_enc(&val, encbuf);
3972 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3973 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3974 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3975 if (!env->me_rmutex) goto fail_errno;
3976 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3977 if (!env->me_wmutex) goto fail_errno;
3978 #elif defined(MDB_USE_POSIX_SEM)
3987 #if defined(__NetBSD__)
3988 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3990 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3991 idbuf.dev = stbuf.st_dev;
3992 idbuf.ino = stbuf.st_ino;
3993 val.mv_data = &idbuf;
3994 val.mv_size = sizeof(idbuf);
3995 mdb_hash_enc(&val, encbuf);
3996 #ifdef MDB_SHORT_SEMNAMES
3997 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3999 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4000 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4001 /* Clean up after a previous run, if needed: Try to
4002 * remove both semaphores before doing anything else.
4004 sem_unlink(env->me_txns->mti_rmname);
4005 sem_unlink(env->me_txns->mti_wmname);
4006 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4007 O_CREAT|O_EXCL, mode, 1);
4008 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4009 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4010 O_CREAT|O_EXCL, mode, 1);
4011 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4012 #else /* MDB_USE_POSIX_SEM */
4013 pthread_mutexattr_t mattr;
4015 if ((rc = pthread_mutexattr_init(&mattr))
4016 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4017 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4018 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4020 pthread_mutexattr_destroy(&mattr);
4021 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4023 env->me_txns->mti_magic = MDB_MAGIC;
4024 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4025 env->me_txns->mti_txnid = 0;
4026 env->me_txns->mti_numreaders = 0;
4029 if (env->me_txns->mti_magic != MDB_MAGIC) {
4030 DPUTS("lock region has invalid magic");
4034 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4035 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4036 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4037 rc = MDB_VERSION_MISMATCH;
4041 if (rc && rc != EACCES && rc != EAGAIN) {
4045 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4046 if (!env->me_rmutex) goto fail_errno;
4047 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4048 if (!env->me_wmutex) goto fail_errno;
4049 #elif defined(MDB_USE_POSIX_SEM)
4050 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4051 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4052 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4053 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4064 /** The name of the lock file in the DB environment */
4065 #define LOCKNAME "/lock.mdb"
4066 /** The name of the data file in the DB environment */
4067 #define DATANAME "/data.mdb"
4068 /** The suffix of the lock file when no subdir is used */
4069 #define LOCKSUFF "-lock"
4070 /** Only a subset of the @ref mdb_env flags can be changed
4071 * at runtime. Changing other flags requires closing the
4072 * environment and re-opening it with the new flags.
4074 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4075 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4076 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4078 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4079 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4083 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4085 int oflags, rc, len, excl = -1;
4086 char *lpath, *dpath;
4088 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4092 if (flags & MDB_NOSUBDIR) {
4093 rc = len + sizeof(LOCKSUFF) + len + 1;
4095 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4100 if (flags & MDB_NOSUBDIR) {
4101 dpath = lpath + len + sizeof(LOCKSUFF);
4102 sprintf(lpath, "%s" LOCKSUFF, path);
4103 strcpy(dpath, path);
4105 dpath = lpath + len + sizeof(LOCKNAME);
4106 sprintf(lpath, "%s" LOCKNAME, path);
4107 sprintf(dpath, "%s" DATANAME, path);
4111 flags |= env->me_flags;
4112 if (flags & MDB_RDONLY) {
4113 /* silently ignore WRITEMAP when we're only getting read access */
4114 flags &= ~MDB_WRITEMAP;
4116 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4117 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4120 env->me_flags = flags |= MDB_ENV_ACTIVE;
4124 env->me_path = strdup(path);
4125 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4126 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4127 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4132 /* For RDONLY, get lockfile after we know datafile exists */
4133 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4134 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4140 if (F_ISSET(flags, MDB_RDONLY)) {
4141 oflags = GENERIC_READ;
4142 len = OPEN_EXISTING;
4144 oflags = GENERIC_READ|GENERIC_WRITE;
4147 mode = FILE_ATTRIBUTE_NORMAL;
4148 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4149 NULL, len, mode, NULL);
4151 if (F_ISSET(flags, MDB_RDONLY))
4154 oflags = O_RDWR | O_CREAT;
4156 env->me_fd = open(dpath, oflags, mode);
4158 if (env->me_fd == INVALID_HANDLE_VALUE) {
4163 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4164 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4169 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4170 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4171 env->me_mfd = env->me_fd;
4173 /* Synchronous fd for meta writes. Needed even with
4174 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4177 len = OPEN_EXISTING;
4178 env->me_mfd = CreateFile(dpath, oflags,
4179 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4180 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4183 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4185 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4190 DPRINTF(("opened dbenv %p", (void *) env));
4192 rc = mdb_env_share_locks(env, &excl);
4196 if (!((flags & MDB_RDONLY) ||
4197 (env->me_pbuf = calloc(1, env->me_psize))))
4203 mdb_env_close0(env, excl);
4209 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4211 mdb_env_close0(MDB_env *env, int excl)
4215 if (!(env->me_flags & MDB_ENV_ACTIVE))
4218 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4219 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4220 free(env->me_dbxs[i].md_name.mv_data);
4223 free(env->me_dbflags);
4226 free(env->me_dirty_list);
4227 mdb_midl_free(env->me_free_pgs);
4229 if (env->me_flags & MDB_ENV_TXKEY) {
4230 pthread_key_delete(env->me_txkey);
4232 /* Delete our key from the global list */
4233 for (i=0; i<mdb_tls_nkeys; i++)
4234 if (mdb_tls_keys[i] == env->me_txkey) {
4235 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4243 munmap(env->me_map, env->me_mapsize);
4245 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4246 (void) close(env->me_mfd);
4247 if (env->me_fd != INVALID_HANDLE_VALUE)
4248 (void) close(env->me_fd);
4250 MDB_PID_T pid = env->me_pid;
4251 /* Clearing readers is done in this function because
4252 * me_txkey with its destructor must be disabled first.
4254 for (i = env->me_numreaders; --i >= 0; )
4255 if (env->me_txns->mti_readers[i].mr_pid == pid)
4256 env->me_txns->mti_readers[i].mr_pid = 0;
4258 if (env->me_rmutex) {
4259 CloseHandle(env->me_rmutex);
4260 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4262 /* Windows automatically destroys the mutexes when
4263 * the last handle closes.
4265 #elif defined(MDB_USE_POSIX_SEM)
4266 if (env->me_rmutex != SEM_FAILED) {
4267 sem_close(env->me_rmutex);
4268 if (env->me_wmutex != SEM_FAILED)
4269 sem_close(env->me_wmutex);
4270 /* If we have the filelock: If we are the
4271 * only remaining user, clean up semaphores.
4274 mdb_env_excl_lock(env, &excl);
4276 sem_unlink(env->me_txns->mti_rmname);
4277 sem_unlink(env->me_txns->mti_wmname);
4281 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4283 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4286 /* Unlock the lockfile. Windows would have unlocked it
4287 * after closing anyway, but not necessarily at once.
4289 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4292 (void) close(env->me_lfd);
4295 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4299 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4301 MDB_txn *txn = NULL;
4307 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4311 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4314 /* Do the lock/unlock of the reader mutex before starting the
4315 * write txn. Otherwise other read txns could block writers.
4317 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4322 /* We must start the actual read txn after blocking writers */
4323 mdb_txn_reset0(txn, "reset-stage1");
4325 /* Temporarily block writers until we snapshot the meta pages */
4328 rc = mdb_txn_renew0(txn);
4330 UNLOCK_MUTEX_W(env);
4335 wsize = env->me_psize * 2;
4339 DO_WRITE(rc, fd, ptr, w2, len);
4343 } else if (len > 0) {
4349 /* Non-blocking or async handles are not supported */
4355 UNLOCK_MUTEX_W(env);
4360 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4362 if (wsize > MAX_WRITE)
4366 DO_WRITE(rc, fd, ptr, w2, len);
4370 } else if (len > 0) {
4387 mdb_env_copy(MDB_env *env, const char *path)
4391 HANDLE newfd = INVALID_HANDLE_VALUE;
4393 if (env->me_flags & MDB_NOSUBDIR) {
4394 lpath = (char *)path;
4397 len += sizeof(DATANAME);
4398 lpath = malloc(len);
4401 sprintf(lpath, "%s" DATANAME, path);
4404 /* The destination path must exist, but the destination file must not.
4405 * We don't want the OS to cache the writes, since the source data is
4406 * already in the OS cache.
4409 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4410 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4412 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4414 if (newfd == INVALID_HANDLE_VALUE) {
4420 /* Set O_DIRECT if the file system supports it */
4421 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4422 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4424 #ifdef F_NOCACHE /* __APPLE__ */
4425 rc = fcntl(newfd, F_NOCACHE, 1);
4432 rc = mdb_env_copyfd(env, newfd);
4435 if (!(env->me_flags & MDB_NOSUBDIR))
4437 if (newfd != INVALID_HANDLE_VALUE)
4438 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4445 mdb_env_close(MDB_env *env)
4452 VGMEMP_DESTROY(env);
4453 while ((dp = env->me_dpages) != NULL) {
4454 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4455 env->me_dpages = dp->mp_next;
4459 mdb_env_close0(env, 0);
4463 /** Compare two items pointing at aligned size_t's */
4465 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4467 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4468 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4471 /** Compare two items pointing at aligned unsigned int's */
4473 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4475 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4476 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4479 /** Compare two items pointing at unsigned ints of unknown alignment.
4480 * Nodes and keys are guaranteed to be 2-byte aligned.
4483 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4485 #if BYTE_ORDER == LITTLE_ENDIAN
4486 unsigned short *u, *c;
4489 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4490 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4493 } while(!x && u > (unsigned short *)a->mv_data);
4496 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4500 /** Compare two items lexically */
4502 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4509 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4515 diff = memcmp(a->mv_data, b->mv_data, len);
4516 return diff ? diff : len_diff<0 ? -1 : len_diff;
4519 /** Compare two items in reverse byte order */
4521 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4523 const unsigned char *p1, *p2, *p1_lim;
4527 p1_lim = (const unsigned char *)a->mv_data;
4528 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4529 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4531 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4537 while (p1 > p1_lim) {
4538 diff = *--p1 - *--p2;
4542 return len_diff<0 ? -1 : len_diff;
4545 /** Search for key within a page, using binary search.
4546 * Returns the smallest entry larger or equal to the key.
4547 * If exactp is non-null, stores whether the found entry was an exact match
4548 * in *exactp (1 or 0).
4549 * Updates the cursor index with the index of the found entry.
4550 * If no entry larger or equal to the key is found, returns NULL.
4553 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4555 unsigned int i = 0, nkeys;
4558 MDB_page *mp = mc->mc_pg[mc->mc_top];
4559 MDB_node *node = NULL;
4564 nkeys = NUMKEYS(mp);
4566 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4567 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4570 low = IS_LEAF(mp) ? 0 : 1;
4572 cmp = mc->mc_dbx->md_cmp;
4574 /* Branch pages have no data, so if using integer keys,
4575 * alignment is guaranteed. Use faster mdb_cmp_int.
4577 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4578 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4585 nodekey.mv_size = mc->mc_db->md_pad;
4586 node = NODEPTR(mp, 0); /* fake */
4587 while (low <= high) {
4588 i = (low + high) >> 1;
4589 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4590 rc = cmp(key, &nodekey);
4591 DPRINTF(("found leaf index %u [%s], rc = %i",
4592 i, DKEY(&nodekey), rc));
4601 while (low <= high) {
4602 i = (low + high) >> 1;
4604 node = NODEPTR(mp, i);
4605 nodekey.mv_size = NODEKSZ(node);
4606 nodekey.mv_data = NODEKEY(node);
4608 rc = cmp(key, &nodekey);
4611 DPRINTF(("found leaf index %u [%s], rc = %i",
4612 i, DKEY(&nodekey), rc));
4614 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4615 i, DKEY(&nodekey), NODEPGNO(node), rc));
4626 if (rc > 0) { /* Found entry is less than the key. */
4627 i++; /* Skip to get the smallest entry larger than key. */
4629 node = NODEPTR(mp, i);
4632 *exactp = (rc == 0 && nkeys > 0);
4633 /* store the key index */
4634 mc->mc_ki[mc->mc_top] = i;
4636 /* There is no entry larger or equal to the key. */
4639 /* nodeptr is fake for LEAF2 */
4645 mdb_cursor_adjust(MDB_cursor *mc, func)
4649 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4650 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4657 /** Pop a page off the top of the cursor's stack. */
4659 mdb_cursor_pop(MDB_cursor *mc)
4663 MDB_page *top = mc->mc_pg[mc->mc_top];
4669 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4670 DDBI(mc), (void *) mc));
4674 /** Push a page onto the top of the cursor's stack. */
4676 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4678 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4679 DDBI(mc), (void *) mc));
4681 if (mc->mc_snum >= CURSOR_STACK) {
4682 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4683 return MDB_CURSOR_FULL;
4686 mc->mc_top = mc->mc_snum++;
4687 mc->mc_pg[mc->mc_top] = mp;
4688 mc->mc_ki[mc->mc_top] = 0;
4693 /** Find the address of the page corresponding to a given page number.
4694 * @param[in] txn the transaction for this access.
4695 * @param[in] pgno the page number for the page to retrieve.
4696 * @param[out] ret address of a pointer where the page's address will be stored.
4697 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4698 * @return 0 on success, non-zero on failure.
4701 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4703 MDB_env *env = txn->mt_env;
4707 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4711 MDB_ID2L dl = tx2->mt_u.dirty_list;
4713 /* Spilled pages were dirtied in this txn and flushed
4714 * because the dirty list got full. Bring this page
4715 * back in from the map (but don't unspill it here,
4716 * leave that unless page_touch happens again).
4718 if (tx2->mt_spill_pgs) {
4719 MDB_ID pn = pgno << 1;
4720 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4721 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4722 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4727 unsigned x = mdb_mid2l_search(dl, pgno);
4728 if (x <= dl[0].mid && dl[x].mid == pgno) {
4734 } while ((tx2 = tx2->mt_parent) != NULL);
4737 if (pgno < txn->mt_next_pgno) {
4739 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4741 DPRINTF(("page %"Z"u not found", pgno));
4742 txn->mt_flags |= MDB_TXN_ERROR;
4743 return MDB_PAGE_NOTFOUND;
4753 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4754 * The cursor is at the root page, set up the rest of it.
4757 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4759 MDB_page *mp = mc->mc_pg[mc->mc_top];
4763 while (IS_BRANCH(mp)) {
4767 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4768 mdb_cassert(mc, NUMKEYS(mp) > 1);
4769 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4771 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4773 if (flags & MDB_PS_LAST)
4774 i = NUMKEYS(mp) - 1;
4777 node = mdb_node_search(mc, key, &exact);
4779 i = NUMKEYS(mp) - 1;
4781 i = mc->mc_ki[mc->mc_top];
4783 mdb_cassert(mc, i > 0);
4787 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4790 mdb_cassert(mc, i < NUMKEYS(mp));
4791 node = NODEPTR(mp, i);
4793 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4796 mc->mc_ki[mc->mc_top] = i;
4797 if ((rc = mdb_cursor_push(mc, mp)))
4800 if (flags & MDB_PS_MODIFY) {
4801 if ((rc = mdb_page_touch(mc)) != 0)
4803 mp = mc->mc_pg[mc->mc_top];
4808 DPRINTF(("internal error, index points to a %02X page!?",
4810 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4811 return MDB_CORRUPTED;
4814 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4815 key ? DKEY(key) : "null"));
4816 mc->mc_flags |= C_INITIALIZED;
4817 mc->mc_flags &= ~C_EOF;
4822 /** Search for the lowest key under the current branch page.
4823 * This just bypasses a NUMKEYS check in the current page
4824 * before calling mdb_page_search_root(), because the callers
4825 * are all in situations where the current page is known to
4829 mdb_page_search_lowest(MDB_cursor *mc)
4831 MDB_page *mp = mc->mc_pg[mc->mc_top];
4832 MDB_node *node = NODEPTR(mp, 0);
4835 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4838 mc->mc_ki[mc->mc_top] = 0;
4839 if ((rc = mdb_cursor_push(mc, mp)))
4841 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4844 /** Search for the page a given key should be in.
4845 * Push it and its parent pages on the cursor stack.
4846 * @param[in,out] mc the cursor for this operation.
4847 * @param[in] key the key to search for, or NULL for first/last page.
4848 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4849 * are touched (updated with new page numbers).
4850 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4851 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4852 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4853 * @return 0 on success, non-zero on failure.
4856 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4861 /* Make sure the txn is still viable, then find the root from
4862 * the txn's db table and set it as the root of the cursor's stack.
4864 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4865 DPUTS("transaction has failed, must abort");
4868 /* Make sure we're using an up-to-date root */
4869 if (*mc->mc_dbflag & DB_STALE) {
4871 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4872 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4879 MDB_node *leaf = mdb_node_search(&mc2,
4880 &mc->mc_dbx->md_name, &exact);
4882 return MDB_NOTFOUND;
4883 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4886 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4888 /* The txn may not know this DBI, or another process may
4889 * have dropped and recreated the DB with other flags.
4891 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4892 return MDB_INCOMPATIBLE;
4893 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4895 *mc->mc_dbflag &= ~DB_STALE;
4897 root = mc->mc_db->md_root;
4899 if (root == P_INVALID) { /* Tree is empty. */
4900 DPUTS("tree is empty");
4901 return MDB_NOTFOUND;
4905 mdb_cassert(mc, root > 1);
4906 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4907 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4913 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4914 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4916 if (flags & MDB_PS_MODIFY) {
4917 if ((rc = mdb_page_touch(mc)))
4921 if (flags & MDB_PS_ROOTONLY)
4924 return mdb_page_search_root(mc, key, flags);
4928 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4930 MDB_txn *txn = mc->mc_txn;
4931 pgno_t pg = mp->mp_pgno;
4932 unsigned x = 0, ovpages = mp->mp_pages;
4933 MDB_env *env = txn->mt_env;
4934 MDB_IDL sl = txn->mt_spill_pgs;
4935 MDB_ID pn = pg << 1;
4938 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4939 /* If the page is dirty or on the spill list we just acquired it,
4940 * so we should give it back to our current free list, if any.
4941 * Otherwise put it onto the list of pages we freed in this txn.
4943 * Won't create me_pghead: me_pglast must be inited along with it.
4944 * Unsupported in nested txns: They would need to hide the page
4945 * range in ancestor txns' dirty and spilled lists.
4947 if (env->me_pghead &&
4949 ((mp->mp_flags & P_DIRTY) ||
4950 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4954 MDB_ID2 *dl, ix, iy;
4955 rc = mdb_midl_need(&env->me_pghead, ovpages);
4958 if (!(mp->mp_flags & P_DIRTY)) {
4959 /* This page is no longer spilled */
4966 /* Remove from dirty list */
4967 dl = txn->mt_u.dirty_list;
4969 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4975 mdb_cassert(mc, x > 1);
4977 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4978 txn->mt_flags |= MDB_TXN_ERROR;
4979 return MDB_CORRUPTED;
4982 if (!(env->me_flags & MDB_WRITEMAP))
4983 mdb_dpage_free(env, mp);
4985 /* Insert in me_pghead */
4986 mop = env->me_pghead;
4987 j = mop[0] + ovpages;
4988 for (i = mop[0]; i && mop[i] < pg; i--)
4994 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4998 mc->mc_db->md_overflow_pages -= ovpages;
5002 /** Return the data associated with a given node.
5003 * @param[in] txn The transaction for this operation.
5004 * @param[in] leaf The node being read.
5005 * @param[out] data Updated to point to the node's data.
5006 * @return 0 on success, non-zero on failure.
5009 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5011 MDB_page *omp; /* overflow page */
5015 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5016 data->mv_size = NODEDSZ(leaf);
5017 data->mv_data = NODEDATA(leaf);
5021 /* Read overflow data.
5023 data->mv_size = NODEDSZ(leaf);
5024 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5025 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5026 DPRINTF(("read overflow page %"Z"u failed", pgno));
5029 data->mv_data = METADATA(omp);
5035 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5036 MDB_val *key, MDB_val *data)
5043 if (key == NULL || data == NULL)
5046 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5048 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5051 if (txn->mt_flags & MDB_TXN_ERROR)
5054 mdb_cursor_init(&mc, txn, dbi, &mx);
5055 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5058 /** Find a sibling for a page.
5059 * Replaces the page at the top of the cursor's stack with the
5060 * specified sibling, if one exists.
5061 * @param[in] mc The cursor for this operation.
5062 * @param[in] move_right Non-zero if the right sibling is requested,
5063 * otherwise the left sibling.
5064 * @return 0 on success, non-zero on failure.
5067 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5073 if (mc->mc_snum < 2) {
5074 return MDB_NOTFOUND; /* root has no siblings */
5078 DPRINTF(("parent page is page %"Z"u, index %u",
5079 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5081 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5082 : (mc->mc_ki[mc->mc_top] == 0)) {
5083 DPRINTF(("no more keys left, moving to %s sibling",
5084 move_right ? "right" : "left"));
5085 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5086 /* undo cursor_pop before returning */
5093 mc->mc_ki[mc->mc_top]++;
5095 mc->mc_ki[mc->mc_top]--;
5096 DPRINTF(("just moving to %s index key %u",
5097 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5099 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5101 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5102 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5103 /* mc will be inconsistent if caller does mc_snum++ as above */
5104 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5108 mdb_cursor_push(mc, mp);
5110 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5115 /** Move the cursor to the next data item. */
5117 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5123 if (mc->mc_flags & C_EOF) {
5124 return MDB_NOTFOUND;
5127 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5129 mp = mc->mc_pg[mc->mc_top];
5131 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5132 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5133 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5134 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5135 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5136 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5137 if (rc == MDB_SUCCESS)
5138 MDB_GET_KEY(leaf, key);
5143 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5144 if (op == MDB_NEXT_DUP)
5145 return MDB_NOTFOUND;
5149 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5150 mdb_dbg_pgno(mp), (void *) mc));
5151 if (mc->mc_flags & C_DEL)
5154 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5155 DPUTS("=====> move to next sibling page");
5156 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5157 mc->mc_flags |= C_EOF;
5160 mp = mc->mc_pg[mc->mc_top];
5161 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5163 mc->mc_ki[mc->mc_top]++;
5166 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5167 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5170 key->mv_size = mc->mc_db->md_pad;
5171 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5175 mdb_cassert(mc, IS_LEAF(mp));
5176 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5178 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5179 mdb_xcursor_init1(mc, leaf);
5182 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5185 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5186 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5187 if (rc != MDB_SUCCESS)
5192 MDB_GET_KEY(leaf, key);
5196 /** Move the cursor to the previous data item. */
5198 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5204 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5206 mp = mc->mc_pg[mc->mc_top];
5208 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5209 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5210 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5211 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5212 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5213 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5214 if (rc == MDB_SUCCESS)
5215 MDB_GET_KEY(leaf, key);
5219 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5220 if (op == MDB_PREV_DUP)
5221 return MDB_NOTFOUND;
5226 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5227 mdb_dbg_pgno(mp), (void *) mc));
5229 if (mc->mc_ki[mc->mc_top] == 0) {
5230 DPUTS("=====> move to prev sibling page");
5231 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5234 mp = mc->mc_pg[mc->mc_top];
5235 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5236 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5238 mc->mc_ki[mc->mc_top]--;
5240 mc->mc_flags &= ~C_EOF;
5242 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5243 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5246 key->mv_size = mc->mc_db->md_pad;
5247 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5251 mdb_cassert(mc, IS_LEAF(mp));
5252 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5254 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5255 mdb_xcursor_init1(mc, leaf);
5258 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5261 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5262 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5263 if (rc != MDB_SUCCESS)
5268 MDB_GET_KEY(leaf, key);
5272 /** Set the cursor on a specific data item. */
5274 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5275 MDB_cursor_op op, int *exactp)
5279 MDB_node *leaf = NULL;
5283 mdb_cassert(mc, key);
5284 if (key->mv_size == 0)
5285 return MDB_BAD_VALSIZE;
5288 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5290 /* See if we're already on the right page */
5291 if (mc->mc_flags & C_INITIALIZED) {
5294 mp = mc->mc_pg[mc->mc_top];
5296 mc->mc_ki[mc->mc_top] = 0;
5297 return MDB_NOTFOUND;
5299 if (mp->mp_flags & P_LEAF2) {
5300 nodekey.mv_size = mc->mc_db->md_pad;
5301 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5303 leaf = NODEPTR(mp, 0);
5304 MDB_GET_KEY2(leaf, nodekey);
5306 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5308 /* Probably happens rarely, but first node on the page
5309 * was the one we wanted.
5311 mc->mc_ki[mc->mc_top] = 0;
5318 unsigned int nkeys = NUMKEYS(mp);
5320 if (mp->mp_flags & P_LEAF2) {
5321 nodekey.mv_data = LEAF2KEY(mp,
5322 nkeys-1, nodekey.mv_size);
5324 leaf = NODEPTR(mp, nkeys-1);
5325 MDB_GET_KEY2(leaf, nodekey);
5327 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5329 /* last node was the one we wanted */
5330 mc->mc_ki[mc->mc_top] = nkeys-1;
5336 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5337 /* This is definitely the right page, skip search_page */
5338 if (mp->mp_flags & P_LEAF2) {
5339 nodekey.mv_data = LEAF2KEY(mp,
5340 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5342 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5343 MDB_GET_KEY2(leaf, nodekey);
5345 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5347 /* current node was the one we wanted */
5357 /* If any parents have right-sibs, search.
5358 * Otherwise, there's nothing further.
5360 for (i=0; i<mc->mc_top; i++)
5362 NUMKEYS(mc->mc_pg[i])-1)
5364 if (i == mc->mc_top) {
5365 /* There are no other pages */
5366 mc->mc_ki[mc->mc_top] = nkeys;
5367 return MDB_NOTFOUND;
5371 /* There are no other pages */
5372 mc->mc_ki[mc->mc_top] = 0;
5373 if (op == MDB_SET_RANGE) {
5377 return MDB_NOTFOUND;
5381 rc = mdb_page_search(mc, key, 0);
5382 if (rc != MDB_SUCCESS)
5385 mp = mc->mc_pg[mc->mc_top];
5386 mdb_cassert(mc, IS_LEAF(mp));
5389 leaf = mdb_node_search(mc, key, exactp);
5390 if (exactp != NULL && !*exactp) {
5391 /* MDB_SET specified and not an exact match. */
5392 return MDB_NOTFOUND;
5396 DPUTS("===> inexact leaf not found, goto sibling");
5397 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5398 return rc; /* no entries matched */
5399 mp = mc->mc_pg[mc->mc_top];
5400 mdb_cassert(mc, IS_LEAF(mp));
5401 leaf = NODEPTR(mp, 0);
5405 mc->mc_flags |= C_INITIALIZED;
5406 mc->mc_flags &= ~C_EOF;
5409 key->mv_size = mc->mc_db->md_pad;
5410 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5414 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5415 mdb_xcursor_init1(mc, leaf);
5418 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5419 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5420 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5423 if (op == MDB_GET_BOTH) {
5429 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5430 if (rc != MDB_SUCCESS)
5433 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5435 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5437 rc = mc->mc_dbx->md_dcmp(data, &d2);
5439 if (op == MDB_GET_BOTH || rc > 0)
5440 return MDB_NOTFOUND;
5447 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5448 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5453 /* The key already matches in all other cases */
5454 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5455 MDB_GET_KEY(leaf, key);
5456 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5461 /** Move the cursor to the first item in the database. */
5463 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5469 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5471 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5472 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5473 if (rc != MDB_SUCCESS)
5476 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5478 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5479 mc->mc_flags |= C_INITIALIZED;
5480 mc->mc_flags &= ~C_EOF;
5482 mc->mc_ki[mc->mc_top] = 0;
5484 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5485 key->mv_size = mc->mc_db->md_pad;
5486 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5491 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5492 mdb_xcursor_init1(mc, leaf);
5493 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5497 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5501 MDB_GET_KEY(leaf, key);
5505 /** Move the cursor to the last item in the database. */
5507 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5513 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5515 if (!(mc->mc_flags & C_EOF)) {
5517 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5518 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5519 if (rc != MDB_SUCCESS)
5522 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5525 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5526 mc->mc_flags |= C_INITIALIZED|C_EOF;
5527 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5529 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5530 key->mv_size = mc->mc_db->md_pad;
5531 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5536 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5537 mdb_xcursor_init1(mc, leaf);
5538 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5542 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5547 MDB_GET_KEY(leaf, key);
5552 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5557 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5562 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5566 case MDB_GET_CURRENT:
5567 if (!(mc->mc_flags & C_INITIALIZED)) {
5570 MDB_page *mp = mc->mc_pg[mc->mc_top];
5571 int nkeys = NUMKEYS(mp);
5572 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5573 mc->mc_ki[mc->mc_top] = nkeys;
5579 key->mv_size = mc->mc_db->md_pad;
5580 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5582 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5583 MDB_GET_KEY(leaf, key);
5585 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5586 if (mc->mc_flags & C_DEL)
5587 mdb_xcursor_init1(mc, leaf);
5588 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5590 rc = mdb_node_read(mc->mc_txn, leaf, data);
5597 case MDB_GET_BOTH_RANGE:
5602 if (mc->mc_xcursor == NULL) {
5603 rc = MDB_INCOMPATIBLE;
5613 rc = mdb_cursor_set(mc, key, data, op,
5614 op == MDB_SET_RANGE ? NULL : &exact);
5617 case MDB_GET_MULTIPLE:
5618 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5622 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5623 rc = MDB_INCOMPATIBLE;
5627 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5628 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5631 case MDB_NEXT_MULTIPLE:
5636 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5637 rc = MDB_INCOMPATIBLE;
5640 if (!(mc->mc_flags & C_INITIALIZED))
5641 rc = mdb_cursor_first(mc, key, data);
5643 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5644 if (rc == MDB_SUCCESS) {
5645 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5648 mx = &mc->mc_xcursor->mx_cursor;
5649 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5651 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5652 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5660 case MDB_NEXT_NODUP:
5661 if (!(mc->mc_flags & C_INITIALIZED))
5662 rc = mdb_cursor_first(mc, key, data);
5664 rc = mdb_cursor_next(mc, key, data, op);
5668 case MDB_PREV_NODUP:
5669 if (!(mc->mc_flags & C_INITIALIZED)) {
5670 rc = mdb_cursor_last(mc, key, data);
5673 mc->mc_flags |= C_INITIALIZED;
5674 mc->mc_ki[mc->mc_top]++;
5676 rc = mdb_cursor_prev(mc, key, data, op);
5679 rc = mdb_cursor_first(mc, key, data);
5682 mfunc = mdb_cursor_first;
5684 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5688 if (mc->mc_xcursor == NULL) {
5689 rc = MDB_INCOMPATIBLE;
5692 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5696 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5699 rc = mdb_cursor_last(mc, key, data);
5702 mfunc = mdb_cursor_last;
5705 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5710 if (mc->mc_flags & C_DEL)
5711 mc->mc_flags ^= C_DEL;
5716 /** Touch all the pages in the cursor stack. Set mc_top.
5717 * Makes sure all the pages are writable, before attempting a write operation.
5718 * @param[in] mc The cursor to operate on.
5721 mdb_cursor_touch(MDB_cursor *mc)
5723 int rc = MDB_SUCCESS;
5725 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5728 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5729 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5732 *mc->mc_dbflag |= DB_DIRTY;
5737 rc = mdb_page_touch(mc);
5738 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5739 mc->mc_top = mc->mc_snum-1;
5744 /** Do not spill pages to disk if txn is getting full, may fail instead */
5745 #define MDB_NOSPILL 0x8000
5748 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5751 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5752 MDB_env *env = mc->mc_txn->mt_env;
5753 MDB_node *leaf = NULL;
5756 MDB_val xdata, *rdata, dkey, olddata;
5758 int do_sub = 0, insert;
5759 unsigned int mcount = 0, dcount = 0, nospill;
5762 unsigned int nflags;
5765 /* Check this first so counter will always be zero on any
5768 if (flags & MDB_MULTIPLE) {
5769 dcount = data[1].mv_size;
5770 data[1].mv_size = 0;
5771 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5772 return MDB_INCOMPATIBLE;
5775 nospill = flags & MDB_NOSPILL;
5776 flags &= ~MDB_NOSPILL;
5778 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5779 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5781 if (flags != MDB_CURRENT && key->mv_size-1 >= ENV_MAXKEY(env))
5782 return MDB_BAD_VALSIZE;
5784 #if SIZE_MAX > MAXDATASIZE
5785 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5786 return MDB_BAD_VALSIZE;
5788 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5789 return MDB_BAD_VALSIZE;
5792 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5793 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5797 if (flags == MDB_CURRENT) {
5798 if (!(mc->mc_flags & C_INITIALIZED))
5801 } else if (mc->mc_db->md_root == P_INVALID) {
5802 /* new database, cursor has nothing to point to */
5805 mc->mc_flags &= ~C_INITIALIZED;
5810 if (flags & MDB_APPEND) {
5812 rc = mdb_cursor_last(mc, &k2, &d2);
5814 rc = mc->mc_dbx->md_cmp(key, &k2);
5817 mc->mc_ki[mc->mc_top]++;
5819 /* new key is <= last key */
5824 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5826 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5827 DPRINTF(("duplicate key [%s]", DKEY(key)));
5829 return MDB_KEYEXIST;
5831 if (rc && rc != MDB_NOTFOUND)
5835 if (mc->mc_flags & C_DEL)
5836 mc->mc_flags ^= C_DEL;
5838 /* Cursor is positioned, check for room in the dirty list */
5840 if (flags & MDB_MULTIPLE) {
5842 xdata.mv_size = data->mv_size * dcount;
5846 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5850 if (rc == MDB_NO_ROOT) {
5852 /* new database, write a root leaf page */
5853 DPUTS("allocating new root leaf page");
5854 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5857 mdb_cursor_push(mc, np);
5858 mc->mc_db->md_root = np->mp_pgno;
5859 mc->mc_db->md_depth++;
5860 *mc->mc_dbflag |= DB_DIRTY;
5861 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5863 np->mp_flags |= P_LEAF2;
5864 mc->mc_flags |= C_INITIALIZED;
5866 /* make sure all cursor pages are writable */
5867 rc2 = mdb_cursor_touch(mc);
5874 /* The key does not exist */
5875 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5876 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5877 LEAFSIZE(key, data) > env->me_nodemax)
5879 /* Too big for a node, insert in sub-DB */
5880 fp_flags = P_LEAF|P_DIRTY;
5882 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5883 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5887 /* there's only a key anyway, so this is a no-op */
5888 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5889 unsigned int ksize = mc->mc_db->md_pad;
5890 if (key->mv_size != ksize)
5891 return MDB_BAD_VALSIZE;
5892 if (flags == MDB_CURRENT) {
5893 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5894 memcpy(ptr, key->mv_data, ksize);
5900 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5901 olddata.mv_size = NODEDSZ(leaf);
5902 olddata.mv_data = NODEDATA(leaf);
5905 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5906 /* Prepare (sub-)page/sub-DB to accept the new item,
5907 * if needed. fp: old sub-page or a header faking
5908 * it. mp: new (sub-)page. offset: growth in page
5909 * size. xdata: node data with new page or DB.
5911 ssize_t i, offset = 0;
5912 mp = fp = xdata.mv_data = env->me_pbuf;
5913 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5915 /* Was a single item before, must convert now */
5916 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5917 /* Just overwrite the current item */
5918 if (flags == MDB_CURRENT)
5921 #if UINT_MAX < SIZE_MAX
5922 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5923 #ifdef MISALIGNED_OK
5924 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5926 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5929 /* if data matches, skip it */
5930 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5931 if (flags & MDB_NODUPDATA)
5933 else if (flags & MDB_MULTIPLE)
5940 /* Back up original data item */
5941 dkey.mv_size = olddata.mv_size;
5942 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5944 /* Make sub-page header for the dup items, with dummy body */
5945 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5946 fp->mp_lower = PAGEHDRSZ;
5947 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5948 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5949 fp->mp_flags |= P_LEAF2;
5950 fp->mp_pad = data->mv_size;
5951 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5953 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5954 (dkey.mv_size & 1) + (data->mv_size & 1);
5956 fp->mp_upper = xdata.mv_size;
5957 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
5958 } else if (leaf->mn_flags & F_SUBDATA) {
5959 /* Data is on sub-DB, just store it */
5960 flags |= F_DUPDATA|F_SUBDATA;
5963 /* Data is on sub-page */
5964 fp = olddata.mv_data;
5967 i = -(ssize_t)SIZELEFT(fp);
5968 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5969 offset = i += (ssize_t) EVEN(
5970 sizeof(indx_t) + NODESIZE + data->mv_size);
5972 i += offset = fp->mp_pad;
5973 offset *= 4; /* space for 4 more */
5977 /* FALLTHRU: Sub-page is big enough */
5979 fp->mp_flags |= P_DIRTY;
5980 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
5981 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5985 xdata.mv_size = olddata.mv_size + offset;
5988 fp_flags = fp->mp_flags;
5989 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
5990 /* Too big for a sub-page, convert to sub-DB */
5991 fp_flags &= ~P_SUBP;
5993 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5994 fp_flags |= P_LEAF2;
5995 dummy.md_pad = fp->mp_pad;
5996 dummy.md_flags = MDB_DUPFIXED;
5997 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5998 dummy.md_flags |= MDB_INTEGERKEY;
6004 dummy.md_branch_pages = 0;
6005 dummy.md_leaf_pages = 1;
6006 dummy.md_overflow_pages = 0;
6007 dummy.md_entries = NUMKEYS(fp);
6008 xdata.mv_size = sizeof(MDB_db);
6009 xdata.mv_data = &dummy;
6010 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6012 offset = env->me_psize - olddata.mv_size;
6013 flags |= F_DUPDATA|F_SUBDATA;
6014 dummy.md_root = mp->mp_pgno;
6017 mp->mp_flags = fp_flags | P_DIRTY;
6018 mp->mp_pad = fp->mp_pad;
6019 mp->mp_lower = fp->mp_lower;
6020 mp->mp_upper = fp->mp_upper + offset;
6021 if (fp_flags & P_LEAF2) {
6022 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6024 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6025 olddata.mv_size - fp->mp_upper);
6026 for (i = NUMKEYS(fp); --i >= 0; )
6027 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6035 mdb_node_del(mc, 0);
6039 /* overflow page overwrites need special handling */
6040 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6043 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6045 memcpy(&pg, olddata.mv_data, sizeof(pg));
6046 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6048 ovpages = omp->mp_pages;
6050 /* Is the ov page large enough? */
6051 if (ovpages >= dpages) {
6052 if (!(omp->mp_flags & P_DIRTY) &&
6053 (level || (env->me_flags & MDB_WRITEMAP)))
6055 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6058 level = 0; /* dirty in this txn or clean */
6061 if (omp->mp_flags & P_DIRTY) {
6062 /* yes, overwrite it. Note in this case we don't
6063 * bother to try shrinking the page if the new data
6064 * is smaller than the overflow threshold.
6067 /* It is writable only in a parent txn */
6068 size_t sz = (size_t) env->me_psize * ovpages, off;
6069 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6075 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6076 if (!(flags & MDB_RESERVE)) {
6077 /* Copy end of page, adjusting alignment so
6078 * compiler may copy words instead of bytes.
6080 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6081 memcpy((size_t *)((char *)np + off),
6082 (size_t *)((char *)omp + off), sz - off);
6085 memcpy(np, omp, sz); /* Copy beginning of page */
6088 SETDSZ(leaf, data->mv_size);
6089 if (F_ISSET(flags, MDB_RESERVE))
6090 data->mv_data = METADATA(omp);
6092 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6096 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6098 } else if (data->mv_size == olddata.mv_size) {
6099 /* same size, just replace it. Note that we could
6100 * also reuse this node if the new data is smaller,
6101 * but instead we opt to shrink the node in that case.
6103 if (F_ISSET(flags, MDB_RESERVE))
6104 data->mv_data = olddata.mv_data;
6105 else if (data->mv_size)
6106 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6108 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6111 mdb_node_del(mc, 0);
6112 mc->mc_db->md_entries--;
6118 nflags = flags & NODE_ADD_FLAGS;
6119 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6120 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6121 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6122 nflags &= ~MDB_APPEND;
6124 nflags |= MDB_SPLIT_REPLACE;
6125 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6127 /* There is room already in this leaf page. */
6128 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6129 if (rc == 0 && !do_sub && insert) {
6130 /* Adjust other cursors pointing to mp */
6131 MDB_cursor *m2, *m3;
6132 MDB_dbi dbi = mc->mc_dbi;
6133 unsigned i = mc->mc_top;
6134 MDB_page *mp = mc->mc_pg[i];
6136 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6137 if (mc->mc_flags & C_SUB)
6138 m3 = &m2->mc_xcursor->mx_cursor;
6141 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6142 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6149 if (rc != MDB_SUCCESS)
6150 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6152 /* Now store the actual data in the child DB. Note that we're
6153 * storing the user data in the keys field, so there are strict
6154 * size limits on dupdata. The actual data fields of the child
6155 * DB are all zero size.
6162 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6163 if (flags & MDB_CURRENT) {
6164 xflags = MDB_CURRENT|MDB_NOSPILL;
6166 mdb_xcursor_init1(mc, leaf);
6167 xflags = (flags & MDB_NODUPDATA) ?
6168 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6170 /* converted, write the original data first */
6172 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6176 /* Adjust other cursors pointing to mp */
6178 unsigned i = mc->mc_top;
6179 MDB_page *mp = mc->mc_pg[i];
6181 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6182 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6183 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6184 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6185 mdb_xcursor_init1(m2, leaf);
6189 /* we've done our job */
6192 if (flags & MDB_APPENDDUP)
6193 xflags |= MDB_APPEND;
6194 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6195 if (flags & F_SUBDATA) {
6196 void *db = NODEDATA(leaf);
6197 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6200 /* sub-writes might have failed so check rc again.
6201 * Don't increment count if we just replaced an existing item.
6203 if (!rc && !(flags & MDB_CURRENT))
6204 mc->mc_db->md_entries++;
6205 if (flags & MDB_MULTIPLE) {
6209 /* let caller know how many succeeded, if any */
6210 data[1].mv_size = mcount;
6211 if (mcount < dcount) {
6212 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6219 /* If we succeeded and the key didn't exist before, make sure
6220 * the cursor is marked valid.
6223 mc->mc_flags |= C_INITIALIZED;
6228 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6234 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6235 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6237 if (!(mc->mc_flags & C_INITIALIZED))
6240 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6241 return MDB_NOTFOUND;
6243 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6246 rc = mdb_cursor_touch(mc);
6250 mp = mc->mc_pg[mc->mc_top];
6251 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6253 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6254 if (!(flags & MDB_NODUPDATA)) {
6255 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6256 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6258 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6259 /* If sub-DB still has entries, we're done */
6260 if (mc->mc_xcursor->mx_db.md_entries) {
6261 if (leaf->mn_flags & F_SUBDATA) {
6262 /* update subDB info */
6263 void *db = NODEDATA(leaf);
6264 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6267 /* shrink fake page */
6268 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6269 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6270 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6271 /* fix other sub-DB cursors pointed at this fake page */
6272 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6273 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6274 if (m2->mc_pg[mc->mc_top] == mp &&
6275 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6276 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6279 mc->mc_db->md_entries--;
6280 mc->mc_flags |= C_DEL;
6283 /* otherwise fall thru and delete the sub-DB */
6286 if (leaf->mn_flags & F_SUBDATA) {
6287 /* add all the child DB's pages to the free list */
6288 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6289 if (rc == MDB_SUCCESS) {
6290 mc->mc_db->md_entries -=
6291 mc->mc_xcursor->mx_db.md_entries;
6296 return mdb_cursor_del0(mc, leaf);
6299 /** Allocate and initialize new pages for a database.
6300 * @param[in] mc a cursor on the database being added to.
6301 * @param[in] flags flags defining what type of page is being allocated.
6302 * @param[in] num the number of pages to allocate. This is usually 1,
6303 * unless allocating overflow pages for a large record.
6304 * @param[out] mp Address of a page, or NULL on failure.
6305 * @return 0 on success, non-zero on failure.
6308 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6313 if ((rc = mdb_page_alloc(mc, num, &np)))
6315 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6316 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6317 np->mp_flags = flags | P_DIRTY;
6318 np->mp_lower = PAGEHDRSZ;
6319 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6322 mc->mc_db->md_branch_pages++;
6323 else if (IS_LEAF(np))
6324 mc->mc_db->md_leaf_pages++;
6325 else if (IS_OVERFLOW(np)) {
6326 mc->mc_db->md_overflow_pages += num;
6334 /** Calculate the size of a leaf node.
6335 * The size depends on the environment's page size; if a data item
6336 * is too large it will be put onto an overflow page and the node
6337 * size will only include the key and not the data. Sizes are always
6338 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6339 * of the #MDB_node headers.
6340 * @param[in] env The environment handle.
6341 * @param[in] key The key for the node.
6342 * @param[in] data The data for the node.
6343 * @return The number of bytes needed to store the node.
6346 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6350 sz = LEAFSIZE(key, data);
6351 if (sz > env->me_nodemax) {
6352 /* put on overflow page */
6353 sz -= data->mv_size - sizeof(pgno_t);
6356 return EVEN(sz + sizeof(indx_t));
6359 /** Calculate the size of a branch node.
6360 * The size should depend on the environment's page size but since
6361 * we currently don't support spilling large keys onto overflow
6362 * pages, it's simply the size of the #MDB_node header plus the
6363 * size of the key. Sizes are always rounded up to an even number
6364 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6365 * @param[in] env The environment handle.
6366 * @param[in] key The key for the node.
6367 * @return The number of bytes needed to store the node.
6370 mdb_branch_size(MDB_env *env, MDB_val *key)
6375 if (sz > env->me_nodemax) {
6376 /* put on overflow page */
6377 /* not implemented */
6378 /* sz -= key->size - sizeof(pgno_t); */
6381 return sz + sizeof(indx_t);
6384 /** Add a node to the page pointed to by the cursor.
6385 * @param[in] mc The cursor for this operation.
6386 * @param[in] indx The index on the page where the new node should be added.
6387 * @param[in] key The key for the new node.
6388 * @param[in] data The data for the new node, if any.
6389 * @param[in] pgno The page number, if adding a branch node.
6390 * @param[in] flags Flags for the node.
6391 * @return 0 on success, non-zero on failure. Possible errors are:
6393 * <li>ENOMEM - failed to allocate overflow pages for the node.
6394 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6395 * should never happen since all callers already calculate the
6396 * page's free space before calling this function.
6400 mdb_node_add(MDB_cursor *mc, indx_t indx,
6401 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6404 size_t node_size = NODESIZE;
6408 MDB_page *mp = mc->mc_pg[mc->mc_top];
6409 MDB_page *ofp = NULL; /* overflow page */
6412 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6414 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6415 IS_LEAF(mp) ? "leaf" : "branch",
6416 IS_SUBP(mp) ? "sub-" : "",
6417 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6418 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6421 /* Move higher keys up one slot. */
6422 int ksize = mc->mc_db->md_pad, dif;
6423 char *ptr = LEAF2KEY(mp, indx, ksize);
6424 dif = NUMKEYS(mp) - indx;
6426 memmove(ptr+ksize, ptr, dif*ksize);
6427 /* insert new key */
6428 memcpy(ptr, key->mv_data, ksize);
6430 /* Just using these for counting */
6431 mp->mp_lower += sizeof(indx_t);
6432 mp->mp_upper -= ksize - sizeof(indx_t);
6436 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6438 node_size += key->mv_size;
6440 mdb_cassert(mc, data);
6441 if (F_ISSET(flags, F_BIGDATA)) {
6442 /* Data already on overflow page. */
6443 node_size += sizeof(pgno_t);
6444 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6445 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6447 /* Put data on overflow page. */
6448 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6449 data->mv_size, node_size+data->mv_size));
6450 node_size = EVEN(node_size + sizeof(pgno_t));
6451 if ((ssize_t)node_size > room)
6453 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6455 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6459 node_size += data->mv_size;
6462 node_size = EVEN(node_size);
6463 if ((ssize_t)node_size > room)
6467 /* Move higher pointers up one slot. */
6468 for (i = NUMKEYS(mp); i > indx; i--)
6469 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6471 /* Adjust free space offsets. */
6472 ofs = mp->mp_upper - node_size;
6473 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6474 mp->mp_ptrs[indx] = ofs;
6476 mp->mp_lower += sizeof(indx_t);
6478 /* Write the node data. */
6479 node = NODEPTR(mp, indx);
6480 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6481 node->mn_flags = flags;
6483 SETDSZ(node,data->mv_size);
6488 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6491 mdb_cassert(mc, key);
6493 if (F_ISSET(flags, F_BIGDATA))
6494 memcpy(node->mn_data + key->mv_size, data->mv_data,
6496 else if (F_ISSET(flags, MDB_RESERVE))
6497 data->mv_data = node->mn_data + key->mv_size;
6499 memcpy(node->mn_data + key->mv_size, data->mv_data,
6502 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6504 if (F_ISSET(flags, MDB_RESERVE))
6505 data->mv_data = METADATA(ofp);
6507 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6514 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6515 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6516 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6517 DPRINTF(("node size = %"Z"u", node_size));
6518 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6519 return MDB_PAGE_FULL;
6522 /** Delete the specified node from a page.
6523 * @param[in] mp The page to operate on.
6524 * @param[in] indx The index of the node to delete.
6525 * @param[in] ksize The size of a node. Only used if the page is
6526 * part of a #MDB_DUPFIXED database.
6529 mdb_node_del(MDB_cursor *mc, int ksize)
6531 MDB_page *mp = mc->mc_pg[mc->mc_top];
6532 indx_t indx = mc->mc_ki[mc->mc_top];
6534 indx_t i, j, numkeys, ptr;
6538 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6539 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6540 numkeys = NUMKEYS(mp);
6541 mdb_cassert(mc, indx < numkeys);
6544 int x = numkeys - 1 - indx;
6545 base = LEAF2KEY(mp, indx, ksize);
6547 memmove(base, base + ksize, x * ksize);
6548 mp->mp_lower -= sizeof(indx_t);
6549 mp->mp_upper += ksize - sizeof(indx_t);
6553 node = NODEPTR(mp, indx);
6554 sz = NODESIZE + node->mn_ksize;
6556 if (F_ISSET(node->mn_flags, F_BIGDATA))
6557 sz += sizeof(pgno_t);
6559 sz += NODEDSZ(node);
6563 ptr = mp->mp_ptrs[indx];
6564 for (i = j = 0; i < numkeys; i++) {
6566 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6567 if (mp->mp_ptrs[i] < ptr)
6568 mp->mp_ptrs[j] += sz;
6573 base = (char *)mp + mp->mp_upper;
6574 memmove(base + sz, base, ptr - mp->mp_upper);
6576 mp->mp_lower -= sizeof(indx_t);
6580 /** Compact the main page after deleting a node on a subpage.
6581 * @param[in] mp The main page to operate on.
6582 * @param[in] indx The index of the subpage on the main page.
6585 mdb_node_shrink(MDB_page *mp, indx_t indx)
6591 indx_t i, numkeys, ptr;
6593 node = NODEPTR(mp, indx);
6594 sp = (MDB_page *)NODEDATA(node);
6595 delta = SIZELEFT(sp);
6596 xp = (MDB_page *)((char *)sp + delta);
6598 /* shift subpage upward */
6600 nsize = NUMKEYS(sp) * sp->mp_pad;
6602 return; /* do not make the node uneven-sized */
6603 memmove(METADATA(xp), METADATA(sp), nsize);
6606 numkeys = NUMKEYS(sp);
6607 for (i=numkeys-1; i>=0; i--)
6608 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6610 xp->mp_upper = sp->mp_lower;
6611 xp->mp_lower = sp->mp_lower;
6612 xp->mp_flags = sp->mp_flags;
6613 xp->mp_pad = sp->mp_pad;
6614 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6616 nsize = NODEDSZ(node) - delta;
6617 SETDSZ(node, nsize);
6619 /* shift lower nodes upward */
6620 ptr = mp->mp_ptrs[indx];
6621 numkeys = NUMKEYS(mp);
6622 for (i = 0; i < numkeys; i++) {
6623 if (mp->mp_ptrs[i] <= ptr)
6624 mp->mp_ptrs[i] += delta;
6627 base = (char *)mp + mp->mp_upper;
6628 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6629 mp->mp_upper += delta;
6632 /** Initial setup of a sorted-dups cursor.
6633 * Sorted duplicates are implemented as a sub-database for the given key.
6634 * The duplicate data items are actually keys of the sub-database.
6635 * Operations on the duplicate data items are performed using a sub-cursor
6636 * initialized when the sub-database is first accessed. This function does
6637 * the preliminary setup of the sub-cursor, filling in the fields that
6638 * depend only on the parent DB.
6639 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6642 mdb_xcursor_init0(MDB_cursor *mc)
6644 MDB_xcursor *mx = mc->mc_xcursor;
6646 mx->mx_cursor.mc_xcursor = NULL;
6647 mx->mx_cursor.mc_txn = mc->mc_txn;
6648 mx->mx_cursor.mc_db = &mx->mx_db;
6649 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6650 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6651 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6652 mx->mx_cursor.mc_snum = 0;
6653 mx->mx_cursor.mc_top = 0;
6654 mx->mx_cursor.mc_flags = C_SUB;
6655 mx->mx_dbx.md_name.mv_size = 0;
6656 mx->mx_dbx.md_name.mv_data = NULL;
6657 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6658 mx->mx_dbx.md_dcmp = NULL;
6659 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6662 /** Final setup of a sorted-dups cursor.
6663 * Sets up the fields that depend on the data from the main cursor.
6664 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6665 * @param[in] node The data containing the #MDB_db record for the
6666 * sorted-dup database.
6669 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6671 MDB_xcursor *mx = mc->mc_xcursor;
6673 if (node->mn_flags & F_SUBDATA) {
6674 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6675 mx->mx_cursor.mc_pg[0] = 0;
6676 mx->mx_cursor.mc_snum = 0;
6677 mx->mx_cursor.mc_top = 0;
6678 mx->mx_cursor.mc_flags = C_SUB;
6680 MDB_page *fp = NODEDATA(node);
6681 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6682 mx->mx_db.md_flags = 0;
6683 mx->mx_db.md_depth = 1;
6684 mx->mx_db.md_branch_pages = 0;
6685 mx->mx_db.md_leaf_pages = 1;
6686 mx->mx_db.md_overflow_pages = 0;
6687 mx->mx_db.md_entries = NUMKEYS(fp);
6688 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6689 mx->mx_cursor.mc_snum = 1;
6690 mx->mx_cursor.mc_top = 0;
6691 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6692 mx->mx_cursor.mc_pg[0] = fp;
6693 mx->mx_cursor.mc_ki[0] = 0;
6694 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6695 mx->mx_db.md_flags = MDB_DUPFIXED;
6696 mx->mx_db.md_pad = fp->mp_pad;
6697 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6698 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6701 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6702 mx->mx_db.md_root));
6703 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6704 #if UINT_MAX < SIZE_MAX
6705 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6706 #ifdef MISALIGNED_OK
6707 mx->mx_dbx.md_cmp = mdb_cmp_long;
6709 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6714 /** Initialize a cursor for a given transaction and database. */
6716 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6719 mc->mc_backup = NULL;
6722 mc->mc_db = &txn->mt_dbs[dbi];
6723 mc->mc_dbx = &txn->mt_dbxs[dbi];
6724 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6729 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6730 mdb_tassert(txn, mx != NULL);
6731 mc->mc_xcursor = mx;
6732 mdb_xcursor_init0(mc);
6734 mc->mc_xcursor = NULL;
6736 if (*mc->mc_dbflag & DB_STALE) {
6737 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6742 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6745 size_t size = sizeof(MDB_cursor);
6747 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6750 if (txn->mt_flags & MDB_TXN_ERROR)
6753 /* Allow read access to the freelist */
6754 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6757 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6758 size += sizeof(MDB_xcursor);
6760 if ((mc = malloc(size)) != NULL) {
6761 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6762 if (txn->mt_cursors) {
6763 mc->mc_next = txn->mt_cursors[dbi];
6764 txn->mt_cursors[dbi] = mc;
6765 mc->mc_flags |= C_UNTRACK;
6777 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6779 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6782 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6785 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6789 /* Return the count of duplicate data items for the current key */
6791 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6795 if (mc == NULL || countp == NULL)
6798 if (mc->mc_xcursor == NULL)
6799 return MDB_INCOMPATIBLE;
6801 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6802 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6805 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6808 *countp = mc->mc_xcursor->mx_db.md_entries;
6814 mdb_cursor_close(MDB_cursor *mc)
6816 if (mc && !mc->mc_backup) {
6817 /* remove from txn, if tracked */
6818 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6819 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6820 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6822 *prev = mc->mc_next;
6829 mdb_cursor_txn(MDB_cursor *mc)
6831 if (!mc) return NULL;
6836 mdb_cursor_dbi(MDB_cursor *mc)
6842 /** Replace the key for a branch node with a new key.
6843 * @param[in] mc Cursor pointing to the node to operate on.
6844 * @param[in] key The new key to use.
6845 * @return 0 on success, non-zero on failure.
6848 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6854 int delta, ksize, oksize;
6855 indx_t ptr, i, numkeys, indx;
6858 indx = mc->mc_ki[mc->mc_top];
6859 mp = mc->mc_pg[mc->mc_top];
6860 node = NODEPTR(mp, indx);
6861 ptr = mp->mp_ptrs[indx];
6865 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6866 k2.mv_data = NODEKEY(node);
6867 k2.mv_size = node->mn_ksize;
6868 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6870 mdb_dkey(&k2, kbuf2),
6876 /* Sizes must be 2-byte aligned. */
6877 ksize = EVEN(key->mv_size);
6878 oksize = EVEN(node->mn_ksize);
6879 delta = ksize - oksize;
6881 /* Shift node contents if EVEN(key length) changed. */
6883 if (delta > 0 && SIZELEFT(mp) < delta) {
6885 /* not enough space left, do a delete and split */
6886 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6887 pgno = NODEPGNO(node);
6888 mdb_node_del(mc, 0);
6889 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6892 numkeys = NUMKEYS(mp);
6893 for (i = 0; i < numkeys; i++) {
6894 if (mp->mp_ptrs[i] <= ptr)
6895 mp->mp_ptrs[i] -= delta;
6898 base = (char *)mp + mp->mp_upper;
6899 len = ptr - mp->mp_upper + NODESIZE;
6900 memmove(base - delta, base, len);
6901 mp->mp_upper -= delta;
6903 node = NODEPTR(mp, indx);
6906 /* But even if no shift was needed, update ksize */
6907 if (node->mn_ksize != key->mv_size)
6908 node->mn_ksize = key->mv_size;
6911 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6917 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6919 /** Move a node from csrc to cdst.
6922 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6929 unsigned short flags;
6933 /* Mark src and dst as dirty. */
6934 if ((rc = mdb_page_touch(csrc)) ||
6935 (rc = mdb_page_touch(cdst)))
6938 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6939 key.mv_size = csrc->mc_db->md_pad;
6940 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6942 data.mv_data = NULL;
6946 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6947 mdb_cassert(csrc, !((size_t)srcnode & 1));
6948 srcpg = NODEPGNO(srcnode);
6949 flags = srcnode->mn_flags;
6950 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6951 unsigned int snum = csrc->mc_snum;
6953 /* must find the lowest key below src */
6954 mdb_page_search_lowest(csrc);
6955 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6956 key.mv_size = csrc->mc_db->md_pad;
6957 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6959 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6960 key.mv_size = NODEKSZ(s2);
6961 key.mv_data = NODEKEY(s2);
6963 csrc->mc_snum = snum--;
6964 csrc->mc_top = snum;
6966 key.mv_size = NODEKSZ(srcnode);
6967 key.mv_data = NODEKEY(srcnode);
6969 data.mv_size = NODEDSZ(srcnode);
6970 data.mv_data = NODEDATA(srcnode);
6972 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6973 unsigned int snum = cdst->mc_snum;
6976 /* must find the lowest key below dst */
6977 mdb_page_search_lowest(cdst);
6978 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6979 bkey.mv_size = cdst->mc_db->md_pad;
6980 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6982 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6983 bkey.mv_size = NODEKSZ(s2);
6984 bkey.mv_data = NODEKEY(s2);
6986 cdst->mc_snum = snum--;
6987 cdst->mc_top = snum;
6988 mdb_cursor_copy(cdst, &mn);
6990 rc = mdb_update_key(&mn, &bkey);
6995 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6996 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6997 csrc->mc_ki[csrc->mc_top],
6999 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7000 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7002 /* Add the node to the destination page.
7004 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7005 if (rc != MDB_SUCCESS)
7008 /* Delete the node from the source page.
7010 mdb_node_del(csrc, key.mv_size);
7013 /* Adjust other cursors pointing to mp */
7014 MDB_cursor *m2, *m3;
7015 MDB_dbi dbi = csrc->mc_dbi;
7016 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7018 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7019 if (csrc->mc_flags & C_SUB)
7020 m3 = &m2->mc_xcursor->mx_cursor;
7023 if (m3 == csrc) continue;
7024 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7025 csrc->mc_ki[csrc->mc_top]) {
7026 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7027 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7032 /* Update the parent separators.
7034 if (csrc->mc_ki[csrc->mc_top] == 0) {
7035 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7036 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7037 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7039 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7040 key.mv_size = NODEKSZ(srcnode);
7041 key.mv_data = NODEKEY(srcnode);
7043 DPRINTF(("update separator for source page %"Z"u to [%s]",
7044 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7045 mdb_cursor_copy(csrc, &mn);
7048 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7051 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7053 indx_t ix = csrc->mc_ki[csrc->mc_top];
7054 nullkey.mv_size = 0;
7055 csrc->mc_ki[csrc->mc_top] = 0;
7056 rc = mdb_update_key(csrc, &nullkey);
7057 csrc->mc_ki[csrc->mc_top] = ix;
7058 mdb_cassert(csrc, rc == MDB_SUCCESS);
7062 if (cdst->mc_ki[cdst->mc_top] == 0) {
7063 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7064 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7065 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7067 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7068 key.mv_size = NODEKSZ(srcnode);
7069 key.mv_data = NODEKEY(srcnode);
7071 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7072 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7073 mdb_cursor_copy(cdst, &mn);
7076 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7079 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7081 indx_t ix = cdst->mc_ki[cdst->mc_top];
7082 nullkey.mv_size = 0;
7083 cdst->mc_ki[cdst->mc_top] = 0;
7084 rc = mdb_update_key(cdst, &nullkey);
7085 cdst->mc_ki[cdst->mc_top] = ix;
7086 mdb_cassert(csrc, rc == MDB_SUCCESS);
7093 /** Merge one page into another.
7094 * The nodes from the page pointed to by \b csrc will
7095 * be copied to the page pointed to by \b cdst and then
7096 * the \b csrc page will be freed.
7097 * @param[in] csrc Cursor pointing to the source page.
7098 * @param[in] cdst Cursor pointing to the destination page.
7101 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7109 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7110 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7112 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7113 mdb_cassert(csrc, cdst->mc_snum > 1);
7115 /* Mark dst as dirty. */
7116 if ((rc = mdb_page_touch(cdst)))
7119 /* Move all nodes from src to dst.
7121 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7122 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7123 key.mv_size = csrc->mc_db->md_pad;
7124 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7125 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7126 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7127 if (rc != MDB_SUCCESS)
7129 key.mv_data = (char *)key.mv_data + key.mv_size;
7132 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7133 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7134 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7135 unsigned int snum = csrc->mc_snum;
7137 /* must find the lowest key below src */
7138 mdb_page_search_lowest(csrc);
7139 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7140 key.mv_size = csrc->mc_db->md_pad;
7141 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7143 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7144 key.mv_size = NODEKSZ(s2);
7145 key.mv_data = NODEKEY(s2);
7147 csrc->mc_snum = snum--;
7148 csrc->mc_top = snum;
7150 key.mv_size = srcnode->mn_ksize;
7151 key.mv_data = NODEKEY(srcnode);
7154 data.mv_size = NODEDSZ(srcnode);
7155 data.mv_data = NODEDATA(srcnode);
7156 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7157 if (rc != MDB_SUCCESS)
7162 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7163 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7164 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7166 /* Unlink the src page from parent and add to free list.
7169 mdb_node_del(csrc, 0);
7170 if (csrc->mc_ki[csrc->mc_top] == 0) {
7172 rc = mdb_update_key(csrc, &key);
7180 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7181 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7184 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7185 csrc->mc_db->md_leaf_pages--;
7187 csrc->mc_db->md_branch_pages--;
7189 /* Adjust other cursors pointing to mp */
7190 MDB_cursor *m2, *m3;
7191 MDB_dbi dbi = csrc->mc_dbi;
7192 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7194 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7195 if (csrc->mc_flags & C_SUB)
7196 m3 = &m2->mc_xcursor->mx_cursor;
7199 if (m3 == csrc) continue;
7200 if (m3->mc_snum < csrc->mc_snum) continue;
7201 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7202 m3->mc_pg[csrc->mc_top] = mp;
7203 m3->mc_ki[csrc->mc_top] += nkeys;
7207 mdb_cursor_pop(csrc);
7209 return mdb_rebalance(csrc);
7212 /** Copy the contents of a cursor.
7213 * @param[in] csrc The cursor to copy from.
7214 * @param[out] cdst The cursor to copy to.
7217 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7221 cdst->mc_txn = csrc->mc_txn;
7222 cdst->mc_dbi = csrc->mc_dbi;
7223 cdst->mc_db = csrc->mc_db;
7224 cdst->mc_dbx = csrc->mc_dbx;
7225 cdst->mc_snum = csrc->mc_snum;
7226 cdst->mc_top = csrc->mc_top;
7227 cdst->mc_flags = csrc->mc_flags;
7229 for (i=0; i<csrc->mc_snum; i++) {
7230 cdst->mc_pg[i] = csrc->mc_pg[i];
7231 cdst->mc_ki[i] = csrc->mc_ki[i];
7235 /** Rebalance the tree after a delete operation.
7236 * @param[in] mc Cursor pointing to the page where rebalancing
7238 * @return 0 on success, non-zero on failure.
7241 mdb_rebalance(MDB_cursor *mc)
7245 unsigned int ptop, minkeys;
7248 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7249 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7250 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7251 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7252 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7254 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7255 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7256 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7257 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7261 if (mc->mc_snum < 2) {
7262 MDB_page *mp = mc->mc_pg[0];
7264 DPUTS("Can't rebalance a subpage, ignoring");
7267 if (NUMKEYS(mp) == 0) {
7268 DPUTS("tree is completely empty");
7269 mc->mc_db->md_root = P_INVALID;
7270 mc->mc_db->md_depth = 0;
7271 mc->mc_db->md_leaf_pages = 0;
7272 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7275 /* Adjust cursors pointing to mp */
7278 mc->mc_flags &= ~C_INITIALIZED;
7280 MDB_cursor *m2, *m3;
7281 MDB_dbi dbi = mc->mc_dbi;
7283 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7284 if (mc->mc_flags & C_SUB)
7285 m3 = &m2->mc_xcursor->mx_cursor;
7288 if (m3->mc_snum < mc->mc_snum) continue;
7289 if (m3->mc_pg[0] == mp) {
7292 m3->mc_flags &= ~C_INITIALIZED;
7296 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7297 DPUTS("collapsing root page!");
7298 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7301 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7302 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7305 mc->mc_db->md_depth--;
7306 mc->mc_db->md_branch_pages--;
7307 mc->mc_ki[0] = mc->mc_ki[1];
7309 /* Adjust other cursors pointing to mp */
7310 MDB_cursor *m2, *m3;
7311 MDB_dbi dbi = mc->mc_dbi;
7313 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7314 if (mc->mc_flags & C_SUB)
7315 m3 = &m2->mc_xcursor->mx_cursor;
7318 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7319 if (m3->mc_pg[0] == mp) {
7323 for (i=0; i<m3->mc_snum; i++) {
7324 m3->mc_pg[i] = m3->mc_pg[i+1];
7325 m3->mc_ki[i] = m3->mc_ki[i+1];
7331 DPUTS("root page doesn't need rebalancing");
7335 /* The parent (branch page) must have at least 2 pointers,
7336 * otherwise the tree is invalid.
7338 ptop = mc->mc_top-1;
7339 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7341 /* Leaf page fill factor is below the threshold.
7342 * Try to move keys from left or right neighbor, or
7343 * merge with a neighbor page.
7348 mdb_cursor_copy(mc, &mn);
7349 mn.mc_xcursor = NULL;
7351 if (mc->mc_ki[ptop] == 0) {
7352 /* We're the leftmost leaf in our parent.
7354 DPUTS("reading right neighbor");
7356 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7357 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7360 mn.mc_ki[mn.mc_top] = 0;
7361 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7363 /* There is at least one neighbor to the left.
7365 DPUTS("reading left neighbor");
7367 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7368 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7371 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7372 mc->mc_ki[mc->mc_top] = 0;
7375 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7376 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7377 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7379 /* If the neighbor page is above threshold and has enough keys,
7380 * move one key from it. Otherwise we should try to merge them.
7381 * (A branch page must never have less than 2 keys.)
7383 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7384 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7385 return mdb_node_move(&mn, mc);
7387 if (mc->mc_ki[ptop] == 0)
7388 rc = mdb_page_merge(&mn, mc);
7390 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7391 rc = mdb_page_merge(mc, &mn);
7392 mdb_cursor_copy(&mn, mc);
7394 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7399 /** Complete a delete operation started by #mdb_cursor_del(). */
7401 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7408 mp = mc->mc_pg[mc->mc_top];
7409 ki = mc->mc_ki[mc->mc_top];
7411 /* add overflow pages to free list */
7412 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7416 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7417 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7418 (rc = mdb_ovpage_free(mc, omp)))
7421 mdb_node_del(mc, mc->mc_db->md_pad);
7422 mc->mc_db->md_entries--;
7423 rc = mdb_rebalance(mc);
7424 if (rc != MDB_SUCCESS)
7425 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7427 MDB_cursor *m2, *m3;
7428 MDB_dbi dbi = mc->mc_dbi;
7430 mp = mc->mc_pg[mc->mc_top];
7431 nkeys = NUMKEYS(mp);
7433 /* if mc points past last node in page, find next sibling */
7434 if (mc->mc_ki[mc->mc_top] >= nkeys)
7435 mdb_cursor_sibling(mc, 1);
7437 /* Adjust other cursors pointing to mp */
7438 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7439 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7440 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7442 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7444 if (m3->mc_pg[mc->mc_top] == mp) {
7445 if (m3->mc_ki[mc->mc_top] >= ki) {
7446 m3->mc_flags |= C_DEL;
7447 if (m3->mc_ki[mc->mc_top] > ki)
7448 m3->mc_ki[mc->mc_top]--;
7450 if (m3->mc_ki[mc->mc_top] >= nkeys)
7451 mdb_cursor_sibling(m3, 1);
7454 mc->mc_flags |= C_DEL;
7461 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7462 MDB_val *key, MDB_val *data)
7467 MDB_val rdata, *xdata;
7474 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7476 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7479 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7480 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7482 mdb_cursor_init(&mc, txn, dbi, &mx);
7485 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7486 /* must ignore any data */
7497 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7499 /* let mdb_page_split know about this cursor if needed:
7500 * delete will trigger a rebalance; if it needs to move
7501 * a node from one page to another, it will have to
7502 * update the parent's separator key(s). If the new sepkey
7503 * is larger than the current one, the parent page may
7504 * run out of space, triggering a split. We need this
7505 * cursor to be consistent until the end of the rebalance.
7507 mc.mc_flags |= C_UNTRACK;
7508 mc.mc_next = txn->mt_cursors[dbi];
7509 txn->mt_cursors[dbi] = &mc;
7510 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7511 txn->mt_cursors[dbi] = mc.mc_next;
7516 /** Split a page and insert a new node.
7517 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7518 * The cursor will be updated to point to the actual page and index where
7519 * the node got inserted after the split.
7520 * @param[in] newkey The key for the newly inserted node.
7521 * @param[in] newdata The data for the newly inserted node.
7522 * @param[in] newpgno The page number, if the new node is a branch node.
7523 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7524 * @return 0 on success, non-zero on failure.
7527 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7528 unsigned int nflags)
7531 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7534 int i, j, split_indx, nkeys, pmax;
7535 MDB_env *env = mc->mc_txn->mt_env;
7537 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7538 MDB_page *copy = NULL;
7539 MDB_page *mp, *rp, *pp;
7544 mp = mc->mc_pg[mc->mc_top];
7545 newindx = mc->mc_ki[mc->mc_top];
7546 nkeys = NUMKEYS(mp);
7548 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7549 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7550 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7552 /* Create a right sibling. */
7553 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7555 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7557 if (mc->mc_snum < 2) {
7558 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7560 /* shift current top to make room for new parent */
7561 mc->mc_pg[1] = mc->mc_pg[0];
7562 mc->mc_ki[1] = mc->mc_ki[0];
7565 mc->mc_db->md_root = pp->mp_pgno;
7566 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7567 mc->mc_db->md_depth++;
7570 /* Add left (implicit) pointer. */
7571 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7572 /* undo the pre-push */
7573 mc->mc_pg[0] = mc->mc_pg[1];
7574 mc->mc_ki[0] = mc->mc_ki[1];
7575 mc->mc_db->md_root = mp->mp_pgno;
7576 mc->mc_db->md_depth--;
7583 ptop = mc->mc_top-1;
7584 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7587 mc->mc_flags |= C_SPLITTING;
7588 mdb_cursor_copy(mc, &mn);
7589 mn.mc_pg[mn.mc_top] = rp;
7590 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7592 if (nflags & MDB_APPEND) {
7593 mn.mc_ki[mn.mc_top] = 0;
7595 split_indx = newindx;
7599 split_indx = (nkeys+1) / 2;
7604 unsigned int lsize, rsize, ksize;
7605 /* Move half of the keys to the right sibling */
7607 x = mc->mc_ki[mc->mc_top] - split_indx;
7608 ksize = mc->mc_db->md_pad;
7609 split = LEAF2KEY(mp, split_indx, ksize);
7610 rsize = (nkeys - split_indx) * ksize;
7611 lsize = (nkeys - split_indx) * sizeof(indx_t);
7612 mp->mp_lower -= lsize;
7613 rp->mp_lower += lsize;
7614 mp->mp_upper += rsize - lsize;
7615 rp->mp_upper -= rsize - lsize;
7616 sepkey.mv_size = ksize;
7617 if (newindx == split_indx) {
7618 sepkey.mv_data = newkey->mv_data;
7620 sepkey.mv_data = split;
7623 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7624 memcpy(rp->mp_ptrs, split, rsize);
7625 sepkey.mv_data = rp->mp_ptrs;
7626 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7627 memcpy(ins, newkey->mv_data, ksize);
7628 mp->mp_lower += sizeof(indx_t);
7629 mp->mp_upper -= ksize - sizeof(indx_t);
7632 memcpy(rp->mp_ptrs, split, x * ksize);
7633 ins = LEAF2KEY(rp, x, ksize);
7634 memcpy(ins, newkey->mv_data, ksize);
7635 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7636 rp->mp_lower += sizeof(indx_t);
7637 rp->mp_upper -= ksize - sizeof(indx_t);
7638 mc->mc_ki[mc->mc_top] = x;
7639 mc->mc_pg[mc->mc_top] = rp;
7642 int psize, nsize, k;
7643 /* Maximum free space in an empty page */
7644 pmax = env->me_psize - PAGEHDRSZ;
7646 nsize = mdb_leaf_size(env, newkey, newdata);
7648 nsize = mdb_branch_size(env, newkey);
7649 nsize = EVEN(nsize);
7651 /* grab a page to hold a temporary copy */
7652 copy = mdb_page_malloc(mc->mc_txn, 1);
7655 copy->mp_pgno = mp->mp_pgno;
7656 copy->mp_flags = mp->mp_flags;
7657 copy->mp_lower = PAGEHDRSZ;
7658 copy->mp_upper = env->me_psize;
7660 /* prepare to insert */
7661 for (i=0, j=0; i<nkeys; i++) {
7663 copy->mp_ptrs[j++] = 0;
7665 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7668 /* When items are relatively large the split point needs
7669 * to be checked, because being off-by-one will make the
7670 * difference between success or failure in mdb_node_add.
7672 * It's also relevant if a page happens to be laid out
7673 * such that one half of its nodes are all "small" and
7674 * the other half of its nodes are "large." If the new
7675 * item is also "large" and falls on the half with
7676 * "large" nodes, it also may not fit.
7678 * As a final tweak, if the new item goes on the last
7679 * spot on the page (and thus, onto the new page), bias
7680 * the split so the new page is emptier than the old page.
7681 * This yields better packing during sequential inserts.
7683 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7684 /* Find split point */
7686 if (newindx <= split_indx || newindx >= nkeys) {
7688 k = newindx >= nkeys ? nkeys : split_indx+2;
7693 for (; i!=k; i+=j) {
7698 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7699 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7701 if (F_ISSET(node->mn_flags, F_BIGDATA))
7702 psize += sizeof(pgno_t);
7704 psize += NODEDSZ(node);
7706 psize = EVEN(psize);
7708 if (psize > pmax || i == k-j) {
7709 split_indx = i + (j<0);
7714 if (split_indx == newindx) {
7715 sepkey.mv_size = newkey->mv_size;
7716 sepkey.mv_data = newkey->mv_data;
7718 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7719 sepkey.mv_size = node->mn_ksize;
7720 sepkey.mv_data = NODEKEY(node);
7725 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7727 /* Copy separator key to the parent.
7729 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7733 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7736 if (mn.mc_snum == mc->mc_snum) {
7737 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7738 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7739 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7740 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7745 /* Right page might now have changed parent.
7746 * Check if left page also changed parent.
7748 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7749 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7750 for (i=0; i<ptop; i++) {
7751 mc->mc_pg[i] = mn.mc_pg[i];
7752 mc->mc_ki[i] = mn.mc_ki[i];
7754 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7755 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7759 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7762 mc->mc_flags ^= C_SPLITTING;
7763 if (rc != MDB_SUCCESS) {
7766 if (nflags & MDB_APPEND) {
7767 mc->mc_pg[mc->mc_top] = rp;
7768 mc->mc_ki[mc->mc_top] = 0;
7769 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7772 for (i=0; i<mc->mc_top; i++)
7773 mc->mc_ki[i] = mn.mc_ki[i];
7774 } else if (!IS_LEAF2(mp)) {
7776 mc->mc_pg[mc->mc_top] = rp;
7781 rkey.mv_data = newkey->mv_data;
7782 rkey.mv_size = newkey->mv_size;
7788 /* Update index for the new key. */
7789 mc->mc_ki[mc->mc_top] = j;
7791 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7792 rkey.mv_data = NODEKEY(node);
7793 rkey.mv_size = node->mn_ksize;
7795 xdata.mv_data = NODEDATA(node);
7796 xdata.mv_size = NODEDSZ(node);
7799 pgno = NODEPGNO(node);
7800 flags = node->mn_flags;
7803 if (!IS_LEAF(mp) && j == 0) {
7804 /* First branch index doesn't need key data. */
7808 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7810 /* return tmp page to freelist */
7811 mdb_page_free(env, copy);
7817 mc->mc_pg[mc->mc_top] = copy;
7822 } while (i != split_indx);
7824 nkeys = NUMKEYS(copy);
7825 for (i=0; i<nkeys; i++)
7826 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7827 mp->mp_lower = copy->mp_lower;
7828 mp->mp_upper = copy->mp_upper;
7829 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7830 env->me_psize - copy->mp_upper);
7832 /* reset back to original page */
7833 if (newindx < split_indx) {
7834 mc->mc_pg[mc->mc_top] = mp;
7835 if (nflags & MDB_RESERVE) {
7836 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7837 if (!(node->mn_flags & F_BIGDATA))
7838 newdata->mv_data = NODEDATA(node);
7841 mc->mc_pg[mc->mc_top] = rp;
7843 /* Make sure mc_ki is still valid.
7845 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7846 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7847 for (i=0; i<ptop; i++) {
7848 mc->mc_pg[i] = mn.mc_pg[i];
7849 mc->mc_ki[i] = mn.mc_ki[i];
7851 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7852 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7855 /* return tmp page to freelist */
7856 mdb_page_free(env, copy);
7860 /* Adjust other cursors pointing to mp */
7861 MDB_cursor *m2, *m3;
7862 MDB_dbi dbi = mc->mc_dbi;
7863 int fixup = NUMKEYS(mp);
7865 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7866 if (mc->mc_flags & C_SUB)
7867 m3 = &m2->mc_xcursor->mx_cursor;
7872 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7874 if (m3->mc_flags & C_SPLITTING)
7879 for (k=m3->mc_top; k>=0; k--) {
7880 m3->mc_ki[k+1] = m3->mc_ki[k];
7881 m3->mc_pg[k+1] = m3->mc_pg[k];
7883 if (m3->mc_ki[0] >= split_indx) {
7888 m3->mc_pg[0] = mc->mc_pg[0];
7892 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7893 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7894 m3->mc_ki[mc->mc_top]++;
7895 if (m3->mc_ki[mc->mc_top] >= fixup) {
7896 m3->mc_pg[mc->mc_top] = rp;
7897 m3->mc_ki[mc->mc_top] -= fixup;
7898 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7900 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7901 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7906 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7911 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7912 MDB_val *key, MDB_val *data, unsigned int flags)
7917 if (key == NULL || data == NULL)
7920 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7923 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7926 mdb_cursor_init(&mc, txn, dbi, &mx);
7927 return mdb_cursor_put(&mc, key, data, flags);
7931 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7933 if ((flag & CHANGEABLE) != flag)
7936 env->me_flags |= flag;
7938 env->me_flags &= ~flag;
7943 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7948 *arg = env->me_flags;
7953 mdb_env_set_userctx(MDB_env *env, void *ctx)
7957 env->me_userctx = ctx;
7962 mdb_env_get_userctx(MDB_env *env)
7964 return env ? env->me_userctx : NULL;
7968 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
7973 env->me_assert_func = func;
7979 mdb_env_get_path(MDB_env *env, const char **arg)
7984 *arg = env->me_path;
7989 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
7998 /** Common code for #mdb_stat() and #mdb_env_stat().
7999 * @param[in] env the environment to operate in.
8000 * @param[in] db the #MDB_db record containing the stats to return.
8001 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8002 * @return 0, this function always succeeds.
8005 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8007 arg->ms_psize = env->me_psize;
8008 arg->ms_depth = db->md_depth;
8009 arg->ms_branch_pages = db->md_branch_pages;
8010 arg->ms_leaf_pages = db->md_leaf_pages;
8011 arg->ms_overflow_pages = db->md_overflow_pages;
8012 arg->ms_entries = db->md_entries;
8017 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8021 if (env == NULL || arg == NULL)
8024 toggle = mdb_env_pick_meta(env);
8026 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8030 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8034 if (env == NULL || arg == NULL)
8037 toggle = mdb_env_pick_meta(env);
8038 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
8039 arg->me_mapsize = env->me_mapsize;
8040 arg->me_maxreaders = env->me_maxreaders;
8042 /* me_numreaders may be zero if this process never used any readers. Use
8043 * the shared numreader count if it exists.
8045 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8047 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8048 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8052 /** Set the default comparison functions for a database.
8053 * Called immediately after a database is opened to set the defaults.
8054 * The user can then override them with #mdb_set_compare() or
8055 * #mdb_set_dupsort().
8056 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8057 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8060 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8062 uint16_t f = txn->mt_dbs[dbi].md_flags;
8064 txn->mt_dbxs[dbi].md_cmp =
8065 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8066 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8068 txn->mt_dbxs[dbi].md_dcmp =
8069 !(f & MDB_DUPSORT) ? 0 :
8070 ((f & MDB_INTEGERDUP)
8071 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8072 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8075 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8080 int rc, dbflag, exact;
8081 unsigned int unused = 0;
8084 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8085 mdb_default_cmp(txn, FREE_DBI);
8088 if ((flags & VALID_FLAGS) != flags)
8090 if (txn->mt_flags & MDB_TXN_ERROR)
8096 if (flags & PERSISTENT_FLAGS) {
8097 uint16_t f2 = flags & PERSISTENT_FLAGS;
8098 /* make sure flag changes get committed */
8099 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8100 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8101 txn->mt_flags |= MDB_TXN_DIRTY;
8104 mdb_default_cmp(txn, MAIN_DBI);
8108 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8109 mdb_default_cmp(txn, MAIN_DBI);
8112 /* Is the DB already open? */
8114 for (i=2; i<txn->mt_numdbs; i++) {
8115 if (!txn->mt_dbxs[i].md_name.mv_size) {
8116 /* Remember this free slot */
8117 if (!unused) unused = i;
8120 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8121 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8127 /* If no free slot and max hit, fail */
8128 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8129 return MDB_DBS_FULL;
8131 /* Cannot mix named databases with some mainDB flags */
8132 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8133 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8135 /* Find the DB info */
8136 dbflag = DB_NEW|DB_VALID;
8139 key.mv_data = (void *)name;
8140 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8141 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8142 if (rc == MDB_SUCCESS) {
8143 /* make sure this is actually a DB */
8144 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8145 if (!(node->mn_flags & F_SUBDATA))
8146 return MDB_INCOMPATIBLE;
8147 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8148 /* Create if requested */
8150 data.mv_size = sizeof(MDB_db);
8151 data.mv_data = &dummy;
8152 memset(&dummy, 0, sizeof(dummy));
8153 dummy.md_root = P_INVALID;
8154 dummy.md_flags = flags & PERSISTENT_FLAGS;
8155 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8159 /* OK, got info, add to table */
8160 if (rc == MDB_SUCCESS) {
8161 unsigned int slot = unused ? unused : txn->mt_numdbs;
8162 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8163 txn->mt_dbxs[slot].md_name.mv_size = len;
8164 txn->mt_dbxs[slot].md_rel = NULL;
8165 txn->mt_dbflags[slot] = dbflag;
8166 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8168 mdb_default_cmp(txn, slot);
8177 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8179 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8182 if (txn->mt_dbflags[dbi] & DB_STALE) {
8185 /* Stale, must read the DB's root. cursor_init does it for us. */
8186 mdb_cursor_init(&mc, txn, dbi, &mx);
8188 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8191 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8194 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8196 ptr = env->me_dbxs[dbi].md_name.mv_data;
8197 env->me_dbxs[dbi].md_name.mv_data = NULL;
8198 env->me_dbxs[dbi].md_name.mv_size = 0;
8199 env->me_dbflags[dbi] = 0;
8203 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8205 /* We could return the flags for the FREE_DBI too but what's the point? */
8206 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8208 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8212 /** Add all the DB's pages to the free list.
8213 * @param[in] mc Cursor on the DB to free.
8214 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8215 * @return 0 on success, non-zero on failure.
8218 mdb_drop0(MDB_cursor *mc, int subs)
8222 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8223 if (rc == MDB_SUCCESS) {
8224 MDB_txn *txn = mc->mc_txn;
8229 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8230 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8233 mdb_cursor_copy(mc, &mx);
8234 while (mc->mc_snum > 0) {
8235 MDB_page *mp = mc->mc_pg[mc->mc_top];
8236 unsigned n = NUMKEYS(mp);
8238 for (i=0; i<n; i++) {
8239 ni = NODEPTR(mp, i);
8240 if (ni->mn_flags & F_BIGDATA) {
8243 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8244 rc = mdb_page_get(txn, pg, &omp, NULL);
8247 mdb_cassert(mc, IS_OVERFLOW(omp));
8248 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8252 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8253 mdb_xcursor_init1(mc, ni);
8254 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8260 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8262 for (i=0; i<n; i++) {
8264 ni = NODEPTR(mp, i);
8267 mdb_midl_xappend(txn->mt_free_pgs, pg);
8272 mc->mc_ki[mc->mc_top] = i;
8273 rc = mdb_cursor_sibling(mc, 1);
8275 /* no more siblings, go back to beginning
8276 * of previous level.
8280 for (i=1; i<mc->mc_snum; i++) {
8282 mc->mc_pg[i] = mx.mc_pg[i];
8287 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8288 } else if (rc == MDB_NOTFOUND) {
8294 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8296 MDB_cursor *mc, *m2;
8299 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8302 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8305 rc = mdb_cursor_open(txn, dbi, &mc);
8309 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8310 /* Invalidate the dropped DB's cursors */
8311 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8312 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8316 /* Can't delete the main DB */
8317 if (del && dbi > MAIN_DBI) {
8318 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8320 txn->mt_dbflags[dbi] = DB_STALE;
8321 mdb_dbi_close(txn->mt_env, dbi);
8324 /* reset the DB record, mark it dirty */
8325 txn->mt_dbflags[dbi] |= DB_DIRTY;
8326 txn->mt_dbs[dbi].md_depth = 0;
8327 txn->mt_dbs[dbi].md_branch_pages = 0;
8328 txn->mt_dbs[dbi].md_leaf_pages = 0;
8329 txn->mt_dbs[dbi].md_overflow_pages = 0;
8330 txn->mt_dbs[dbi].md_entries = 0;
8331 txn->mt_dbs[dbi].md_root = P_INVALID;
8333 txn->mt_flags |= MDB_TXN_DIRTY;
8336 mdb_cursor_close(mc);
8340 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8342 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8345 txn->mt_dbxs[dbi].md_cmp = cmp;
8349 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8351 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8354 txn->mt_dbxs[dbi].md_dcmp = cmp;
8358 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8360 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8363 txn->mt_dbxs[dbi].md_rel = rel;
8367 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8369 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8372 txn->mt_dbxs[dbi].md_relctx = ctx;
8376 int mdb_env_get_maxkeysize(MDB_env *env)
8378 return ENV_MAXKEY(env);
8381 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8383 unsigned int i, rdrs;
8386 int rc = 0, first = 1;
8390 if (!env->me_txns) {
8391 return func("(no reader locks)\n", ctx);
8393 rdrs = env->me_txns->mti_numreaders;
8394 mr = env->me_txns->mti_readers;
8395 for (i=0; i<rdrs; i++) {
8397 txnid_t txnid = mr[i].mr_txnid;
8398 sprintf(buf, txnid == (txnid_t)-1 ?
8399 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
8400 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
8403 rc = func(" pid thread txnid\n", ctx);
8407 rc = func(buf, ctx);
8413 rc = func("(no active readers)\n", ctx);
8418 /** Insert pid into list if not already present.
8419 * return -1 if already present.
8421 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8423 /* binary search of pid in list */
8425 unsigned cursor = 1;
8427 unsigned n = ids[0];
8430 unsigned pivot = n >> 1;
8431 cursor = base + pivot + 1;
8432 val = pid - ids[cursor];
8437 } else if ( val > 0 ) {
8442 /* found, so it's a duplicate */
8451 for (n = ids[0]; n > cursor; n--)
8457 int mdb_reader_check(MDB_env *env, int *dead)
8459 unsigned int i, j, rdrs;
8461 MDB_PID_T *pids, pid;
8470 rdrs = env->me_txns->mti_numreaders;
8471 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8475 mr = env->me_txns->mti_readers;
8476 for (i=0; i<rdrs; i++) {
8477 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8479 if (mdb_pid_insert(pids, pid) == 0) {
8480 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8482 /* Recheck, a new process may have reused pid */
8483 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8484 for (j=i; j<rdrs; j++)
8485 if (mr[j].mr_pid == pid) {
8486 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8487 (unsigned) pid, mr[j].mr_txnid));
8492 UNLOCK_MUTEX_R(env);