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
78 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
79 #include <netinet/in.h>
80 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
83 #if defined(__APPLE__) || defined (BSD)
84 # define MDB_USE_POSIX_SEM 1
85 # define MDB_FDATASYNC fsync
86 #elif defined(ANDROID)
87 # define MDB_FDATASYNC fsync
92 #ifdef MDB_USE_POSIX_SEM
93 # define MDB_USE_HASH 1
94 #include <semaphore.h>
99 #include <valgrind/memcheck.h>
100 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
101 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
102 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
103 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
104 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
106 #define VGMEMP_CREATE(h,r,z)
107 #define VGMEMP_ALLOC(h,a,s)
108 #define VGMEMP_FREE(h,a)
109 #define VGMEMP_DESTROY(h)
110 #define VGMEMP_DEFINED(a,s)
114 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
115 /* Solaris just defines one or the other */
116 # define LITTLE_ENDIAN 1234
117 # define BIG_ENDIAN 4321
118 # ifdef _LITTLE_ENDIAN
119 # define BYTE_ORDER LITTLE_ENDIAN
121 # define BYTE_ORDER BIG_ENDIAN
124 # define BYTE_ORDER __BYTE_ORDER
128 #ifndef LITTLE_ENDIAN
129 #define LITTLE_ENDIAN __LITTLE_ENDIAN
132 #define BIG_ENDIAN __BIG_ENDIAN
135 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
136 #define MISALIGNED_OK 1
142 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
143 # error "Unknown or unsupported endianness (BYTE_ORDER)"
144 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
145 # error "Two's complement, reasonably sized integer types, please"
148 /** @defgroup internal MDB Internals
151 /** @defgroup compat Compatibility Macros
152 * A bunch of macros to minimize the amount of platform-specific ifdefs
153 * needed throughout the rest of the code. When the features this library
154 * needs are similar enough to POSIX to be hidden in a one-or-two line
155 * replacement, this macro approach is used.
159 /** Wrapper around __func__, which is a C99 feature */
160 #if __STDC_VERSION__ >= 199901L
161 # define mdb_func_ __func__
162 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
163 # define mdb_func_ __FUNCTION__
165 /* If a debug message says <mdb_unknown>(), update the #if statements above */
166 # define mdb_func_ "<mdb_unknown>"
170 #define MDB_USE_HASH 1
171 #define MDB_PIDLOCK 0
172 #define pthread_t DWORD
173 #define pthread_mutex_t HANDLE
174 #define pthread_key_t DWORD
175 #define pthread_self() GetCurrentThreadId()
176 #define pthread_key_create(x,y) \
177 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
178 #define pthread_key_delete(x) TlsFree(x)
179 #define pthread_getspecific(x) TlsGetValue(x)
180 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
181 #define pthread_mutex_unlock(x) ReleaseMutex(x)
182 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
183 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
184 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
185 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
186 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
187 #define getpid() GetCurrentProcessId()
188 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
189 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
190 #define ErrCode() GetLastError()
191 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
192 #define close(fd) (CloseHandle(fd) ? 0 : -1)
193 #define munmap(ptr,len) UnmapViewOfFile(ptr)
194 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
195 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
197 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
202 #define Z "z" /**< printf format modifier for size_t */
204 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
205 #define MDB_PIDLOCK 1
207 #ifdef MDB_USE_POSIX_SEM
209 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
210 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
211 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
212 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
215 mdb_sem_wait(sem_t *sem)
218 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
223 /** Lock the reader mutex.
225 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
226 /** Unlock the reader mutex.
228 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
230 /** Lock the writer mutex.
231 * Only a single write transaction is allowed at a time. Other writers
232 * will block waiting for this mutex.
234 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
235 /** Unlock the writer mutex.
237 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
238 #endif /* MDB_USE_POSIX_SEM */
240 /** Get the error code for the last failed system function.
242 #define ErrCode() errno
244 /** An abstraction for a file handle.
245 * On POSIX systems file handles are small integers. On Windows
246 * they're opaque pointers.
250 /** A value for an invalid file handle.
251 * Mainly used to initialize file variables and signify that they are
254 #define INVALID_HANDLE_VALUE (-1)
256 /** Get the size of a memory page for the system.
257 * This is the basic size that the platform's memory manager uses, and is
258 * fundamental to the use of memory-mapped files.
260 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
263 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
266 #define MNAME_LEN (sizeof(pthread_mutex_t))
272 /** A flag for opening a file and requesting synchronous data writes.
273 * This is only used when writing a meta page. It's not strictly needed;
274 * we could just do a normal write and then immediately perform a flush.
275 * But if this flag is available it saves us an extra system call.
277 * @note If O_DSYNC is undefined but exists in /usr/include,
278 * preferably set some compiler flag to get the definition.
279 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
282 # define MDB_DSYNC O_DSYNC
286 /** Function for flushing the data of a file. Define this to fsync
287 * if fdatasync() is not supported.
289 #ifndef MDB_FDATASYNC
290 # define MDB_FDATASYNC fdatasync
294 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
305 /** A page number in the database.
306 * Note that 64 bit page numbers are overkill, since pages themselves
307 * already represent 12-13 bits of addressable memory, and the OS will
308 * always limit applications to a maximum of 63 bits of address space.
310 * @note In the #MDB_node structure, we only store 48 bits of this value,
311 * which thus limits us to only 60 bits of addressable data.
313 typedef MDB_ID pgno_t;
315 /** A transaction ID.
316 * See struct MDB_txn.mt_txnid for details.
318 typedef MDB_ID txnid_t;
320 /** @defgroup debug Debug Macros
324 /** Enable debug output. Needs variable argument macros (a C99 feature).
325 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
326 * read from and written to the database (used for free space management).
332 static int mdb_debug;
333 static txnid_t mdb_debug_start;
335 /** Print a debug message with printf formatting.
336 * Requires double parenthesis around 2 or more args.
338 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
339 # define DPRINTF0(fmt, ...) \
340 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
342 # define DPRINTF(args) ((void) 0)
344 /** Print a debug string.
345 * The string is printed literally, with no format processing.
347 #define DPUTS(arg) DPRINTF(("%s", arg))
348 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
350 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
353 /** @brief The maximum size of a database page.
355 * This is 32k, since it must fit in #MDB_page.#mp_upper.
357 * LMDB will use database pages < OS pages if needed.
358 * That causes more I/O in write transactions: The OS must
359 * know (read) the whole page before writing a partial page.
361 * Note that we don't currently support Huge pages. On Linux,
362 * regular data files cannot use Huge pages, and in general
363 * Huge pages aren't actually pageable. We rely on the OS
364 * demand-pager to read our data and page it out when memory
365 * pressure from other processes is high. So until OSs have
366 * actual paging support for Huge pages, they're not viable.
368 #define MAX_PAGESIZE 0x8000
370 /** The minimum number of keys required in a database page.
371 * Setting this to a larger value will place a smaller bound on the
372 * maximum size of a data item. Data items larger than this size will
373 * be pushed into overflow pages instead of being stored directly in
374 * the B-tree node. This value used to default to 4. With a page size
375 * of 4096 bytes that meant that any item larger than 1024 bytes would
376 * go into an overflow page. That also meant that on average 2-3KB of
377 * each overflow page was wasted space. The value cannot be lower than
378 * 2 because then there would no longer be a tree structure. With this
379 * value, items larger than 2KB will go into overflow pages, and on
380 * average only 1KB will be wasted.
382 #define MDB_MINKEYS 2
384 /** A stamp that identifies a file as an MDB file.
385 * There's nothing special about this value other than that it is easily
386 * recognizable, and it will reflect any byte order mismatches.
388 #define MDB_MAGIC 0xBEEFC0DE
390 /** The version number for a database's datafile format. */
391 #define MDB_DATA_VERSION 1
392 /** The version number for a database's lockfile format. */
393 #define MDB_LOCK_VERSION 1
395 /** @brief The max size of a key we can write, or 0 for dynamic max.
397 * Define this as 0 to compute the max from the page size. 511
398 * is default for backwards compat: liblmdb <= 0.9.10 can break
399 * when modifying a DB with keys/dupsort data bigger than its max.
401 * Data items in an #MDB_DUPSORT database are also limited to
402 * this size, since they're actually keys of a sub-DB. Keys and
403 * #MDB_DUPSORT data items must fit on a node in a regular page.
405 #ifndef MDB_MAXKEYSIZE
406 #define MDB_MAXKEYSIZE 511
409 /** The maximum size of a key we can write to the environment. */
411 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
413 #define ENV_MAXKEY(env) ((env)->me_maxkey)
416 /** @brief The maximum size of a data item.
418 * We only store a 32 bit value for node sizes.
420 #define MAXDATASIZE 0xffffffffUL
423 /** Key size which fits in a #DKBUF.
426 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
429 * This is used for printing a hex dump of a key's contents.
431 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
432 /** Display a key in hex.
434 * Invoke a function to display a key in hex.
436 #define DKEY(x) mdb_dkey(x, kbuf)
442 /** An invalid page number.
443 * Mainly used to denote an empty tree.
445 #define P_INVALID (~(pgno_t)0)
447 /** Test if the flags \b f are set in a flag word \b w. */
448 #define F_ISSET(w, f) (((w) & (f)) == (f))
450 /** Round \b n up to an even number. */
451 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
453 /** Used for offsets within a single page.
454 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
457 typedef uint16_t indx_t;
459 /** Default size of memory map.
460 * This is certainly too small for any actual applications. Apps should always set
461 * the size explicitly using #mdb_env_set_mapsize().
463 #define DEFAULT_MAPSIZE 1048576
465 /** @defgroup readers Reader Lock Table
466 * Readers don't acquire any locks for their data access. Instead, they
467 * simply record their transaction ID in the reader table. The reader
468 * mutex is needed just to find an empty slot in the reader table. The
469 * slot's address is saved in thread-specific data so that subsequent read
470 * transactions started by the same thread need no further locking to proceed.
472 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
474 * No reader table is used if the database is on a read-only filesystem, or
475 * if #MDB_NOLOCK is set.
477 * Since the database uses multi-version concurrency control, readers don't
478 * actually need any locking. This table is used to keep track of which
479 * readers are using data from which old transactions, so that we'll know
480 * when a particular old transaction is no longer in use. Old transactions
481 * that have discarded any data pages can then have those pages reclaimed
482 * for use by a later write transaction.
484 * The lock table is constructed such that reader slots are aligned with the
485 * processor's cache line size. Any slot is only ever used by one thread.
486 * This alignment guarantees that there will be no contention or cache
487 * thrashing as threads update their own slot info, and also eliminates
488 * any need for locking when accessing a slot.
490 * A writer thread will scan every slot in the table to determine the oldest
491 * outstanding reader transaction. Any freed pages older than this will be
492 * reclaimed by the writer. The writer doesn't use any locks when scanning
493 * this table. This means that there's no guarantee that the writer will
494 * see the most up-to-date reader info, but that's not required for correct
495 * operation - all we need is to know the upper bound on the oldest reader,
496 * we don't care at all about the newest reader. So the only consequence of
497 * reading stale information here is that old pages might hang around a
498 * while longer before being reclaimed. That's actually good anyway, because
499 * the longer we delay reclaiming old pages, the more likely it is that a
500 * string of contiguous pages can be found after coalescing old pages from
501 * many old transactions together.
504 /** Number of slots in the reader table.
505 * This value was chosen somewhat arbitrarily. 126 readers plus a
506 * couple mutexes fit exactly into 8KB on my development machine.
507 * Applications should set the table size using #mdb_env_set_maxreaders().
509 #define DEFAULT_READERS 126
511 /** The size of a CPU cache line in bytes. We want our lock structures
512 * aligned to this size to avoid false cache line sharing in the
514 * This value works for most CPUs. For Itanium this should be 128.
520 /** The information we store in a single slot of the reader table.
521 * In addition to a transaction ID, we also record the process and
522 * thread ID that owns a slot, so that we can detect stale information,
523 * e.g. threads or processes that went away without cleaning up.
524 * @note We currently don't check for stale records. We simply re-init
525 * the table when we know that we're the only process opening the
528 typedef struct MDB_rxbody {
529 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
530 * Multiple readers that start at the same time will probably have the
531 * same ID here. Again, it's not important to exclude them from
532 * anything; all we need to know is which version of the DB they
533 * started from so we can avoid overwriting any data used in that
534 * particular version.
537 /** The process ID of the process owning this reader txn. */
539 /** The thread ID of the thread owning this txn. */
543 /** The actual reader record, with cacheline padding. */
544 typedef struct MDB_reader {
547 /** shorthand for mrb_txnid */
548 #define mr_txnid mru.mrx.mrb_txnid
549 #define mr_pid mru.mrx.mrb_pid
550 #define mr_tid mru.mrx.mrb_tid
551 /** cache line alignment */
552 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
556 /** The header for the reader table.
557 * The table resides in a memory-mapped file. (This is a different file
558 * than is used for the main database.)
560 * For POSIX the actual mutexes reside in the shared memory of this
561 * mapped file. On Windows, mutexes are named objects allocated by the
562 * kernel; we store the mutex names in this mapped file so that other
563 * processes can grab them. This same approach is also used on
564 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
565 * process-shared POSIX mutexes. For these cases where a named object
566 * is used, the object name is derived from a 64 bit FNV hash of the
567 * environment pathname. As such, naming collisions are extremely
568 * unlikely. If a collision occurs, the results are unpredictable.
570 typedef struct MDB_txbody {
571 /** Stamp identifying this as an MDB file. It must be set
574 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
576 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
577 char mtb_rmname[MNAME_LEN];
579 /** Mutex protecting access to this table.
580 * This is the reader lock that #LOCK_MUTEX_R acquires.
582 pthread_mutex_t mtb_mutex;
584 /** The ID of the last transaction committed to the database.
585 * This is recorded here only for convenience; the value can always
586 * be determined by reading the main database meta pages.
589 /** The number of slots that have been used in the reader table.
590 * This always records the maximum count, it is not decremented
591 * when readers release their slots.
593 unsigned mtb_numreaders;
596 /** The actual reader table definition. */
597 typedef struct MDB_txninfo {
600 #define mti_magic mt1.mtb.mtb_magic
601 #define mti_format mt1.mtb.mtb_format
602 #define mti_mutex mt1.mtb.mtb_mutex
603 #define mti_rmname mt1.mtb.mtb_rmname
604 #define mti_txnid mt1.mtb.mtb_txnid
605 #define mti_numreaders mt1.mtb.mtb_numreaders
606 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
609 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
610 char mt2_wmname[MNAME_LEN];
611 #define mti_wmname mt2.mt2_wmname
613 pthread_mutex_t mt2_wmutex;
614 #define mti_wmutex mt2.mt2_wmutex
616 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
618 MDB_reader mti_readers[1];
621 /** Lockfile format signature: version, features and field layout */
622 #define MDB_LOCK_FORMAT \
624 ((MDB_LOCK_VERSION) \
625 /* Flags which describe functionality */ \
626 + (((MDB_PIDLOCK) != 0) << 16)))
629 /** Common header for all page types.
630 * Overflow records occupy a number of contiguous pages with no
631 * headers on any page after the first.
633 typedef struct MDB_page {
634 #define mp_pgno mp_p.p_pgno
635 #define mp_next mp_p.p_next
637 pgno_t p_pgno; /**< page number */
638 void * p_next; /**< for in-memory list of freed structs */
641 /** @defgroup mdb_page Page Flags
643 * Flags for the page headers.
646 #define P_BRANCH 0x01 /**< branch page */
647 #define P_LEAF 0x02 /**< leaf page */
648 #define P_OVERFLOW 0x04 /**< overflow page */
649 #define P_META 0x08 /**< meta page */
650 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
651 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
652 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
653 #define P_KEEP 0x8000 /**< leave this page alone during spill */
655 uint16_t mp_flags; /**< @ref mdb_page */
656 #define mp_lower mp_pb.pb.pb_lower
657 #define mp_upper mp_pb.pb.pb_upper
658 #define mp_pages mp_pb.pb_pages
661 indx_t pb_lower; /**< lower bound of free space */
662 indx_t pb_upper; /**< upper bound of free space */
664 uint32_t pb_pages; /**< number of overflow pages */
666 indx_t mp_ptrs[1]; /**< dynamic size */
669 /** Size of the page header, excluding dynamic data at the end */
670 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
672 /** Address of first usable data byte in a page, after the header */
673 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
675 /** Number of nodes on a page */
676 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
678 /** The amount of space remaining in the page */
679 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
681 /** The percentage of space used in the page, in tenths of a percent. */
682 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
683 ((env)->me_psize - PAGEHDRSZ))
684 /** The minimum page fill factor, in tenths of a percent.
685 * Pages emptier than this are candidates for merging.
687 #define FILL_THRESHOLD 250
689 /** Test if a page is a leaf page */
690 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
691 /** Test if a page is a LEAF2 page */
692 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
693 /** Test if a page is a branch page */
694 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
695 /** Test if a page is an overflow page */
696 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
697 /** Test if a page is a sub page */
698 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
700 /** The number of overflow pages needed to store the given size. */
701 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
703 /** Header for a single key/data pair within a page.
704 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
705 * We guarantee 2-byte alignment for 'MDB_node's.
707 typedef struct MDB_node {
708 /** lo and hi are used for data size on leaf nodes and for
709 * child pgno on branch nodes. On 64 bit platforms, flags
710 * is also used for pgno. (Branch nodes have no flags).
711 * They are in host byte order in case that lets some
712 * accesses be optimized into a 32-bit word access.
714 #if BYTE_ORDER == LITTLE_ENDIAN
715 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
717 unsigned short mn_hi, mn_lo;
719 /** @defgroup mdb_node Node Flags
721 * Flags for node headers.
724 #define F_BIGDATA 0x01 /**< data put on overflow page */
725 #define F_SUBDATA 0x02 /**< data is a sub-database */
726 #define F_DUPDATA 0x04 /**< data has duplicates */
728 /** valid flags for #mdb_node_add() */
729 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
732 unsigned short mn_flags; /**< @ref mdb_node */
733 unsigned short mn_ksize; /**< key size */
734 char mn_data[1]; /**< key and data are appended here */
737 /** Size of the node header, excluding dynamic data at the end */
738 #define NODESIZE offsetof(MDB_node, mn_data)
740 /** Bit position of top word in page number, for shifting mn_flags */
741 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
743 /** Size of a node in a branch page with a given key.
744 * This is just the node header plus the key, there is no data.
746 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
748 /** Size of a node in a leaf page with a given key and data.
749 * This is node header plus key plus data size.
751 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
753 /** Address of node \b i in page \b p */
754 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
756 /** Address of the key for the node */
757 #define NODEKEY(node) (void *)((node)->mn_data)
759 /** Address of the data for a node */
760 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
762 /** Get the page number pointed to by a branch node */
763 #define NODEPGNO(node) \
764 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
765 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
766 /** Set the page number in a branch node */
767 #define SETPGNO(node,pgno) do { \
768 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
769 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
771 /** Get the size of the data in a leaf node */
772 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
773 /** Set the size of the data for a leaf node */
774 #define SETDSZ(node,size) do { \
775 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
776 /** The size of a key in a node */
777 #define NODEKSZ(node) ((node)->mn_ksize)
779 /** Copy a page number from src to dst */
781 #define COPY_PGNO(dst,src) dst = src
783 #if SIZE_MAX > 4294967295UL
784 #define COPY_PGNO(dst,src) do { \
785 unsigned short *s, *d; \
786 s = (unsigned short *)&(src); \
787 d = (unsigned short *)&(dst); \
794 #define COPY_PGNO(dst,src) do { \
795 unsigned short *s, *d; \
796 s = (unsigned short *)&(src); \
797 d = (unsigned short *)&(dst); \
803 /** The address of a key in a LEAF2 page.
804 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
805 * There are no node headers, keys are stored contiguously.
807 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
809 /** Set the \b node's key into \b keyptr, if requested. */
810 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
811 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
813 /** Set the \b node's key into \b key. */
814 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
816 /** Information about a single database in the environment. */
817 typedef struct MDB_db {
818 uint32_t md_pad; /**< also ksize for LEAF2 pages */
819 uint16_t md_flags; /**< @ref mdb_dbi_open */
820 uint16_t md_depth; /**< depth of this tree */
821 pgno_t md_branch_pages; /**< number of internal pages */
822 pgno_t md_leaf_pages; /**< number of leaf pages */
823 pgno_t md_overflow_pages; /**< number of overflow pages */
824 size_t md_entries; /**< number of data items */
825 pgno_t md_root; /**< the root page of this tree */
828 /** mdb_dbi_open flags */
829 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
830 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
831 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
832 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
834 /** Handle for the DB used to track free pages. */
836 /** Handle for the default DB. */
839 /** Meta page content.
840 * A meta page is the start point for accessing a database snapshot.
841 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
843 typedef struct MDB_meta {
844 /** Stamp identifying this as an MDB file. It must be set
847 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
849 void *mm_address; /**< address for fixed mapping */
850 size_t mm_mapsize; /**< size of mmap region */
851 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
852 /** The size of pages used in this DB */
853 #define mm_psize mm_dbs[0].md_pad
854 /** Any persistent environment flags. @ref mdb_env */
855 #define mm_flags mm_dbs[0].md_flags
856 pgno_t mm_last_pg; /**< last used page in file */
857 txnid_t mm_txnid; /**< txnid that committed this page */
860 /** Buffer for a stack-allocated meta page.
861 * The members define size and alignment, and silence type
862 * aliasing warnings. They are not used directly; that could
863 * mean incorrectly using several union members in parallel.
865 typedef union MDB_metabuf {
868 char mm_pad[PAGEHDRSZ];
873 /** Auxiliary DB info.
874 * The information here is mostly static/read-only. There is
875 * only a single copy of this record in the environment.
877 typedef struct MDB_dbx {
878 MDB_val md_name; /**< name of the database */
879 MDB_cmp_func *md_cmp; /**< function for comparing keys */
880 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
881 MDB_rel_func *md_rel; /**< user relocate function */
882 void *md_relctx; /**< user-provided context for md_rel */
885 /** A database transaction.
886 * Every operation requires a transaction handle.
889 MDB_txn *mt_parent; /**< parent of a nested txn */
890 MDB_txn *mt_child; /**< nested txn under this txn */
891 pgno_t mt_next_pgno; /**< next unallocated page */
892 /** The ID of this transaction. IDs are integers incrementing from 1.
893 * Only committed write transactions increment the ID. If a transaction
894 * aborts, the ID may be re-used by the next writer.
897 MDB_env *mt_env; /**< the DB environment */
898 /** The list of pages that became unused during this transaction.
901 /** The sorted list of dirty pages we temporarily wrote to disk
902 * because the dirty list was full. page numbers in here are
903 * shifted left by 1, deleted slots have the LSB set.
905 MDB_IDL mt_spill_pgs;
907 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
909 /** For read txns: This thread/txn's reader table slot, or NULL. */
912 /** Array of records for each DB known in the environment. */
914 /** Array of MDB_db records for each known DB */
916 /** @defgroup mt_dbflag Transaction DB Flags
920 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
921 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
922 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
923 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
925 /** In write txns, array of cursors for each DB */
926 MDB_cursor **mt_cursors;
927 /** Array of flags for each DB */
928 unsigned char *mt_dbflags;
929 /** Number of DB records in use. This number only ever increments;
930 * we don't decrement it when individual DB handles are closed.
934 /** @defgroup mdb_txn Transaction Flags
938 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
939 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
940 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
941 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
943 unsigned int mt_flags; /**< @ref mdb_txn */
944 /** dirty_list room: Array size - #dirty pages visible to this txn.
945 * Includes ancestor txns' dirty pages not hidden by other txns'
946 * dirty/spilled pages. Thus commit(nested txn) has room to merge
947 * dirty_list into mt_parent after freeing hidden mt_parent pages.
949 unsigned int mt_dirty_room;
952 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
953 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
954 * raise this on a 64 bit machine.
956 #define CURSOR_STACK 32
960 /** Cursors are used for all DB operations.
961 * A cursor holds a path of (page pointer, key index) from the DB
962 * root to a position in the DB, plus other state. #MDB_DUPSORT
963 * cursors include an xcursor to the current data item. Write txns
964 * track their cursors and keep them up to date when data moves.
965 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
966 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
969 /** Next cursor on this DB in this txn */
971 /** Backup of the original cursor if this cursor is a shadow */
972 MDB_cursor *mc_backup;
973 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
974 struct MDB_xcursor *mc_xcursor;
975 /** The transaction that owns this cursor */
977 /** The database handle this cursor operates on */
979 /** The database record for this cursor */
981 /** The database auxiliary record for this cursor */
983 /** The @ref mt_dbflag for this database */
984 unsigned char *mc_dbflag;
985 unsigned short mc_snum; /**< number of pushed pages */
986 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
987 /** @defgroup mdb_cursor Cursor Flags
989 * Cursor state flags.
992 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
993 #define C_EOF 0x02 /**< No more data */
994 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
995 #define C_DEL 0x08 /**< last op was a cursor_del */
996 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
997 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
999 unsigned int mc_flags; /**< @ref mdb_cursor */
1000 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1001 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1004 /** Context for sorted-dup records.
1005 * We could have gone to a fully recursive design, with arbitrarily
1006 * deep nesting of sub-databases. But for now we only handle these
1007 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1009 typedef struct MDB_xcursor {
1010 /** A sub-cursor for traversing the Dup DB */
1011 MDB_cursor mx_cursor;
1012 /** The database record for this Dup DB */
1014 /** The auxiliary DB record for this Dup DB */
1016 /** The @ref mt_dbflag for this Dup DB */
1017 unsigned char mx_dbflag;
1020 /** State of FreeDB old pages, stored in the MDB_env */
1021 typedef struct MDB_pgstate {
1022 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1023 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1026 /** The database environment. */
1028 HANDLE me_fd; /**< The main data file */
1029 HANDLE me_lfd; /**< The lock file */
1030 HANDLE me_mfd; /**< just for writing the meta pages */
1031 /** Failed to update the meta page. Probably an I/O error. */
1032 #define MDB_FATAL_ERROR 0x80000000U
1033 /** Some fields are initialized. */
1034 #define MDB_ENV_ACTIVE 0x20000000U
1035 /** me_txkey is set */
1036 #define MDB_ENV_TXKEY 0x10000000U
1037 /** Have liveness lock in reader table */
1038 #define MDB_LIVE_READER 0x08000000U
1039 uint32_t me_flags; /**< @ref mdb_env */
1040 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1041 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1042 unsigned int me_maxreaders; /**< size of the reader table */
1043 unsigned int me_numreaders; /**< max numreaders set by this env */
1044 MDB_dbi me_numdbs; /**< number of DBs opened */
1045 MDB_dbi me_maxdbs; /**< size of the DB table */
1046 MDB_PID_T me_pid; /**< process ID of this env */
1047 char *me_path; /**< path to the DB files */
1048 char *me_map; /**< the memory map of the data file */
1049 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1050 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1051 void *me_pbuf; /**< scratch area for DUPSORT put() */
1052 MDB_txn *me_txn; /**< current write transaction */
1053 size_t me_mapsize; /**< size of the data memory map */
1054 off_t me_size; /**< current file size */
1055 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1056 MDB_dbx *me_dbxs; /**< array of static DB info */
1057 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1058 pthread_key_t me_txkey; /**< thread-key for readers */
1059 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1060 # define me_pglast me_pgstate.mf_pglast
1061 # define me_pghead me_pgstate.mf_pghead
1062 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1063 /** IDL of pages that became unused in a write txn */
1064 MDB_IDL me_free_pgs;
1065 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1066 MDB_ID2L me_dirty_list;
1067 /** Max number of freelist items that can fit in a single overflow page */
1069 /** Max size of a node on a page */
1070 unsigned int me_nodemax;
1071 #if !(MDB_MAXKEYSIZE)
1072 unsigned int me_maxkey; /**< max size of a key */
1075 int me_pidquery; /**< Used in OpenProcess */
1076 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1078 #elif defined(MDB_USE_POSIX_SEM)
1079 sem_t *me_rmutex; /* Shared mutexes are not supported */
1082 void *me_userctx; /**< User-settable context */
1083 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1086 /** Nested transaction */
1087 typedef struct MDB_ntxn {
1088 MDB_txn mnt_txn; /**< the transaction */
1089 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1092 /** max number of pages to commit in one writev() call */
1093 #define MDB_COMMIT_PAGES 64
1094 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1095 #undef MDB_COMMIT_PAGES
1096 #define MDB_COMMIT_PAGES IOV_MAX
1099 /* max bytes to write in one call */
1100 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1102 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1103 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1104 static int mdb_page_touch(MDB_cursor *mc);
1106 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1107 static int mdb_page_search_root(MDB_cursor *mc,
1108 MDB_val *key, int modify);
1109 #define MDB_PS_MODIFY 1
1110 #define MDB_PS_ROOTONLY 2
1111 #define MDB_PS_FIRST 4
1112 #define MDB_PS_LAST 8
1113 static int mdb_page_search(MDB_cursor *mc,
1114 MDB_val *key, int flags);
1115 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1117 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1118 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1119 pgno_t newpgno, unsigned int nflags);
1121 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1122 static int mdb_env_pick_meta(const MDB_env *env);
1123 static int mdb_env_write_meta(MDB_txn *txn);
1124 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1125 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1127 static void mdb_env_close0(MDB_env *env, int excl);
1129 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1130 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1131 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1132 static void mdb_node_del(MDB_cursor *mc, int ksize);
1133 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1134 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1135 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1136 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1137 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1139 static int mdb_rebalance(MDB_cursor *mc);
1140 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1142 static void mdb_cursor_pop(MDB_cursor *mc);
1143 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1145 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1146 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1147 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1148 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1149 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1151 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1152 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1154 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1155 static void mdb_xcursor_init0(MDB_cursor *mc);
1156 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1158 static int mdb_drop0(MDB_cursor *mc, int subs);
1159 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1162 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1166 static SECURITY_DESCRIPTOR mdb_null_sd;
1167 static SECURITY_ATTRIBUTES mdb_all_sa;
1168 static int mdb_sec_inited;
1171 /** Return the library version info. */
1173 mdb_version(int *major, int *minor, int *patch)
1175 if (major) *major = MDB_VERSION_MAJOR;
1176 if (minor) *minor = MDB_VERSION_MINOR;
1177 if (patch) *patch = MDB_VERSION_PATCH;
1178 return MDB_VERSION_STRING;
1181 /** Table of descriptions for MDB @ref errors */
1182 static char *const mdb_errstr[] = {
1183 "MDB_KEYEXIST: Key/data pair already exists",
1184 "MDB_NOTFOUND: No matching key/data pair found",
1185 "MDB_PAGE_NOTFOUND: Requested page not found",
1186 "MDB_CORRUPTED: Located page was wrong type",
1187 "MDB_PANIC: Update of meta page failed",
1188 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1189 "MDB_INVALID: File is not an MDB file",
1190 "MDB_MAP_FULL: Environment mapsize limit reached",
1191 "MDB_DBS_FULL: Environment maxdbs limit reached",
1192 "MDB_READERS_FULL: Environment maxreaders limit reached",
1193 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1194 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1195 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1196 "MDB_PAGE_FULL: Internal error - page has no more space",
1197 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1198 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1199 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1200 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1201 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1205 mdb_strerror(int err)
1209 return ("Successful return: 0");
1211 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1212 i = err - MDB_KEYEXIST;
1213 return mdb_errstr[i];
1216 return strerror(err);
1219 /** assert(3) variant in cursor context */
1220 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1221 /** assert(3) variant in transaction context */
1222 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1223 /** assert(3) variant in environment context */
1224 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1227 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1228 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1231 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1232 const char *func, const char *file, int line)
1235 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1236 file, line, expr_txt, func);
1237 if (env->me_assert_func)
1238 env->me_assert_func(env, buf);
1239 fprintf(stderr, "%s\n", buf);
1243 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1247 /** Return the page number of \b mp which may be sub-page, for debug output */
1249 mdb_dbg_pgno(MDB_page *mp)
1252 COPY_PGNO(ret, mp->mp_pgno);
1256 /** Display a key in hexadecimal and return the address of the result.
1257 * @param[in] key the key to display
1258 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1259 * @return The key in hexadecimal form.
1262 mdb_dkey(MDB_val *key, char *buf)
1265 unsigned char *c = key->mv_data;
1271 if (key->mv_size > DKBUF_MAXKEYSIZE)
1272 return "MDB_MAXKEYSIZE";
1273 /* may want to make this a dynamic check: if the key is mostly
1274 * printable characters, print it as-is instead of converting to hex.
1278 for (i=0; i<key->mv_size; i++)
1279 ptr += sprintf(ptr, "%02x", *c++);
1281 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1286 /** Display all the keys in the page. */
1288 mdb_page_list(MDB_page *mp)
1291 unsigned int i, nkeys, nsize, total = 0;
1295 nkeys = NUMKEYS(mp);
1296 fprintf(stderr, "Page %"Z"u numkeys %d\n", mdb_dbg_pgno(mp), nkeys);
1297 for (i=0; i<nkeys; i++) {
1298 node = NODEPTR(mp, i);
1299 key.mv_size = node->mn_ksize;
1300 key.mv_data = node->mn_data;
1301 nsize = NODESIZE + key.mv_size;
1302 if (IS_BRANCH(mp)) {
1303 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1307 if (F_ISSET(node->mn_flags, F_BIGDATA))
1308 nsize += sizeof(pgno_t);
1310 nsize += NODEDSZ(node);
1312 nsize += sizeof(indx_t);
1313 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1315 total = EVEN(total);
1317 fprintf(stderr, "Total: %d\n", total);
1321 mdb_cursor_chk(MDB_cursor *mc)
1327 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1328 for (i=0; i<mc->mc_top; i++) {
1330 node = NODEPTR(mp, mc->mc_ki[i]);
1331 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1334 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1340 /** Count all the pages in each DB and in the freelist
1341 * and make sure it matches the actual number of pages
1344 static void mdb_audit(MDB_txn *txn)
1348 MDB_ID freecount, count;
1353 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1354 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1355 freecount += *(MDB_ID *)data.mv_data;
1358 for (i = 0; i<txn->mt_numdbs; i++) {
1360 mdb_cursor_init(&mc, txn, i, &mx);
1361 if (txn->mt_dbs[i].md_root == P_INVALID)
1363 count += txn->mt_dbs[i].md_branch_pages +
1364 txn->mt_dbs[i].md_leaf_pages +
1365 txn->mt_dbs[i].md_overflow_pages;
1366 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1367 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1371 mp = mc.mc_pg[mc.mc_top];
1372 for (j=0; j<NUMKEYS(mp); j++) {
1373 MDB_node *leaf = NODEPTR(mp, j);
1374 if (leaf->mn_flags & F_SUBDATA) {
1376 memcpy(&db, NODEDATA(leaf), sizeof(db));
1377 count += db.md_branch_pages + db.md_leaf_pages +
1378 db.md_overflow_pages;
1382 while (mdb_cursor_sibling(&mc, 1) == 0);
1385 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1386 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1387 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1393 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1395 return txn->mt_dbxs[dbi].md_cmp(a, b);
1399 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1401 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1404 /** Allocate memory for a page.
1405 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1408 mdb_page_malloc(MDB_txn *txn, unsigned num)
1410 MDB_env *env = txn->mt_env;
1411 MDB_page *ret = env->me_dpages;
1412 size_t psize = env->me_psize, sz = psize, off;
1413 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1414 * For a single page alloc, we init everything after the page header.
1415 * For multi-page, we init the final page; if the caller needed that
1416 * many pages they will be filling in at least up to the last page.
1420 VGMEMP_ALLOC(env, ret, sz);
1421 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1422 env->me_dpages = ret->mp_next;
1425 psize -= off = PAGEHDRSZ;
1430 if ((ret = malloc(sz)) != NULL) {
1431 VGMEMP_ALLOC(env, ret, sz);
1432 if (!(env->me_flags & MDB_NOMEMINIT)) {
1433 memset((char *)ret + off, 0, psize);
1437 txn->mt_flags |= MDB_TXN_ERROR;
1442 /** Free a single page.
1443 * Saves single pages to a list, for future reuse.
1444 * (This is not used for multi-page overflow pages.)
1447 mdb_page_free(MDB_env *env, MDB_page *mp)
1449 mp->mp_next = env->me_dpages;
1450 VGMEMP_FREE(env, mp);
1451 env->me_dpages = mp;
1454 /** Free a dirty page */
1456 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1458 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1459 mdb_page_free(env, dp);
1461 /* large pages just get freed directly */
1462 VGMEMP_FREE(env, dp);
1467 /** Return all dirty pages to dpage list */
1469 mdb_dlist_free(MDB_txn *txn)
1471 MDB_env *env = txn->mt_env;
1472 MDB_ID2L dl = txn->mt_u.dirty_list;
1473 unsigned i, n = dl[0].mid;
1475 for (i = 1; i <= n; i++) {
1476 mdb_dpage_free(env, dl[i].mptr);
1481 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1482 * @param[in] mc A cursor handle for the current operation.
1483 * @param[in] pflags Flags of the pages to update:
1484 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1485 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1486 * @return 0 on success, non-zero on failure.
1489 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1491 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1492 MDB_txn *txn = mc->mc_txn;
1498 int rc = MDB_SUCCESS, level;
1500 /* Mark pages seen by cursors */
1501 if (mc->mc_flags & C_UNTRACK)
1502 mc = NULL; /* will find mc in mt_cursors */
1503 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1504 for (; mc; mc=mc->mc_next) {
1505 if (!(mc->mc_flags & C_INITIALIZED))
1507 for (m3 = mc;; m3 = &mx->mx_cursor) {
1509 for (j=0; j<m3->mc_snum; j++) {
1511 if ((mp->mp_flags & Mask) == pflags)
1512 mp->mp_flags ^= P_KEEP;
1514 mx = m3->mc_xcursor;
1515 /* Proceed to mx if it is at a sub-database */
1516 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1518 if (! (mp && (mp->mp_flags & P_LEAF)))
1520 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1521 if (!(leaf->mn_flags & F_SUBDATA))
1530 /* Mark dirty root pages */
1531 for (i=0; i<txn->mt_numdbs; i++) {
1532 if (txn->mt_dbflags[i] & DB_DIRTY) {
1533 pgno_t pgno = txn->mt_dbs[i].md_root;
1534 if (pgno == P_INVALID)
1536 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1538 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1539 dp->mp_flags ^= P_KEEP;
1547 static int mdb_page_flush(MDB_txn *txn, int keep);
1549 /** Spill pages from the dirty list back to disk.
1550 * This is intended to prevent running into #MDB_TXN_FULL situations,
1551 * but note that they may still occur in a few cases:
1552 * 1) our estimate of the txn size could be too small. Currently this
1553 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1554 * 2) child txns may run out of space if their parents dirtied a
1555 * lot of pages and never spilled them. TODO: we probably should do
1556 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1557 * the parent's dirty_room is below a given threshold.
1559 * Otherwise, if not using nested txns, it is expected that apps will
1560 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1561 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1562 * If the txn never references them again, they can be left alone.
1563 * If the txn only reads them, they can be used without any fuss.
1564 * If the txn writes them again, they can be dirtied immediately without
1565 * going thru all of the work of #mdb_page_touch(). Such references are
1566 * handled by #mdb_page_unspill().
1568 * Also note, we never spill DB root pages, nor pages of active cursors,
1569 * because we'll need these back again soon anyway. And in nested txns,
1570 * we can't spill a page in a child txn if it was already spilled in a
1571 * parent txn. That would alter the parent txns' data even though
1572 * the child hasn't committed yet, and we'd have no way to undo it if
1573 * the child aborted.
1575 * @param[in] m0 cursor A cursor handle identifying the transaction and
1576 * database for which we are checking space.
1577 * @param[in] key For a put operation, the key being stored.
1578 * @param[in] data For a put operation, the data being stored.
1579 * @return 0 on success, non-zero on failure.
1582 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1584 MDB_txn *txn = m0->mc_txn;
1586 MDB_ID2L dl = txn->mt_u.dirty_list;
1587 unsigned int i, j, need;
1590 if (m0->mc_flags & C_SUB)
1593 /* Estimate how much space this op will take */
1594 i = m0->mc_db->md_depth;
1595 /* Named DBs also dirty the main DB */
1596 if (m0->mc_dbi > MAIN_DBI)
1597 i += txn->mt_dbs[MAIN_DBI].md_depth;
1598 /* For puts, roughly factor in the key+data size */
1600 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1601 i += i; /* double it for good measure */
1604 if (txn->mt_dirty_room > i)
1607 if (!txn->mt_spill_pgs) {
1608 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1609 if (!txn->mt_spill_pgs)
1612 /* purge deleted slots */
1613 MDB_IDL sl = txn->mt_spill_pgs;
1614 unsigned int num = sl[0];
1616 for (i=1; i<=num; i++) {
1623 /* Preserve pages which may soon be dirtied again */
1624 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1627 /* Less aggressive spill - we originally spilled the entire dirty list,
1628 * with a few exceptions for cursor pages and DB root pages. But this
1629 * turns out to be a lot of wasted effort because in a large txn many
1630 * of those pages will need to be used again. So now we spill only 1/8th
1631 * of the dirty pages. Testing revealed this to be a good tradeoff,
1632 * better than 1/2, 1/4, or 1/10.
1634 if (need < MDB_IDL_UM_MAX / 8)
1635 need = MDB_IDL_UM_MAX / 8;
1637 /* Save the page IDs of all the pages we're flushing */
1638 /* flush from the tail forward, this saves a lot of shifting later on. */
1639 for (i=dl[0].mid; i && need; i--) {
1640 MDB_ID pn = dl[i].mid << 1;
1642 if (dp->mp_flags & P_KEEP)
1644 /* Can't spill twice, make sure it's not already in a parent's
1647 if (txn->mt_parent) {
1649 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1650 if (tx2->mt_spill_pgs) {
1651 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1652 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1653 dp->mp_flags |= P_KEEP;
1661 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1665 mdb_midl_sort(txn->mt_spill_pgs);
1667 /* Flush the spilled part of dirty list */
1668 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1671 /* Reset any dirty pages we kept that page_flush didn't see */
1672 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1675 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1679 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1681 mdb_find_oldest(MDB_txn *txn)
1684 txnid_t mr, oldest = txn->mt_txnid - 1;
1685 if (txn->mt_env->me_txns) {
1686 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1687 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1698 /** Add a page to the txn's dirty list */
1700 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1703 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1705 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1706 insert = mdb_mid2l_append;
1708 insert = mdb_mid2l_insert;
1710 mid.mid = mp->mp_pgno;
1712 rc = insert(txn->mt_u.dirty_list, &mid);
1713 mdb_tassert(txn, rc == 0);
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 rc = mdb_mid2l_insert(dl, &mid);
2036 mdb_cassert(mc, rc == 0);
2041 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2043 np->mp_flags |= P_DIRTY;
2046 /* Adjust cursors pointing to mp */
2047 mc->mc_pg[mc->mc_top] = np;
2048 m2 = txn->mt_cursors[mc->mc_dbi];
2049 if (mc->mc_flags & C_SUB) {
2050 for (; m2; m2=m2->mc_next) {
2051 m3 = &m2->mc_xcursor->mx_cursor;
2052 if (m3->mc_snum < mc->mc_snum) continue;
2053 if (m3->mc_pg[mc->mc_top] == mp)
2054 m3->mc_pg[mc->mc_top] = np;
2057 for (; m2; m2=m2->mc_next) {
2058 if (m2->mc_snum < mc->mc_snum) continue;
2059 if (m2->mc_pg[mc->mc_top] == mp) {
2060 m2->mc_pg[mc->mc_top] = np;
2061 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2062 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2064 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2065 if (!(leaf->mn_flags & F_SUBDATA))
2066 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2074 txn->mt_flags |= MDB_TXN_ERROR;
2079 mdb_env_sync(MDB_env *env, int force)
2082 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2083 if (env->me_flags & MDB_WRITEMAP) {
2084 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2085 ? MS_ASYNC : MS_SYNC;
2086 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2089 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2093 if (MDB_FDATASYNC(env->me_fd))
2100 /** Back up parent txn's cursors, then grab the originals for tracking */
2102 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2104 MDB_cursor *mc, *bk;
2109 for (i = src->mt_numdbs; --i >= 0; ) {
2110 if ((mc = src->mt_cursors[i]) != NULL) {
2111 size = sizeof(MDB_cursor);
2113 size += sizeof(MDB_xcursor);
2114 for (; mc; mc = bk->mc_next) {
2120 mc->mc_db = &dst->mt_dbs[i];
2121 /* Kill pointers into src - and dst to reduce abuse: The
2122 * user may not use mc until dst ends. Otherwise we'd...
2124 mc->mc_txn = NULL; /* ...set this to dst */
2125 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2126 if ((mx = mc->mc_xcursor) != NULL) {
2127 *(MDB_xcursor *)(bk+1) = *mx;
2128 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2130 mc->mc_next = dst->mt_cursors[i];
2131 dst->mt_cursors[i] = mc;
2138 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2139 * @param[in] txn the transaction handle.
2140 * @param[in] merge true to keep changes to parent cursors, false to revert.
2141 * @return 0 on success, non-zero on failure.
2144 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2146 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2150 for (i = txn->mt_numdbs; --i >= 0; ) {
2151 for (mc = cursors[i]; mc; mc = next) {
2153 if ((bk = mc->mc_backup) != NULL) {
2155 /* Commit changes to parent txn */
2156 mc->mc_next = bk->mc_next;
2157 mc->mc_backup = bk->mc_backup;
2158 mc->mc_txn = bk->mc_txn;
2159 mc->mc_db = bk->mc_db;
2160 mc->mc_dbflag = bk->mc_dbflag;
2161 if ((mx = mc->mc_xcursor) != NULL)
2162 mx->mx_cursor.mc_txn = bk->mc_txn;
2164 /* Abort nested txn */
2166 if ((mx = mc->mc_xcursor) != NULL)
2167 *mx = *(MDB_xcursor *)(bk+1);
2171 /* Only malloced cursors are permanently tracked. */
2179 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2182 mdb_txn_reset0(MDB_txn *txn, const char *act);
2184 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2190 Pidset = F_SETLK, Pidcheck = F_GETLK
2194 /** Set or check a pid lock. Set returns 0 on success.
2195 * Check returns 0 if the process is certainly dead, nonzero if it may
2196 * be alive (the lock exists or an error happened so we do not know).
2198 * On Windows Pidset is a no-op, we merely check for the existence
2199 * of the process with the given pid. On POSIX we use a single byte
2200 * lock on the lockfile, set at an offset equal to the pid.
2203 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2205 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2208 if (op == Pidcheck) {
2209 h = OpenProcess(env->me_pidquery, FALSE, pid);
2210 /* No documented "no such process" code, but other program use this: */
2212 return ErrCode() != ERROR_INVALID_PARAMETER;
2213 /* A process exists until all handles to it close. Has it exited? */
2214 ret = WaitForSingleObject(h, 0) != 0;
2221 struct flock lock_info;
2222 memset(&lock_info, 0, sizeof(lock_info));
2223 lock_info.l_type = F_WRLCK;
2224 lock_info.l_whence = SEEK_SET;
2225 lock_info.l_start = pid;
2226 lock_info.l_len = 1;
2227 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2228 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2230 } else if ((rc = ErrCode()) == EINTR) {
2238 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2239 * @param[in] txn the transaction handle to initialize
2240 * @return 0 on success, non-zero on failure.
2243 mdb_txn_renew0(MDB_txn *txn)
2245 MDB_env *env = txn->mt_env;
2246 MDB_txninfo *ti = env->me_txns;
2250 int rc, new_notls = 0;
2253 txn->mt_numdbs = env->me_numdbs;
2254 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2256 if (txn->mt_flags & MDB_TXN_RDONLY) {
2258 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2259 txn->mt_txnid = meta->mm_txnid;
2260 txn->mt_u.reader = NULL;
2262 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2263 pthread_getspecific(env->me_txkey);
2265 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2266 return MDB_BAD_RSLOT;
2268 MDB_PID_T pid = env->me_pid;
2269 pthread_t tid = pthread_self();
2271 if (!(env->me_flags & MDB_LIVE_READER)) {
2272 rc = mdb_reader_pid(env, Pidset, pid);
2275 env->me_flags |= MDB_LIVE_READER;
2279 nr = ti->mti_numreaders;
2280 for (i=0; i<nr; i++)
2281 if (ti->mti_readers[i].mr_pid == 0)
2283 if (i == env->me_maxreaders) {
2284 UNLOCK_MUTEX_R(env);
2285 return MDB_READERS_FULL;
2287 ti->mti_readers[i].mr_pid = pid;
2288 ti->mti_readers[i].mr_tid = tid;
2290 ti->mti_numreaders = ++nr;
2291 /* Save numreaders for un-mutexed mdb_env_close() */
2292 env->me_numreaders = nr;
2293 UNLOCK_MUTEX_R(env);
2295 r = &ti->mti_readers[i];
2296 new_notls = (env->me_flags & MDB_NOTLS);
2297 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2302 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2303 txn->mt_u.reader = r;
2304 meta = env->me_metas[txn->mt_txnid & 1];
2310 txn->mt_txnid = ti->mti_txnid;
2311 meta = env->me_metas[txn->mt_txnid & 1];
2313 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2314 txn->mt_txnid = meta->mm_txnid;
2318 if (txn->mt_txnid == mdb_debug_start)
2321 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2322 txn->mt_u.dirty_list = env->me_dirty_list;
2323 txn->mt_u.dirty_list[0].mid = 0;
2324 txn->mt_free_pgs = env->me_free_pgs;
2325 txn->mt_free_pgs[0] = 0;
2326 txn->mt_spill_pgs = NULL;
2330 /* Copy the DB info and flags */
2331 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2333 /* Moved to here to avoid a data race in read TXNs */
2334 txn->mt_next_pgno = meta->mm_last_pg+1;
2336 for (i=2; i<txn->mt_numdbs; i++) {
2337 x = env->me_dbflags[i];
2338 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2339 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2341 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2343 if (env->me_maxpg < txn->mt_next_pgno) {
2344 mdb_txn_reset0(txn, "renew0-mapfail");
2346 txn->mt_u.reader->mr_pid = 0;
2347 txn->mt_u.reader = NULL;
2349 return MDB_MAP_RESIZED;
2356 mdb_txn_renew(MDB_txn *txn)
2360 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2363 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2364 DPUTS("environment had fatal error, must shutdown!");
2368 rc = mdb_txn_renew0(txn);
2369 if (rc == MDB_SUCCESS) {
2370 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2371 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2372 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2378 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2382 int rc, size, tsize = sizeof(MDB_txn);
2384 if (env->me_flags & MDB_FATAL_ERROR) {
2385 DPUTS("environment had fatal error, must shutdown!");
2388 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2391 /* Nested transactions: Max 1 child, write txns only, no writemap */
2392 if (parent->mt_child ||
2393 (flags & MDB_RDONLY) ||
2394 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2395 (env->me_flags & MDB_WRITEMAP))
2397 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2399 tsize = sizeof(MDB_ntxn);
2401 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2402 if (!(flags & MDB_RDONLY))
2403 size += env->me_maxdbs * sizeof(MDB_cursor *);
2405 if ((txn = calloc(1, size)) == NULL) {
2406 DPRINTF(("calloc: %s", strerror(ErrCode())));
2409 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2410 if (flags & MDB_RDONLY) {
2411 txn->mt_flags |= MDB_TXN_RDONLY;
2412 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2414 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2415 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2421 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2422 if (!txn->mt_u.dirty_list ||
2423 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2425 free(txn->mt_u.dirty_list);
2429 txn->mt_txnid = parent->mt_txnid;
2430 txn->mt_dirty_room = parent->mt_dirty_room;
2431 txn->mt_u.dirty_list[0].mid = 0;
2432 txn->mt_spill_pgs = NULL;
2433 txn->mt_next_pgno = parent->mt_next_pgno;
2434 parent->mt_child = txn;
2435 txn->mt_parent = parent;
2436 txn->mt_numdbs = parent->mt_numdbs;
2437 txn->mt_flags = parent->mt_flags;
2438 txn->mt_dbxs = parent->mt_dbxs;
2439 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2440 /* Copy parent's mt_dbflags, but clear DB_NEW */
2441 for (i=0; i<txn->mt_numdbs; i++)
2442 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2444 ntxn = (MDB_ntxn *)txn;
2445 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2446 if (env->me_pghead) {
2447 size = MDB_IDL_SIZEOF(env->me_pghead);
2448 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2450 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2455 rc = mdb_cursor_shadow(parent, txn);
2457 mdb_txn_reset0(txn, "beginchild-fail");
2459 rc = mdb_txn_renew0(txn);
2465 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2466 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2467 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2474 mdb_txn_env(MDB_txn *txn)
2476 if(!txn) return NULL;
2480 /** Export or close DBI handles opened in this txn. */
2482 mdb_dbis_update(MDB_txn *txn, int keep)
2485 MDB_dbi n = txn->mt_numdbs;
2486 MDB_env *env = txn->mt_env;
2487 unsigned char *tdbflags = txn->mt_dbflags;
2489 for (i = n; --i >= 2;) {
2490 if (tdbflags[i] & DB_NEW) {
2492 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2494 char *ptr = env->me_dbxs[i].md_name.mv_data;
2495 env->me_dbxs[i].md_name.mv_data = NULL;
2496 env->me_dbxs[i].md_name.mv_size = 0;
2497 env->me_dbflags[i] = 0;
2502 if (keep && env->me_numdbs < n)
2506 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2507 * May be called twice for readonly txns: First reset it, then abort.
2508 * @param[in] txn the transaction handle to reset
2509 * @param[in] act why the transaction is being reset
2512 mdb_txn_reset0(MDB_txn *txn, const char *act)
2514 MDB_env *env = txn->mt_env;
2516 /* Close any DBI handles opened in this txn */
2517 mdb_dbis_update(txn, 0);
2519 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2520 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2521 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2523 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2524 if (txn->mt_u.reader) {
2525 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2526 if (!(env->me_flags & MDB_NOTLS))
2527 txn->mt_u.reader = NULL; /* txn does not own reader */
2529 txn->mt_numdbs = 0; /* close nothing if called again */
2530 txn->mt_dbxs = NULL; /* mark txn as reset */
2532 mdb_cursors_close(txn, 0);
2534 if (!(env->me_flags & MDB_WRITEMAP)) {
2535 mdb_dlist_free(txn);
2537 mdb_midl_free(env->me_pghead);
2539 if (txn->mt_parent) {
2540 txn->mt_parent->mt_child = NULL;
2541 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2542 mdb_midl_free(txn->mt_free_pgs);
2543 mdb_midl_free(txn->mt_spill_pgs);
2544 free(txn->mt_u.dirty_list);
2548 if (mdb_midl_shrink(&txn->mt_free_pgs))
2549 env->me_free_pgs = txn->mt_free_pgs;
2550 env->me_pghead = NULL;
2554 /* The writer mutex was locked in mdb_txn_begin. */
2556 UNLOCK_MUTEX_W(env);
2561 mdb_txn_reset(MDB_txn *txn)
2566 /* This call is only valid for read-only txns */
2567 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2570 mdb_txn_reset0(txn, "reset");
2574 mdb_txn_abort(MDB_txn *txn)
2580 mdb_txn_abort(txn->mt_child);
2582 mdb_txn_reset0(txn, "abort");
2583 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2584 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2585 txn->mt_u.reader->mr_pid = 0;
2590 /** Save the freelist as of this transaction to the freeDB.
2591 * This changes the freelist. Keep trying until it stabilizes.
2594 mdb_freelist_save(MDB_txn *txn)
2596 /* env->me_pghead[] can grow and shrink during this call.
2597 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2598 * Page numbers cannot disappear from txn->mt_free_pgs[].
2601 MDB_env *env = txn->mt_env;
2602 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2603 txnid_t pglast = 0, head_id = 0;
2604 pgno_t freecnt = 0, *free_pgs, *mop;
2605 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2607 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2609 if (env->me_pghead) {
2610 /* Make sure first page of freeDB is touched and on freelist */
2611 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2612 if (rc && rc != MDB_NOTFOUND)
2616 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2617 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2618 ? SSIZE_MAX : maxfree_1pg;
2621 /* Come back here after each Put() in case freelist changed */
2626 /* If using records from freeDB which we have not yet
2627 * deleted, delete them and any we reserved for me_pghead.
2629 while (pglast < env->me_pglast) {
2630 rc = mdb_cursor_first(&mc, &key, NULL);
2633 pglast = head_id = *(txnid_t *)key.mv_data;
2634 total_room = head_room = 0;
2635 mdb_tassert(txn, pglast <= env->me_pglast);
2636 rc = mdb_cursor_del(&mc, 0);
2641 /* Save the IDL of pages freed by this txn, to a single record */
2642 if (freecnt < txn->mt_free_pgs[0]) {
2644 /* Make sure last page of freeDB is touched and on freelist */
2645 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2646 if (rc && rc != MDB_NOTFOUND)
2649 free_pgs = txn->mt_free_pgs;
2650 /* Write to last page of freeDB */
2651 key.mv_size = sizeof(txn->mt_txnid);
2652 key.mv_data = &txn->mt_txnid;
2654 freecnt = free_pgs[0];
2655 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2656 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2659 /* Retry if mt_free_pgs[] grew during the Put() */
2660 free_pgs = txn->mt_free_pgs;
2661 } while (freecnt < free_pgs[0]);
2662 mdb_midl_sort(free_pgs);
2663 memcpy(data.mv_data, free_pgs, data.mv_size);
2666 unsigned int i = free_pgs[0];
2667 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2668 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2670 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2676 mop = env->me_pghead;
2677 mop_len = mop ? mop[0] : 0;
2679 /* Reserve records for me_pghead[]. Split it if multi-page,
2680 * to avoid searching freeDB for a page range. Use keys in
2681 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2683 if (total_room >= mop_len) {
2684 if (total_room == mop_len || --more < 0)
2686 } else if (head_room >= maxfree_1pg && head_id > 1) {
2687 /* Keep current record (overflow page), add a new one */
2691 /* (Re)write {key = head_id, IDL length = head_room} */
2692 total_room -= head_room;
2693 head_room = mop_len - total_room;
2694 if (head_room > maxfree_1pg && head_id > 1) {
2695 /* Overflow multi-page for part of me_pghead */
2696 head_room /= head_id; /* amortize page sizes */
2697 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2698 } else if (head_room < 0) {
2699 /* Rare case, not bothering to delete this record */
2702 key.mv_size = sizeof(head_id);
2703 key.mv_data = &head_id;
2704 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2705 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2708 /* IDL is initially empty, zero out at least the length */
2709 pgs = (pgno_t *)data.mv_data;
2710 j = head_room > clean_limit ? head_room : 0;
2714 total_room += head_room;
2717 /* Fill in the reserved me_pghead records */
2723 rc = mdb_cursor_first(&mc, &key, &data);
2724 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2725 unsigned flags = MDB_CURRENT;
2726 txnid_t id = *(txnid_t *)key.mv_data;
2727 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2730 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2732 if (len > mop_len) {
2734 data.mv_size = (len + 1) * sizeof(MDB_ID);
2737 data.mv_data = mop -= len;
2740 rc = mdb_cursor_put(&mc, &key, &data, flags);
2742 if (rc || !(mop_len -= len))
2749 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2750 * @param[in] txn the transaction that's being committed
2751 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2752 * @return 0 on success, non-zero on failure.
2755 mdb_page_flush(MDB_txn *txn, int keep)
2757 MDB_env *env = txn->mt_env;
2758 MDB_ID2L dl = txn->mt_u.dirty_list;
2759 unsigned psize = env->me_psize, j;
2760 int i, pagecount = dl[0].mid, rc;
2761 size_t size = 0, pos = 0;
2763 MDB_page *dp = NULL;
2767 struct iovec iov[MDB_COMMIT_PAGES];
2768 ssize_t wpos = 0, wsize = 0, wres;
2769 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2775 if (env->me_flags & MDB_WRITEMAP) {
2776 /* Clear dirty flags */
2777 while (++i <= pagecount) {
2779 /* Don't flush this page yet */
2780 if (dp->mp_flags & P_KEEP) {
2781 dp->mp_flags ^= P_KEEP;
2785 dp->mp_flags &= ~P_DIRTY;
2790 /* Write the pages */
2792 if (++i <= pagecount) {
2794 /* Don't flush this page yet */
2795 if (dp->mp_flags & P_KEEP) {
2796 dp->mp_flags ^= P_KEEP;
2801 /* clear dirty flag */
2802 dp->mp_flags &= ~P_DIRTY;
2805 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2810 /* Windows actually supports scatter/gather I/O, but only on
2811 * unbuffered file handles. Since we're relying on the OS page
2812 * cache for all our data, that's self-defeating. So we just
2813 * write pages one at a time. We use the ov structure to set
2814 * the write offset, to at least save the overhead of a Seek
2817 DPRINTF(("committing page %"Z"u", pgno));
2818 memset(&ov, 0, sizeof(ov));
2819 ov.Offset = pos & 0xffffffff;
2820 ov.OffsetHigh = pos >> 16 >> 16;
2821 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2823 DPRINTF(("WriteFile: %d", rc));
2827 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2828 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2830 /* Write previous page(s) */
2831 #ifdef MDB_USE_PWRITEV
2832 wres = pwritev(env->me_fd, iov, n, wpos);
2835 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2837 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2839 DPRINTF(("lseek: %s", strerror(rc)));
2842 wres = writev(env->me_fd, iov, n);
2845 if (wres != wsize) {
2848 DPRINTF(("Write error: %s", strerror(rc)));
2850 rc = EIO; /* TODO: Use which error code? */
2851 DPUTS("short write, filesystem full?");
2862 DPRINTF(("committing page %"Z"u", pgno));
2863 next_pos = pos + size;
2864 iov[n].iov_len = size;
2865 iov[n].iov_base = (char *)dp;
2871 for (i = keep; ++i <= pagecount; ) {
2873 /* This is a page we skipped above */
2876 dl[j].mid = dp->mp_pgno;
2879 mdb_dpage_free(env, dp);
2884 txn->mt_dirty_room += i - j;
2890 mdb_txn_commit(MDB_txn *txn)
2896 if (txn == NULL || txn->mt_env == NULL)
2899 if (txn->mt_child) {
2900 rc = mdb_txn_commit(txn->mt_child);
2901 txn->mt_child = NULL;
2908 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2909 mdb_dbis_update(txn, 1);
2910 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2915 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2916 DPUTS("error flag is set, can't commit");
2918 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2923 if (txn->mt_parent) {
2924 MDB_txn *parent = txn->mt_parent;
2927 unsigned x, y, len, ps_len;
2929 /* Append our free list to parent's */
2930 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2933 mdb_midl_free(txn->mt_free_pgs);
2934 /* Failures after this must either undo the changes
2935 * to the parent or set MDB_TXN_ERROR in the parent.
2938 parent->mt_next_pgno = txn->mt_next_pgno;
2939 parent->mt_flags = txn->mt_flags;
2941 /* Merge our cursors into parent's and close them */
2942 mdb_cursors_close(txn, 1);
2944 /* Update parent's DB table. */
2945 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2946 parent->mt_numdbs = txn->mt_numdbs;
2947 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2948 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2949 for (i=2; i<txn->mt_numdbs; i++) {
2950 /* preserve parent's DB_NEW status */
2951 x = parent->mt_dbflags[i] & DB_NEW;
2952 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2955 dst = parent->mt_u.dirty_list;
2956 src = txn->mt_u.dirty_list;
2957 /* Remove anything in our dirty list from parent's spill list */
2958 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2960 pspill[0] = (pgno_t)-1;
2961 /* Mark our dirty pages as deleted in parent spill list */
2962 for (i=0, len=src[0].mid; ++i <= len; ) {
2963 MDB_ID pn = src[i].mid << 1;
2964 while (pn > pspill[x])
2966 if (pn == pspill[x]) {
2971 /* Squash deleted pagenums if we deleted any */
2972 for (x=y; ++x <= ps_len; )
2973 if (!(pspill[x] & 1))
2974 pspill[++y] = pspill[x];
2978 /* Find len = length of merging our dirty list with parent's */
2980 dst[0].mid = 0; /* simplify loops */
2981 if (parent->mt_parent) {
2982 len = x + src[0].mid;
2983 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2984 for (i = x; y && i; y--) {
2985 pgno_t yp = src[y].mid;
2986 while (yp < dst[i].mid)
2988 if (yp == dst[i].mid) {
2993 } else { /* Simplify the above for single-ancestor case */
2994 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2996 /* Merge our dirty list with parent's */
2998 for (i = len; y; dst[i--] = src[y--]) {
2999 pgno_t yp = src[y].mid;
3000 while (yp < dst[x].mid)
3001 dst[i--] = dst[x--];
3002 if (yp == dst[x].mid)
3003 free(dst[x--].mptr);
3005 mdb_tassert(txn, i == x);
3007 free(txn->mt_u.dirty_list);
3008 parent->mt_dirty_room = txn->mt_dirty_room;
3009 if (txn->mt_spill_pgs) {
3010 if (parent->mt_spill_pgs) {
3011 /* TODO: Prevent failure here, so parent does not fail */
3012 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3014 parent->mt_flags |= MDB_TXN_ERROR;
3015 mdb_midl_free(txn->mt_spill_pgs);
3016 mdb_midl_sort(parent->mt_spill_pgs);
3018 parent->mt_spill_pgs = txn->mt_spill_pgs;
3022 parent->mt_child = NULL;
3023 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3028 if (txn != env->me_txn) {
3029 DPUTS("attempt to commit unknown transaction");
3034 mdb_cursors_close(txn, 0);
3036 if (!txn->mt_u.dirty_list[0].mid &&
3037 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3040 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3041 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3043 /* Update DB root pointers */
3044 if (txn->mt_numdbs > 2) {
3048 data.mv_size = sizeof(MDB_db);
3050 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3051 for (i = 2; i < txn->mt_numdbs; i++) {
3052 if (txn->mt_dbflags[i] & DB_DIRTY) {
3053 data.mv_data = &txn->mt_dbs[i];
3054 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3061 rc = mdb_freelist_save(txn);
3065 mdb_midl_free(env->me_pghead);
3066 env->me_pghead = NULL;
3067 if (mdb_midl_shrink(&txn->mt_free_pgs))
3068 env->me_free_pgs = txn->mt_free_pgs;
3074 if ((rc = mdb_page_flush(txn, 0)) ||
3075 (rc = mdb_env_sync(env, 0)) ||
3076 (rc = mdb_env_write_meta(txn)))
3082 mdb_dbis_update(txn, 1);
3085 UNLOCK_MUTEX_W(env);
3095 /** Read the environment parameters of a DB environment before
3096 * mapping it into memory.
3097 * @param[in] env the environment handle
3098 * @param[out] meta address of where to store the meta information
3099 * @return 0 on success, non-zero on failure.
3102 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3108 enum { Size = sizeof(pbuf) };
3110 /* We don't know the page size yet, so use a minimum value.
3111 * Read both meta pages so we can use the latest one.
3114 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3118 memset(&ov, 0, sizeof(ov));
3120 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3121 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3124 rc = pread(env->me_fd, &pbuf, Size, off);
3127 if (rc == 0 && off == 0)
3129 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3130 DPRINTF(("read: %s", mdb_strerror(rc)));
3134 p = (MDB_page *)&pbuf;
3136 if (!F_ISSET(p->mp_flags, P_META)) {
3137 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3142 if (m->mm_magic != MDB_MAGIC) {
3143 DPUTS("meta has invalid magic");
3147 if (m->mm_version != MDB_DATA_VERSION) {
3148 DPRINTF(("database is version %u, expected version %u",
3149 m->mm_version, MDB_DATA_VERSION));
3150 return MDB_VERSION_MISMATCH;
3153 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3159 /** Write the environment parameters of a freshly created DB environment.
3160 * @param[in] env the environment handle
3161 * @param[out] meta address of where to store the meta information
3162 * @return 0 on success, non-zero on failure.
3165 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3173 memset(&ov, 0, sizeof(ov));
3174 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3176 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3179 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3180 len = pwrite(fd, ptr, size, pos); \
3181 rc = (len >= 0); } while(0)
3184 DPUTS("writing new meta page");
3186 psize = env->me_psize;
3188 meta->mm_magic = MDB_MAGIC;
3189 meta->mm_version = MDB_DATA_VERSION;
3190 meta->mm_mapsize = env->me_mapsize;
3191 meta->mm_psize = psize;
3192 meta->mm_last_pg = 1;
3193 meta->mm_flags = env->me_flags & 0xffff;
3194 meta->mm_flags |= MDB_INTEGERKEY;
3195 meta->mm_dbs[0].md_root = P_INVALID;
3196 meta->mm_dbs[1].md_root = P_INVALID;
3198 p = calloc(2, psize);
3200 p->mp_flags = P_META;
3201 *(MDB_meta *)METADATA(p) = *meta;
3203 q = (MDB_page *)((char *)p + psize);
3205 q->mp_flags = P_META;
3206 *(MDB_meta *)METADATA(q) = *meta;
3208 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3211 else if ((unsigned) len == psize * 2)
3219 /** Update the environment info to commit a transaction.
3220 * @param[in] txn the transaction that's being committed
3221 * @return 0 on success, non-zero on failure.
3224 mdb_env_write_meta(MDB_txn *txn)
3227 MDB_meta meta, metab, *mp;
3229 int rc, len, toggle;
3238 toggle = txn->mt_txnid & 1;
3239 DPRINTF(("writing meta page %d for root page %"Z"u",
3240 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3243 mp = env->me_metas[toggle];
3245 if (env->me_flags & MDB_WRITEMAP) {
3246 /* Persist any increases of mapsize config */
3247 if (env->me_mapsize > mp->mm_mapsize)
3248 mp->mm_mapsize = env->me_mapsize;
3249 mp->mm_dbs[0] = txn->mt_dbs[0];
3250 mp->mm_dbs[1] = txn->mt_dbs[1];
3251 mp->mm_last_pg = txn->mt_next_pgno - 1;
3252 mp->mm_txnid = txn->mt_txnid;
3253 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3254 unsigned meta_size = env->me_psize;
3255 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3258 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3259 if (meta_size < env->me_os_psize)
3260 meta_size += meta_size;
3265 if (MDB_MSYNC(ptr, meta_size, rc)) {
3272 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3273 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3275 ptr = (char *)&meta;
3276 if (env->me_mapsize > mp->mm_mapsize) {
3277 /* Persist any increases of mapsize config */
3278 meta.mm_mapsize = env->me_mapsize;
3279 off = offsetof(MDB_meta, mm_mapsize);
3281 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3283 len = sizeof(MDB_meta) - off;
3286 meta.mm_dbs[0] = txn->mt_dbs[0];
3287 meta.mm_dbs[1] = txn->mt_dbs[1];
3288 meta.mm_last_pg = txn->mt_next_pgno - 1;
3289 meta.mm_txnid = txn->mt_txnid;
3292 off += env->me_psize;
3295 /* Write to the SYNC fd */
3296 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3297 env->me_fd : env->me_mfd;
3300 memset(&ov, 0, sizeof(ov));
3302 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3306 rc = pwrite(mfd, ptr, len, off);
3309 rc = rc < 0 ? ErrCode() : EIO;
3310 DPUTS("write failed, disk error?");
3311 /* On a failure, the pagecache still contains the new data.
3312 * Write some old data back, to prevent it from being used.
3313 * Use the non-SYNC fd; we know it will fail anyway.
3315 meta.mm_last_pg = metab.mm_last_pg;
3316 meta.mm_txnid = metab.mm_txnid;
3318 memset(&ov, 0, sizeof(ov));
3320 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3322 r2 = pwrite(env->me_fd, ptr, len, off);
3323 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3326 env->me_flags |= MDB_FATAL_ERROR;
3330 /* Memory ordering issues are irrelevant; since the entire writer
3331 * is wrapped by wmutex, all of these changes will become visible
3332 * after the wmutex is unlocked. Since the DB is multi-version,
3333 * readers will get consistent data regardless of how fresh or
3334 * how stale their view of these values is.
3337 env->me_txns->mti_txnid = txn->mt_txnid;
3342 /** Check both meta pages to see which one is newer.
3343 * @param[in] env the environment handle
3344 * @return meta toggle (0 or 1).
3347 mdb_env_pick_meta(const MDB_env *env)
3349 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3353 mdb_env_create(MDB_env **env)
3357 e = calloc(1, sizeof(MDB_env));
3361 e->me_maxreaders = DEFAULT_READERS;
3362 e->me_maxdbs = e->me_numdbs = 2;
3363 e->me_fd = INVALID_HANDLE_VALUE;
3364 e->me_lfd = INVALID_HANDLE_VALUE;
3365 e->me_mfd = INVALID_HANDLE_VALUE;
3366 #ifdef MDB_USE_POSIX_SEM
3367 e->me_rmutex = SEM_FAILED;
3368 e->me_wmutex = SEM_FAILED;
3370 e->me_pid = getpid();
3371 GET_PAGESIZE(e->me_os_psize);
3372 VGMEMP_CREATE(e,0,0);
3378 mdb_env_map(MDB_env *env, void *addr, int newsize)
3381 unsigned int flags = env->me_flags;
3385 LONG sizelo, sizehi;
3386 sizelo = env->me_mapsize & 0xffffffff;
3387 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3389 /* Windows won't create mappings for zero length files.
3390 * Just allocate the maxsize right now.
3393 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3394 || !SetEndOfFile(env->me_fd)
3395 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3398 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3399 PAGE_READWRITE : PAGE_READONLY,
3400 sizehi, sizelo, NULL);
3403 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3404 FILE_MAP_WRITE : FILE_MAP_READ,
3405 0, 0, env->me_mapsize, addr);
3406 rc = env->me_map ? 0 : ErrCode();
3411 int prot = PROT_READ;
3412 if (flags & MDB_WRITEMAP) {
3414 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3417 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3419 if (env->me_map == MAP_FAILED) {
3424 if (flags & MDB_NORDAHEAD) {
3425 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3427 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3429 #ifdef POSIX_MADV_RANDOM
3430 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3431 #endif /* POSIX_MADV_RANDOM */
3432 #endif /* MADV_RANDOM */
3436 /* Can happen because the address argument to mmap() is just a
3437 * hint. mmap() can pick another, e.g. if the range is in use.
3438 * The MAP_FIXED flag would prevent that, but then mmap could
3439 * instead unmap existing pages to make room for the new map.
3441 if (addr && env->me_map != addr)
3442 return EBUSY; /* TODO: Make a new MDB_* error code? */
3444 p = (MDB_page *)env->me_map;
3445 env->me_metas[0] = METADATA(p);
3446 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3452 mdb_env_set_mapsize(MDB_env *env, size_t size)
3454 /* If env is already open, caller is responsible for making
3455 * sure there are no active txns.
3463 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3464 else if (size < env->me_mapsize) {
3465 /* If the configured size is smaller, make sure it's
3466 * still big enough. Silently round up to minimum if not.
3468 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3472 munmap(env->me_map, env->me_mapsize);
3473 env->me_mapsize = size;
3474 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3475 rc = mdb_env_map(env, old, 1);
3479 env->me_mapsize = size;
3481 env->me_maxpg = env->me_mapsize / env->me_psize;
3486 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3490 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3495 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3497 if (env->me_map || readers < 1)
3499 env->me_maxreaders = readers;
3504 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3506 if (!env || !readers)
3508 *readers = env->me_maxreaders;
3512 /** Further setup required for opening an MDB environment
3515 mdb_env_open2(MDB_env *env)
3517 unsigned int flags = env->me_flags;
3518 int i, newenv = 0, rc;
3522 /* See if we should use QueryLimited */
3524 if ((rc & 0xff) > 5)
3525 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3527 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3530 memset(&meta, 0, sizeof(meta));
3532 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3535 DPUTS("new mdbenv");
3537 env->me_psize = env->me_os_psize;
3538 if (env->me_psize > MAX_PAGESIZE)
3539 env->me_psize = MAX_PAGESIZE;
3541 env->me_psize = meta.mm_psize;
3544 /* Was a mapsize configured? */
3545 if (!env->me_mapsize) {
3546 /* If this is a new environment, take the default,
3547 * else use the size recorded in the existing env.
3549 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3550 } else if (env->me_mapsize < meta.mm_mapsize) {
3551 /* If the configured size is smaller, make sure it's
3552 * still big enough. Silently round up to minimum if not.
3554 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3555 if (env->me_mapsize < minsize)
3556 env->me_mapsize = minsize;
3559 rc = mdb_env_map(env, meta.mm_address, newenv);
3564 if (flags & MDB_FIXEDMAP)
3565 meta.mm_address = env->me_map;
3566 i = mdb_env_init_meta(env, &meta);
3567 if (i != MDB_SUCCESS) {
3572 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3573 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3575 #if !(MDB_MAXKEYSIZE)
3576 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3578 env->me_maxpg = env->me_mapsize / env->me_psize;
3582 int toggle = mdb_env_pick_meta(env);
3583 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3585 DPRINTF(("opened database version %u, pagesize %u",
3586 env->me_metas[0]->mm_version, env->me_psize));
3587 DPRINTF(("using meta page %d", toggle));
3588 DPRINTF(("depth: %u", db->md_depth));
3589 DPRINTF(("entries: %"Z"u", db->md_entries));
3590 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3591 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3592 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3593 DPRINTF(("root: %"Z"u", db->md_root));
3601 /** Release a reader thread's slot in the reader lock table.
3602 * This function is called automatically when a thread exits.
3603 * @param[in] ptr This points to the slot in the reader lock table.
3606 mdb_env_reader_dest(void *ptr)
3608 MDB_reader *reader = ptr;
3614 /** Junk for arranging thread-specific callbacks on Windows. This is
3615 * necessarily platform and compiler-specific. Windows supports up
3616 * to 1088 keys. Let's assume nobody opens more than 64 environments
3617 * in a single process, for now. They can override this if needed.
3619 #ifndef MAX_TLS_KEYS
3620 #define MAX_TLS_KEYS 64
3622 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3623 static int mdb_tls_nkeys;
3625 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3629 case DLL_PROCESS_ATTACH: break;
3630 case DLL_THREAD_ATTACH: break;
3631 case DLL_THREAD_DETACH:
3632 for (i=0; i<mdb_tls_nkeys; i++) {
3633 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3634 mdb_env_reader_dest(r);
3637 case DLL_PROCESS_DETACH: break;
3642 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3644 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3648 /* Force some symbol references.
3649 * _tls_used forces the linker to create the TLS directory if not already done
3650 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3652 #pragma comment(linker, "/INCLUDE:_tls_used")
3653 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3654 #pragma const_seg(".CRT$XLB")
3655 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3656 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3659 #pragma comment(linker, "/INCLUDE:__tls_used")
3660 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3661 #pragma data_seg(".CRT$XLB")
3662 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3664 #endif /* WIN 32/64 */
3665 #endif /* !__GNUC__ */
3668 /** Downgrade the exclusive lock on the region back to shared */
3670 mdb_env_share_locks(MDB_env *env, int *excl)
3672 int rc = 0, toggle = mdb_env_pick_meta(env);
3674 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3679 /* First acquire a shared lock. The Unlock will
3680 * then release the existing exclusive lock.
3682 memset(&ov, 0, sizeof(ov));
3683 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3686 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3692 struct flock lock_info;
3693 /* The shared lock replaces the existing lock */
3694 memset((void *)&lock_info, 0, sizeof(lock_info));
3695 lock_info.l_type = F_RDLCK;
3696 lock_info.l_whence = SEEK_SET;
3697 lock_info.l_start = 0;
3698 lock_info.l_len = 1;
3699 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3700 (rc = ErrCode()) == EINTR) ;
3701 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3708 /** Try to get exlusive lock, otherwise shared.
3709 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3712 mdb_env_excl_lock(MDB_env *env, int *excl)
3716 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3720 memset(&ov, 0, sizeof(ov));
3721 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3728 struct flock lock_info;
3729 memset((void *)&lock_info, 0, sizeof(lock_info));
3730 lock_info.l_type = F_WRLCK;
3731 lock_info.l_whence = SEEK_SET;
3732 lock_info.l_start = 0;
3733 lock_info.l_len = 1;
3734 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3735 (rc = ErrCode()) == EINTR) ;
3739 # ifdef MDB_USE_POSIX_SEM
3740 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3743 lock_info.l_type = F_RDLCK;
3744 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3745 (rc = ErrCode()) == EINTR) ;
3755 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3757 * @(#) $Revision: 5.1 $
3758 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3759 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3761 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3765 * Please do not copyright this code. This code is in the public domain.
3767 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3768 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3769 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3770 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3771 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3772 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3773 * PERFORMANCE OF THIS SOFTWARE.
3776 * chongo <Landon Curt Noll> /\oo/\
3777 * http://www.isthe.com/chongo/
3779 * Share and Enjoy! :-)
3782 typedef unsigned long long mdb_hash_t;
3783 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3785 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3786 * @param[in] val value to hash
3787 * @param[in] hval initial value for hash
3788 * @return 64 bit hash
3790 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3791 * hval arg on the first call.
3794 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3796 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3797 unsigned char *end = s + val->mv_size;
3799 * FNV-1a hash each octet of the string
3802 /* xor the bottom with the current octet */
3803 hval ^= (mdb_hash_t)*s++;
3805 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3806 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3807 (hval << 7) + (hval << 8) + (hval << 40);
3809 /* return our new hash value */
3813 /** Hash the string and output the encoded hash.
3814 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3815 * very short name limits. We don't care about the encoding being reversible,
3816 * we just want to preserve as many bits of the input as possible in a
3817 * small printable string.
3818 * @param[in] str string to hash
3819 * @param[out] encbuf an array of 11 chars to hold the hash
3821 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3824 mdb_pack85(unsigned long l, char *out)
3828 for (i=0; i<5; i++) {
3829 *out++ = mdb_a85[l % 85];
3835 mdb_hash_enc(MDB_val *val, char *encbuf)
3837 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3839 mdb_pack85(h, encbuf);
3840 mdb_pack85(h>>32, encbuf+5);
3845 /** Open and/or initialize the lock region for the environment.
3846 * @param[in] env The MDB environment.
3847 * @param[in] lpath The pathname of the file used for the lock region.
3848 * @param[in] mode The Unix permissions for the file, if we create it.
3849 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3850 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3851 * @return 0 on success, non-zero on failure.
3854 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3857 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3859 # define MDB_ERRCODE_ROFS EROFS
3860 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3861 # define MDB_CLOEXEC O_CLOEXEC
3864 # define MDB_CLOEXEC 0
3871 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3872 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3873 FILE_ATTRIBUTE_NORMAL, NULL);
3875 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3877 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3879 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3884 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3885 /* Lose record locks when exec*() */
3886 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3887 fcntl(env->me_lfd, F_SETFD, fdflags);
3890 if (!(env->me_flags & MDB_NOTLS)) {
3891 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3894 env->me_flags |= MDB_ENV_TXKEY;
3896 /* Windows TLS callbacks need help finding their TLS info. */
3897 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3901 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3905 /* Try to get exclusive lock. If we succeed, then
3906 * nobody is using the lock region and we should initialize it.
3908 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3911 size = GetFileSize(env->me_lfd, NULL);
3913 size = lseek(env->me_lfd, 0, SEEK_END);
3914 if (size == -1) goto fail_errno;
3916 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3917 if (size < rsize && *excl > 0) {
3919 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3920 || !SetEndOfFile(env->me_lfd))
3923 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3927 size = rsize - sizeof(MDB_txninfo);
3928 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3933 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3935 if (!mh) goto fail_errno;
3936 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3938 if (!env->me_txns) goto fail_errno;
3940 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3942 if (m == MAP_FAILED) goto fail_errno;
3948 BY_HANDLE_FILE_INFORMATION stbuf;
3957 if (!mdb_sec_inited) {
3958 InitializeSecurityDescriptor(&mdb_null_sd,
3959 SECURITY_DESCRIPTOR_REVISION);
3960 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3961 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3962 mdb_all_sa.bInheritHandle = FALSE;
3963 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3966 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3967 idbuf.volume = stbuf.dwVolumeSerialNumber;
3968 idbuf.nhigh = stbuf.nFileIndexHigh;
3969 idbuf.nlow = stbuf.nFileIndexLow;
3970 val.mv_data = &idbuf;
3971 val.mv_size = sizeof(idbuf);
3972 mdb_hash_enc(&val, encbuf);
3973 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3974 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3975 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3976 if (!env->me_rmutex) goto fail_errno;
3977 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3978 if (!env->me_wmutex) goto fail_errno;
3979 #elif defined(MDB_USE_POSIX_SEM)
3988 #if defined(__NetBSD__)
3989 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3991 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3992 idbuf.dev = stbuf.st_dev;
3993 idbuf.ino = stbuf.st_ino;
3994 val.mv_data = &idbuf;
3995 val.mv_size = sizeof(idbuf);
3996 mdb_hash_enc(&val, encbuf);
3997 #ifdef MDB_SHORT_SEMNAMES
3998 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4000 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4001 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4002 /* Clean up after a previous run, if needed: Try to
4003 * remove both semaphores before doing anything else.
4005 sem_unlink(env->me_txns->mti_rmname);
4006 sem_unlink(env->me_txns->mti_wmname);
4007 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4008 O_CREAT|O_EXCL, mode, 1);
4009 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4010 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4011 O_CREAT|O_EXCL, mode, 1);
4012 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4013 #else /* MDB_USE_POSIX_SEM */
4014 pthread_mutexattr_t mattr;
4016 if ((rc = pthread_mutexattr_init(&mattr))
4017 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4018 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4019 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4021 pthread_mutexattr_destroy(&mattr);
4022 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4024 env->me_txns->mti_magic = MDB_MAGIC;
4025 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4026 env->me_txns->mti_txnid = 0;
4027 env->me_txns->mti_numreaders = 0;
4030 if (env->me_txns->mti_magic != MDB_MAGIC) {
4031 DPUTS("lock region has invalid magic");
4035 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4036 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4037 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4038 rc = MDB_VERSION_MISMATCH;
4042 if (rc && rc != EACCES && rc != EAGAIN) {
4046 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4047 if (!env->me_rmutex) goto fail_errno;
4048 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4049 if (!env->me_wmutex) goto fail_errno;
4050 #elif defined(MDB_USE_POSIX_SEM)
4051 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4052 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4053 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4054 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4065 /** The name of the lock file in the DB environment */
4066 #define LOCKNAME "/lock.mdb"
4067 /** The name of the data file in the DB environment */
4068 #define DATANAME "/data.mdb"
4069 /** The suffix of the lock file when no subdir is used */
4070 #define LOCKSUFF "-lock"
4071 /** Only a subset of the @ref mdb_env flags can be changed
4072 * at runtime. Changing other flags requires closing the
4073 * environment and re-opening it with the new flags.
4075 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4076 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4077 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4079 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4080 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4084 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4086 int oflags, rc, len, excl = -1;
4087 char *lpath, *dpath;
4089 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4093 if (flags & MDB_NOSUBDIR) {
4094 rc = len + sizeof(LOCKSUFF) + len + 1;
4096 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4101 if (flags & MDB_NOSUBDIR) {
4102 dpath = lpath + len + sizeof(LOCKSUFF);
4103 sprintf(lpath, "%s" LOCKSUFF, path);
4104 strcpy(dpath, path);
4106 dpath = lpath + len + sizeof(LOCKNAME);
4107 sprintf(lpath, "%s" LOCKNAME, path);
4108 sprintf(dpath, "%s" DATANAME, path);
4112 flags |= env->me_flags;
4113 if (flags & MDB_RDONLY) {
4114 /* silently ignore WRITEMAP when we're only getting read access */
4115 flags &= ~MDB_WRITEMAP;
4117 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4118 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4121 env->me_flags = flags |= MDB_ENV_ACTIVE;
4125 env->me_path = strdup(path);
4126 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4127 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4128 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4133 /* For RDONLY, get lockfile after we know datafile exists */
4134 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4135 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4141 if (F_ISSET(flags, MDB_RDONLY)) {
4142 oflags = GENERIC_READ;
4143 len = OPEN_EXISTING;
4145 oflags = GENERIC_READ|GENERIC_WRITE;
4148 mode = FILE_ATTRIBUTE_NORMAL;
4149 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4150 NULL, len, mode, NULL);
4152 if (F_ISSET(flags, MDB_RDONLY))
4155 oflags = O_RDWR | O_CREAT;
4157 env->me_fd = open(dpath, oflags, mode);
4159 if (env->me_fd == INVALID_HANDLE_VALUE) {
4164 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4165 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4170 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4171 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4172 env->me_mfd = env->me_fd;
4174 /* Synchronous fd for meta writes. Needed even with
4175 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4178 len = OPEN_EXISTING;
4179 env->me_mfd = CreateFile(dpath, oflags,
4180 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4181 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4184 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4186 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4191 DPRINTF(("opened dbenv %p", (void *) env));
4193 rc = mdb_env_share_locks(env, &excl);
4197 if (!((flags & MDB_RDONLY) ||
4198 (env->me_pbuf = calloc(1, env->me_psize))))
4204 mdb_env_close0(env, excl);
4210 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4212 mdb_env_close0(MDB_env *env, int excl)
4216 if (!(env->me_flags & MDB_ENV_ACTIVE))
4219 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4220 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4221 free(env->me_dbxs[i].md_name.mv_data);
4224 free(env->me_dbflags);
4227 free(env->me_dirty_list);
4228 mdb_midl_free(env->me_free_pgs);
4230 if (env->me_flags & MDB_ENV_TXKEY) {
4231 pthread_key_delete(env->me_txkey);
4233 /* Delete our key from the global list */
4234 for (i=0; i<mdb_tls_nkeys; i++)
4235 if (mdb_tls_keys[i] == env->me_txkey) {
4236 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4244 munmap(env->me_map, env->me_mapsize);
4246 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4247 (void) close(env->me_mfd);
4248 if (env->me_fd != INVALID_HANDLE_VALUE)
4249 (void) close(env->me_fd);
4251 MDB_PID_T pid = env->me_pid;
4252 /* Clearing readers is done in this function because
4253 * me_txkey with its destructor must be disabled first.
4255 for (i = env->me_numreaders; --i >= 0; )
4256 if (env->me_txns->mti_readers[i].mr_pid == pid)
4257 env->me_txns->mti_readers[i].mr_pid = 0;
4259 if (env->me_rmutex) {
4260 CloseHandle(env->me_rmutex);
4261 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4263 /* Windows automatically destroys the mutexes when
4264 * the last handle closes.
4266 #elif defined(MDB_USE_POSIX_SEM)
4267 if (env->me_rmutex != SEM_FAILED) {
4268 sem_close(env->me_rmutex);
4269 if (env->me_wmutex != SEM_FAILED)
4270 sem_close(env->me_wmutex);
4271 /* If we have the filelock: If we are the
4272 * only remaining user, clean up semaphores.
4275 mdb_env_excl_lock(env, &excl);
4277 sem_unlink(env->me_txns->mti_rmname);
4278 sem_unlink(env->me_txns->mti_wmname);
4282 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4284 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4287 /* Unlock the lockfile. Windows would have unlocked it
4288 * after closing anyway, but not necessarily at once.
4290 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4293 (void) close(env->me_lfd);
4296 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4300 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4302 MDB_txn *txn = NULL;
4308 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4312 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4315 /* Do the lock/unlock of the reader mutex before starting the
4316 * write txn. Otherwise other read txns could block writers.
4318 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4323 /* We must start the actual read txn after blocking writers */
4324 mdb_txn_reset0(txn, "reset-stage1");
4326 /* Temporarily block writers until we snapshot the meta pages */
4329 rc = mdb_txn_renew0(txn);
4331 UNLOCK_MUTEX_W(env);
4336 wsize = env->me_psize * 2;
4340 DO_WRITE(rc, fd, ptr, w2, len);
4344 } else if (len > 0) {
4350 /* Non-blocking or async handles are not supported */
4356 UNLOCK_MUTEX_W(env);
4361 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4363 if (wsize > MAX_WRITE)
4367 DO_WRITE(rc, fd, ptr, w2, len);
4371 } else if (len > 0) {
4388 mdb_env_copy(MDB_env *env, const char *path)
4392 HANDLE newfd = INVALID_HANDLE_VALUE;
4394 if (env->me_flags & MDB_NOSUBDIR) {
4395 lpath = (char *)path;
4398 len += sizeof(DATANAME);
4399 lpath = malloc(len);
4402 sprintf(lpath, "%s" DATANAME, path);
4405 /* The destination path must exist, but the destination file must not.
4406 * We don't want the OS to cache the writes, since the source data is
4407 * already in the OS cache.
4410 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4411 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4413 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4415 if (newfd == INVALID_HANDLE_VALUE) {
4421 /* Set O_DIRECT if the file system supports it */
4422 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4423 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4425 #ifdef F_NOCACHE /* __APPLE__ */
4426 rc = fcntl(newfd, F_NOCACHE, 1);
4433 rc = mdb_env_copyfd(env, newfd);
4436 if (!(env->me_flags & MDB_NOSUBDIR))
4438 if (newfd != INVALID_HANDLE_VALUE)
4439 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4446 mdb_env_close(MDB_env *env)
4453 VGMEMP_DESTROY(env);
4454 while ((dp = env->me_dpages) != NULL) {
4455 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4456 env->me_dpages = dp->mp_next;
4460 mdb_env_close0(env, 0);
4464 /** Compare two items pointing at aligned size_t's */
4466 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4468 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4469 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4472 /** Compare two items pointing at aligned unsigned int's */
4474 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4476 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4477 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4480 /** Compare two items pointing at unsigned ints of unknown alignment.
4481 * Nodes and keys are guaranteed to be 2-byte aligned.
4484 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4486 #if BYTE_ORDER == LITTLE_ENDIAN
4487 unsigned short *u, *c;
4490 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4491 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4494 } while(!x && u > (unsigned short *)a->mv_data);
4497 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4501 /** Compare two items lexically */
4503 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4510 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4516 diff = memcmp(a->mv_data, b->mv_data, len);
4517 return diff ? diff : len_diff<0 ? -1 : len_diff;
4520 /** Compare two items in reverse byte order */
4522 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4524 const unsigned char *p1, *p2, *p1_lim;
4528 p1_lim = (const unsigned char *)a->mv_data;
4529 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4530 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4532 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4538 while (p1 > p1_lim) {
4539 diff = *--p1 - *--p2;
4543 return len_diff<0 ? -1 : len_diff;
4546 /** Search for key within a page, using binary search.
4547 * Returns the smallest entry larger or equal to the key.
4548 * If exactp is non-null, stores whether the found entry was an exact match
4549 * in *exactp (1 or 0).
4550 * Updates the cursor index with the index of the found entry.
4551 * If no entry larger or equal to the key is found, returns NULL.
4554 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4556 unsigned int i = 0, nkeys;
4559 MDB_page *mp = mc->mc_pg[mc->mc_top];
4560 MDB_node *node = NULL;
4565 nkeys = NUMKEYS(mp);
4567 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4568 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4571 low = IS_LEAF(mp) ? 0 : 1;
4573 cmp = mc->mc_dbx->md_cmp;
4575 /* Branch pages have no data, so if using integer keys,
4576 * alignment is guaranteed. Use faster mdb_cmp_int.
4578 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4579 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4586 nodekey.mv_size = mc->mc_db->md_pad;
4587 node = NODEPTR(mp, 0); /* fake */
4588 while (low <= high) {
4589 i = (low + high) >> 1;
4590 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4591 rc = cmp(key, &nodekey);
4592 DPRINTF(("found leaf index %u [%s], rc = %i",
4593 i, DKEY(&nodekey), rc));
4602 while (low <= high) {
4603 i = (low + high) >> 1;
4605 node = NODEPTR(mp, i);
4606 nodekey.mv_size = NODEKSZ(node);
4607 nodekey.mv_data = NODEKEY(node);
4609 rc = cmp(key, &nodekey);
4612 DPRINTF(("found leaf index %u [%s], rc = %i",
4613 i, DKEY(&nodekey), rc));
4615 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4616 i, DKEY(&nodekey), NODEPGNO(node), rc));
4627 if (rc > 0) { /* Found entry is less than the key. */
4628 i++; /* Skip to get the smallest entry larger than key. */
4630 node = NODEPTR(mp, i);
4633 *exactp = (rc == 0 && nkeys > 0);
4634 /* store the key index */
4635 mc->mc_ki[mc->mc_top] = i;
4637 /* There is no entry larger or equal to the key. */
4640 /* nodeptr is fake for LEAF2 */
4646 mdb_cursor_adjust(MDB_cursor *mc, func)
4650 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4651 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4658 /** Pop a page off the top of the cursor's stack. */
4660 mdb_cursor_pop(MDB_cursor *mc)
4664 MDB_page *top = mc->mc_pg[mc->mc_top];
4670 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4671 DDBI(mc), (void *) mc));
4675 /** Push a page onto the top of the cursor's stack. */
4677 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4679 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4680 DDBI(mc), (void *) mc));
4682 if (mc->mc_snum >= CURSOR_STACK) {
4683 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4684 return MDB_CURSOR_FULL;
4687 mc->mc_top = mc->mc_snum++;
4688 mc->mc_pg[mc->mc_top] = mp;
4689 mc->mc_ki[mc->mc_top] = 0;
4694 /** Find the address of the page corresponding to a given page number.
4695 * @param[in] txn the transaction for this access.
4696 * @param[in] pgno the page number for the page to retrieve.
4697 * @param[out] ret address of a pointer where the page's address will be stored.
4698 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4699 * @return 0 on success, non-zero on failure.
4702 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4704 MDB_env *env = txn->mt_env;
4708 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4712 MDB_ID2L dl = tx2->mt_u.dirty_list;
4714 /* Spilled pages were dirtied in this txn and flushed
4715 * because the dirty list got full. Bring this page
4716 * back in from the map (but don't unspill it here,
4717 * leave that unless page_touch happens again).
4719 if (tx2->mt_spill_pgs) {
4720 MDB_ID pn = pgno << 1;
4721 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4722 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4723 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4728 unsigned x = mdb_mid2l_search(dl, pgno);
4729 if (x <= dl[0].mid && dl[x].mid == pgno) {
4735 } while ((tx2 = tx2->mt_parent) != NULL);
4738 if (pgno < txn->mt_next_pgno) {
4740 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4742 DPRINTF(("page %"Z"u not found", pgno));
4743 txn->mt_flags |= MDB_TXN_ERROR;
4744 return MDB_PAGE_NOTFOUND;
4754 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4755 * The cursor is at the root page, set up the rest of it.
4758 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4760 MDB_page *mp = mc->mc_pg[mc->mc_top];
4764 while (IS_BRANCH(mp)) {
4768 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4769 mdb_cassert(mc, NUMKEYS(mp) > 1);
4770 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4772 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4774 if (flags & MDB_PS_LAST)
4775 i = NUMKEYS(mp) - 1;
4778 node = mdb_node_search(mc, key, &exact);
4780 i = NUMKEYS(mp) - 1;
4782 i = mc->mc_ki[mc->mc_top];
4784 mdb_cassert(mc, i > 0);
4788 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4791 mdb_cassert(mc, i < NUMKEYS(mp));
4792 node = NODEPTR(mp, i);
4794 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4797 mc->mc_ki[mc->mc_top] = i;
4798 if ((rc = mdb_cursor_push(mc, mp)))
4801 if (flags & MDB_PS_MODIFY) {
4802 if ((rc = mdb_page_touch(mc)) != 0)
4804 mp = mc->mc_pg[mc->mc_top];
4809 DPRINTF(("internal error, index points to a %02X page!?",
4811 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4812 return MDB_CORRUPTED;
4815 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4816 key ? DKEY(key) : "null"));
4817 mc->mc_flags |= C_INITIALIZED;
4818 mc->mc_flags &= ~C_EOF;
4823 /** Search for the lowest key under the current branch page.
4824 * This just bypasses a NUMKEYS check in the current page
4825 * before calling mdb_page_search_root(), because the callers
4826 * are all in situations where the current page is known to
4830 mdb_page_search_lowest(MDB_cursor *mc)
4832 MDB_page *mp = mc->mc_pg[mc->mc_top];
4833 MDB_node *node = NODEPTR(mp, 0);
4836 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4839 mc->mc_ki[mc->mc_top] = 0;
4840 if ((rc = mdb_cursor_push(mc, mp)))
4842 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4845 /** Search for the page a given key should be in.
4846 * Push it and its parent pages on the cursor stack.
4847 * @param[in,out] mc the cursor for this operation.
4848 * @param[in] key the key to search for, or NULL for first/last page.
4849 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4850 * are touched (updated with new page numbers).
4851 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4852 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4853 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4854 * @return 0 on success, non-zero on failure.
4857 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4862 /* Make sure the txn is still viable, then find the root from
4863 * the txn's db table and set it as the root of the cursor's stack.
4865 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4866 DPUTS("transaction has failed, must abort");
4869 /* Make sure we're using an up-to-date root */
4870 if (*mc->mc_dbflag & DB_STALE) {
4872 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4873 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4880 MDB_node *leaf = mdb_node_search(&mc2,
4881 &mc->mc_dbx->md_name, &exact);
4883 return MDB_NOTFOUND;
4884 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4887 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4889 /* The txn may not know this DBI, or another process may
4890 * have dropped and recreated the DB with other flags.
4892 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4893 return MDB_INCOMPATIBLE;
4894 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4896 *mc->mc_dbflag &= ~DB_STALE;
4898 root = mc->mc_db->md_root;
4900 if (root == P_INVALID) { /* Tree is empty. */
4901 DPUTS("tree is empty");
4902 return MDB_NOTFOUND;
4906 mdb_cassert(mc, root > 1);
4907 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4908 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4914 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4915 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4917 if (flags & MDB_PS_MODIFY) {
4918 if ((rc = mdb_page_touch(mc)))
4922 if (flags & MDB_PS_ROOTONLY)
4925 return mdb_page_search_root(mc, key, flags);
4929 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4931 MDB_txn *txn = mc->mc_txn;
4932 pgno_t pg = mp->mp_pgno;
4933 unsigned x = 0, ovpages = mp->mp_pages;
4934 MDB_env *env = txn->mt_env;
4935 MDB_IDL sl = txn->mt_spill_pgs;
4936 MDB_ID pn = pg << 1;
4939 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4940 /* If the page is dirty or on the spill list we just acquired it,
4941 * so we should give it back to our current free list, if any.
4942 * Otherwise put it onto the list of pages we freed in this txn.
4944 * Won't create me_pghead: me_pglast must be inited along with it.
4945 * Unsupported in nested txns: They would need to hide the page
4946 * range in ancestor txns' dirty and spilled lists.
4948 if (env->me_pghead &&
4950 ((mp->mp_flags & P_DIRTY) ||
4951 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4955 MDB_ID2 *dl, ix, iy;
4956 rc = mdb_midl_need(&env->me_pghead, ovpages);
4959 if (!(mp->mp_flags & P_DIRTY)) {
4960 /* This page is no longer spilled */
4967 /* Remove from dirty list */
4968 dl = txn->mt_u.dirty_list;
4970 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4976 mdb_cassert(mc, x > 1);
4978 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4979 txn->mt_flags |= MDB_TXN_ERROR;
4980 return MDB_CORRUPTED;
4983 if (!(env->me_flags & MDB_WRITEMAP))
4984 mdb_dpage_free(env, mp);
4986 /* Insert in me_pghead */
4987 mop = env->me_pghead;
4988 j = mop[0] + ovpages;
4989 for (i = mop[0]; i && mop[i] < pg; i--)
4995 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4999 mc->mc_db->md_overflow_pages -= ovpages;
5003 /** Return the data associated with a given node.
5004 * @param[in] txn The transaction for this operation.
5005 * @param[in] leaf The node being read.
5006 * @param[out] data Updated to point to the node's data.
5007 * @return 0 on success, non-zero on failure.
5010 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5012 MDB_page *omp; /* overflow page */
5016 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5017 data->mv_size = NODEDSZ(leaf);
5018 data->mv_data = NODEDATA(leaf);
5022 /* Read overflow data.
5024 data->mv_size = NODEDSZ(leaf);
5025 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5026 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5027 DPRINTF(("read overflow page %"Z"u failed", pgno));
5030 data->mv_data = METADATA(omp);
5036 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5037 MDB_val *key, MDB_val *data)
5044 if (key == NULL || data == NULL)
5047 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5049 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5052 if (txn->mt_flags & MDB_TXN_ERROR)
5055 mdb_cursor_init(&mc, txn, dbi, &mx);
5056 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5059 /** Find a sibling for a page.
5060 * Replaces the page at the top of the cursor's stack with the
5061 * specified sibling, if one exists.
5062 * @param[in] mc The cursor for this operation.
5063 * @param[in] move_right Non-zero if the right sibling is requested,
5064 * otherwise the left sibling.
5065 * @return 0 on success, non-zero on failure.
5068 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5074 if (mc->mc_snum < 2) {
5075 return MDB_NOTFOUND; /* root has no siblings */
5079 DPRINTF(("parent page is page %"Z"u, index %u",
5080 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5082 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5083 : (mc->mc_ki[mc->mc_top] == 0)) {
5084 DPRINTF(("no more keys left, moving to %s sibling",
5085 move_right ? "right" : "left"));
5086 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5087 /* undo cursor_pop before returning */
5094 mc->mc_ki[mc->mc_top]++;
5096 mc->mc_ki[mc->mc_top]--;
5097 DPRINTF(("just moving to %s index key %u",
5098 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5100 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5102 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5103 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5104 /* mc will be inconsistent if caller does mc_snum++ as above */
5105 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5109 mdb_cursor_push(mc, mp);
5111 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5116 /** Move the cursor to the next data item. */
5118 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5124 if (mc->mc_flags & C_EOF) {
5125 return MDB_NOTFOUND;
5128 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5130 mp = mc->mc_pg[mc->mc_top];
5132 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5133 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5134 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5135 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5136 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5137 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5138 if (rc == MDB_SUCCESS)
5139 MDB_GET_KEY(leaf, key);
5144 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5145 if (op == MDB_NEXT_DUP)
5146 return MDB_NOTFOUND;
5150 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5151 mdb_dbg_pgno(mp), (void *) mc));
5152 if (mc->mc_flags & C_DEL)
5155 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5156 DPUTS("=====> move to next sibling page");
5157 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5158 mc->mc_flags |= C_EOF;
5161 mp = mc->mc_pg[mc->mc_top];
5162 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5164 mc->mc_ki[mc->mc_top]++;
5167 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5168 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5171 key->mv_size = mc->mc_db->md_pad;
5172 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5176 mdb_cassert(mc, IS_LEAF(mp));
5177 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5179 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5180 mdb_xcursor_init1(mc, leaf);
5183 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5186 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5187 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5188 if (rc != MDB_SUCCESS)
5193 MDB_GET_KEY(leaf, key);
5197 /** Move the cursor to the previous data item. */
5199 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5205 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5207 mp = mc->mc_pg[mc->mc_top];
5209 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5210 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5211 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5212 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5213 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5214 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5215 if (rc == MDB_SUCCESS)
5216 MDB_GET_KEY(leaf, key);
5220 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5221 if (op == MDB_PREV_DUP)
5222 return MDB_NOTFOUND;
5227 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5228 mdb_dbg_pgno(mp), (void *) mc));
5230 if (mc->mc_ki[mc->mc_top] == 0) {
5231 DPUTS("=====> move to prev sibling page");
5232 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5235 mp = mc->mc_pg[mc->mc_top];
5236 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5237 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5239 mc->mc_ki[mc->mc_top]--;
5241 mc->mc_flags &= ~C_EOF;
5243 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5244 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5247 key->mv_size = mc->mc_db->md_pad;
5248 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5252 mdb_cassert(mc, IS_LEAF(mp));
5253 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5255 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5256 mdb_xcursor_init1(mc, leaf);
5259 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5262 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5263 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5264 if (rc != MDB_SUCCESS)
5269 MDB_GET_KEY(leaf, key);
5273 /** Set the cursor on a specific data item. */
5275 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5276 MDB_cursor_op op, int *exactp)
5280 MDB_node *leaf = NULL;
5283 if (key->mv_size == 0)
5284 return MDB_BAD_VALSIZE;
5287 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5289 /* See if we're already on the right page */
5290 if (mc->mc_flags & C_INITIALIZED) {
5293 mp = mc->mc_pg[mc->mc_top];
5295 mc->mc_ki[mc->mc_top] = 0;
5296 return MDB_NOTFOUND;
5298 if (mp->mp_flags & P_LEAF2) {
5299 nodekey.mv_size = mc->mc_db->md_pad;
5300 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5302 leaf = NODEPTR(mp, 0);
5303 MDB_GET_KEY2(leaf, nodekey);
5305 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5307 /* Probably happens rarely, but first node on the page
5308 * was the one we wanted.
5310 mc->mc_ki[mc->mc_top] = 0;
5317 unsigned int nkeys = NUMKEYS(mp);
5319 if (mp->mp_flags & P_LEAF2) {
5320 nodekey.mv_data = LEAF2KEY(mp,
5321 nkeys-1, nodekey.mv_size);
5323 leaf = NODEPTR(mp, nkeys-1);
5324 MDB_GET_KEY2(leaf, nodekey);
5326 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5328 /* last node was the one we wanted */
5329 mc->mc_ki[mc->mc_top] = nkeys-1;
5335 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5336 /* This is definitely the right page, skip search_page */
5337 if (mp->mp_flags & P_LEAF2) {
5338 nodekey.mv_data = LEAF2KEY(mp,
5339 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5341 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5342 MDB_GET_KEY2(leaf, nodekey);
5344 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5346 /* current node was the one we wanted */
5356 /* If any parents have right-sibs, search.
5357 * Otherwise, there's nothing further.
5359 for (i=0; i<mc->mc_top; i++)
5361 NUMKEYS(mc->mc_pg[i])-1)
5363 if (i == mc->mc_top) {
5364 /* There are no other pages */
5365 mc->mc_ki[mc->mc_top] = nkeys;
5366 return MDB_NOTFOUND;
5370 /* There are no other pages */
5371 mc->mc_ki[mc->mc_top] = 0;
5372 if (op == MDB_SET_RANGE) {
5376 return MDB_NOTFOUND;
5380 rc = mdb_page_search(mc, key, 0);
5381 if (rc != MDB_SUCCESS)
5384 mp = mc->mc_pg[mc->mc_top];
5385 mdb_cassert(mc, IS_LEAF(mp));
5388 leaf = mdb_node_search(mc, key, exactp);
5389 if (exactp != NULL && !*exactp) {
5390 /* MDB_SET specified and not an exact match. */
5391 return MDB_NOTFOUND;
5395 DPUTS("===> inexact leaf not found, goto sibling");
5396 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5397 return rc; /* no entries matched */
5398 mp = mc->mc_pg[mc->mc_top];
5399 mdb_cassert(mc, IS_LEAF(mp));
5400 leaf = NODEPTR(mp, 0);
5404 mc->mc_flags |= C_INITIALIZED;
5405 mc->mc_flags &= ~C_EOF;
5408 key->mv_size = mc->mc_db->md_pad;
5409 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5413 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5414 mdb_xcursor_init1(mc, leaf);
5417 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5418 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5419 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5422 if (op == MDB_GET_BOTH) {
5428 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5429 if (rc != MDB_SUCCESS)
5432 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5434 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5436 rc = mc->mc_dbx->md_dcmp(data, &d2);
5438 if (op == MDB_GET_BOTH || rc > 0)
5439 return MDB_NOTFOUND;
5446 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5447 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5452 /* The key already matches in all other cases */
5453 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5454 MDB_GET_KEY(leaf, key);
5455 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5460 /** Move the cursor to the first item in the database. */
5462 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5468 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5470 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5471 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5472 if (rc != MDB_SUCCESS)
5475 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5477 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5478 mc->mc_flags |= C_INITIALIZED;
5479 mc->mc_flags &= ~C_EOF;
5481 mc->mc_ki[mc->mc_top] = 0;
5483 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5484 key->mv_size = mc->mc_db->md_pad;
5485 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5490 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5491 mdb_xcursor_init1(mc, leaf);
5492 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5496 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5500 MDB_GET_KEY(leaf, key);
5504 /** Move the cursor to the last item in the database. */
5506 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5512 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5514 if (!(mc->mc_flags & C_EOF)) {
5516 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5517 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5518 if (rc != MDB_SUCCESS)
5521 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5524 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5525 mc->mc_flags |= C_INITIALIZED|C_EOF;
5526 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5528 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5529 key->mv_size = mc->mc_db->md_pad;
5530 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5535 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5536 mdb_xcursor_init1(mc, leaf);
5537 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5541 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5546 MDB_GET_KEY(leaf, key);
5551 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5556 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5561 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5565 case MDB_GET_CURRENT:
5566 if (!(mc->mc_flags & C_INITIALIZED)) {
5569 MDB_page *mp = mc->mc_pg[mc->mc_top];
5570 int nkeys = NUMKEYS(mp);
5571 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5572 mc->mc_ki[mc->mc_top] = nkeys;
5578 key->mv_size = mc->mc_db->md_pad;
5579 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5581 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5582 MDB_GET_KEY(leaf, key);
5584 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5585 if (mc->mc_flags & C_DEL)
5586 mdb_xcursor_init1(mc, leaf);
5587 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5589 rc = mdb_node_read(mc->mc_txn, leaf, data);
5596 case MDB_GET_BOTH_RANGE:
5601 if (mc->mc_xcursor == NULL) {
5602 rc = MDB_INCOMPATIBLE;
5612 rc = mdb_cursor_set(mc, key, data, op,
5613 op == MDB_SET_RANGE ? NULL : &exact);
5616 case MDB_GET_MULTIPLE:
5617 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5621 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5622 rc = MDB_INCOMPATIBLE;
5626 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5627 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5630 case MDB_NEXT_MULTIPLE:
5635 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5636 rc = MDB_INCOMPATIBLE;
5639 if (!(mc->mc_flags & C_INITIALIZED))
5640 rc = mdb_cursor_first(mc, key, data);
5642 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5643 if (rc == MDB_SUCCESS) {
5644 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5647 mx = &mc->mc_xcursor->mx_cursor;
5648 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5650 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5651 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5659 case MDB_NEXT_NODUP:
5660 if (!(mc->mc_flags & C_INITIALIZED))
5661 rc = mdb_cursor_first(mc, key, data);
5663 rc = mdb_cursor_next(mc, key, data, op);
5667 case MDB_PREV_NODUP:
5668 if (!(mc->mc_flags & C_INITIALIZED)) {
5669 rc = mdb_cursor_last(mc, key, data);
5672 mc->mc_flags |= C_INITIALIZED;
5673 mc->mc_ki[mc->mc_top]++;
5675 rc = mdb_cursor_prev(mc, key, data, op);
5678 rc = mdb_cursor_first(mc, key, data);
5681 mfunc = mdb_cursor_first;
5683 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5687 if (mc->mc_xcursor == NULL) {
5688 rc = MDB_INCOMPATIBLE;
5691 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5695 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5698 rc = mdb_cursor_last(mc, key, data);
5701 mfunc = mdb_cursor_last;
5704 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5709 if (mc->mc_flags & C_DEL)
5710 mc->mc_flags ^= C_DEL;
5715 /** Touch all the pages in the cursor stack. Set mc_top.
5716 * Makes sure all the pages are writable, before attempting a write operation.
5717 * @param[in] mc The cursor to operate on.
5720 mdb_cursor_touch(MDB_cursor *mc)
5722 int rc = MDB_SUCCESS;
5724 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5727 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5728 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5731 *mc->mc_dbflag |= DB_DIRTY;
5736 rc = mdb_page_touch(mc);
5737 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5738 mc->mc_top = mc->mc_snum-1;
5743 /** Do not spill pages to disk if txn is getting full, may fail instead */
5744 #define MDB_NOSPILL 0x8000
5747 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5750 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5752 MDB_node *leaf = NULL;
5755 MDB_val xdata, *rdata, dkey, olddata;
5757 int do_sub = 0, insert;
5758 unsigned int mcount = 0, dcount = 0, nospill;
5761 unsigned int nflags;
5764 if (mc == NULL || key == NULL)
5767 env = mc->mc_txn->mt_env;
5769 /* Check this first so counter will always be zero on any
5772 if (flags & MDB_MULTIPLE) {
5773 dcount = data[1].mv_size;
5774 data[1].mv_size = 0;
5775 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5776 return MDB_INCOMPATIBLE;
5779 nospill = flags & MDB_NOSPILL;
5780 flags &= ~MDB_NOSPILL;
5782 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5783 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5785 if (flags != MDB_CURRENT && key->mv_size-1 >= ENV_MAXKEY(env))
5786 return MDB_BAD_VALSIZE;
5788 #if SIZE_MAX > MAXDATASIZE
5789 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5790 return MDB_BAD_VALSIZE;
5792 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5793 return MDB_BAD_VALSIZE;
5796 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5797 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5801 if (flags == MDB_CURRENT) {
5802 if (!(mc->mc_flags & C_INITIALIZED))
5805 } else if (mc->mc_db->md_root == P_INVALID) {
5806 /* new database, cursor has nothing to point to */
5809 mc->mc_flags &= ~C_INITIALIZED;
5814 if (flags & MDB_APPEND) {
5816 rc = mdb_cursor_last(mc, &k2, &d2);
5818 rc = mc->mc_dbx->md_cmp(key, &k2);
5821 mc->mc_ki[mc->mc_top]++;
5823 /* new key is <= last key */
5828 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5830 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5831 DPRINTF(("duplicate key [%s]", DKEY(key)));
5833 return MDB_KEYEXIST;
5835 if (rc && rc != MDB_NOTFOUND)
5839 if (mc->mc_flags & C_DEL)
5840 mc->mc_flags ^= C_DEL;
5842 /* Cursor is positioned, check for room in the dirty list */
5844 if (flags & MDB_MULTIPLE) {
5846 xdata.mv_size = data->mv_size * dcount;
5850 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5854 if (rc == MDB_NO_ROOT) {
5856 /* new database, write a root leaf page */
5857 DPUTS("allocating new root leaf page");
5858 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5861 mdb_cursor_push(mc, np);
5862 mc->mc_db->md_root = np->mp_pgno;
5863 mc->mc_db->md_depth++;
5864 *mc->mc_dbflag |= DB_DIRTY;
5865 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5867 np->mp_flags |= P_LEAF2;
5868 mc->mc_flags |= C_INITIALIZED;
5870 /* make sure all cursor pages are writable */
5871 rc2 = mdb_cursor_touch(mc);
5878 /* The key does not exist */
5879 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5880 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5881 LEAFSIZE(key, data) > env->me_nodemax)
5883 /* Too big for a node, insert in sub-DB */
5884 fp_flags = P_LEAF|P_DIRTY;
5886 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5887 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5891 /* there's only a key anyway, so this is a no-op */
5892 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5893 unsigned int ksize = mc->mc_db->md_pad;
5894 if (key->mv_size != ksize)
5895 return MDB_BAD_VALSIZE;
5896 if (flags == MDB_CURRENT) {
5897 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5898 memcpy(ptr, key->mv_data, ksize);
5904 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5905 olddata.mv_size = NODEDSZ(leaf);
5906 olddata.mv_data = NODEDATA(leaf);
5909 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5910 /* Prepare (sub-)page/sub-DB to accept the new item,
5911 * if needed. fp: old sub-page or a header faking
5912 * it. mp: new (sub-)page. offset: growth in page
5913 * size. xdata: node data with new page or DB.
5915 ssize_t i, offset = 0;
5916 mp = fp = xdata.mv_data = env->me_pbuf;
5917 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5919 /* Was a single item before, must convert now */
5920 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5921 /* Just overwrite the current item */
5922 if (flags == MDB_CURRENT)
5925 #if UINT_MAX < SIZE_MAX
5926 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5927 #ifdef MISALIGNED_OK
5928 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5930 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5933 /* if data matches, skip it */
5934 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5935 if (flags & MDB_NODUPDATA)
5937 else if (flags & MDB_MULTIPLE)
5944 /* Back up original data item */
5945 dkey.mv_size = olddata.mv_size;
5946 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5948 /* Make sub-page header for the dup items, with dummy body */
5949 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5950 fp->mp_lower = PAGEHDRSZ;
5951 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5952 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5953 fp->mp_flags |= P_LEAF2;
5954 fp->mp_pad = data->mv_size;
5955 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5957 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5958 (dkey.mv_size & 1) + (data->mv_size & 1);
5960 fp->mp_upper = xdata.mv_size;
5961 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
5962 } else if (leaf->mn_flags & F_SUBDATA) {
5963 /* Data is on sub-DB, just store it */
5964 flags |= F_DUPDATA|F_SUBDATA;
5967 /* Data is on sub-page */
5968 fp = olddata.mv_data;
5971 i = -(ssize_t)SIZELEFT(fp);
5972 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5973 offset = i += (ssize_t) EVEN(
5974 sizeof(indx_t) + NODESIZE + data->mv_size);
5976 i += offset = fp->mp_pad;
5977 offset *= 4; /* space for 4 more */
5981 /* FALLTHRU: Sub-page is big enough */
5983 fp->mp_flags |= P_DIRTY;
5984 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
5985 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5989 xdata.mv_size = olddata.mv_size + offset;
5992 fp_flags = fp->mp_flags;
5993 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
5994 /* Too big for a sub-page, convert to sub-DB */
5995 fp_flags &= ~P_SUBP;
5997 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5998 fp_flags |= P_LEAF2;
5999 dummy.md_pad = fp->mp_pad;
6000 dummy.md_flags = MDB_DUPFIXED;
6001 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6002 dummy.md_flags |= MDB_INTEGERKEY;
6008 dummy.md_branch_pages = 0;
6009 dummy.md_leaf_pages = 1;
6010 dummy.md_overflow_pages = 0;
6011 dummy.md_entries = NUMKEYS(fp);
6012 xdata.mv_size = sizeof(MDB_db);
6013 xdata.mv_data = &dummy;
6014 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6016 offset = env->me_psize - olddata.mv_size;
6017 flags |= F_DUPDATA|F_SUBDATA;
6018 dummy.md_root = mp->mp_pgno;
6021 mp->mp_flags = fp_flags | P_DIRTY;
6022 mp->mp_pad = fp->mp_pad;
6023 mp->mp_lower = fp->mp_lower;
6024 mp->mp_upper = fp->mp_upper + offset;
6025 if (fp_flags & P_LEAF2) {
6026 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6028 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6029 olddata.mv_size - fp->mp_upper);
6030 for (i = NUMKEYS(fp); --i >= 0; )
6031 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6039 mdb_node_del(mc, 0);
6043 /* overflow page overwrites need special handling */
6044 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6047 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6049 memcpy(&pg, olddata.mv_data, sizeof(pg));
6050 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6052 ovpages = omp->mp_pages;
6054 /* Is the ov page large enough? */
6055 if (ovpages >= dpages) {
6056 if (!(omp->mp_flags & P_DIRTY) &&
6057 (level || (env->me_flags & MDB_WRITEMAP)))
6059 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6062 level = 0; /* dirty in this txn or clean */
6065 if (omp->mp_flags & P_DIRTY) {
6066 /* yes, overwrite it. Note in this case we don't
6067 * bother to try shrinking the page if the new data
6068 * is smaller than the overflow threshold.
6071 /* It is writable only in a parent txn */
6072 size_t sz = (size_t) env->me_psize * ovpages, off;
6073 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6079 rc = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6080 mdb_cassert(mc, rc == 0);
6081 if (!(flags & MDB_RESERVE)) {
6082 /* Copy end of page, adjusting alignment so
6083 * compiler may copy words instead of bytes.
6085 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6086 memcpy((size_t *)((char *)np + off),
6087 (size_t *)((char *)omp + off), sz - off);
6090 memcpy(np, omp, sz); /* Copy beginning of page */
6093 SETDSZ(leaf, data->mv_size);
6094 if (F_ISSET(flags, MDB_RESERVE))
6095 data->mv_data = METADATA(omp);
6097 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6101 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6103 } else if (data->mv_size == olddata.mv_size) {
6104 /* same size, just replace it. Note that we could
6105 * also reuse this node if the new data is smaller,
6106 * but instead we opt to shrink the node in that case.
6108 if (F_ISSET(flags, MDB_RESERVE))
6109 data->mv_data = olddata.mv_data;
6110 else if (data->mv_size)
6111 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6113 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6116 mdb_node_del(mc, 0);
6117 mc->mc_db->md_entries--;
6123 nflags = flags & NODE_ADD_FLAGS;
6124 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6125 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6126 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6127 nflags &= ~MDB_APPEND;
6129 nflags |= MDB_SPLIT_REPLACE;
6130 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6132 /* There is room already in this leaf page. */
6133 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6134 if (rc == 0 && !do_sub && insert) {
6135 /* Adjust other cursors pointing to mp */
6136 MDB_cursor *m2, *m3;
6137 MDB_dbi dbi = mc->mc_dbi;
6138 unsigned i = mc->mc_top;
6139 MDB_page *mp = mc->mc_pg[i];
6141 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6142 if (mc->mc_flags & C_SUB)
6143 m3 = &m2->mc_xcursor->mx_cursor;
6146 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6147 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6154 if (rc != MDB_SUCCESS)
6155 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6157 /* Now store the actual data in the child DB. Note that we're
6158 * storing the user data in the keys field, so there are strict
6159 * size limits on dupdata. The actual data fields of the child
6160 * DB are all zero size.
6167 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6168 if (flags & MDB_CURRENT) {
6169 xflags = MDB_CURRENT|MDB_NOSPILL;
6171 mdb_xcursor_init1(mc, leaf);
6172 xflags = (flags & MDB_NODUPDATA) ?
6173 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6175 /* converted, write the original data first */
6177 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6181 /* Adjust other cursors pointing to mp */
6183 unsigned i = mc->mc_top;
6184 MDB_page *mp = mc->mc_pg[i];
6186 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6187 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6188 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6189 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6190 mdb_xcursor_init1(m2, leaf);
6194 /* we've done our job */
6197 if (flags & MDB_APPENDDUP)
6198 xflags |= MDB_APPEND;
6199 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6200 if (flags & F_SUBDATA) {
6201 void *db = NODEDATA(leaf);
6202 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6205 /* sub-writes might have failed so check rc again.
6206 * Don't increment count if we just replaced an existing item.
6208 if (!rc && !(flags & MDB_CURRENT))
6209 mc->mc_db->md_entries++;
6210 if (flags & MDB_MULTIPLE) {
6214 /* let caller know how many succeeded, if any */
6215 data[1].mv_size = mcount;
6216 if (mcount < dcount) {
6217 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6224 /* If we succeeded and the key didn't exist before, make sure
6225 * the cursor is marked valid.
6228 mc->mc_flags |= C_INITIALIZED;
6233 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6239 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6240 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6242 if (!(mc->mc_flags & C_INITIALIZED))
6245 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6246 return MDB_NOTFOUND;
6248 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6251 rc = mdb_cursor_touch(mc);
6255 mp = mc->mc_pg[mc->mc_top];
6256 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6258 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6259 if (!(flags & MDB_NODUPDATA)) {
6260 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6261 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6263 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6264 /* If sub-DB still has entries, we're done */
6265 if (mc->mc_xcursor->mx_db.md_entries) {
6266 if (leaf->mn_flags & F_SUBDATA) {
6267 /* update subDB info */
6268 void *db = NODEDATA(leaf);
6269 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6272 /* shrink fake page */
6273 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6274 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6275 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6276 /* fix other sub-DB cursors pointed at this fake page */
6277 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6278 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6279 if (m2->mc_pg[mc->mc_top] == mp &&
6280 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6281 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6284 mc->mc_db->md_entries--;
6285 mc->mc_flags |= C_DEL;
6288 /* otherwise fall thru and delete the sub-DB */
6291 if (leaf->mn_flags & F_SUBDATA) {
6292 /* add all the child DB's pages to the free list */
6293 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6294 if (rc == MDB_SUCCESS) {
6295 mc->mc_db->md_entries -=
6296 mc->mc_xcursor->mx_db.md_entries;
6301 return mdb_cursor_del0(mc, leaf);
6304 /** Allocate and initialize new pages for a database.
6305 * @param[in] mc a cursor on the database being added to.
6306 * @param[in] flags flags defining what type of page is being allocated.
6307 * @param[in] num the number of pages to allocate. This is usually 1,
6308 * unless allocating overflow pages for a large record.
6309 * @param[out] mp Address of a page, or NULL on failure.
6310 * @return 0 on success, non-zero on failure.
6313 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6318 if ((rc = mdb_page_alloc(mc, num, &np)))
6320 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6321 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6322 np->mp_flags = flags | P_DIRTY;
6323 np->mp_lower = PAGEHDRSZ;
6324 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6327 mc->mc_db->md_branch_pages++;
6328 else if (IS_LEAF(np))
6329 mc->mc_db->md_leaf_pages++;
6330 else if (IS_OVERFLOW(np)) {
6331 mc->mc_db->md_overflow_pages += num;
6339 /** Calculate the size of a leaf node.
6340 * The size depends on the environment's page size; if a data item
6341 * is too large it will be put onto an overflow page and the node
6342 * size will only include the key and not the data. Sizes are always
6343 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6344 * of the #MDB_node headers.
6345 * @param[in] env The environment handle.
6346 * @param[in] key The key for the node.
6347 * @param[in] data The data for the node.
6348 * @return The number of bytes needed to store the node.
6351 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6355 sz = LEAFSIZE(key, data);
6356 if (sz > env->me_nodemax) {
6357 /* put on overflow page */
6358 sz -= data->mv_size - sizeof(pgno_t);
6361 return EVEN(sz + sizeof(indx_t));
6364 /** Calculate the size of a branch node.
6365 * The size should depend on the environment's page size but since
6366 * we currently don't support spilling large keys onto overflow
6367 * pages, it's simply the size of the #MDB_node header plus the
6368 * size of the key. Sizes are always rounded up to an even number
6369 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6370 * @param[in] env The environment handle.
6371 * @param[in] key The key for the node.
6372 * @return The number of bytes needed to store the node.
6375 mdb_branch_size(MDB_env *env, MDB_val *key)
6380 if (sz > env->me_nodemax) {
6381 /* put on overflow page */
6382 /* not implemented */
6383 /* sz -= key->size - sizeof(pgno_t); */
6386 return sz + sizeof(indx_t);
6389 /** Add a node to the page pointed to by the cursor.
6390 * @param[in] mc The cursor for this operation.
6391 * @param[in] indx The index on the page where the new node should be added.
6392 * @param[in] key The key for the new node.
6393 * @param[in] data The data for the new node, if any.
6394 * @param[in] pgno The page number, if adding a branch node.
6395 * @param[in] flags Flags for the node.
6396 * @return 0 on success, non-zero on failure. Possible errors are:
6398 * <li>ENOMEM - failed to allocate overflow pages for the node.
6399 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6400 * should never happen since all callers already calculate the
6401 * page's free space before calling this function.
6405 mdb_node_add(MDB_cursor *mc, indx_t indx,
6406 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6409 size_t node_size = NODESIZE;
6413 MDB_page *mp = mc->mc_pg[mc->mc_top];
6414 MDB_page *ofp = NULL; /* overflow page */
6417 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6419 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6420 IS_LEAF(mp) ? "leaf" : "branch",
6421 IS_SUBP(mp) ? "sub-" : "",
6422 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6423 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6426 /* Move higher keys up one slot. */
6427 int ksize = mc->mc_db->md_pad, dif;
6428 char *ptr = LEAF2KEY(mp, indx, ksize);
6429 dif = NUMKEYS(mp) - indx;
6431 memmove(ptr+ksize, ptr, dif*ksize);
6432 /* insert new key */
6433 memcpy(ptr, key->mv_data, ksize);
6435 /* Just using these for counting */
6436 mp->mp_lower += sizeof(indx_t);
6437 mp->mp_upper -= ksize - sizeof(indx_t);
6441 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6443 node_size += key->mv_size;
6445 mdb_cassert(mc, data);
6446 if (F_ISSET(flags, F_BIGDATA)) {
6447 /* Data already on overflow page. */
6448 node_size += sizeof(pgno_t);
6449 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6450 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6452 /* Put data on overflow page. */
6453 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6454 data->mv_size, node_size+data->mv_size));
6455 node_size = EVEN(node_size + sizeof(pgno_t));
6456 if ((ssize_t)node_size > room)
6458 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6460 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6464 node_size += data->mv_size;
6467 node_size = EVEN(node_size);
6468 if ((ssize_t)node_size > room)
6472 /* Move higher pointers up one slot. */
6473 for (i = NUMKEYS(mp); i > indx; i--)
6474 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6476 /* Adjust free space offsets. */
6477 ofs = mp->mp_upper - node_size;
6478 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6479 mp->mp_ptrs[indx] = ofs;
6481 mp->mp_lower += sizeof(indx_t);
6483 /* Write the node data. */
6484 node = NODEPTR(mp, indx);
6485 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6486 node->mn_flags = flags;
6488 SETDSZ(node,data->mv_size);
6493 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6496 mdb_cassert(mc, key);
6498 if (F_ISSET(flags, F_BIGDATA))
6499 memcpy(node->mn_data + key->mv_size, data->mv_data,
6501 else if (F_ISSET(flags, MDB_RESERVE))
6502 data->mv_data = node->mn_data + key->mv_size;
6504 memcpy(node->mn_data + key->mv_size, data->mv_data,
6507 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6509 if (F_ISSET(flags, MDB_RESERVE))
6510 data->mv_data = METADATA(ofp);
6512 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6519 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6520 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6521 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6522 DPRINTF(("node size = %"Z"u", node_size));
6523 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6524 return MDB_PAGE_FULL;
6527 /** Delete the specified node from a page.
6528 * @param[in] mp The page to operate on.
6529 * @param[in] indx The index of the node to delete.
6530 * @param[in] ksize The size of a node. Only used if the page is
6531 * part of a #MDB_DUPFIXED database.
6534 mdb_node_del(MDB_cursor *mc, int ksize)
6536 MDB_page *mp = mc->mc_pg[mc->mc_top];
6537 indx_t indx = mc->mc_ki[mc->mc_top];
6539 indx_t i, j, numkeys, ptr;
6543 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6544 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6545 numkeys = NUMKEYS(mp);
6546 mdb_cassert(mc, indx < numkeys);
6549 int x = numkeys - 1 - indx;
6550 base = LEAF2KEY(mp, indx, ksize);
6552 memmove(base, base + ksize, x * ksize);
6553 mp->mp_lower -= sizeof(indx_t);
6554 mp->mp_upper += ksize - sizeof(indx_t);
6558 node = NODEPTR(mp, indx);
6559 sz = NODESIZE + node->mn_ksize;
6561 if (F_ISSET(node->mn_flags, F_BIGDATA))
6562 sz += sizeof(pgno_t);
6564 sz += NODEDSZ(node);
6568 ptr = mp->mp_ptrs[indx];
6569 for (i = j = 0; i < numkeys; i++) {
6571 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6572 if (mp->mp_ptrs[i] < ptr)
6573 mp->mp_ptrs[j] += sz;
6578 base = (char *)mp + mp->mp_upper;
6579 memmove(base + sz, base, ptr - mp->mp_upper);
6581 mp->mp_lower -= sizeof(indx_t);
6585 /** Compact the main page after deleting a node on a subpage.
6586 * @param[in] mp The main page to operate on.
6587 * @param[in] indx The index of the subpage on the main page.
6590 mdb_node_shrink(MDB_page *mp, indx_t indx)
6596 indx_t i, numkeys, ptr;
6598 node = NODEPTR(mp, indx);
6599 sp = (MDB_page *)NODEDATA(node);
6600 delta = SIZELEFT(sp);
6601 xp = (MDB_page *)((char *)sp + delta);
6603 /* shift subpage upward */
6605 nsize = NUMKEYS(sp) * sp->mp_pad;
6607 return; /* do not make the node uneven-sized */
6608 memmove(METADATA(xp), METADATA(sp), nsize);
6611 numkeys = NUMKEYS(sp);
6612 for (i=numkeys-1; i>=0; i--)
6613 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6615 xp->mp_upper = sp->mp_lower;
6616 xp->mp_lower = sp->mp_lower;
6617 xp->mp_flags = sp->mp_flags;
6618 xp->mp_pad = sp->mp_pad;
6619 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6621 nsize = NODEDSZ(node) - delta;
6622 SETDSZ(node, nsize);
6624 /* shift lower nodes upward */
6625 ptr = mp->mp_ptrs[indx];
6626 numkeys = NUMKEYS(mp);
6627 for (i = 0; i < numkeys; i++) {
6628 if (mp->mp_ptrs[i] <= ptr)
6629 mp->mp_ptrs[i] += delta;
6632 base = (char *)mp + mp->mp_upper;
6633 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6634 mp->mp_upper += delta;
6637 /** Initial setup of a sorted-dups cursor.
6638 * Sorted duplicates are implemented as a sub-database for the given key.
6639 * The duplicate data items are actually keys of the sub-database.
6640 * Operations on the duplicate data items are performed using a sub-cursor
6641 * initialized when the sub-database is first accessed. This function does
6642 * the preliminary setup of the sub-cursor, filling in the fields that
6643 * depend only on the parent DB.
6644 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6647 mdb_xcursor_init0(MDB_cursor *mc)
6649 MDB_xcursor *mx = mc->mc_xcursor;
6651 mx->mx_cursor.mc_xcursor = NULL;
6652 mx->mx_cursor.mc_txn = mc->mc_txn;
6653 mx->mx_cursor.mc_db = &mx->mx_db;
6654 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6655 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6656 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6657 mx->mx_cursor.mc_snum = 0;
6658 mx->mx_cursor.mc_top = 0;
6659 mx->mx_cursor.mc_flags = C_SUB;
6660 mx->mx_dbx.md_name.mv_size = 0;
6661 mx->mx_dbx.md_name.mv_data = NULL;
6662 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6663 mx->mx_dbx.md_dcmp = NULL;
6664 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6667 /** Final setup of a sorted-dups cursor.
6668 * Sets up the fields that depend on the data from the main cursor.
6669 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6670 * @param[in] node The data containing the #MDB_db record for the
6671 * sorted-dup database.
6674 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6676 MDB_xcursor *mx = mc->mc_xcursor;
6678 if (node->mn_flags & F_SUBDATA) {
6679 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6680 mx->mx_cursor.mc_pg[0] = 0;
6681 mx->mx_cursor.mc_snum = 0;
6682 mx->mx_cursor.mc_top = 0;
6683 mx->mx_cursor.mc_flags = C_SUB;
6685 MDB_page *fp = NODEDATA(node);
6686 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6687 mx->mx_db.md_flags = 0;
6688 mx->mx_db.md_depth = 1;
6689 mx->mx_db.md_branch_pages = 0;
6690 mx->mx_db.md_leaf_pages = 1;
6691 mx->mx_db.md_overflow_pages = 0;
6692 mx->mx_db.md_entries = NUMKEYS(fp);
6693 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6694 mx->mx_cursor.mc_snum = 1;
6695 mx->mx_cursor.mc_top = 0;
6696 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6697 mx->mx_cursor.mc_pg[0] = fp;
6698 mx->mx_cursor.mc_ki[0] = 0;
6699 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6700 mx->mx_db.md_flags = MDB_DUPFIXED;
6701 mx->mx_db.md_pad = fp->mp_pad;
6702 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6703 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6706 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6707 mx->mx_db.md_root));
6708 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6709 #if UINT_MAX < SIZE_MAX
6710 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6711 #ifdef MISALIGNED_OK
6712 mx->mx_dbx.md_cmp = mdb_cmp_long;
6714 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6719 /** Initialize a cursor for a given transaction and database. */
6721 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6724 mc->mc_backup = NULL;
6727 mc->mc_db = &txn->mt_dbs[dbi];
6728 mc->mc_dbx = &txn->mt_dbxs[dbi];
6729 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6734 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6735 mdb_tassert(txn, mx != NULL);
6736 mc->mc_xcursor = mx;
6737 mdb_xcursor_init0(mc);
6739 mc->mc_xcursor = NULL;
6741 if (*mc->mc_dbflag & DB_STALE) {
6742 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6747 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6750 size_t size = sizeof(MDB_cursor);
6752 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6755 if (txn->mt_flags & MDB_TXN_ERROR)
6758 /* Allow read access to the freelist */
6759 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6762 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6763 size += sizeof(MDB_xcursor);
6765 if ((mc = malloc(size)) != NULL) {
6766 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6767 if (txn->mt_cursors) {
6768 mc->mc_next = txn->mt_cursors[dbi];
6769 txn->mt_cursors[dbi] = mc;
6770 mc->mc_flags |= C_UNTRACK;
6782 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6784 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6787 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6790 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6794 /* Return the count of duplicate data items for the current key */
6796 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6800 if (mc == NULL || countp == NULL)
6803 if (mc->mc_xcursor == NULL)
6804 return MDB_INCOMPATIBLE;
6806 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6807 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6810 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6813 *countp = mc->mc_xcursor->mx_db.md_entries;
6819 mdb_cursor_close(MDB_cursor *mc)
6821 if (mc && !mc->mc_backup) {
6822 /* remove from txn, if tracked */
6823 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6824 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6825 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6827 *prev = mc->mc_next;
6834 mdb_cursor_txn(MDB_cursor *mc)
6836 if (!mc) return NULL;
6841 mdb_cursor_dbi(MDB_cursor *mc)
6846 /** Replace the key for a branch node with a new key.
6847 * @param[in] mc Cursor pointing to the node to operate on.
6848 * @param[in] key The new key to use.
6849 * @return 0 on success, non-zero on failure.
6852 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6858 int delta, ksize, oksize;
6859 indx_t ptr, i, numkeys, indx;
6862 indx = mc->mc_ki[mc->mc_top];
6863 mp = mc->mc_pg[mc->mc_top];
6864 node = NODEPTR(mp, indx);
6865 ptr = mp->mp_ptrs[indx];
6869 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6870 k2.mv_data = NODEKEY(node);
6871 k2.mv_size = node->mn_ksize;
6872 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6874 mdb_dkey(&k2, kbuf2),
6880 /* Sizes must be 2-byte aligned. */
6881 ksize = EVEN(key->mv_size);
6882 oksize = EVEN(node->mn_ksize);
6883 delta = ksize - oksize;
6885 /* Shift node contents if EVEN(key length) changed. */
6887 if (delta > 0 && SIZELEFT(mp) < delta) {
6889 /* not enough space left, do a delete and split */
6890 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6891 pgno = NODEPGNO(node);
6892 mdb_node_del(mc, 0);
6893 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6896 numkeys = NUMKEYS(mp);
6897 for (i = 0; i < numkeys; i++) {
6898 if (mp->mp_ptrs[i] <= ptr)
6899 mp->mp_ptrs[i] -= delta;
6902 base = (char *)mp + mp->mp_upper;
6903 len = ptr - mp->mp_upper + NODESIZE;
6904 memmove(base - delta, base, len);
6905 mp->mp_upper -= delta;
6907 node = NODEPTR(mp, indx);
6910 /* But even if no shift was needed, update ksize */
6911 if (node->mn_ksize != key->mv_size)
6912 node->mn_ksize = key->mv_size;
6915 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6921 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6923 /** Move a node from csrc to cdst.
6926 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6933 unsigned short flags;
6937 /* Mark src and dst as dirty. */
6938 if ((rc = mdb_page_touch(csrc)) ||
6939 (rc = mdb_page_touch(cdst)))
6942 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6943 key.mv_size = csrc->mc_db->md_pad;
6944 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6946 data.mv_data = NULL;
6950 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6951 mdb_cassert(csrc, !((size_t)srcnode & 1));
6952 srcpg = NODEPGNO(srcnode);
6953 flags = srcnode->mn_flags;
6954 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6955 unsigned int snum = csrc->mc_snum;
6957 /* must find the lowest key below src */
6958 mdb_page_search_lowest(csrc);
6959 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6960 key.mv_size = csrc->mc_db->md_pad;
6961 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6963 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6964 key.mv_size = NODEKSZ(s2);
6965 key.mv_data = NODEKEY(s2);
6967 csrc->mc_snum = snum--;
6968 csrc->mc_top = snum;
6970 key.mv_size = NODEKSZ(srcnode);
6971 key.mv_data = NODEKEY(srcnode);
6973 data.mv_size = NODEDSZ(srcnode);
6974 data.mv_data = NODEDATA(srcnode);
6976 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6977 unsigned int snum = cdst->mc_snum;
6980 /* must find the lowest key below dst */
6981 mdb_page_search_lowest(cdst);
6982 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6983 bkey.mv_size = cdst->mc_db->md_pad;
6984 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6986 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6987 bkey.mv_size = NODEKSZ(s2);
6988 bkey.mv_data = NODEKEY(s2);
6990 cdst->mc_snum = snum--;
6991 cdst->mc_top = snum;
6992 mdb_cursor_copy(cdst, &mn);
6994 rc = mdb_update_key(&mn, &bkey);
6999 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7000 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7001 csrc->mc_ki[csrc->mc_top],
7003 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7004 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7006 /* Add the node to the destination page.
7008 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7009 if (rc != MDB_SUCCESS)
7012 /* Delete the node from the source page.
7014 mdb_node_del(csrc, key.mv_size);
7017 /* Adjust other cursors pointing to mp */
7018 MDB_cursor *m2, *m3;
7019 MDB_dbi dbi = csrc->mc_dbi;
7020 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7022 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7023 if (csrc->mc_flags & C_SUB)
7024 m3 = &m2->mc_xcursor->mx_cursor;
7027 if (m3 == csrc) continue;
7028 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7029 csrc->mc_ki[csrc->mc_top]) {
7030 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7031 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7036 /* Update the parent separators.
7038 if (csrc->mc_ki[csrc->mc_top] == 0) {
7039 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7040 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7041 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7043 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7044 key.mv_size = NODEKSZ(srcnode);
7045 key.mv_data = NODEKEY(srcnode);
7047 DPRINTF(("update separator for source page %"Z"u to [%s]",
7048 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7049 mdb_cursor_copy(csrc, &mn);
7052 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7055 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7057 indx_t ix = csrc->mc_ki[csrc->mc_top];
7058 nullkey.mv_size = 0;
7059 csrc->mc_ki[csrc->mc_top] = 0;
7060 rc = mdb_update_key(csrc, &nullkey);
7061 csrc->mc_ki[csrc->mc_top] = ix;
7062 mdb_cassert(csrc, rc == MDB_SUCCESS);
7066 if (cdst->mc_ki[cdst->mc_top] == 0) {
7067 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7068 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7069 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7071 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7072 key.mv_size = NODEKSZ(srcnode);
7073 key.mv_data = NODEKEY(srcnode);
7075 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7076 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7077 mdb_cursor_copy(cdst, &mn);
7080 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7083 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7085 indx_t ix = cdst->mc_ki[cdst->mc_top];
7086 nullkey.mv_size = 0;
7087 cdst->mc_ki[cdst->mc_top] = 0;
7088 rc = mdb_update_key(cdst, &nullkey);
7089 cdst->mc_ki[cdst->mc_top] = ix;
7090 mdb_cassert(csrc, rc == MDB_SUCCESS);
7097 /** Merge one page into another.
7098 * The nodes from the page pointed to by \b csrc will
7099 * be copied to the page pointed to by \b cdst and then
7100 * the \b csrc page will be freed.
7101 * @param[in] csrc Cursor pointing to the source page.
7102 * @param[in] cdst Cursor pointing to the destination page.
7105 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7113 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7114 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7116 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7117 mdb_cassert(csrc, cdst->mc_snum > 1);
7119 /* Mark dst as dirty. */
7120 if ((rc = mdb_page_touch(cdst)))
7123 /* Move all nodes from src to dst.
7125 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7126 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7127 key.mv_size = csrc->mc_db->md_pad;
7128 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7129 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7130 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7131 if (rc != MDB_SUCCESS)
7133 key.mv_data = (char *)key.mv_data + key.mv_size;
7136 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7137 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7138 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7139 unsigned int snum = csrc->mc_snum;
7141 /* must find the lowest key below src */
7142 mdb_page_search_lowest(csrc);
7143 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7144 key.mv_size = csrc->mc_db->md_pad;
7145 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7147 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7148 key.mv_size = NODEKSZ(s2);
7149 key.mv_data = NODEKEY(s2);
7151 csrc->mc_snum = snum--;
7152 csrc->mc_top = snum;
7154 key.mv_size = srcnode->mn_ksize;
7155 key.mv_data = NODEKEY(srcnode);
7158 data.mv_size = NODEDSZ(srcnode);
7159 data.mv_data = NODEDATA(srcnode);
7160 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7161 if (rc != MDB_SUCCESS)
7166 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7167 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7168 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7170 /* Unlink the src page from parent and add to free list.
7173 mdb_node_del(csrc, 0);
7174 if (csrc->mc_ki[csrc->mc_top] == 0) {
7176 rc = mdb_update_key(csrc, &key);
7184 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7185 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7188 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7189 csrc->mc_db->md_leaf_pages--;
7191 csrc->mc_db->md_branch_pages--;
7193 /* Adjust other cursors pointing to mp */
7194 MDB_cursor *m2, *m3;
7195 MDB_dbi dbi = csrc->mc_dbi;
7196 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7198 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7199 if (csrc->mc_flags & C_SUB)
7200 m3 = &m2->mc_xcursor->mx_cursor;
7203 if (m3 == csrc) continue;
7204 if (m3->mc_snum < csrc->mc_snum) continue;
7205 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7206 m3->mc_pg[csrc->mc_top] = mp;
7207 m3->mc_ki[csrc->mc_top] += nkeys;
7211 mdb_cursor_pop(csrc);
7213 return mdb_rebalance(csrc);
7216 /** Copy the contents of a cursor.
7217 * @param[in] csrc The cursor to copy from.
7218 * @param[out] cdst The cursor to copy to.
7221 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7225 cdst->mc_txn = csrc->mc_txn;
7226 cdst->mc_dbi = csrc->mc_dbi;
7227 cdst->mc_db = csrc->mc_db;
7228 cdst->mc_dbx = csrc->mc_dbx;
7229 cdst->mc_snum = csrc->mc_snum;
7230 cdst->mc_top = csrc->mc_top;
7231 cdst->mc_flags = csrc->mc_flags;
7233 for (i=0; i<csrc->mc_snum; i++) {
7234 cdst->mc_pg[i] = csrc->mc_pg[i];
7235 cdst->mc_ki[i] = csrc->mc_ki[i];
7239 /** Rebalance the tree after a delete operation.
7240 * @param[in] mc Cursor pointing to the page where rebalancing
7242 * @return 0 on success, non-zero on failure.
7245 mdb_rebalance(MDB_cursor *mc)
7249 unsigned int ptop, minkeys;
7252 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7253 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7254 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7255 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7256 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7258 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7259 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7260 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7261 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7265 if (mc->mc_snum < 2) {
7266 MDB_page *mp = mc->mc_pg[0];
7268 DPUTS("Can't rebalance a subpage, ignoring");
7271 if (NUMKEYS(mp) == 0) {
7272 DPUTS("tree is completely empty");
7273 mc->mc_db->md_root = P_INVALID;
7274 mc->mc_db->md_depth = 0;
7275 mc->mc_db->md_leaf_pages = 0;
7276 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7279 /* Adjust cursors pointing to mp */
7282 mc->mc_flags &= ~C_INITIALIZED;
7284 MDB_cursor *m2, *m3;
7285 MDB_dbi dbi = mc->mc_dbi;
7287 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7288 if (mc->mc_flags & C_SUB)
7289 m3 = &m2->mc_xcursor->mx_cursor;
7292 if (m3->mc_snum < mc->mc_snum) continue;
7293 if (m3->mc_pg[0] == mp) {
7296 m3->mc_flags &= ~C_INITIALIZED;
7300 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7301 DPUTS("collapsing root page!");
7302 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7305 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7306 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7309 mc->mc_db->md_depth--;
7310 mc->mc_db->md_branch_pages--;
7311 mc->mc_ki[0] = mc->mc_ki[1];
7313 /* Adjust other cursors pointing to mp */
7314 MDB_cursor *m2, *m3;
7315 MDB_dbi dbi = mc->mc_dbi;
7317 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7318 if (mc->mc_flags & C_SUB)
7319 m3 = &m2->mc_xcursor->mx_cursor;
7322 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7323 if (m3->mc_pg[0] == mp) {
7327 for (i=0; i<m3->mc_snum; i++) {
7328 m3->mc_pg[i] = m3->mc_pg[i+1];
7329 m3->mc_ki[i] = m3->mc_ki[i+1];
7335 DPUTS("root page doesn't need rebalancing");
7339 /* The parent (branch page) must have at least 2 pointers,
7340 * otherwise the tree is invalid.
7342 ptop = mc->mc_top-1;
7343 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7345 /* Leaf page fill factor is below the threshold.
7346 * Try to move keys from left or right neighbor, or
7347 * merge with a neighbor page.
7352 mdb_cursor_copy(mc, &mn);
7353 mn.mc_xcursor = NULL;
7355 if (mc->mc_ki[ptop] == 0) {
7356 /* We're the leftmost leaf in our parent.
7358 DPUTS("reading right neighbor");
7360 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7361 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7364 mn.mc_ki[mn.mc_top] = 0;
7365 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7367 /* There is at least one neighbor to the left.
7369 DPUTS("reading left neighbor");
7371 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7372 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7375 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7376 mc->mc_ki[mc->mc_top] = 0;
7379 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7380 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7381 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7383 /* If the neighbor page is above threshold and has enough keys,
7384 * move one key from it. Otherwise we should try to merge them.
7385 * (A branch page must never have less than 2 keys.)
7387 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7388 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7389 return mdb_node_move(&mn, mc);
7391 if (mc->mc_ki[ptop] == 0)
7392 rc = mdb_page_merge(&mn, mc);
7394 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7395 rc = mdb_page_merge(mc, &mn);
7396 mdb_cursor_copy(&mn, mc);
7398 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7403 /** Complete a delete operation started by #mdb_cursor_del(). */
7405 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7412 mp = mc->mc_pg[mc->mc_top];
7413 ki = mc->mc_ki[mc->mc_top];
7415 /* add overflow pages to free list */
7416 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7420 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7421 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7422 (rc = mdb_ovpage_free(mc, omp)))
7425 mdb_node_del(mc, mc->mc_db->md_pad);
7426 mc->mc_db->md_entries--;
7427 rc = mdb_rebalance(mc);
7428 if (rc != MDB_SUCCESS)
7429 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7431 MDB_cursor *m2, *m3;
7432 MDB_dbi dbi = mc->mc_dbi;
7434 mp = mc->mc_pg[mc->mc_top];
7435 nkeys = NUMKEYS(mp);
7437 /* if mc points past last node in page, find next sibling */
7438 if (mc->mc_ki[mc->mc_top] >= nkeys)
7439 mdb_cursor_sibling(mc, 1);
7441 /* Adjust other cursors pointing to mp */
7442 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7443 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7444 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7446 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7448 if (m3->mc_pg[mc->mc_top] == mp) {
7449 if (m3->mc_ki[mc->mc_top] >= ki) {
7450 m3->mc_flags |= C_DEL;
7451 if (m3->mc_ki[mc->mc_top] > ki)
7452 m3->mc_ki[mc->mc_top]--;
7454 if (m3->mc_ki[mc->mc_top] >= nkeys)
7455 mdb_cursor_sibling(m3, 1);
7458 mc->mc_flags |= C_DEL;
7465 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7466 MDB_val *key, MDB_val *data)
7471 MDB_val rdata, *xdata;
7478 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7480 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7483 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7484 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7486 mdb_cursor_init(&mc, txn, dbi, &mx);
7489 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7490 /* must ignore any data */
7501 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7503 /* let mdb_page_split know about this cursor if needed:
7504 * delete will trigger a rebalance; if it needs to move
7505 * a node from one page to another, it will have to
7506 * update the parent's separator key(s). If the new sepkey
7507 * is larger than the current one, the parent page may
7508 * run out of space, triggering a split. We need this
7509 * cursor to be consistent until the end of the rebalance.
7511 mc.mc_flags |= C_UNTRACK;
7512 mc.mc_next = txn->mt_cursors[dbi];
7513 txn->mt_cursors[dbi] = &mc;
7514 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7515 txn->mt_cursors[dbi] = mc.mc_next;
7520 /** Split a page and insert a new node.
7521 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7522 * The cursor will be updated to point to the actual page and index where
7523 * the node got inserted after the split.
7524 * @param[in] newkey The key for the newly inserted node.
7525 * @param[in] newdata The data for the newly inserted node.
7526 * @param[in] newpgno The page number, if the new node is a branch node.
7527 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7528 * @return 0 on success, non-zero on failure.
7531 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7532 unsigned int nflags)
7535 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7538 int i, j, split_indx, nkeys, pmax;
7539 MDB_env *env = mc->mc_txn->mt_env;
7541 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7542 MDB_page *copy = NULL;
7543 MDB_page *mp, *rp, *pp;
7548 mp = mc->mc_pg[mc->mc_top];
7549 newindx = mc->mc_ki[mc->mc_top];
7550 nkeys = NUMKEYS(mp);
7552 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7553 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7554 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7556 /* Create a right sibling. */
7557 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7559 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7561 if (mc->mc_snum < 2) {
7562 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7564 /* shift current top to make room for new parent */
7565 mc->mc_pg[1] = mc->mc_pg[0];
7566 mc->mc_ki[1] = mc->mc_ki[0];
7569 mc->mc_db->md_root = pp->mp_pgno;
7570 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7571 mc->mc_db->md_depth++;
7574 /* Add left (implicit) pointer. */
7575 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7576 /* undo the pre-push */
7577 mc->mc_pg[0] = mc->mc_pg[1];
7578 mc->mc_ki[0] = mc->mc_ki[1];
7579 mc->mc_db->md_root = mp->mp_pgno;
7580 mc->mc_db->md_depth--;
7587 ptop = mc->mc_top-1;
7588 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7591 mc->mc_flags |= C_SPLITTING;
7592 mdb_cursor_copy(mc, &mn);
7593 mn.mc_pg[mn.mc_top] = rp;
7594 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7596 if (nflags & MDB_APPEND) {
7597 mn.mc_ki[mn.mc_top] = 0;
7599 split_indx = newindx;
7603 split_indx = (nkeys+1) / 2;
7608 unsigned int lsize, rsize, ksize;
7609 /* Move half of the keys to the right sibling */
7611 x = mc->mc_ki[mc->mc_top] - split_indx;
7612 ksize = mc->mc_db->md_pad;
7613 split = LEAF2KEY(mp, split_indx, ksize);
7614 rsize = (nkeys - split_indx) * ksize;
7615 lsize = (nkeys - split_indx) * sizeof(indx_t);
7616 mp->mp_lower -= lsize;
7617 rp->mp_lower += lsize;
7618 mp->mp_upper += rsize - lsize;
7619 rp->mp_upper -= rsize - lsize;
7620 sepkey.mv_size = ksize;
7621 if (newindx == split_indx) {
7622 sepkey.mv_data = newkey->mv_data;
7624 sepkey.mv_data = split;
7627 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7628 memcpy(rp->mp_ptrs, split, rsize);
7629 sepkey.mv_data = rp->mp_ptrs;
7630 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7631 memcpy(ins, newkey->mv_data, ksize);
7632 mp->mp_lower += sizeof(indx_t);
7633 mp->mp_upper -= ksize - sizeof(indx_t);
7636 memcpy(rp->mp_ptrs, split, x * ksize);
7637 ins = LEAF2KEY(rp, x, ksize);
7638 memcpy(ins, newkey->mv_data, ksize);
7639 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7640 rp->mp_lower += sizeof(indx_t);
7641 rp->mp_upper -= ksize - sizeof(indx_t);
7642 mc->mc_ki[mc->mc_top] = x;
7643 mc->mc_pg[mc->mc_top] = rp;
7646 int psize, nsize, k;
7647 /* Maximum free space in an empty page */
7648 pmax = env->me_psize - PAGEHDRSZ;
7650 nsize = mdb_leaf_size(env, newkey, newdata);
7652 nsize = mdb_branch_size(env, newkey);
7653 nsize = EVEN(nsize);
7655 /* grab a page to hold a temporary copy */
7656 copy = mdb_page_malloc(mc->mc_txn, 1);
7659 copy->mp_pgno = mp->mp_pgno;
7660 copy->mp_flags = mp->mp_flags;
7661 copy->mp_lower = PAGEHDRSZ;
7662 copy->mp_upper = env->me_psize;
7664 /* prepare to insert */
7665 for (i=0, j=0; i<nkeys; i++) {
7667 copy->mp_ptrs[j++] = 0;
7669 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7672 /* When items are relatively large the split point needs
7673 * to be checked, because being off-by-one will make the
7674 * difference between success or failure in mdb_node_add.
7676 * It's also relevant if a page happens to be laid out
7677 * such that one half of its nodes are all "small" and
7678 * the other half of its nodes are "large." If the new
7679 * item is also "large" and falls on the half with
7680 * "large" nodes, it also may not fit.
7682 * As a final tweak, if the new item goes on the last
7683 * spot on the page (and thus, onto the new page), bias
7684 * the split so the new page is emptier than the old page.
7685 * This yields better packing during sequential inserts.
7687 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7688 /* Find split point */
7690 if (newindx <= split_indx || newindx >= nkeys) {
7692 k = newindx >= nkeys ? nkeys : split_indx+2;
7697 for (; i!=k; i+=j) {
7702 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7703 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7705 if (F_ISSET(node->mn_flags, F_BIGDATA))
7706 psize += sizeof(pgno_t);
7708 psize += NODEDSZ(node);
7710 psize = EVEN(psize);
7712 if (psize > pmax || i == k-j) {
7713 split_indx = i + (j<0);
7718 if (split_indx == newindx) {
7719 sepkey.mv_size = newkey->mv_size;
7720 sepkey.mv_data = newkey->mv_data;
7722 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7723 sepkey.mv_size = node->mn_ksize;
7724 sepkey.mv_data = NODEKEY(node);
7729 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7731 /* Copy separator key to the parent.
7733 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7737 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7740 if (mn.mc_snum == mc->mc_snum) {
7741 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7742 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7743 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7744 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7749 /* Right page might now have changed parent.
7750 * Check if left page also changed parent.
7752 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7753 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7754 for (i=0; i<ptop; i++) {
7755 mc->mc_pg[i] = mn.mc_pg[i];
7756 mc->mc_ki[i] = mn.mc_ki[i];
7758 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7759 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7763 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7766 mc->mc_flags ^= C_SPLITTING;
7767 if (rc != MDB_SUCCESS) {
7770 if (nflags & MDB_APPEND) {
7771 mc->mc_pg[mc->mc_top] = rp;
7772 mc->mc_ki[mc->mc_top] = 0;
7773 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7776 for (i=0; i<mc->mc_top; i++)
7777 mc->mc_ki[i] = mn.mc_ki[i];
7778 } else if (!IS_LEAF2(mp)) {
7780 mc->mc_pg[mc->mc_top] = rp;
7785 rkey.mv_data = newkey->mv_data;
7786 rkey.mv_size = newkey->mv_size;
7792 /* Update index for the new key. */
7793 mc->mc_ki[mc->mc_top] = j;
7795 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7796 rkey.mv_data = NODEKEY(node);
7797 rkey.mv_size = node->mn_ksize;
7799 xdata.mv_data = NODEDATA(node);
7800 xdata.mv_size = NODEDSZ(node);
7803 pgno = NODEPGNO(node);
7804 flags = node->mn_flags;
7807 if (!IS_LEAF(mp) && j == 0) {
7808 /* First branch index doesn't need key data. */
7812 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7814 /* return tmp page to freelist */
7815 mdb_page_free(env, copy);
7821 mc->mc_pg[mc->mc_top] = copy;
7826 } while (i != split_indx);
7828 nkeys = NUMKEYS(copy);
7829 for (i=0; i<nkeys; i++)
7830 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7831 mp->mp_lower = copy->mp_lower;
7832 mp->mp_upper = copy->mp_upper;
7833 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7834 env->me_psize - copy->mp_upper);
7836 /* reset back to original page */
7837 if (newindx < split_indx) {
7838 mc->mc_pg[mc->mc_top] = mp;
7839 if (nflags & MDB_RESERVE) {
7840 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7841 if (!(node->mn_flags & F_BIGDATA))
7842 newdata->mv_data = NODEDATA(node);
7845 mc->mc_pg[mc->mc_top] = rp;
7847 /* Make sure mc_ki is still valid.
7849 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7850 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7851 for (i=0; i<ptop; i++) {
7852 mc->mc_pg[i] = mn.mc_pg[i];
7853 mc->mc_ki[i] = mn.mc_ki[i];
7855 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7856 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7859 /* return tmp page to freelist */
7860 mdb_page_free(env, copy);
7864 /* Adjust other cursors pointing to mp */
7865 MDB_cursor *m2, *m3;
7866 MDB_dbi dbi = mc->mc_dbi;
7867 int fixup = NUMKEYS(mp);
7869 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7870 if (mc->mc_flags & C_SUB)
7871 m3 = &m2->mc_xcursor->mx_cursor;
7876 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7878 if (m3->mc_flags & C_SPLITTING)
7883 for (k=m3->mc_top; k>=0; k--) {
7884 m3->mc_ki[k+1] = m3->mc_ki[k];
7885 m3->mc_pg[k+1] = m3->mc_pg[k];
7887 if (m3->mc_ki[0] >= split_indx) {
7892 m3->mc_pg[0] = mc->mc_pg[0];
7896 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7897 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7898 m3->mc_ki[mc->mc_top]++;
7899 if (m3->mc_ki[mc->mc_top] >= fixup) {
7900 m3->mc_pg[mc->mc_top] = rp;
7901 m3->mc_ki[mc->mc_top] -= fixup;
7902 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7904 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7905 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7910 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7915 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7916 MDB_val *key, MDB_val *data, unsigned int flags)
7921 if (key == NULL || data == NULL)
7924 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7927 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7930 mdb_cursor_init(&mc, txn, dbi, &mx);
7931 return mdb_cursor_put(&mc, key, data, flags);
7935 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7937 if ((flag & CHANGEABLE) != flag)
7940 env->me_flags |= flag;
7942 env->me_flags &= ~flag;
7947 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7952 *arg = env->me_flags;
7957 mdb_env_set_userctx(MDB_env *env, void *ctx)
7961 env->me_userctx = ctx;
7966 mdb_env_get_userctx(MDB_env *env)
7968 return env ? env->me_userctx : NULL;
7972 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
7977 env->me_assert_func = func;
7983 mdb_env_get_path(MDB_env *env, const char **arg)
7988 *arg = env->me_path;
7993 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8002 /** Common code for #mdb_stat() and #mdb_env_stat().
8003 * @param[in] env the environment to operate in.
8004 * @param[in] db the #MDB_db record containing the stats to return.
8005 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8006 * @return 0, this function always succeeds.
8009 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8011 arg->ms_psize = env->me_psize;
8012 arg->ms_depth = db->md_depth;
8013 arg->ms_branch_pages = db->md_branch_pages;
8014 arg->ms_leaf_pages = db->md_leaf_pages;
8015 arg->ms_overflow_pages = db->md_overflow_pages;
8016 arg->ms_entries = db->md_entries;
8021 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8025 if (env == NULL || arg == NULL)
8028 toggle = mdb_env_pick_meta(env);
8030 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8034 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8038 if (env == NULL || arg == NULL)
8041 toggle = mdb_env_pick_meta(env);
8042 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
8043 arg->me_mapsize = env->me_mapsize;
8044 arg->me_maxreaders = env->me_maxreaders;
8046 /* me_numreaders may be zero if this process never used any readers. Use
8047 * the shared numreader count if it exists.
8049 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8051 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8052 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8056 /** Set the default comparison functions for a database.
8057 * Called immediately after a database is opened to set the defaults.
8058 * The user can then override them with #mdb_set_compare() or
8059 * #mdb_set_dupsort().
8060 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8061 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8064 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8066 uint16_t f = txn->mt_dbs[dbi].md_flags;
8068 txn->mt_dbxs[dbi].md_cmp =
8069 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8070 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8072 txn->mt_dbxs[dbi].md_dcmp =
8073 !(f & MDB_DUPSORT) ? 0 :
8074 ((f & MDB_INTEGERDUP)
8075 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8076 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8079 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8084 int rc, dbflag, exact;
8085 unsigned int unused = 0;
8088 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8089 mdb_default_cmp(txn, FREE_DBI);
8092 if ((flags & VALID_FLAGS) != flags)
8094 if (txn->mt_flags & MDB_TXN_ERROR)
8100 if (flags & PERSISTENT_FLAGS) {
8101 uint16_t f2 = flags & PERSISTENT_FLAGS;
8102 /* make sure flag changes get committed */
8103 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8104 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8105 txn->mt_flags |= MDB_TXN_DIRTY;
8108 mdb_default_cmp(txn, MAIN_DBI);
8112 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8113 mdb_default_cmp(txn, MAIN_DBI);
8116 /* Is the DB already open? */
8118 for (i=2; i<txn->mt_numdbs; i++) {
8119 if (!txn->mt_dbxs[i].md_name.mv_size) {
8120 /* Remember this free slot */
8121 if (!unused) unused = i;
8124 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8125 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8131 /* If no free slot and max hit, fail */
8132 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8133 return MDB_DBS_FULL;
8135 /* Cannot mix named databases with some mainDB flags */
8136 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8137 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8139 /* Find the DB info */
8140 dbflag = DB_NEW|DB_VALID;
8143 key.mv_data = (void *)name;
8144 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8145 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8146 if (rc == MDB_SUCCESS) {
8147 /* make sure this is actually a DB */
8148 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8149 if (!(node->mn_flags & F_SUBDATA))
8150 return MDB_INCOMPATIBLE;
8151 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8152 /* Create if requested */
8154 data.mv_size = sizeof(MDB_db);
8155 data.mv_data = &dummy;
8156 memset(&dummy, 0, sizeof(dummy));
8157 dummy.md_root = P_INVALID;
8158 dummy.md_flags = flags & PERSISTENT_FLAGS;
8159 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8163 /* OK, got info, add to table */
8164 if (rc == MDB_SUCCESS) {
8165 unsigned int slot = unused ? unused : txn->mt_numdbs;
8166 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8167 txn->mt_dbxs[slot].md_name.mv_size = len;
8168 txn->mt_dbxs[slot].md_rel = NULL;
8169 txn->mt_dbflags[slot] = dbflag;
8170 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8172 mdb_default_cmp(txn, slot);
8181 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8183 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8186 if (txn->mt_dbflags[dbi] & DB_STALE) {
8189 /* Stale, must read the DB's root. cursor_init does it for us. */
8190 mdb_cursor_init(&mc, txn, dbi, &mx);
8192 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8195 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8198 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8200 ptr = env->me_dbxs[dbi].md_name.mv_data;
8201 env->me_dbxs[dbi].md_name.mv_data = NULL;
8202 env->me_dbxs[dbi].md_name.mv_size = 0;
8203 env->me_dbflags[dbi] = 0;
8207 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8209 /* We could return the flags for the FREE_DBI too but what's the point? */
8210 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8212 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8216 /** Add all the DB's pages to the free list.
8217 * @param[in] mc Cursor on the DB to free.
8218 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8219 * @return 0 on success, non-zero on failure.
8222 mdb_drop0(MDB_cursor *mc, int subs)
8226 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8227 if (rc == MDB_SUCCESS) {
8228 MDB_txn *txn = mc->mc_txn;
8233 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8234 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8237 mdb_cursor_copy(mc, &mx);
8238 while (mc->mc_snum > 0) {
8239 MDB_page *mp = mc->mc_pg[mc->mc_top];
8240 unsigned n = NUMKEYS(mp);
8242 for (i=0; i<n; i++) {
8243 ni = NODEPTR(mp, i);
8244 if (ni->mn_flags & F_BIGDATA) {
8247 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8248 rc = mdb_page_get(txn, pg, &omp, NULL);
8251 mdb_cassert(mc, IS_OVERFLOW(omp));
8252 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8256 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8257 mdb_xcursor_init1(mc, ni);
8258 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8264 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8266 for (i=0; i<n; i++) {
8268 ni = NODEPTR(mp, i);
8271 mdb_midl_xappend(txn->mt_free_pgs, pg);
8276 mc->mc_ki[mc->mc_top] = i;
8277 rc = mdb_cursor_sibling(mc, 1);
8279 /* no more siblings, go back to beginning
8280 * of previous level.
8284 for (i=1; i<mc->mc_snum; i++) {
8286 mc->mc_pg[i] = mx.mc_pg[i];
8291 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8292 } else if (rc == MDB_NOTFOUND) {
8298 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8300 MDB_cursor *mc, *m2;
8303 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8306 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8309 rc = mdb_cursor_open(txn, dbi, &mc);
8313 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8314 /* Invalidate the dropped DB's cursors */
8315 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8316 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8320 /* Can't delete the main DB */
8321 if (del && dbi > MAIN_DBI) {
8322 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8324 txn->mt_dbflags[dbi] = DB_STALE;
8325 mdb_dbi_close(txn->mt_env, dbi);
8328 /* reset the DB record, mark it dirty */
8329 txn->mt_dbflags[dbi] |= DB_DIRTY;
8330 txn->mt_dbs[dbi].md_depth = 0;
8331 txn->mt_dbs[dbi].md_branch_pages = 0;
8332 txn->mt_dbs[dbi].md_leaf_pages = 0;
8333 txn->mt_dbs[dbi].md_overflow_pages = 0;
8334 txn->mt_dbs[dbi].md_entries = 0;
8335 txn->mt_dbs[dbi].md_root = P_INVALID;
8337 txn->mt_flags |= MDB_TXN_DIRTY;
8340 mdb_cursor_close(mc);
8344 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8346 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8349 txn->mt_dbxs[dbi].md_cmp = cmp;
8353 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8355 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8358 txn->mt_dbxs[dbi].md_dcmp = cmp;
8362 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8364 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8367 txn->mt_dbxs[dbi].md_rel = rel;
8371 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8373 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8376 txn->mt_dbxs[dbi].md_relctx = ctx;
8380 int mdb_env_get_maxkeysize(MDB_env *env)
8382 return ENV_MAXKEY(env);
8385 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8387 unsigned int i, rdrs;
8390 int rc = 0, first = 1;
8394 if (!env->me_txns) {
8395 return func("(no reader locks)\n", ctx);
8397 rdrs = env->me_txns->mti_numreaders;
8398 mr = env->me_txns->mti_readers;
8399 for (i=0; i<rdrs; i++) {
8401 txnid_t txnid = mr[i].mr_txnid;
8402 sprintf(buf, txnid == (txnid_t)-1 ?
8403 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
8404 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
8407 rc = func(" pid thread txnid\n", ctx);
8411 rc = func(buf, ctx);
8417 rc = func("(no active readers)\n", ctx);
8422 /** Insert pid into list if not already present.
8423 * return -1 if already present.
8425 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8427 /* binary search of pid in list */
8429 unsigned cursor = 1;
8431 unsigned n = ids[0];
8434 unsigned pivot = n >> 1;
8435 cursor = base + pivot + 1;
8436 val = pid - ids[cursor];
8441 } else if ( val > 0 ) {
8446 /* found, so it's a duplicate */
8455 for (n = ids[0]; n > cursor; n--)
8461 int mdb_reader_check(MDB_env *env, int *dead)
8463 unsigned int i, j, rdrs;
8465 MDB_PID_T *pids, pid;
8474 rdrs = env->me_txns->mti_numreaders;
8475 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8479 mr = env->me_txns->mti_readers;
8480 for (i=0; i<rdrs; i++) {
8481 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8483 if (mdb_pid_insert(pids, pid) == 0) {
8484 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8486 /* Recheck, a new process may have reused pid */
8487 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8488 for (j=i; j<rdrs; j++)
8489 if (mr[j].mr_pid == pid) {
8490 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8491 (unsigned) pid, mr[j].mr_txnid));
8496 UNLOCK_MUTEX_R(env);