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 uint32_t me_flags; /**< @ref mdb_env */
1038 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1039 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1040 unsigned int me_maxreaders; /**< size of the reader table */
1041 unsigned int me_numreaders; /**< max numreaders set by this env */
1042 MDB_dbi me_numdbs; /**< number of DBs opened */
1043 MDB_dbi me_maxdbs; /**< size of the DB table */
1044 MDB_PID_T me_pid; /**< process ID of this env */
1045 char *me_path; /**< path to the DB files */
1046 char *me_map; /**< the memory map of the data file */
1047 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1048 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1049 void *me_pbuf; /**< scratch area for DUPSORT put() */
1050 MDB_txn *me_txn; /**< current write transaction */
1051 size_t me_mapsize; /**< size of the data memory map */
1052 off_t me_size; /**< current file size */
1053 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1054 MDB_dbx *me_dbxs; /**< array of static DB info */
1055 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1056 pthread_key_t me_txkey; /**< thread-key for readers */
1057 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1058 # define me_pglast me_pgstate.mf_pglast
1059 # define me_pghead me_pgstate.mf_pghead
1060 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1061 /** IDL of pages that became unused in a write txn */
1062 MDB_IDL me_free_pgs;
1063 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1064 MDB_ID2L me_dirty_list;
1065 /** Max number of freelist items that can fit in a single overflow page */
1067 /** Max size of a node on a page */
1068 unsigned int me_nodemax;
1069 #if !(MDB_MAXKEYSIZE)
1070 unsigned int me_maxkey; /**< max size of a key */
1072 int me_live_reader; /**< have liveness lock in reader table */
1074 int me_pidquery; /**< Used in OpenProcess */
1075 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1077 #elif defined(MDB_USE_POSIX_SEM)
1078 sem_t *me_rmutex; /* Shared mutexes are not supported */
1081 void *me_userctx; /**< User-settable context */
1082 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1085 /** Nested transaction */
1086 typedef struct MDB_ntxn {
1087 MDB_txn mnt_txn; /**< the transaction */
1088 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1091 /** max number of pages to commit in one writev() call */
1092 #define MDB_COMMIT_PAGES 64
1093 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1094 #undef MDB_COMMIT_PAGES
1095 #define MDB_COMMIT_PAGES IOV_MAX
1098 /** max bytes to write in one call */
1099 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1101 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1102 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1103 static int mdb_page_touch(MDB_cursor *mc);
1105 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1106 static int mdb_page_search_root(MDB_cursor *mc,
1107 MDB_val *key, int modify);
1108 #define MDB_PS_MODIFY 1
1109 #define MDB_PS_ROOTONLY 2
1110 #define MDB_PS_FIRST 4
1111 #define MDB_PS_LAST 8
1112 static int mdb_page_search(MDB_cursor *mc,
1113 MDB_val *key, int flags);
1114 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1116 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1117 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1118 pgno_t newpgno, unsigned int nflags);
1120 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1121 static int mdb_env_pick_meta(const MDB_env *env);
1122 static int mdb_env_write_meta(MDB_txn *txn);
1123 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1124 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1126 static void mdb_env_close0(MDB_env *env, int excl);
1128 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1129 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1130 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1131 static void mdb_node_del(MDB_cursor *mc, int ksize);
1132 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1133 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1134 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1135 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1136 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1138 static int mdb_rebalance(MDB_cursor *mc);
1139 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1141 static void mdb_cursor_pop(MDB_cursor *mc);
1142 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1144 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1145 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1146 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1147 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1148 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1150 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1151 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1153 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1154 static void mdb_xcursor_init0(MDB_cursor *mc);
1155 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1157 static int mdb_drop0(MDB_cursor *mc, int subs);
1158 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1161 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1165 static SECURITY_DESCRIPTOR mdb_null_sd;
1166 static SECURITY_ATTRIBUTES mdb_all_sa;
1167 static int mdb_sec_inited;
1170 /** Return the library version info. */
1172 mdb_version(int *major, int *minor, int *patch)
1174 if (major) *major = MDB_VERSION_MAJOR;
1175 if (minor) *minor = MDB_VERSION_MINOR;
1176 if (patch) *patch = MDB_VERSION_PATCH;
1177 return MDB_VERSION_STRING;
1180 /** Table of descriptions for MDB @ref errors */
1181 static char *const mdb_errstr[] = {
1182 "MDB_KEYEXIST: Key/data pair already exists",
1183 "MDB_NOTFOUND: No matching key/data pair found",
1184 "MDB_PAGE_NOTFOUND: Requested page not found",
1185 "MDB_CORRUPTED: Located page was wrong type",
1186 "MDB_PANIC: Update of meta page failed",
1187 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1188 "MDB_INVALID: File is not an MDB file",
1189 "MDB_MAP_FULL: Environment mapsize limit reached",
1190 "MDB_DBS_FULL: Environment maxdbs limit reached",
1191 "MDB_READERS_FULL: Environment maxreaders limit reached",
1192 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1193 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1194 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1195 "MDB_PAGE_FULL: Internal error - page has no more space",
1196 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1197 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1198 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1199 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1200 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1204 mdb_strerror(int err)
1208 return ("Successful return: 0");
1210 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1211 i = err - MDB_KEYEXIST;
1212 return mdb_errstr[i];
1215 return strerror(err);
1218 /** assert(3) variant in cursor context */
1219 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1220 /** assert(3) variant in transaction context */
1221 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1222 /** assert(3) variant in environment context */
1223 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1226 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1227 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1230 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1231 const char *func, const char *file, int line)
1234 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1235 file, line, expr_txt, func);
1236 if (env->me_assert_func)
1237 env->me_assert_func(env, buf);
1238 fprintf(stderr, "%s\n", buf);
1242 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1246 /** Return the page number of \b mp which may be sub-page, for debug output */
1248 mdb_dbg_pgno(MDB_page *mp)
1251 COPY_PGNO(ret, mp->mp_pgno);
1255 /** Display a key in hexadecimal and return the address of the result.
1256 * @param[in] key the key to display
1257 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1258 * @return The key in hexadecimal form.
1261 mdb_dkey(MDB_val *key, char *buf)
1264 unsigned char *c = key->mv_data;
1270 if (key->mv_size > DKBUF_MAXKEYSIZE)
1271 return "MDB_MAXKEYSIZE";
1272 /* may want to make this a dynamic check: if the key is mostly
1273 * printable characters, print it as-is instead of converting to hex.
1277 for (i=0; i<key->mv_size; i++)
1278 ptr += sprintf(ptr, "%02x", *c++);
1280 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1286 mdb_leafnode_type(MDB_node *n)
1288 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1289 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1290 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1293 /** Display all the keys in the page. */
1295 mdb_page_list(MDB_page *mp)
1297 pgno_t pgno = mdb_dbg_pgno(mp);
1298 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1300 unsigned int i, nkeys, nsize, total = 0;
1304 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1305 case P_BRANCH: type = "Branch page"; break;
1306 case P_LEAF: type = "Leaf page"; break;
1307 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1308 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1309 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1311 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1312 pgno, mp->mp_pages, state);
1315 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1316 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1319 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1323 nkeys = NUMKEYS(mp);
1324 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1326 for (i=0; i<nkeys; i++) {
1327 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1328 key.mv_size = nsize = mp->mp_pad;
1329 key.mv_data = LEAF2KEY(mp, i, nsize);
1331 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1334 node = NODEPTR(mp, i);
1335 key.mv_size = node->mn_ksize;
1336 key.mv_data = node->mn_data;
1337 nsize = NODESIZE + key.mv_size;
1338 if (IS_BRANCH(mp)) {
1339 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1343 if (F_ISSET(node->mn_flags, F_BIGDATA))
1344 nsize += sizeof(pgno_t);
1346 nsize += NODEDSZ(node);
1348 nsize += sizeof(indx_t);
1349 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1350 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1352 total = EVEN(total);
1354 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1355 IS_LEAF2(mp) ? PAGEHDRSZ : mp->mp_lower, total, SIZELEFT(mp));
1359 mdb_cursor_chk(MDB_cursor *mc)
1365 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1366 for (i=0; i<mc->mc_top; i++) {
1368 node = NODEPTR(mp, mc->mc_ki[i]);
1369 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1372 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1378 /** Count all the pages in each DB and in the freelist
1379 * and make sure it matches the actual number of pages
1382 static void mdb_audit(MDB_txn *txn)
1386 MDB_ID freecount, count;
1391 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1392 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1393 freecount += *(MDB_ID *)data.mv_data;
1396 for (i = 0; i<txn->mt_numdbs; i++) {
1398 mdb_cursor_init(&mc, txn, i, &mx);
1399 if (txn->mt_dbs[i].md_root == P_INVALID)
1401 count += txn->mt_dbs[i].md_branch_pages +
1402 txn->mt_dbs[i].md_leaf_pages +
1403 txn->mt_dbs[i].md_overflow_pages;
1404 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1405 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1409 mp = mc.mc_pg[mc.mc_top];
1410 for (j=0; j<NUMKEYS(mp); j++) {
1411 MDB_node *leaf = NODEPTR(mp, j);
1412 if (leaf->mn_flags & F_SUBDATA) {
1414 memcpy(&db, NODEDATA(leaf), sizeof(db));
1415 count += db.md_branch_pages + db.md_leaf_pages +
1416 db.md_overflow_pages;
1420 while (mdb_cursor_sibling(&mc, 1) == 0);
1423 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1424 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1425 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1431 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1433 return txn->mt_dbxs[dbi].md_cmp(a, b);
1437 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1439 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1442 /** Allocate memory for a page.
1443 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1446 mdb_page_malloc(MDB_txn *txn, unsigned num)
1448 MDB_env *env = txn->mt_env;
1449 MDB_page *ret = env->me_dpages;
1450 size_t psize = env->me_psize, sz = psize, off;
1451 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1452 * For a single page alloc, we init everything after the page header.
1453 * For multi-page, we init the final page; if the caller needed that
1454 * many pages they will be filling in at least up to the last page.
1458 VGMEMP_ALLOC(env, ret, sz);
1459 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1460 env->me_dpages = ret->mp_next;
1463 psize -= off = PAGEHDRSZ;
1468 if ((ret = malloc(sz)) != NULL) {
1469 VGMEMP_ALLOC(env, ret, sz);
1470 if (!(env->me_flags & MDB_NOMEMINIT)) {
1471 memset((char *)ret + off, 0, psize);
1475 txn->mt_flags |= MDB_TXN_ERROR;
1480 /** Free a single page.
1481 * Saves single pages to a list, for future reuse.
1482 * (This is not used for multi-page overflow pages.)
1485 mdb_page_free(MDB_env *env, MDB_page *mp)
1487 mp->mp_next = env->me_dpages;
1488 VGMEMP_FREE(env, mp);
1489 env->me_dpages = mp;
1492 /** Free a dirty page */
1494 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1496 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1497 mdb_page_free(env, dp);
1499 /* large pages just get freed directly */
1500 VGMEMP_FREE(env, dp);
1505 /** Return all dirty pages to dpage list */
1507 mdb_dlist_free(MDB_txn *txn)
1509 MDB_env *env = txn->mt_env;
1510 MDB_ID2L dl = txn->mt_u.dirty_list;
1511 unsigned i, n = dl[0].mid;
1513 for (i = 1; i <= n; i++) {
1514 mdb_dpage_free(env, dl[i].mptr);
1519 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1520 * @param[in] mc A cursor handle for the current operation.
1521 * @param[in] pflags Flags of the pages to update:
1522 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1523 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1524 * @return 0 on success, non-zero on failure.
1527 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1529 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1530 MDB_txn *txn = mc->mc_txn;
1536 int rc = MDB_SUCCESS, level;
1538 /* Mark pages seen by cursors */
1539 if (mc->mc_flags & C_UNTRACK)
1540 mc = NULL; /* will find mc in mt_cursors */
1541 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1542 for (; mc; mc=mc->mc_next) {
1543 if (!(mc->mc_flags & C_INITIALIZED))
1545 for (m3 = mc;; m3 = &mx->mx_cursor) {
1547 for (j=0; j<m3->mc_snum; j++) {
1549 if ((mp->mp_flags & Mask) == pflags)
1550 mp->mp_flags ^= P_KEEP;
1552 mx = m3->mc_xcursor;
1553 /* Proceed to mx if it is at a sub-database */
1554 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1556 if (! (mp && (mp->mp_flags & P_LEAF)))
1558 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1559 if (!(leaf->mn_flags & F_SUBDATA))
1568 /* Mark dirty root pages */
1569 for (i=0; i<txn->mt_numdbs; i++) {
1570 if (txn->mt_dbflags[i] & DB_DIRTY) {
1571 pgno_t pgno = txn->mt_dbs[i].md_root;
1572 if (pgno == P_INVALID)
1574 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1576 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1577 dp->mp_flags ^= P_KEEP;
1585 static int mdb_page_flush(MDB_txn *txn, int keep);
1587 /** Spill pages from the dirty list back to disk.
1588 * This is intended to prevent running into #MDB_TXN_FULL situations,
1589 * but note that they may still occur in a few cases:
1590 * 1) our estimate of the txn size could be too small. Currently this
1591 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1592 * 2) child txns may run out of space if their parents dirtied a
1593 * lot of pages and never spilled them. TODO: we probably should do
1594 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1595 * the parent's dirty_room is below a given threshold.
1597 * Otherwise, if not using nested txns, it is expected that apps will
1598 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1599 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1600 * If the txn never references them again, they can be left alone.
1601 * If the txn only reads them, they can be used without any fuss.
1602 * If the txn writes them again, they can be dirtied immediately without
1603 * going thru all of the work of #mdb_page_touch(). Such references are
1604 * handled by #mdb_page_unspill().
1606 * Also note, we never spill DB root pages, nor pages of active cursors,
1607 * because we'll need these back again soon anyway. And in nested txns,
1608 * we can't spill a page in a child txn if it was already spilled in a
1609 * parent txn. That would alter the parent txns' data even though
1610 * the child hasn't committed yet, and we'd have no way to undo it if
1611 * the child aborted.
1613 * @param[in] m0 cursor A cursor handle identifying the transaction and
1614 * database for which we are checking space.
1615 * @param[in] key For a put operation, the key being stored.
1616 * @param[in] data For a put operation, the data being stored.
1617 * @return 0 on success, non-zero on failure.
1620 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1622 MDB_txn *txn = m0->mc_txn;
1624 MDB_ID2L dl = txn->mt_u.dirty_list;
1625 unsigned int i, j, need;
1628 if (m0->mc_flags & C_SUB)
1631 /* Estimate how much space this op will take */
1632 i = m0->mc_db->md_depth;
1633 /* Named DBs also dirty the main DB */
1634 if (m0->mc_dbi > MAIN_DBI)
1635 i += txn->mt_dbs[MAIN_DBI].md_depth;
1636 /* For puts, roughly factor in the key+data size */
1638 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1639 i += i; /* double it for good measure */
1642 if (txn->mt_dirty_room > i)
1645 if (!txn->mt_spill_pgs) {
1646 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1647 if (!txn->mt_spill_pgs)
1650 /* purge deleted slots */
1651 MDB_IDL sl = txn->mt_spill_pgs;
1652 unsigned int num = sl[0];
1654 for (i=1; i<=num; i++) {
1661 /* Preserve pages which may soon be dirtied again */
1662 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1665 /* Less aggressive spill - we originally spilled the entire dirty list,
1666 * with a few exceptions for cursor pages and DB root pages. But this
1667 * turns out to be a lot of wasted effort because in a large txn many
1668 * of those pages will need to be used again. So now we spill only 1/8th
1669 * of the dirty pages. Testing revealed this to be a good tradeoff,
1670 * better than 1/2, 1/4, or 1/10.
1672 if (need < MDB_IDL_UM_MAX / 8)
1673 need = MDB_IDL_UM_MAX / 8;
1675 /* Save the page IDs of all the pages we're flushing */
1676 /* flush from the tail forward, this saves a lot of shifting later on. */
1677 for (i=dl[0].mid; i && need; i--) {
1678 MDB_ID pn = dl[i].mid << 1;
1680 if (dp->mp_flags & P_KEEP)
1682 /* Can't spill twice, make sure it's not already in a parent's
1685 if (txn->mt_parent) {
1687 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1688 if (tx2->mt_spill_pgs) {
1689 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1690 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1691 dp->mp_flags |= P_KEEP;
1699 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1703 mdb_midl_sort(txn->mt_spill_pgs);
1705 /* Flush the spilled part of dirty list */
1706 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1709 /* Reset any dirty pages we kept that page_flush didn't see */
1710 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1713 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1717 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1719 mdb_find_oldest(MDB_txn *txn)
1722 txnid_t mr, oldest = txn->mt_txnid - 1;
1723 if (txn->mt_env->me_txns) {
1724 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1725 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1736 /** Add a page to the txn's dirty list */
1738 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1741 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1743 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1744 insert = mdb_mid2l_append;
1746 insert = mdb_mid2l_insert;
1748 mid.mid = mp->mp_pgno;
1750 rc = insert(txn->mt_u.dirty_list, &mid);
1751 mdb_tassert(txn, rc == 0);
1752 txn->mt_dirty_room--;
1755 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1756 * me_pghead and mt_next_pgno.
1758 * If there are free pages available from older transactions, they
1759 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1760 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1761 * and move me_pglast to say which records were consumed. Only this
1762 * function can create me_pghead and move me_pglast/mt_next_pgno.
1763 * @param[in] mc cursor A cursor handle identifying the transaction and
1764 * database for which we are allocating.
1765 * @param[in] num the number of pages to allocate.
1766 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1767 * will always be satisfied by a single contiguous chunk of memory.
1768 * @return 0 on success, non-zero on failure.
1771 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1773 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1774 /* Get at most <Max_retries> more freeDB records once me_pghead
1775 * has enough pages. If not enough, use new pages from the map.
1776 * If <Paranoid> and mc is updating the freeDB, only get new
1777 * records if me_pghead is empty. Then the freelist cannot play
1778 * catch-up with itself by growing while trying to save it.
1780 enum { Paranoid = 1, Max_retries = 500 };
1782 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1784 int rc, retry = Max_retries;
1785 MDB_txn *txn = mc->mc_txn;
1786 MDB_env *env = txn->mt_env;
1787 pgno_t pgno, *mop = env->me_pghead;
1788 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1790 txnid_t oldest = 0, last;
1796 /* If our dirty list is already full, we can't do anything */
1797 if (txn->mt_dirty_room == 0) {
1802 for (op = MDB_FIRST;; op = MDB_NEXT) {
1805 pgno_t *idl, old_id, new_id;
1807 /* Seek a big enough contiguous page range. Prefer
1808 * pages at the tail, just truncating the list.
1814 if (mop[i-n2] == pgno+n2)
1817 if (Max_retries < INT_MAX && --retry < 0)
1821 if (op == MDB_FIRST) { /* 1st iteration */
1822 /* Prepare to fetch more and coalesce */
1823 oldest = mdb_find_oldest(txn);
1824 last = env->me_pglast;
1825 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1828 key.mv_data = &last; /* will look up last+1 */
1829 key.mv_size = sizeof(last);
1831 if (Paranoid && mc->mc_dbi == FREE_DBI)
1834 if (Paranoid && retry < 0 && mop_len)
1838 /* Do not fetch more if the record will be too recent */
1841 rc = mdb_cursor_get(&m2, &key, NULL, op);
1843 if (rc == MDB_NOTFOUND)
1847 last = *(txnid_t*)key.mv_data;
1850 np = m2.mc_pg[m2.mc_top];
1851 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1852 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1855 idl = (MDB_ID *) data.mv_data;
1858 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1863 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1865 mop = env->me_pghead;
1867 env->me_pglast = last;
1869 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1870 last, txn->mt_dbs[FREE_DBI].md_root, i));
1872 DPRINTF(("IDL %"Z"u", idl[k]));
1874 /* Merge in descending sorted order */
1877 mop[0] = (pgno_t)-1;
1881 for (; old_id < new_id; old_id = mop[--j])
1888 /* Use new pages from the map when nothing suitable in the freeDB */
1890 pgno = txn->mt_next_pgno;
1891 if (pgno + num >= env->me_maxpg) {
1892 DPUTS("DB size maxed out");
1898 if (env->me_flags & MDB_WRITEMAP) {
1899 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1901 if (!(np = mdb_page_malloc(txn, num))) {
1907 mop[0] = mop_len -= num;
1908 /* Move any stragglers down */
1909 for (j = i-num; j < mop_len; )
1910 mop[++j] = mop[++i];
1912 txn->mt_next_pgno = pgno + num;
1915 mdb_page_dirty(txn, np);
1921 txn->mt_flags |= MDB_TXN_ERROR;
1925 /** Copy the used portions of a non-overflow page.
1926 * @param[in] dst page to copy into
1927 * @param[in] src page to copy from
1928 * @param[in] psize size of a page
1931 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1933 enum { Align = sizeof(pgno_t) };
1934 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1936 /* If page isn't full, just copy the used portion. Adjust
1937 * alignment so memcpy may copy words instead of bytes.
1939 if ((unused &= -Align) && !IS_LEAF2(src)) {
1941 memcpy(dst, src, (lower + (Align-1)) & -Align);
1942 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1945 memcpy(dst, src, psize - unused);
1949 /** Pull a page off the txn's spill list, if present.
1950 * If a page being referenced was spilled to disk in this txn, bring
1951 * it back and make it dirty/writable again.
1952 * @param[in] txn the transaction handle.
1953 * @param[in] mp the page being referenced. It must not be dirty.
1954 * @param[out] ret the writable page, if any. ret is unchanged if
1955 * mp wasn't spilled.
1958 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1960 MDB_env *env = txn->mt_env;
1963 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1965 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1966 if (!tx2->mt_spill_pgs)
1968 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1969 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1972 if (txn->mt_dirty_room == 0)
1973 return MDB_TXN_FULL;
1974 if (IS_OVERFLOW(mp))
1978 if (env->me_flags & MDB_WRITEMAP) {
1981 np = mdb_page_malloc(txn, num);
1985 memcpy(np, mp, num * env->me_psize);
1987 mdb_page_copy(np, mp, env->me_psize);
1990 /* If in current txn, this page is no longer spilled.
1991 * If it happens to be the last page, truncate the spill list.
1992 * Otherwise mark it as deleted by setting the LSB.
1994 if (x == txn->mt_spill_pgs[0])
1995 txn->mt_spill_pgs[0]--;
1997 txn->mt_spill_pgs[x] |= 1;
1998 } /* otherwise, if belonging to a parent txn, the
1999 * page remains spilled until child commits
2002 mdb_page_dirty(txn, np);
2003 np->mp_flags |= P_DIRTY;
2011 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2012 * @param[in] mc cursor pointing to the page to be touched
2013 * @return 0 on success, non-zero on failure.
2016 mdb_page_touch(MDB_cursor *mc)
2018 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2019 MDB_txn *txn = mc->mc_txn;
2020 MDB_cursor *m2, *m3;
2024 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2025 if (txn->mt_flags & MDB_TXN_SPILLS) {
2027 rc = mdb_page_unspill(txn, mp, &np);
2033 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2034 (rc = mdb_page_alloc(mc, 1, &np)))
2037 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2038 mp->mp_pgno, pgno));
2039 mdb_cassert(mc, mp->mp_pgno != pgno);
2040 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2041 /* Update the parent page, if any, to point to the new page */
2043 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2044 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2045 SETPGNO(node, pgno);
2047 mc->mc_db->md_root = pgno;
2049 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2050 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2052 /* If txn has a parent, make sure the page is in our
2056 unsigned x = mdb_mid2l_search(dl, pgno);
2057 if (x <= dl[0].mid && dl[x].mid == pgno) {
2058 if (mp != dl[x].mptr) { /* bad cursor? */
2059 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2060 txn->mt_flags |= MDB_TXN_ERROR;
2061 return MDB_CORRUPTED;
2066 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2068 np = mdb_page_malloc(txn, 1);
2073 rc = mdb_mid2l_insert(dl, &mid);
2074 mdb_cassert(mc, rc == 0);
2079 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2081 np->mp_flags |= P_DIRTY;
2084 /* Adjust cursors pointing to mp */
2085 mc->mc_pg[mc->mc_top] = np;
2086 m2 = txn->mt_cursors[mc->mc_dbi];
2087 if (mc->mc_flags & C_SUB) {
2088 for (; m2; m2=m2->mc_next) {
2089 m3 = &m2->mc_xcursor->mx_cursor;
2090 if (m3->mc_snum < mc->mc_snum) continue;
2091 if (m3->mc_pg[mc->mc_top] == mp)
2092 m3->mc_pg[mc->mc_top] = np;
2095 for (; m2; m2=m2->mc_next) {
2096 if (m2->mc_snum < mc->mc_snum) continue;
2097 if (m2->mc_pg[mc->mc_top] == mp) {
2098 m2->mc_pg[mc->mc_top] = np;
2099 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2100 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2102 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2103 if (!(leaf->mn_flags & F_SUBDATA))
2104 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2112 txn->mt_flags |= MDB_TXN_ERROR;
2117 mdb_env_sync(MDB_env *env, int force)
2120 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2121 if (env->me_flags & MDB_WRITEMAP) {
2122 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2123 ? MS_ASYNC : MS_SYNC;
2124 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2127 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2131 if (MDB_FDATASYNC(env->me_fd))
2138 /** Back up parent txn's cursors, then grab the originals for tracking */
2140 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2142 MDB_cursor *mc, *bk;
2147 for (i = src->mt_numdbs; --i >= 0; ) {
2148 if ((mc = src->mt_cursors[i]) != NULL) {
2149 size = sizeof(MDB_cursor);
2151 size += sizeof(MDB_xcursor);
2152 for (; mc; mc = bk->mc_next) {
2158 mc->mc_db = &dst->mt_dbs[i];
2159 /* Kill pointers into src - and dst to reduce abuse: The
2160 * user may not use mc until dst ends. Otherwise we'd...
2162 mc->mc_txn = NULL; /* ...set this to dst */
2163 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2164 if ((mx = mc->mc_xcursor) != NULL) {
2165 *(MDB_xcursor *)(bk+1) = *mx;
2166 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2168 mc->mc_next = dst->mt_cursors[i];
2169 dst->mt_cursors[i] = mc;
2176 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2177 * @param[in] txn the transaction handle.
2178 * @param[in] merge true to keep changes to parent cursors, false to revert.
2179 * @return 0 on success, non-zero on failure.
2182 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2184 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2188 for (i = txn->mt_numdbs; --i >= 0; ) {
2189 for (mc = cursors[i]; mc; mc = next) {
2191 if ((bk = mc->mc_backup) != NULL) {
2193 /* Commit changes to parent txn */
2194 mc->mc_next = bk->mc_next;
2195 mc->mc_backup = bk->mc_backup;
2196 mc->mc_txn = bk->mc_txn;
2197 mc->mc_db = bk->mc_db;
2198 mc->mc_dbflag = bk->mc_dbflag;
2199 if ((mx = mc->mc_xcursor) != NULL)
2200 mx->mx_cursor.mc_txn = bk->mc_txn;
2202 /* Abort nested txn */
2204 if ((mx = mc->mc_xcursor) != NULL)
2205 *mx = *(MDB_xcursor *)(bk+1);
2209 /* Only malloced cursors are permanently tracked. */
2217 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2220 mdb_txn_reset0(MDB_txn *txn, const char *act);
2222 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2228 Pidset = F_SETLK, Pidcheck = F_GETLK
2232 /** Set or check a pid lock. Set returns 0 on success.
2233 * Check returns 0 if the process is certainly dead, nonzero if it may
2234 * be alive (the lock exists or an error happened so we do not know).
2236 * On Windows Pidset is a no-op, we merely check for the existence
2237 * of the process with the given pid. On POSIX we use a single byte
2238 * lock on the lockfile, set at an offset equal to the pid.
2241 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2243 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2246 if (op == Pidcheck) {
2247 h = OpenProcess(env->me_pidquery, FALSE, pid);
2248 /* No documented "no such process" code, but other program use this: */
2250 return ErrCode() != ERROR_INVALID_PARAMETER;
2251 /* A process exists until all handles to it close. Has it exited? */
2252 ret = WaitForSingleObject(h, 0) != 0;
2259 struct flock lock_info;
2260 memset(&lock_info, 0, sizeof(lock_info));
2261 lock_info.l_type = F_WRLCK;
2262 lock_info.l_whence = SEEK_SET;
2263 lock_info.l_start = pid;
2264 lock_info.l_len = 1;
2265 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2266 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2268 } else if ((rc = ErrCode()) == EINTR) {
2276 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2277 * @param[in] txn the transaction handle to initialize
2278 * @return 0 on success, non-zero on failure.
2281 mdb_txn_renew0(MDB_txn *txn)
2283 MDB_env *env = txn->mt_env;
2284 MDB_txninfo *ti = env->me_txns;
2288 int rc, new_notls = 0;
2291 txn->mt_numdbs = env->me_numdbs;
2292 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2294 if (txn->mt_flags & MDB_TXN_RDONLY) {
2296 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2297 txn->mt_txnid = meta->mm_txnid;
2298 txn->mt_u.reader = NULL;
2300 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2301 pthread_getspecific(env->me_txkey);
2303 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2304 return MDB_BAD_RSLOT;
2306 MDB_PID_T pid = env->me_pid;
2307 pthread_t tid = pthread_self();
2309 if (!env->me_live_reader) {
2310 rc = mdb_reader_pid(env, Pidset, pid);
2313 env->me_live_reader = 1;
2317 nr = ti->mti_numreaders;
2318 for (i=0; i<nr; i++)
2319 if (ti->mti_readers[i].mr_pid == 0)
2321 if (i == env->me_maxreaders) {
2322 UNLOCK_MUTEX_R(env);
2323 return MDB_READERS_FULL;
2325 ti->mti_readers[i].mr_pid = pid;
2326 ti->mti_readers[i].mr_tid = tid;
2328 ti->mti_numreaders = ++nr;
2329 /* Save numreaders for un-mutexed mdb_env_close() */
2330 env->me_numreaders = nr;
2331 UNLOCK_MUTEX_R(env);
2333 r = &ti->mti_readers[i];
2334 new_notls = (env->me_flags & MDB_NOTLS);
2335 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2340 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2341 txn->mt_u.reader = r;
2342 meta = env->me_metas[txn->mt_txnid & 1];
2348 txn->mt_txnid = ti->mti_txnid;
2349 meta = env->me_metas[txn->mt_txnid & 1];
2351 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2352 txn->mt_txnid = meta->mm_txnid;
2356 if (txn->mt_txnid == mdb_debug_start)
2359 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2360 txn->mt_u.dirty_list = env->me_dirty_list;
2361 txn->mt_u.dirty_list[0].mid = 0;
2362 txn->mt_free_pgs = env->me_free_pgs;
2363 txn->mt_free_pgs[0] = 0;
2364 txn->mt_spill_pgs = NULL;
2368 /* Copy the DB info and flags */
2369 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2371 /* Moved to here to avoid a data race in read TXNs */
2372 txn->mt_next_pgno = meta->mm_last_pg+1;
2374 for (i=2; i<txn->mt_numdbs; i++) {
2375 x = env->me_dbflags[i];
2376 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2377 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2379 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2381 if (env->me_maxpg < txn->mt_next_pgno) {
2382 mdb_txn_reset0(txn, "renew0-mapfail");
2384 txn->mt_u.reader->mr_pid = 0;
2385 txn->mt_u.reader = NULL;
2387 return MDB_MAP_RESIZED;
2394 mdb_txn_renew(MDB_txn *txn)
2398 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2401 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2402 DPUTS("environment had fatal error, must shutdown!");
2406 rc = mdb_txn_renew0(txn);
2407 if (rc == MDB_SUCCESS) {
2408 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2409 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2410 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2416 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2420 int rc, size, tsize = sizeof(MDB_txn);
2422 if (env->me_flags & MDB_FATAL_ERROR) {
2423 DPUTS("environment had fatal error, must shutdown!");
2426 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2429 /* Nested transactions: Max 1 child, write txns only, no writemap */
2430 if (parent->mt_child ||
2431 (flags & MDB_RDONLY) ||
2432 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2433 (env->me_flags & MDB_WRITEMAP))
2435 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2437 tsize = sizeof(MDB_ntxn);
2439 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2440 if (!(flags & MDB_RDONLY))
2441 size += env->me_maxdbs * sizeof(MDB_cursor *);
2443 if ((txn = calloc(1, size)) == NULL) {
2444 DPRINTF(("calloc: %s", strerror(ErrCode())));
2447 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2448 if (flags & MDB_RDONLY) {
2449 txn->mt_flags |= MDB_TXN_RDONLY;
2450 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2452 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2453 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2459 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2460 if (!txn->mt_u.dirty_list ||
2461 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2463 free(txn->mt_u.dirty_list);
2467 txn->mt_txnid = parent->mt_txnid;
2468 txn->mt_dirty_room = parent->mt_dirty_room;
2469 txn->mt_u.dirty_list[0].mid = 0;
2470 txn->mt_spill_pgs = NULL;
2471 txn->mt_next_pgno = parent->mt_next_pgno;
2472 parent->mt_child = txn;
2473 txn->mt_parent = parent;
2474 txn->mt_numdbs = parent->mt_numdbs;
2475 txn->mt_flags = parent->mt_flags;
2476 txn->mt_dbxs = parent->mt_dbxs;
2477 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2478 /* Copy parent's mt_dbflags, but clear DB_NEW */
2479 for (i=0; i<txn->mt_numdbs; i++)
2480 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2482 ntxn = (MDB_ntxn *)txn;
2483 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2484 if (env->me_pghead) {
2485 size = MDB_IDL_SIZEOF(env->me_pghead);
2486 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2488 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2493 rc = mdb_cursor_shadow(parent, txn);
2495 mdb_txn_reset0(txn, "beginchild-fail");
2497 rc = mdb_txn_renew0(txn);
2503 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2504 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2505 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2512 mdb_txn_env(MDB_txn *txn)
2514 if(!txn) return NULL;
2518 /** Export or close DBI handles opened in this txn. */
2520 mdb_dbis_update(MDB_txn *txn, int keep)
2523 MDB_dbi n = txn->mt_numdbs;
2524 MDB_env *env = txn->mt_env;
2525 unsigned char *tdbflags = txn->mt_dbflags;
2527 for (i = n; --i >= 2;) {
2528 if (tdbflags[i] & DB_NEW) {
2530 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2532 char *ptr = env->me_dbxs[i].md_name.mv_data;
2533 env->me_dbxs[i].md_name.mv_data = NULL;
2534 env->me_dbxs[i].md_name.mv_size = 0;
2535 env->me_dbflags[i] = 0;
2540 if (keep && env->me_numdbs < n)
2544 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2545 * May be called twice for readonly txns: First reset it, then abort.
2546 * @param[in] txn the transaction handle to reset
2547 * @param[in] act why the transaction is being reset
2550 mdb_txn_reset0(MDB_txn *txn, const char *act)
2552 MDB_env *env = txn->mt_env;
2554 /* Close any DBI handles opened in this txn */
2555 mdb_dbis_update(txn, 0);
2557 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2558 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2559 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2561 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2562 if (txn->mt_u.reader) {
2563 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2564 if (!(env->me_flags & MDB_NOTLS))
2565 txn->mt_u.reader = NULL; /* txn does not own reader */
2567 txn->mt_numdbs = 0; /* close nothing if called again */
2568 txn->mt_dbxs = NULL; /* mark txn as reset */
2570 mdb_cursors_close(txn, 0);
2572 if (!(env->me_flags & MDB_WRITEMAP)) {
2573 mdb_dlist_free(txn);
2575 mdb_midl_free(env->me_pghead);
2577 if (txn->mt_parent) {
2578 txn->mt_parent->mt_child = NULL;
2579 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2580 mdb_midl_free(txn->mt_free_pgs);
2581 mdb_midl_free(txn->mt_spill_pgs);
2582 free(txn->mt_u.dirty_list);
2586 if (mdb_midl_shrink(&txn->mt_free_pgs))
2587 env->me_free_pgs = txn->mt_free_pgs;
2588 env->me_pghead = NULL;
2592 /* The writer mutex was locked in mdb_txn_begin. */
2594 UNLOCK_MUTEX_W(env);
2599 mdb_txn_reset(MDB_txn *txn)
2604 /* This call is only valid for read-only txns */
2605 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2608 mdb_txn_reset0(txn, "reset");
2612 mdb_txn_abort(MDB_txn *txn)
2618 mdb_txn_abort(txn->mt_child);
2620 mdb_txn_reset0(txn, "abort");
2621 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2622 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2623 txn->mt_u.reader->mr_pid = 0;
2628 /** Save the freelist as of this transaction to the freeDB.
2629 * This changes the freelist. Keep trying until it stabilizes.
2632 mdb_freelist_save(MDB_txn *txn)
2634 /* env->me_pghead[] can grow and shrink during this call.
2635 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2636 * Page numbers cannot disappear from txn->mt_free_pgs[].
2639 MDB_env *env = txn->mt_env;
2640 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2641 txnid_t pglast = 0, head_id = 0;
2642 pgno_t freecnt = 0, *free_pgs, *mop;
2643 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2645 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2647 if (env->me_pghead) {
2648 /* Make sure first page of freeDB is touched and on freelist */
2649 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2650 if (rc && rc != MDB_NOTFOUND)
2654 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2655 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2656 ? SSIZE_MAX : maxfree_1pg;
2659 /* Come back here after each Put() in case freelist changed */
2664 /* If using records from freeDB which we have not yet
2665 * deleted, delete them and any we reserved for me_pghead.
2667 while (pglast < env->me_pglast) {
2668 rc = mdb_cursor_first(&mc, &key, NULL);
2671 pglast = head_id = *(txnid_t *)key.mv_data;
2672 total_room = head_room = 0;
2673 mdb_tassert(txn, pglast <= env->me_pglast);
2674 rc = mdb_cursor_del(&mc, 0);
2679 /* Save the IDL of pages freed by this txn, to a single record */
2680 if (freecnt < txn->mt_free_pgs[0]) {
2682 /* Make sure last page of freeDB is touched and on freelist */
2683 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2684 if (rc && rc != MDB_NOTFOUND)
2687 free_pgs = txn->mt_free_pgs;
2688 /* Write to last page of freeDB */
2689 key.mv_size = sizeof(txn->mt_txnid);
2690 key.mv_data = &txn->mt_txnid;
2692 freecnt = free_pgs[0];
2693 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2694 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2697 /* Retry if mt_free_pgs[] grew during the Put() */
2698 free_pgs = txn->mt_free_pgs;
2699 } while (freecnt < free_pgs[0]);
2700 mdb_midl_sort(free_pgs);
2701 memcpy(data.mv_data, free_pgs, data.mv_size);
2704 unsigned int i = free_pgs[0];
2705 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2706 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2708 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2714 mop = env->me_pghead;
2715 mop_len = mop ? mop[0] : 0;
2717 /* Reserve records for me_pghead[]. Split it if multi-page,
2718 * to avoid searching freeDB for a page range. Use keys in
2719 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2721 if (total_room >= mop_len) {
2722 if (total_room == mop_len || --more < 0)
2724 } else if (head_room >= maxfree_1pg && head_id > 1) {
2725 /* Keep current record (overflow page), add a new one */
2729 /* (Re)write {key = head_id, IDL length = head_room} */
2730 total_room -= head_room;
2731 head_room = mop_len - total_room;
2732 if (head_room > maxfree_1pg && head_id > 1) {
2733 /* Overflow multi-page for part of me_pghead */
2734 head_room /= head_id; /* amortize page sizes */
2735 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2736 } else if (head_room < 0) {
2737 /* Rare case, not bothering to delete this record */
2740 key.mv_size = sizeof(head_id);
2741 key.mv_data = &head_id;
2742 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2743 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2746 /* IDL is initially empty, zero out at least the length */
2747 pgs = (pgno_t *)data.mv_data;
2748 j = head_room > clean_limit ? head_room : 0;
2752 total_room += head_room;
2755 /* Fill in the reserved me_pghead records */
2761 rc = mdb_cursor_first(&mc, &key, &data);
2762 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2763 unsigned flags = MDB_CURRENT;
2764 txnid_t id = *(txnid_t *)key.mv_data;
2765 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2768 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2769 if (len > mop_len) {
2771 data.mv_size = (len + 1) * sizeof(MDB_ID);
2772 /* Drop MDB_CURRENT when changing the data size */
2776 data.mv_data = mop -= len;
2779 rc = mdb_cursor_put(&mc, &key, &data, flags);
2781 if (rc || !(mop_len -= len))
2788 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2789 * @param[in] txn the transaction that's being committed
2790 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2791 * @return 0 on success, non-zero on failure.
2794 mdb_page_flush(MDB_txn *txn, int keep)
2796 MDB_env *env = txn->mt_env;
2797 MDB_ID2L dl = txn->mt_u.dirty_list;
2798 unsigned psize = env->me_psize, j;
2799 int i, pagecount = dl[0].mid, rc;
2800 size_t size = 0, pos = 0;
2802 MDB_page *dp = NULL;
2806 struct iovec iov[MDB_COMMIT_PAGES];
2807 ssize_t wpos = 0, wsize = 0, wres;
2808 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2814 if (env->me_flags & MDB_WRITEMAP) {
2815 /* Clear dirty flags */
2816 while (++i <= pagecount) {
2818 /* Don't flush this page yet */
2819 if (dp->mp_flags & P_KEEP) {
2820 dp->mp_flags ^= P_KEEP;
2824 dp->mp_flags &= ~P_DIRTY;
2829 /* Write the pages */
2831 if (++i <= pagecount) {
2833 /* Don't flush this page yet */
2834 if (dp->mp_flags & P_KEEP) {
2835 dp->mp_flags ^= P_KEEP;
2840 /* clear dirty flag */
2841 dp->mp_flags &= ~P_DIRTY;
2844 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2849 /* Windows actually supports scatter/gather I/O, but only on
2850 * unbuffered file handles. Since we're relying on the OS page
2851 * cache for all our data, that's self-defeating. So we just
2852 * write pages one at a time. We use the ov structure to set
2853 * the write offset, to at least save the overhead of a Seek
2856 DPRINTF(("committing page %"Z"u", pgno));
2857 memset(&ov, 0, sizeof(ov));
2858 ov.Offset = pos & 0xffffffff;
2859 ov.OffsetHigh = pos >> 16 >> 16;
2860 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2862 DPRINTF(("WriteFile: %d", rc));
2866 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2867 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2869 /* Write previous page(s) */
2870 #ifdef MDB_USE_PWRITEV
2871 wres = pwritev(env->me_fd, iov, n, wpos);
2874 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2876 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2878 DPRINTF(("lseek: %s", strerror(rc)));
2881 wres = writev(env->me_fd, iov, n);
2884 if (wres != wsize) {
2887 DPRINTF(("Write error: %s", strerror(rc)));
2889 rc = EIO; /* TODO: Use which error code? */
2890 DPUTS("short write, filesystem full?");
2901 DPRINTF(("committing page %"Z"u", pgno));
2902 next_pos = pos + size;
2903 iov[n].iov_len = size;
2904 iov[n].iov_base = (char *)dp;
2910 for (i = keep; ++i <= pagecount; ) {
2912 /* This is a page we skipped above */
2915 dl[j].mid = dp->mp_pgno;
2918 mdb_dpage_free(env, dp);
2923 txn->mt_dirty_room += i - j;
2929 mdb_txn_commit(MDB_txn *txn)
2935 if (txn == NULL || txn->mt_env == NULL)
2938 if (txn->mt_child) {
2939 rc = mdb_txn_commit(txn->mt_child);
2940 txn->mt_child = NULL;
2947 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2948 mdb_dbis_update(txn, 1);
2949 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2954 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2955 DPUTS("error flag is set, can't commit");
2957 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2962 if (txn->mt_parent) {
2963 MDB_txn *parent = txn->mt_parent;
2966 unsigned x, y, len, ps_len;
2968 /* Append our free list to parent's */
2969 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2972 mdb_midl_free(txn->mt_free_pgs);
2973 /* Failures after this must either undo the changes
2974 * to the parent or set MDB_TXN_ERROR in the parent.
2977 parent->mt_next_pgno = txn->mt_next_pgno;
2978 parent->mt_flags = txn->mt_flags;
2980 /* Merge our cursors into parent's and close them */
2981 mdb_cursors_close(txn, 1);
2983 /* Update parent's DB table. */
2984 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2985 parent->mt_numdbs = txn->mt_numdbs;
2986 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2987 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2988 for (i=2; i<txn->mt_numdbs; i++) {
2989 /* preserve parent's DB_NEW status */
2990 x = parent->mt_dbflags[i] & DB_NEW;
2991 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2994 dst = parent->mt_u.dirty_list;
2995 src = txn->mt_u.dirty_list;
2996 /* Remove anything in our dirty list from parent's spill list */
2997 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2999 pspill[0] = (pgno_t)-1;
3000 /* Mark our dirty pages as deleted in parent spill list */
3001 for (i=0, len=src[0].mid; ++i <= len; ) {
3002 MDB_ID pn = src[i].mid << 1;
3003 while (pn > pspill[x])
3005 if (pn == pspill[x]) {
3010 /* Squash deleted pagenums if we deleted any */
3011 for (x=y; ++x <= ps_len; )
3012 if (!(pspill[x] & 1))
3013 pspill[++y] = pspill[x];
3017 /* Find len = length of merging our dirty list with parent's */
3019 dst[0].mid = 0; /* simplify loops */
3020 if (parent->mt_parent) {
3021 len = x + src[0].mid;
3022 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3023 for (i = x; y && i; y--) {
3024 pgno_t yp = src[y].mid;
3025 while (yp < dst[i].mid)
3027 if (yp == dst[i].mid) {
3032 } else { /* Simplify the above for single-ancestor case */
3033 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3035 /* Merge our dirty list with parent's */
3037 for (i = len; y; dst[i--] = src[y--]) {
3038 pgno_t yp = src[y].mid;
3039 while (yp < dst[x].mid)
3040 dst[i--] = dst[x--];
3041 if (yp == dst[x].mid)
3042 free(dst[x--].mptr);
3044 mdb_tassert(txn, i == x);
3046 free(txn->mt_u.dirty_list);
3047 parent->mt_dirty_room = txn->mt_dirty_room;
3048 if (txn->mt_spill_pgs) {
3049 if (parent->mt_spill_pgs) {
3050 /* TODO: Prevent failure here, so parent does not fail */
3051 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3053 parent->mt_flags |= MDB_TXN_ERROR;
3054 mdb_midl_free(txn->mt_spill_pgs);
3055 mdb_midl_sort(parent->mt_spill_pgs);
3057 parent->mt_spill_pgs = txn->mt_spill_pgs;
3061 parent->mt_child = NULL;
3062 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3067 if (txn != env->me_txn) {
3068 DPUTS("attempt to commit unknown transaction");
3073 mdb_cursors_close(txn, 0);
3075 if (!txn->mt_u.dirty_list[0].mid &&
3076 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3079 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3080 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3082 /* Update DB root pointers */
3083 if (txn->mt_numdbs > 2) {
3087 data.mv_size = sizeof(MDB_db);
3089 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3090 for (i = 2; i < txn->mt_numdbs; i++) {
3091 if (txn->mt_dbflags[i] & DB_DIRTY) {
3092 data.mv_data = &txn->mt_dbs[i];
3093 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3100 rc = mdb_freelist_save(txn);
3104 mdb_midl_free(env->me_pghead);
3105 env->me_pghead = NULL;
3106 if (mdb_midl_shrink(&txn->mt_free_pgs))
3107 env->me_free_pgs = txn->mt_free_pgs;
3113 if ((rc = mdb_page_flush(txn, 0)) ||
3114 (rc = mdb_env_sync(env, 0)) ||
3115 (rc = mdb_env_write_meta(txn)))
3121 mdb_dbis_update(txn, 1);
3124 UNLOCK_MUTEX_W(env);
3134 /** Read the environment parameters of a DB environment before
3135 * mapping it into memory.
3136 * @param[in] env the environment handle
3137 * @param[out] meta address of where to store the meta information
3138 * @return 0 on success, non-zero on failure.
3141 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3147 enum { Size = sizeof(pbuf) };
3149 /* We don't know the page size yet, so use a minimum value.
3150 * Read both meta pages so we can use the latest one.
3153 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3157 memset(&ov, 0, sizeof(ov));
3159 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3160 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3163 rc = pread(env->me_fd, &pbuf, Size, off);
3166 if (rc == 0 && off == 0)
3168 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3169 DPRINTF(("read: %s", mdb_strerror(rc)));
3173 p = (MDB_page *)&pbuf;
3175 if (!F_ISSET(p->mp_flags, P_META)) {
3176 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3181 if (m->mm_magic != MDB_MAGIC) {
3182 DPUTS("meta has invalid magic");
3186 if (m->mm_version != MDB_DATA_VERSION) {
3187 DPRINTF(("database is version %u, expected version %u",
3188 m->mm_version, MDB_DATA_VERSION));
3189 return MDB_VERSION_MISMATCH;
3192 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3198 /** Write the environment parameters of a freshly created DB environment.
3199 * @param[in] env the environment handle
3200 * @param[out] meta address of where to store the meta information
3201 * @return 0 on success, non-zero on failure.
3204 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3212 memset(&ov, 0, sizeof(ov));
3213 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3215 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3218 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3219 len = pwrite(fd, ptr, size, pos); \
3220 rc = (len >= 0); } while(0)
3223 DPUTS("writing new meta page");
3225 psize = env->me_psize;
3227 meta->mm_magic = MDB_MAGIC;
3228 meta->mm_version = MDB_DATA_VERSION;
3229 meta->mm_mapsize = env->me_mapsize;
3230 meta->mm_psize = psize;
3231 meta->mm_last_pg = 1;
3232 meta->mm_flags = env->me_flags & 0xffff;
3233 meta->mm_flags |= MDB_INTEGERKEY;
3234 meta->mm_dbs[0].md_root = P_INVALID;
3235 meta->mm_dbs[1].md_root = P_INVALID;
3237 p = calloc(2, psize);
3239 p->mp_flags = P_META;
3240 *(MDB_meta *)METADATA(p) = *meta;
3242 q = (MDB_page *)((char *)p + psize);
3244 q->mp_flags = P_META;
3245 *(MDB_meta *)METADATA(q) = *meta;
3247 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3250 else if ((unsigned) len == psize * 2)
3258 /** Update the environment info to commit a transaction.
3259 * @param[in] txn the transaction that's being committed
3260 * @return 0 on success, non-zero on failure.
3263 mdb_env_write_meta(MDB_txn *txn)
3266 MDB_meta meta, metab, *mp;
3268 int rc, len, toggle;
3277 toggle = txn->mt_txnid & 1;
3278 DPRINTF(("writing meta page %d for root page %"Z"u",
3279 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3282 mp = env->me_metas[toggle];
3284 if (env->me_flags & MDB_WRITEMAP) {
3285 /* Persist any increases of mapsize config */
3286 if (env->me_mapsize > mp->mm_mapsize)
3287 mp->mm_mapsize = env->me_mapsize;
3288 mp->mm_dbs[0] = txn->mt_dbs[0];
3289 mp->mm_dbs[1] = txn->mt_dbs[1];
3290 mp->mm_last_pg = txn->mt_next_pgno - 1;
3291 mp->mm_txnid = txn->mt_txnid;
3292 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3293 unsigned meta_size = env->me_psize;
3294 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3297 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3298 if (meta_size < env->me_os_psize)
3299 meta_size += meta_size;
3304 if (MDB_MSYNC(ptr, meta_size, rc)) {
3311 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3312 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3314 ptr = (char *)&meta;
3315 if (env->me_mapsize > mp->mm_mapsize) {
3316 /* Persist any increases of mapsize config */
3317 meta.mm_mapsize = env->me_mapsize;
3318 off = offsetof(MDB_meta, mm_mapsize);
3320 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3322 len = sizeof(MDB_meta) - off;
3325 meta.mm_dbs[0] = txn->mt_dbs[0];
3326 meta.mm_dbs[1] = txn->mt_dbs[1];
3327 meta.mm_last_pg = txn->mt_next_pgno - 1;
3328 meta.mm_txnid = txn->mt_txnid;
3331 off += env->me_psize;
3334 /* Write to the SYNC fd */
3335 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3336 env->me_fd : env->me_mfd;
3339 memset(&ov, 0, sizeof(ov));
3341 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3345 rc = pwrite(mfd, ptr, len, off);
3348 rc = rc < 0 ? ErrCode() : EIO;
3349 DPUTS("write failed, disk error?");
3350 /* On a failure, the pagecache still contains the new data.
3351 * Write some old data back, to prevent it from being used.
3352 * Use the non-SYNC fd; we know it will fail anyway.
3354 meta.mm_last_pg = metab.mm_last_pg;
3355 meta.mm_txnid = metab.mm_txnid;
3357 memset(&ov, 0, sizeof(ov));
3359 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3361 r2 = pwrite(env->me_fd, ptr, len, off);
3362 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3365 env->me_flags |= MDB_FATAL_ERROR;
3369 /* Memory ordering issues are irrelevant; since the entire writer
3370 * is wrapped by wmutex, all of these changes will become visible
3371 * after the wmutex is unlocked. Since the DB is multi-version,
3372 * readers will get consistent data regardless of how fresh or
3373 * how stale their view of these values is.
3376 env->me_txns->mti_txnid = txn->mt_txnid;
3381 /** Check both meta pages to see which one is newer.
3382 * @param[in] env the environment handle
3383 * @return meta toggle (0 or 1).
3386 mdb_env_pick_meta(const MDB_env *env)
3388 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3392 mdb_env_create(MDB_env **env)
3396 e = calloc(1, sizeof(MDB_env));
3400 e->me_maxreaders = DEFAULT_READERS;
3401 e->me_maxdbs = e->me_numdbs = 2;
3402 e->me_fd = INVALID_HANDLE_VALUE;
3403 e->me_lfd = INVALID_HANDLE_VALUE;
3404 e->me_mfd = INVALID_HANDLE_VALUE;
3405 #ifdef MDB_USE_POSIX_SEM
3406 e->me_rmutex = SEM_FAILED;
3407 e->me_wmutex = SEM_FAILED;
3409 e->me_pid = getpid();
3410 GET_PAGESIZE(e->me_os_psize);
3411 VGMEMP_CREATE(e,0,0);
3417 mdb_env_map(MDB_env *env, void *addr, int newsize)
3420 unsigned int flags = env->me_flags;
3424 LONG sizelo, sizehi;
3425 sizelo = env->me_mapsize & 0xffffffff;
3426 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3428 /* Windows won't create mappings for zero length files.
3429 * Just allocate the maxsize right now.
3432 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3433 || !SetEndOfFile(env->me_fd)
3434 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3437 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3438 PAGE_READWRITE : PAGE_READONLY,
3439 sizehi, sizelo, NULL);
3442 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3443 FILE_MAP_WRITE : FILE_MAP_READ,
3444 0, 0, env->me_mapsize, addr);
3445 rc = env->me_map ? 0 : ErrCode();
3450 int prot = PROT_READ;
3451 if (flags & MDB_WRITEMAP) {
3453 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3456 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3458 if (env->me_map == MAP_FAILED) {
3463 if (flags & MDB_NORDAHEAD) {
3464 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3466 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3468 #ifdef POSIX_MADV_RANDOM
3469 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3470 #endif /* POSIX_MADV_RANDOM */
3471 #endif /* MADV_RANDOM */
3475 /* Can happen because the address argument to mmap() is just a
3476 * hint. mmap() can pick another, e.g. if the range is in use.
3477 * The MAP_FIXED flag would prevent that, but then mmap could
3478 * instead unmap existing pages to make room for the new map.
3480 if (addr && env->me_map != addr)
3481 return EBUSY; /* TODO: Make a new MDB_* error code? */
3483 p = (MDB_page *)env->me_map;
3484 env->me_metas[0] = METADATA(p);
3485 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3491 mdb_env_set_mapsize(MDB_env *env, size_t size)
3493 /* If env is already open, caller is responsible for making
3494 * sure there are no active txns.
3502 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3503 else if (size < env->me_mapsize) {
3504 /* If the configured size is smaller, make sure it's
3505 * still big enough. Silently round up to minimum if not.
3507 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3511 munmap(env->me_map, env->me_mapsize);
3512 env->me_mapsize = size;
3513 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3514 rc = mdb_env_map(env, old, 1);
3518 env->me_mapsize = size;
3520 env->me_maxpg = env->me_mapsize / env->me_psize;
3525 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3529 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3534 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3536 if (env->me_map || readers < 1)
3538 env->me_maxreaders = readers;
3543 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3545 if (!env || !readers)
3547 *readers = env->me_maxreaders;
3551 /** Further setup required for opening an MDB environment
3554 mdb_env_open2(MDB_env *env)
3556 unsigned int flags = env->me_flags;
3557 int i, newenv = 0, rc;
3561 /* See if we should use QueryLimited */
3563 if ((rc & 0xff) > 5)
3564 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3566 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3569 memset(&meta, 0, sizeof(meta));
3571 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3574 DPUTS("new mdbenv");
3576 env->me_psize = env->me_os_psize;
3577 if (env->me_psize > MAX_PAGESIZE)
3578 env->me_psize = MAX_PAGESIZE;
3580 env->me_psize = meta.mm_psize;
3583 /* Was a mapsize configured? */
3584 if (!env->me_mapsize) {
3585 /* If this is a new environment, take the default,
3586 * else use the size recorded in the existing env.
3588 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3589 } else if (env->me_mapsize < meta.mm_mapsize) {
3590 /* If the configured size is smaller, make sure it's
3591 * still big enough. Silently round up to minimum if not.
3593 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3594 if (env->me_mapsize < minsize)
3595 env->me_mapsize = minsize;
3598 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3603 if (flags & MDB_FIXEDMAP)
3604 meta.mm_address = env->me_map;
3605 i = mdb_env_init_meta(env, &meta);
3606 if (i != MDB_SUCCESS) {
3611 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3612 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3614 #if !(MDB_MAXKEYSIZE)
3615 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3617 env->me_maxpg = env->me_mapsize / env->me_psize;
3621 int toggle = mdb_env_pick_meta(env);
3622 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3624 DPRINTF(("opened database version %u, pagesize %u",
3625 env->me_metas[0]->mm_version, env->me_psize));
3626 DPRINTF(("using meta page %d", toggle));
3627 DPRINTF(("depth: %u", db->md_depth));
3628 DPRINTF(("entries: %"Z"u", db->md_entries));
3629 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3630 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3631 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3632 DPRINTF(("root: %"Z"u", db->md_root));
3640 /** Release a reader thread's slot in the reader lock table.
3641 * This function is called automatically when a thread exits.
3642 * @param[in] ptr This points to the slot in the reader lock table.
3645 mdb_env_reader_dest(void *ptr)
3647 MDB_reader *reader = ptr;
3653 /** Junk for arranging thread-specific callbacks on Windows. This is
3654 * necessarily platform and compiler-specific. Windows supports up
3655 * to 1088 keys. Let's assume nobody opens more than 64 environments
3656 * in a single process, for now. They can override this if needed.
3658 #ifndef MAX_TLS_KEYS
3659 #define MAX_TLS_KEYS 64
3661 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3662 static int mdb_tls_nkeys;
3664 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3668 case DLL_PROCESS_ATTACH: break;
3669 case DLL_THREAD_ATTACH: break;
3670 case DLL_THREAD_DETACH:
3671 for (i=0; i<mdb_tls_nkeys; i++) {
3672 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3673 mdb_env_reader_dest(r);
3676 case DLL_PROCESS_DETACH: break;
3681 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3683 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3687 /* Force some symbol references.
3688 * _tls_used forces the linker to create the TLS directory if not already done
3689 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3691 #pragma comment(linker, "/INCLUDE:_tls_used")
3692 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3693 #pragma const_seg(".CRT$XLB")
3694 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3695 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3698 #pragma comment(linker, "/INCLUDE:__tls_used")
3699 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3700 #pragma data_seg(".CRT$XLB")
3701 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3703 #endif /* WIN 32/64 */
3704 #endif /* !__GNUC__ */
3707 /** Downgrade the exclusive lock on the region back to shared */
3709 mdb_env_share_locks(MDB_env *env, int *excl)
3711 int rc = 0, toggle = mdb_env_pick_meta(env);
3713 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3718 /* First acquire a shared lock. The Unlock will
3719 * then release the existing exclusive lock.
3721 memset(&ov, 0, sizeof(ov));
3722 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3725 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3731 struct flock lock_info;
3732 /* The shared lock replaces the existing lock */
3733 memset((void *)&lock_info, 0, sizeof(lock_info));
3734 lock_info.l_type = F_RDLCK;
3735 lock_info.l_whence = SEEK_SET;
3736 lock_info.l_start = 0;
3737 lock_info.l_len = 1;
3738 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3739 (rc = ErrCode()) == EINTR) ;
3740 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3747 /** Try to get exlusive lock, otherwise shared.
3748 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3751 mdb_env_excl_lock(MDB_env *env, int *excl)
3755 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3759 memset(&ov, 0, sizeof(ov));
3760 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3767 struct flock lock_info;
3768 memset((void *)&lock_info, 0, sizeof(lock_info));
3769 lock_info.l_type = F_WRLCK;
3770 lock_info.l_whence = SEEK_SET;
3771 lock_info.l_start = 0;
3772 lock_info.l_len = 1;
3773 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3774 (rc = ErrCode()) == EINTR) ;
3778 # ifdef MDB_USE_POSIX_SEM
3779 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3782 lock_info.l_type = F_RDLCK;
3783 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3784 (rc = ErrCode()) == EINTR) ;
3794 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3796 * @(#) $Revision: 5.1 $
3797 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3798 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3800 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3804 * Please do not copyright this code. This code is in the public domain.
3806 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3807 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3808 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3809 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3810 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3811 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3812 * PERFORMANCE OF THIS SOFTWARE.
3815 * chongo <Landon Curt Noll> /\oo/\
3816 * http://www.isthe.com/chongo/
3818 * Share and Enjoy! :-)
3821 typedef unsigned long long mdb_hash_t;
3822 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3824 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3825 * @param[in] val value to hash
3826 * @param[in] hval initial value for hash
3827 * @return 64 bit hash
3829 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3830 * hval arg on the first call.
3833 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3835 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3836 unsigned char *end = s + val->mv_size;
3838 * FNV-1a hash each octet of the string
3841 /* xor the bottom with the current octet */
3842 hval ^= (mdb_hash_t)*s++;
3844 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3845 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3846 (hval << 7) + (hval << 8) + (hval << 40);
3848 /* return our new hash value */
3852 /** Hash the string and output the encoded hash.
3853 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3854 * very short name limits. We don't care about the encoding being reversible,
3855 * we just want to preserve as many bits of the input as possible in a
3856 * small printable string.
3857 * @param[in] str string to hash
3858 * @param[out] encbuf an array of 11 chars to hold the hash
3860 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3863 mdb_pack85(unsigned long l, char *out)
3867 for (i=0; i<5; i++) {
3868 *out++ = mdb_a85[l % 85];
3874 mdb_hash_enc(MDB_val *val, char *encbuf)
3876 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3878 mdb_pack85(h, encbuf);
3879 mdb_pack85(h>>32, encbuf+5);
3884 /** Open and/or initialize the lock region for the environment.
3885 * @param[in] env The MDB environment.
3886 * @param[in] lpath The pathname of the file used for the lock region.
3887 * @param[in] mode The Unix permissions for the file, if we create it.
3888 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3889 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3890 * @return 0 on success, non-zero on failure.
3893 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3896 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3898 # define MDB_ERRCODE_ROFS EROFS
3899 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3900 # define MDB_CLOEXEC O_CLOEXEC
3903 # define MDB_CLOEXEC 0
3910 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3911 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3912 FILE_ATTRIBUTE_NORMAL, NULL);
3914 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3916 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3918 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3923 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3924 /* Lose record locks when exec*() */
3925 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3926 fcntl(env->me_lfd, F_SETFD, fdflags);
3929 if (!(env->me_flags & MDB_NOTLS)) {
3930 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3933 env->me_flags |= MDB_ENV_TXKEY;
3935 /* Windows TLS callbacks need help finding their TLS info. */
3936 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3940 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3944 /* Try to get exclusive lock. If we succeed, then
3945 * nobody is using the lock region and we should initialize it.
3947 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3950 size = GetFileSize(env->me_lfd, NULL);
3952 size = lseek(env->me_lfd, 0, SEEK_END);
3953 if (size == -1) goto fail_errno;
3955 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3956 if (size < rsize && *excl > 0) {
3958 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3959 || !SetEndOfFile(env->me_lfd))
3962 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3966 size = rsize - sizeof(MDB_txninfo);
3967 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3972 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3974 if (!mh) goto fail_errno;
3975 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3977 if (!env->me_txns) goto fail_errno;
3979 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3981 if (m == MAP_FAILED) goto fail_errno;
3987 BY_HANDLE_FILE_INFORMATION stbuf;
3996 if (!mdb_sec_inited) {
3997 InitializeSecurityDescriptor(&mdb_null_sd,
3998 SECURITY_DESCRIPTOR_REVISION);
3999 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4000 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4001 mdb_all_sa.bInheritHandle = FALSE;
4002 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4005 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4006 idbuf.volume = stbuf.dwVolumeSerialNumber;
4007 idbuf.nhigh = stbuf.nFileIndexHigh;
4008 idbuf.nlow = stbuf.nFileIndexLow;
4009 val.mv_data = &idbuf;
4010 val.mv_size = sizeof(idbuf);
4011 mdb_hash_enc(&val, encbuf);
4012 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4013 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4014 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4015 if (!env->me_rmutex) goto fail_errno;
4016 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4017 if (!env->me_wmutex) goto fail_errno;
4018 #elif defined(MDB_USE_POSIX_SEM)
4027 #if defined(__NetBSD__)
4028 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4030 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4031 idbuf.dev = stbuf.st_dev;
4032 idbuf.ino = stbuf.st_ino;
4033 val.mv_data = &idbuf;
4034 val.mv_size = sizeof(idbuf);
4035 mdb_hash_enc(&val, encbuf);
4036 #ifdef MDB_SHORT_SEMNAMES
4037 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4039 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4040 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4041 /* Clean up after a previous run, if needed: Try to
4042 * remove both semaphores before doing anything else.
4044 sem_unlink(env->me_txns->mti_rmname);
4045 sem_unlink(env->me_txns->mti_wmname);
4046 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4047 O_CREAT|O_EXCL, mode, 1);
4048 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4049 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4050 O_CREAT|O_EXCL, mode, 1);
4051 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4052 #else /* MDB_USE_POSIX_SEM */
4053 pthread_mutexattr_t mattr;
4055 if ((rc = pthread_mutexattr_init(&mattr))
4056 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4057 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4058 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4060 pthread_mutexattr_destroy(&mattr);
4061 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4063 env->me_txns->mti_magic = MDB_MAGIC;
4064 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4065 env->me_txns->mti_txnid = 0;
4066 env->me_txns->mti_numreaders = 0;
4069 if (env->me_txns->mti_magic != MDB_MAGIC) {
4070 DPUTS("lock region has invalid magic");
4074 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4075 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4076 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4077 rc = MDB_VERSION_MISMATCH;
4081 if (rc && rc != EACCES && rc != EAGAIN) {
4085 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4086 if (!env->me_rmutex) goto fail_errno;
4087 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4088 if (!env->me_wmutex) goto fail_errno;
4089 #elif defined(MDB_USE_POSIX_SEM)
4090 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4091 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4092 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4093 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4104 /** The name of the lock file in the DB environment */
4105 #define LOCKNAME "/lock.mdb"
4106 /** The name of the data file in the DB environment */
4107 #define DATANAME "/data.mdb"
4108 /** The suffix of the lock file when no subdir is used */
4109 #define LOCKSUFF "-lock"
4110 /** Only a subset of the @ref mdb_env flags can be changed
4111 * at runtime. Changing other flags requires closing the
4112 * environment and re-opening it with the new flags.
4114 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4115 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4116 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4118 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4119 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4123 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4125 int oflags, rc, len, excl = -1;
4126 char *lpath, *dpath;
4128 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4132 if (flags & MDB_NOSUBDIR) {
4133 rc = len + sizeof(LOCKSUFF) + len + 1;
4135 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4140 if (flags & MDB_NOSUBDIR) {
4141 dpath = lpath + len + sizeof(LOCKSUFF);
4142 sprintf(lpath, "%s" LOCKSUFF, path);
4143 strcpy(dpath, path);
4145 dpath = lpath + len + sizeof(LOCKNAME);
4146 sprintf(lpath, "%s" LOCKNAME, path);
4147 sprintf(dpath, "%s" DATANAME, path);
4151 flags |= env->me_flags;
4152 if (flags & MDB_RDONLY) {
4153 /* silently ignore WRITEMAP when we're only getting read access */
4154 flags &= ~MDB_WRITEMAP;
4156 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4157 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4160 env->me_flags = flags |= MDB_ENV_ACTIVE;
4164 env->me_path = strdup(path);
4165 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4166 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4167 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4172 /* For RDONLY, get lockfile after we know datafile exists */
4173 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4174 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4180 if (F_ISSET(flags, MDB_RDONLY)) {
4181 oflags = GENERIC_READ;
4182 len = OPEN_EXISTING;
4184 oflags = GENERIC_READ|GENERIC_WRITE;
4187 mode = FILE_ATTRIBUTE_NORMAL;
4188 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4189 NULL, len, mode, NULL);
4191 if (F_ISSET(flags, MDB_RDONLY))
4194 oflags = O_RDWR | O_CREAT;
4196 env->me_fd = open(dpath, oflags, mode);
4198 if (env->me_fd == INVALID_HANDLE_VALUE) {
4203 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4204 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4209 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4210 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4211 env->me_mfd = env->me_fd;
4213 /* Synchronous fd for meta writes. Needed even with
4214 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4217 len = OPEN_EXISTING;
4218 env->me_mfd = CreateFile(dpath, oflags,
4219 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4220 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4223 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4225 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4230 DPRINTF(("opened dbenv %p", (void *) env));
4232 rc = mdb_env_share_locks(env, &excl);
4236 if (!((flags & MDB_RDONLY) ||
4237 (env->me_pbuf = calloc(1, env->me_psize))))
4243 mdb_env_close0(env, excl);
4249 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4251 mdb_env_close0(MDB_env *env, int excl)
4255 if (!(env->me_flags & MDB_ENV_ACTIVE))
4258 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4259 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4260 free(env->me_dbxs[i].md_name.mv_data);
4263 free(env->me_dbflags);
4266 free(env->me_dirty_list);
4267 mdb_midl_free(env->me_free_pgs);
4269 if (env->me_flags & MDB_ENV_TXKEY) {
4270 pthread_key_delete(env->me_txkey);
4272 /* Delete our key from the global list */
4273 for (i=0; i<mdb_tls_nkeys; i++)
4274 if (mdb_tls_keys[i] == env->me_txkey) {
4275 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4283 munmap(env->me_map, env->me_mapsize);
4285 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4286 (void) close(env->me_mfd);
4287 if (env->me_fd != INVALID_HANDLE_VALUE)
4288 (void) close(env->me_fd);
4290 MDB_PID_T pid = env->me_pid;
4291 /* Clearing readers is done in this function because
4292 * me_txkey with its destructor must be disabled first.
4294 for (i = env->me_numreaders; --i >= 0; )
4295 if (env->me_txns->mti_readers[i].mr_pid == pid)
4296 env->me_txns->mti_readers[i].mr_pid = 0;
4298 if (env->me_rmutex) {
4299 CloseHandle(env->me_rmutex);
4300 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4302 /* Windows automatically destroys the mutexes when
4303 * the last handle closes.
4305 #elif defined(MDB_USE_POSIX_SEM)
4306 if (env->me_rmutex != SEM_FAILED) {
4307 sem_close(env->me_rmutex);
4308 if (env->me_wmutex != SEM_FAILED)
4309 sem_close(env->me_wmutex);
4310 /* If we have the filelock: If we are the
4311 * only remaining user, clean up semaphores.
4314 mdb_env_excl_lock(env, &excl);
4316 sem_unlink(env->me_txns->mti_rmname);
4317 sem_unlink(env->me_txns->mti_wmname);
4321 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4323 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4326 /* Unlock the lockfile. Windows would have unlocked it
4327 * after closing anyway, but not necessarily at once.
4329 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4332 (void) close(env->me_lfd);
4335 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4339 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4341 MDB_txn *txn = NULL;
4347 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4351 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4354 /* Do the lock/unlock of the reader mutex before starting the
4355 * write txn. Otherwise other read txns could block writers.
4357 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4362 /* We must start the actual read txn after blocking writers */
4363 mdb_txn_reset0(txn, "reset-stage1");
4365 /* Temporarily block writers until we snapshot the meta pages */
4368 rc = mdb_txn_renew0(txn);
4370 UNLOCK_MUTEX_W(env);
4375 wsize = env->me_psize * 2;
4379 DO_WRITE(rc, fd, ptr, w2, len);
4383 } else if (len > 0) {
4389 /* Non-blocking or async handles are not supported */
4395 UNLOCK_MUTEX_W(env);
4400 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4402 if (wsize > MAX_WRITE)
4406 DO_WRITE(rc, fd, ptr, w2, len);
4410 } else if (len > 0) {
4427 mdb_env_copy(MDB_env *env, const char *path)
4431 HANDLE newfd = INVALID_HANDLE_VALUE;
4433 if (env->me_flags & MDB_NOSUBDIR) {
4434 lpath = (char *)path;
4437 len += sizeof(DATANAME);
4438 lpath = malloc(len);
4441 sprintf(lpath, "%s" DATANAME, path);
4444 /* The destination path must exist, but the destination file must not.
4445 * We don't want the OS to cache the writes, since the source data is
4446 * already in the OS cache.
4449 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4450 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4452 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4454 if (newfd == INVALID_HANDLE_VALUE) {
4460 /* Set O_DIRECT if the file system supports it */
4461 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4462 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4464 #ifdef F_NOCACHE /* __APPLE__ */
4465 rc = fcntl(newfd, F_NOCACHE, 1);
4472 rc = mdb_env_copyfd(env, newfd);
4475 if (!(env->me_flags & MDB_NOSUBDIR))
4477 if (newfd != INVALID_HANDLE_VALUE)
4478 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4485 mdb_env_close(MDB_env *env)
4492 VGMEMP_DESTROY(env);
4493 while ((dp = env->me_dpages) != NULL) {
4494 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4495 env->me_dpages = dp->mp_next;
4499 mdb_env_close0(env, 0);
4503 /** Compare two items pointing at aligned size_t's */
4505 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4507 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4508 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4511 /** Compare two items pointing at aligned unsigned int's */
4513 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4515 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4516 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4519 /** Compare two items pointing at unsigned ints of unknown alignment.
4520 * Nodes and keys are guaranteed to be 2-byte aligned.
4523 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4525 #if BYTE_ORDER == LITTLE_ENDIAN
4526 unsigned short *u, *c;
4529 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4530 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4533 } while(!x && u > (unsigned short *)a->mv_data);
4536 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4540 /** Compare two items lexically */
4542 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4549 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4555 diff = memcmp(a->mv_data, b->mv_data, len);
4556 return diff ? diff : len_diff<0 ? -1 : len_diff;
4559 /** Compare two items in reverse byte order */
4561 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4563 const unsigned char *p1, *p2, *p1_lim;
4567 p1_lim = (const unsigned char *)a->mv_data;
4568 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4569 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4571 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4577 while (p1 > p1_lim) {
4578 diff = *--p1 - *--p2;
4582 return len_diff<0 ? -1 : len_diff;
4585 /** Search for key within a page, using binary search.
4586 * Returns the smallest entry larger or equal to the key.
4587 * If exactp is non-null, stores whether the found entry was an exact match
4588 * in *exactp (1 or 0).
4589 * Updates the cursor index with the index of the found entry.
4590 * If no entry larger or equal to the key is found, returns NULL.
4593 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4595 unsigned int i = 0, nkeys;
4598 MDB_page *mp = mc->mc_pg[mc->mc_top];
4599 MDB_node *node = NULL;
4604 nkeys = NUMKEYS(mp);
4606 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4607 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4610 low = IS_LEAF(mp) ? 0 : 1;
4612 cmp = mc->mc_dbx->md_cmp;
4614 /* Branch pages have no data, so if using integer keys,
4615 * alignment is guaranteed. Use faster mdb_cmp_int.
4617 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4618 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4625 nodekey.mv_size = mc->mc_db->md_pad;
4626 node = NODEPTR(mp, 0); /* fake */
4627 while (low <= high) {
4628 i = (low + high) >> 1;
4629 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4630 rc = cmp(key, &nodekey);
4631 DPRINTF(("found leaf index %u [%s], rc = %i",
4632 i, DKEY(&nodekey), rc));
4641 while (low <= high) {
4642 i = (low + high) >> 1;
4644 node = NODEPTR(mp, i);
4645 nodekey.mv_size = NODEKSZ(node);
4646 nodekey.mv_data = NODEKEY(node);
4648 rc = cmp(key, &nodekey);
4651 DPRINTF(("found leaf index %u [%s], rc = %i",
4652 i, DKEY(&nodekey), rc));
4654 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4655 i, DKEY(&nodekey), NODEPGNO(node), rc));
4666 if (rc > 0) { /* Found entry is less than the key. */
4667 i++; /* Skip to get the smallest entry larger than key. */
4669 node = NODEPTR(mp, i);
4672 *exactp = (rc == 0 && nkeys > 0);
4673 /* store the key index */
4674 mc->mc_ki[mc->mc_top] = i;
4676 /* There is no entry larger or equal to the key. */
4679 /* nodeptr is fake for LEAF2 */
4685 mdb_cursor_adjust(MDB_cursor *mc, func)
4689 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4690 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4697 /** Pop a page off the top of the cursor's stack. */
4699 mdb_cursor_pop(MDB_cursor *mc)
4703 MDB_page *top = mc->mc_pg[mc->mc_top];
4709 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4710 DDBI(mc), (void *) mc));
4714 /** Push a page onto the top of the cursor's stack. */
4716 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4718 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4719 DDBI(mc), (void *) mc));
4721 if (mc->mc_snum >= CURSOR_STACK) {
4722 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4723 return MDB_CURSOR_FULL;
4726 mc->mc_top = mc->mc_snum++;
4727 mc->mc_pg[mc->mc_top] = mp;
4728 mc->mc_ki[mc->mc_top] = 0;
4733 /** Find the address of the page corresponding to a given page number.
4734 * @param[in] txn the transaction for this access.
4735 * @param[in] pgno the page number for the page to retrieve.
4736 * @param[out] ret address of a pointer where the page's address will be stored.
4737 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4738 * @return 0 on success, non-zero on failure.
4741 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4743 MDB_env *env = txn->mt_env;
4747 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4751 MDB_ID2L dl = tx2->mt_u.dirty_list;
4753 /* Spilled pages were dirtied in this txn and flushed
4754 * because the dirty list got full. Bring this page
4755 * back in from the map (but don't unspill it here,
4756 * leave that unless page_touch happens again).
4758 if (tx2->mt_spill_pgs) {
4759 MDB_ID pn = pgno << 1;
4760 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4761 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4762 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4767 unsigned x = mdb_mid2l_search(dl, pgno);
4768 if (x <= dl[0].mid && dl[x].mid == pgno) {
4774 } while ((tx2 = tx2->mt_parent) != NULL);
4777 if (pgno < txn->mt_next_pgno) {
4779 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4781 DPRINTF(("page %"Z"u not found", pgno));
4782 txn->mt_flags |= MDB_TXN_ERROR;
4783 return MDB_PAGE_NOTFOUND;
4793 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4794 * The cursor is at the root page, set up the rest of it.
4797 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4799 MDB_page *mp = mc->mc_pg[mc->mc_top];
4803 while (IS_BRANCH(mp)) {
4807 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4808 mdb_cassert(mc, NUMKEYS(mp) > 1);
4809 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4811 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4813 if (flags & MDB_PS_LAST)
4814 i = NUMKEYS(mp) - 1;
4817 node = mdb_node_search(mc, key, &exact);
4819 i = NUMKEYS(mp) - 1;
4821 i = mc->mc_ki[mc->mc_top];
4823 mdb_cassert(mc, i > 0);
4827 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4830 mdb_cassert(mc, i < NUMKEYS(mp));
4831 node = NODEPTR(mp, i);
4833 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4836 mc->mc_ki[mc->mc_top] = i;
4837 if ((rc = mdb_cursor_push(mc, mp)))
4840 if (flags & MDB_PS_MODIFY) {
4841 if ((rc = mdb_page_touch(mc)) != 0)
4843 mp = mc->mc_pg[mc->mc_top];
4848 DPRINTF(("internal error, index points to a %02X page!?",
4850 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4851 return MDB_CORRUPTED;
4854 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4855 key ? DKEY(key) : "null"));
4856 mc->mc_flags |= C_INITIALIZED;
4857 mc->mc_flags &= ~C_EOF;
4862 /** Search for the lowest key under the current branch page.
4863 * This just bypasses a NUMKEYS check in the current page
4864 * before calling mdb_page_search_root(), because the callers
4865 * are all in situations where the current page is known to
4869 mdb_page_search_lowest(MDB_cursor *mc)
4871 MDB_page *mp = mc->mc_pg[mc->mc_top];
4872 MDB_node *node = NODEPTR(mp, 0);
4875 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4878 mc->mc_ki[mc->mc_top] = 0;
4879 if ((rc = mdb_cursor_push(mc, mp)))
4881 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4884 /** Search for the page a given key should be in.
4885 * Push it and its parent pages on the cursor stack.
4886 * @param[in,out] mc the cursor for this operation.
4887 * @param[in] key the key to search for, or NULL for first/last page.
4888 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4889 * are touched (updated with new page numbers).
4890 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4891 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4892 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4893 * @return 0 on success, non-zero on failure.
4896 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4901 /* Make sure the txn is still viable, then find the root from
4902 * the txn's db table and set it as the root of the cursor's stack.
4904 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4905 DPUTS("transaction has failed, must abort");
4908 /* Make sure we're using an up-to-date root */
4909 if (*mc->mc_dbflag & DB_STALE) {
4911 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4912 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4919 MDB_node *leaf = mdb_node_search(&mc2,
4920 &mc->mc_dbx->md_name, &exact);
4922 return MDB_NOTFOUND;
4923 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4926 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4928 /* The txn may not know this DBI, or another process may
4929 * have dropped and recreated the DB with other flags.
4931 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4932 return MDB_INCOMPATIBLE;
4933 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4935 *mc->mc_dbflag &= ~DB_STALE;
4937 root = mc->mc_db->md_root;
4939 if (root == P_INVALID) { /* Tree is empty. */
4940 DPUTS("tree is empty");
4941 return MDB_NOTFOUND;
4945 mdb_cassert(mc, root > 1);
4946 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4947 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4953 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4954 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4956 if (flags & MDB_PS_MODIFY) {
4957 if ((rc = mdb_page_touch(mc)))
4961 if (flags & MDB_PS_ROOTONLY)
4964 return mdb_page_search_root(mc, key, flags);
4968 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4970 MDB_txn *txn = mc->mc_txn;
4971 pgno_t pg = mp->mp_pgno;
4972 unsigned x = 0, ovpages = mp->mp_pages;
4973 MDB_env *env = txn->mt_env;
4974 MDB_IDL sl = txn->mt_spill_pgs;
4975 MDB_ID pn = pg << 1;
4978 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4979 /* If the page is dirty or on the spill list we just acquired it,
4980 * so we should give it back to our current free list, if any.
4981 * Otherwise put it onto the list of pages we freed in this txn.
4983 * Won't create me_pghead: me_pglast must be inited along with it.
4984 * Unsupported in nested txns: They would need to hide the page
4985 * range in ancestor txns' dirty and spilled lists.
4987 if (env->me_pghead &&
4989 ((mp->mp_flags & P_DIRTY) ||
4990 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4994 MDB_ID2 *dl, ix, iy;
4995 rc = mdb_midl_need(&env->me_pghead, ovpages);
4998 if (!(mp->mp_flags & P_DIRTY)) {
4999 /* This page is no longer spilled */
5006 /* Remove from dirty list */
5007 dl = txn->mt_u.dirty_list;
5009 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5015 mdb_cassert(mc, x > 1);
5017 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5018 txn->mt_flags |= MDB_TXN_ERROR;
5019 return MDB_CORRUPTED;
5022 if (!(env->me_flags & MDB_WRITEMAP))
5023 mdb_dpage_free(env, mp);
5025 /* Insert in me_pghead */
5026 mop = env->me_pghead;
5027 j = mop[0] + ovpages;
5028 for (i = mop[0]; i && mop[i] < pg; i--)
5034 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5038 mc->mc_db->md_overflow_pages -= ovpages;
5042 /** Return the data associated with a given node.
5043 * @param[in] txn The transaction for this operation.
5044 * @param[in] leaf The node being read.
5045 * @param[out] data Updated to point to the node's data.
5046 * @return 0 on success, non-zero on failure.
5049 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5051 MDB_page *omp; /* overflow page */
5055 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5056 data->mv_size = NODEDSZ(leaf);
5057 data->mv_data = NODEDATA(leaf);
5061 /* Read overflow data.
5063 data->mv_size = NODEDSZ(leaf);
5064 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5065 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5066 DPRINTF(("read overflow page %"Z"u failed", pgno));
5069 data->mv_data = METADATA(omp);
5075 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5076 MDB_val *key, MDB_val *data)
5083 if (key == NULL || data == NULL)
5086 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5088 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5091 if (txn->mt_flags & MDB_TXN_ERROR)
5094 mdb_cursor_init(&mc, txn, dbi, &mx);
5095 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5098 /** Find a sibling for a page.
5099 * Replaces the page at the top of the cursor's stack with the
5100 * specified sibling, if one exists.
5101 * @param[in] mc The cursor for this operation.
5102 * @param[in] move_right Non-zero if the right sibling is requested,
5103 * otherwise the left sibling.
5104 * @return 0 on success, non-zero on failure.
5107 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5113 if (mc->mc_snum < 2) {
5114 return MDB_NOTFOUND; /* root has no siblings */
5118 DPRINTF(("parent page is page %"Z"u, index %u",
5119 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5121 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5122 : (mc->mc_ki[mc->mc_top] == 0)) {
5123 DPRINTF(("no more keys left, moving to %s sibling",
5124 move_right ? "right" : "left"));
5125 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5126 /* undo cursor_pop before returning */
5133 mc->mc_ki[mc->mc_top]++;
5135 mc->mc_ki[mc->mc_top]--;
5136 DPRINTF(("just moving to %s index key %u",
5137 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5139 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5141 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5142 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5143 /* mc will be inconsistent if caller does mc_snum++ as above */
5144 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5148 mdb_cursor_push(mc, mp);
5150 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5155 /** Move the cursor to the next data item. */
5157 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5163 if (mc->mc_flags & C_EOF) {
5164 return MDB_NOTFOUND;
5167 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5169 mp = mc->mc_pg[mc->mc_top];
5171 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5172 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5173 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5174 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5175 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5176 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5177 if (rc == MDB_SUCCESS)
5178 MDB_GET_KEY(leaf, key);
5183 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5184 if (op == MDB_NEXT_DUP)
5185 return MDB_NOTFOUND;
5189 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5190 mdb_dbg_pgno(mp), (void *) mc));
5191 if (mc->mc_flags & C_DEL)
5194 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5195 DPUTS("=====> move to next sibling page");
5196 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5197 mc->mc_flags |= C_EOF;
5200 mp = mc->mc_pg[mc->mc_top];
5201 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5203 mc->mc_ki[mc->mc_top]++;
5206 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5207 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5210 key->mv_size = mc->mc_db->md_pad;
5211 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5215 mdb_cassert(mc, IS_LEAF(mp));
5216 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5218 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5219 mdb_xcursor_init1(mc, leaf);
5222 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5225 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5226 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5227 if (rc != MDB_SUCCESS)
5232 MDB_GET_KEY(leaf, key);
5236 /** Move the cursor to the previous data item. */
5238 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5244 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5246 mp = mc->mc_pg[mc->mc_top];
5248 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5249 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5250 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5251 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5252 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5253 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5254 if (rc == MDB_SUCCESS)
5255 MDB_GET_KEY(leaf, key);
5259 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5260 if (op == MDB_PREV_DUP)
5261 return MDB_NOTFOUND;
5266 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5267 mdb_dbg_pgno(mp), (void *) mc));
5269 if (mc->mc_ki[mc->mc_top] == 0) {
5270 DPUTS("=====> move to prev sibling page");
5271 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5274 mp = mc->mc_pg[mc->mc_top];
5275 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5276 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5278 mc->mc_ki[mc->mc_top]--;
5280 mc->mc_flags &= ~C_EOF;
5282 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5283 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5286 key->mv_size = mc->mc_db->md_pad;
5287 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5291 mdb_cassert(mc, IS_LEAF(mp));
5292 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5294 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5295 mdb_xcursor_init1(mc, leaf);
5298 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5301 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5302 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5303 if (rc != MDB_SUCCESS)
5308 MDB_GET_KEY(leaf, key);
5312 /** Set the cursor on a specific data item. */
5314 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5315 MDB_cursor_op op, int *exactp)
5319 MDB_node *leaf = NULL;
5322 if (key->mv_size == 0)
5323 return MDB_BAD_VALSIZE;
5326 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5328 /* See if we're already on the right page */
5329 if (mc->mc_flags & C_INITIALIZED) {
5332 mp = mc->mc_pg[mc->mc_top];
5334 mc->mc_ki[mc->mc_top] = 0;
5335 return MDB_NOTFOUND;
5337 if (mp->mp_flags & P_LEAF2) {
5338 nodekey.mv_size = mc->mc_db->md_pad;
5339 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5341 leaf = NODEPTR(mp, 0);
5342 MDB_GET_KEY2(leaf, nodekey);
5344 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5346 /* Probably happens rarely, but first node on the page
5347 * was the one we wanted.
5349 mc->mc_ki[mc->mc_top] = 0;
5356 unsigned int nkeys = NUMKEYS(mp);
5358 if (mp->mp_flags & P_LEAF2) {
5359 nodekey.mv_data = LEAF2KEY(mp,
5360 nkeys-1, nodekey.mv_size);
5362 leaf = NODEPTR(mp, nkeys-1);
5363 MDB_GET_KEY2(leaf, nodekey);
5365 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5367 /* last node was the one we wanted */
5368 mc->mc_ki[mc->mc_top] = nkeys-1;
5374 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5375 /* This is definitely the right page, skip search_page */
5376 if (mp->mp_flags & P_LEAF2) {
5377 nodekey.mv_data = LEAF2KEY(mp,
5378 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5380 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5381 MDB_GET_KEY2(leaf, nodekey);
5383 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5385 /* current node was the one we wanted */
5395 /* If any parents have right-sibs, search.
5396 * Otherwise, there's nothing further.
5398 for (i=0; i<mc->mc_top; i++)
5400 NUMKEYS(mc->mc_pg[i])-1)
5402 if (i == mc->mc_top) {
5403 /* There are no other pages */
5404 mc->mc_ki[mc->mc_top] = nkeys;
5405 return MDB_NOTFOUND;
5409 /* There are no other pages */
5410 mc->mc_ki[mc->mc_top] = 0;
5411 if (op == MDB_SET_RANGE) {
5415 return MDB_NOTFOUND;
5419 rc = mdb_page_search(mc, key, 0);
5420 if (rc != MDB_SUCCESS)
5423 mp = mc->mc_pg[mc->mc_top];
5424 mdb_cassert(mc, IS_LEAF(mp));
5427 leaf = mdb_node_search(mc, key, exactp);
5428 if (exactp != NULL && !*exactp) {
5429 /* MDB_SET specified and not an exact match. */
5430 return MDB_NOTFOUND;
5434 DPUTS("===> inexact leaf not found, goto sibling");
5435 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5436 return rc; /* no entries matched */
5437 mp = mc->mc_pg[mc->mc_top];
5438 mdb_cassert(mc, IS_LEAF(mp));
5439 leaf = NODEPTR(mp, 0);
5443 mc->mc_flags |= C_INITIALIZED;
5444 mc->mc_flags &= ~C_EOF;
5447 key->mv_size = mc->mc_db->md_pad;
5448 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5452 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5453 mdb_xcursor_init1(mc, leaf);
5456 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5457 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5458 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5461 if (op == MDB_GET_BOTH) {
5467 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5468 if (rc != MDB_SUCCESS)
5471 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5473 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5475 rc = mc->mc_dbx->md_dcmp(data, &d2);
5477 if (op == MDB_GET_BOTH || rc > 0)
5478 return MDB_NOTFOUND;
5485 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5486 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5491 /* The key already matches in all other cases */
5492 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5493 MDB_GET_KEY(leaf, key);
5494 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5499 /** Move the cursor to the first item in the database. */
5501 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5507 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5509 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5510 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5511 if (rc != MDB_SUCCESS)
5514 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5516 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5517 mc->mc_flags |= C_INITIALIZED;
5518 mc->mc_flags &= ~C_EOF;
5520 mc->mc_ki[mc->mc_top] = 0;
5522 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5523 key->mv_size = mc->mc_db->md_pad;
5524 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5529 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5530 mdb_xcursor_init1(mc, leaf);
5531 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5535 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5539 MDB_GET_KEY(leaf, key);
5543 /** Move the cursor to the last item in the database. */
5545 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5551 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5553 if (!(mc->mc_flags & C_EOF)) {
5555 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5556 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5557 if (rc != MDB_SUCCESS)
5560 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5563 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5564 mc->mc_flags |= C_INITIALIZED|C_EOF;
5565 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5567 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5568 key->mv_size = mc->mc_db->md_pad;
5569 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5574 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5575 mdb_xcursor_init1(mc, leaf);
5576 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5580 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5585 MDB_GET_KEY(leaf, key);
5590 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5595 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5600 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5604 case MDB_GET_CURRENT:
5605 if (!(mc->mc_flags & C_INITIALIZED)) {
5608 MDB_page *mp = mc->mc_pg[mc->mc_top];
5609 int nkeys = NUMKEYS(mp);
5610 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5611 mc->mc_ki[mc->mc_top] = nkeys;
5617 key->mv_size = mc->mc_db->md_pad;
5618 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5620 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5621 MDB_GET_KEY(leaf, key);
5623 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5624 if (mc->mc_flags & C_DEL)
5625 mdb_xcursor_init1(mc, leaf);
5626 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5628 rc = mdb_node_read(mc->mc_txn, leaf, data);
5635 case MDB_GET_BOTH_RANGE:
5640 if (mc->mc_xcursor == NULL) {
5641 rc = MDB_INCOMPATIBLE;
5651 rc = mdb_cursor_set(mc, key, data, op,
5652 op == MDB_SET_RANGE ? NULL : &exact);
5655 case MDB_GET_MULTIPLE:
5656 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5660 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5661 rc = MDB_INCOMPATIBLE;
5665 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5666 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5669 case MDB_NEXT_MULTIPLE:
5674 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5675 rc = MDB_INCOMPATIBLE;
5678 if (!(mc->mc_flags & C_INITIALIZED))
5679 rc = mdb_cursor_first(mc, key, data);
5681 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5682 if (rc == MDB_SUCCESS) {
5683 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5686 mx = &mc->mc_xcursor->mx_cursor;
5687 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5689 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5690 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5698 case MDB_NEXT_NODUP:
5699 if (!(mc->mc_flags & C_INITIALIZED))
5700 rc = mdb_cursor_first(mc, key, data);
5702 rc = mdb_cursor_next(mc, key, data, op);
5706 case MDB_PREV_NODUP:
5707 if (!(mc->mc_flags & C_INITIALIZED)) {
5708 rc = mdb_cursor_last(mc, key, data);
5711 mc->mc_flags |= C_INITIALIZED;
5712 mc->mc_ki[mc->mc_top]++;
5714 rc = mdb_cursor_prev(mc, key, data, op);
5717 rc = mdb_cursor_first(mc, key, data);
5720 mfunc = mdb_cursor_first;
5722 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5726 if (mc->mc_xcursor == NULL) {
5727 rc = MDB_INCOMPATIBLE;
5730 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5734 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5737 rc = mdb_cursor_last(mc, key, data);
5740 mfunc = mdb_cursor_last;
5743 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5748 if (mc->mc_flags & C_DEL)
5749 mc->mc_flags ^= C_DEL;
5754 /** Touch all the pages in the cursor stack. Set mc_top.
5755 * Makes sure all the pages are writable, before attempting a write operation.
5756 * @param[in] mc The cursor to operate on.
5759 mdb_cursor_touch(MDB_cursor *mc)
5761 int rc = MDB_SUCCESS;
5763 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5766 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5767 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5770 *mc->mc_dbflag |= DB_DIRTY;
5775 rc = mdb_page_touch(mc);
5776 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5777 mc->mc_top = mc->mc_snum-1;
5782 /** Do not spill pages to disk if txn is getting full, may fail instead */
5783 #define MDB_NOSPILL 0x8000
5786 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5789 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5791 MDB_node *leaf = NULL;
5794 MDB_val xdata, *rdata, dkey, olddata;
5796 int do_sub = 0, insert;
5797 unsigned int mcount = 0, dcount = 0, nospill;
5800 unsigned int nflags;
5806 env = mc->mc_txn->mt_env;
5808 /* Check this first so counter will always be zero on any
5811 if (flags & MDB_MULTIPLE) {
5812 dcount = data[1].mv_size;
5813 data[1].mv_size = 0;
5814 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5815 return MDB_INCOMPATIBLE;
5818 nospill = flags & MDB_NOSPILL;
5819 flags &= ~MDB_NOSPILL;
5821 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5822 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5824 if (flags != MDB_CURRENT) {
5827 if (key->mv_size-1 >= ENV_MAXKEY(env))
5828 return MDB_BAD_VALSIZE;
5830 /* Ignore key except in sub-cursor, where key holds the data */
5831 if (!(mc->mc_flags & C_SUB))
5835 #if SIZE_MAX > MAXDATASIZE
5836 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5837 return MDB_BAD_VALSIZE;
5839 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5840 return MDB_BAD_VALSIZE;
5843 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5844 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5848 if (flags == MDB_CURRENT) {
5849 if (!(mc->mc_flags & C_INITIALIZED))
5852 } else if (mc->mc_db->md_root == P_INVALID) {
5853 /* new database, cursor has nothing to point to */
5856 mc->mc_flags &= ~C_INITIALIZED;
5861 if (flags & MDB_APPEND) {
5863 rc = mdb_cursor_last(mc, &k2, &d2);
5865 rc = mc->mc_dbx->md_cmp(key, &k2);
5868 mc->mc_ki[mc->mc_top]++;
5870 /* new key is <= last key */
5875 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5877 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5878 DPRINTF(("duplicate key [%s]", DKEY(key)));
5880 return MDB_KEYEXIST;
5882 if (rc && rc != MDB_NOTFOUND)
5886 if (mc->mc_flags & C_DEL)
5887 mc->mc_flags ^= C_DEL;
5889 /* Cursor is positioned, check for room in the dirty list */
5891 if (flags & MDB_MULTIPLE) {
5893 xdata.mv_size = data->mv_size * dcount;
5897 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5901 if (rc == MDB_NO_ROOT) {
5903 /* new database, write a root leaf page */
5904 DPUTS("allocating new root leaf page");
5905 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5908 mdb_cursor_push(mc, np);
5909 mc->mc_db->md_root = np->mp_pgno;
5910 mc->mc_db->md_depth++;
5911 *mc->mc_dbflag |= DB_DIRTY;
5912 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5914 np->mp_flags |= P_LEAF2;
5915 mc->mc_flags |= C_INITIALIZED;
5917 /* make sure all cursor pages are writable */
5918 rc2 = mdb_cursor_touch(mc);
5925 /* The key does not exist */
5926 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5927 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5928 LEAFSIZE(key, data) > env->me_nodemax)
5930 /* Too big for a node, insert in sub-DB */
5931 fp_flags = P_LEAF|P_DIRTY;
5933 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5934 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5938 /* there's only a key anyway, so this is a no-op */
5939 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5940 unsigned int ksize = mc->mc_db->md_pad;
5941 if (key->mv_size != ksize)
5942 return MDB_BAD_VALSIZE;
5943 if (flags == MDB_CURRENT) {
5944 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5945 memcpy(ptr, key->mv_data, ksize);
5951 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5952 olddata.mv_size = NODEDSZ(leaf);
5953 olddata.mv_data = NODEDATA(leaf);
5956 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5957 /* Prepare (sub-)page/sub-DB to accept the new item,
5958 * if needed. fp: old sub-page or a header faking
5959 * it. mp: new (sub-)page. offset: growth in page
5960 * size. xdata: node data with new page or DB.
5962 unsigned i, offset = 0;
5963 mp = fp = xdata.mv_data = env->me_pbuf;
5964 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5966 /* Was a single item before, must convert now */
5967 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5968 /* Just overwrite the current item */
5969 if (flags == MDB_CURRENT)
5972 #if UINT_MAX < SIZE_MAX
5973 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5974 #ifdef MISALIGNED_OK
5975 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5977 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5980 /* if data matches, skip it */
5981 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5982 if (flags & MDB_NODUPDATA)
5984 else if (flags & MDB_MULTIPLE)
5991 /* Back up original data item */
5992 dkey.mv_size = olddata.mv_size;
5993 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5995 /* Make sub-page header for the dup items, with dummy body */
5996 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5997 fp->mp_lower = PAGEHDRSZ;
5998 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5999 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6000 fp->mp_flags |= P_LEAF2;
6001 fp->mp_pad = data->mv_size;
6002 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6004 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6005 (dkey.mv_size & 1) + (data->mv_size & 1);
6007 fp->mp_upper = xdata.mv_size;
6008 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
6009 } else if (leaf->mn_flags & F_SUBDATA) {
6010 /* Data is on sub-DB, just store it */
6011 flags |= F_DUPDATA|F_SUBDATA;
6014 /* Data is on sub-page */
6015 fp = olddata.mv_data;
6018 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6019 offset = EVEN(NODESIZE + sizeof(indx_t) +
6023 offset = fp->mp_pad;
6024 if (SIZELEFT(fp) < offset) {
6025 offset *= 4; /* space for 4 more */
6028 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6030 fp->mp_flags |= P_DIRTY;
6031 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6032 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6036 xdata.mv_size = olddata.mv_size + offset;
6039 fp_flags = fp->mp_flags;
6040 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6041 /* Too big for a sub-page, convert to sub-DB */
6042 fp_flags &= ~P_SUBP;
6044 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6045 fp_flags |= P_LEAF2;
6046 dummy.md_pad = fp->mp_pad;
6047 dummy.md_flags = MDB_DUPFIXED;
6048 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6049 dummy.md_flags |= MDB_INTEGERKEY;
6055 dummy.md_branch_pages = 0;
6056 dummy.md_leaf_pages = 1;
6057 dummy.md_overflow_pages = 0;
6058 dummy.md_entries = NUMKEYS(fp);
6059 xdata.mv_size = sizeof(MDB_db);
6060 xdata.mv_data = &dummy;
6061 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6063 offset = env->me_psize - olddata.mv_size;
6064 flags |= F_DUPDATA|F_SUBDATA;
6065 dummy.md_root = mp->mp_pgno;
6068 mp->mp_flags = fp_flags | P_DIRTY;
6069 mp->mp_pad = fp->mp_pad;
6070 mp->mp_lower = fp->mp_lower;
6071 mp->mp_upper = fp->mp_upper + offset;
6072 if (fp_flags & P_LEAF2) {
6073 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6075 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6076 olddata.mv_size - fp->mp_upper);
6077 for (i=0; i<NUMKEYS(fp); i++)
6078 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6086 mdb_node_del(mc, 0);
6090 /* overflow page overwrites need special handling */
6091 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6094 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6096 memcpy(&pg, olddata.mv_data, sizeof(pg));
6097 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6099 ovpages = omp->mp_pages;
6101 /* Is the ov page large enough? */
6102 if (ovpages >= dpages) {
6103 if (!(omp->mp_flags & P_DIRTY) &&
6104 (level || (env->me_flags & MDB_WRITEMAP)))
6106 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6109 level = 0; /* dirty in this txn or clean */
6112 if (omp->mp_flags & P_DIRTY) {
6113 /* yes, overwrite it. Note in this case we don't
6114 * bother to try shrinking the page if the new data
6115 * is smaller than the overflow threshold.
6118 /* It is writable only in a parent txn */
6119 size_t sz = (size_t) env->me_psize * ovpages, off;
6120 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6126 rc = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6127 mdb_cassert(mc, rc == 0);
6128 if (!(flags & MDB_RESERVE)) {
6129 /* Copy end of page, adjusting alignment so
6130 * compiler may copy words instead of bytes.
6132 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6133 memcpy((size_t *)((char *)np + off),
6134 (size_t *)((char *)omp + off), sz - off);
6137 memcpy(np, omp, sz); /* Copy beginning of page */
6140 SETDSZ(leaf, data->mv_size);
6141 if (F_ISSET(flags, MDB_RESERVE))
6142 data->mv_data = METADATA(omp);
6144 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6148 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6150 } else if (data->mv_size == olddata.mv_size) {
6151 /* same size, just replace it. Note that we could
6152 * also reuse this node if the new data is smaller,
6153 * but instead we opt to shrink the node in that case.
6155 if (F_ISSET(flags, MDB_RESERVE))
6156 data->mv_data = olddata.mv_data;
6157 else if (!(mc->mc_flags & C_SUB))
6158 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6160 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6163 mdb_node_del(mc, 0);
6164 mc->mc_db->md_entries--;
6170 nflags = flags & NODE_ADD_FLAGS;
6171 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6172 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6173 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6174 nflags &= ~MDB_APPEND;
6176 nflags |= MDB_SPLIT_REPLACE;
6177 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6179 /* There is room already in this leaf page. */
6180 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6181 if (rc == 0 && !do_sub && insert) {
6182 /* Adjust other cursors pointing to mp */
6183 MDB_cursor *m2, *m3;
6184 MDB_dbi dbi = mc->mc_dbi;
6185 unsigned i = mc->mc_top;
6186 MDB_page *mp = mc->mc_pg[i];
6188 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6189 if (mc->mc_flags & C_SUB)
6190 m3 = &m2->mc_xcursor->mx_cursor;
6193 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6194 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6201 if (rc != MDB_SUCCESS)
6202 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6204 /* Now store the actual data in the child DB. Note that we're
6205 * storing the user data in the keys field, so there are strict
6206 * size limits on dupdata. The actual data fields of the child
6207 * DB are all zero size.
6214 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6215 if (flags & MDB_CURRENT) {
6216 xflags = MDB_CURRENT|MDB_NOSPILL;
6218 mdb_xcursor_init1(mc, leaf);
6219 xflags = (flags & MDB_NODUPDATA) ?
6220 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6222 /* converted, write the original data first */
6224 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6228 /* Adjust other cursors pointing to mp */
6230 unsigned i = mc->mc_top;
6231 MDB_page *mp = mc->mc_pg[i];
6233 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6234 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6235 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6236 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6237 mdb_xcursor_init1(m2, leaf);
6241 /* we've done our job */
6244 if (flags & MDB_APPENDDUP)
6245 xflags |= MDB_APPEND;
6246 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6247 if (flags & F_SUBDATA) {
6248 void *db = NODEDATA(leaf);
6249 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6252 /* sub-writes might have failed so check rc again.
6253 * Don't increment count if we just replaced an existing item.
6255 if (!rc && !(flags & MDB_CURRENT))
6256 mc->mc_db->md_entries++;
6257 if (flags & MDB_MULTIPLE) {
6261 /* let caller know how many succeeded, if any */
6262 data[1].mv_size = mcount;
6263 if (mcount < dcount) {
6264 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6271 /* If we succeeded and the key didn't exist before, make sure
6272 * the cursor is marked valid.
6275 mc->mc_flags |= C_INITIALIZED;
6280 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6286 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6287 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6289 if (!(mc->mc_flags & C_INITIALIZED))
6292 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6293 return MDB_NOTFOUND;
6295 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6298 rc = mdb_cursor_touch(mc);
6302 mp = mc->mc_pg[mc->mc_top];
6303 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6305 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6306 if (!(flags & MDB_NODUPDATA)) {
6307 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6308 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6310 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6311 /* If sub-DB still has entries, we're done */
6312 if (mc->mc_xcursor->mx_db.md_entries) {
6313 if (leaf->mn_flags & F_SUBDATA) {
6314 /* update subDB info */
6315 void *db = NODEDATA(leaf);
6316 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6319 /* shrink fake page */
6320 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6321 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6322 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6323 /* fix other sub-DB cursors pointed at this fake page */
6324 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6325 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6326 if (m2->mc_pg[mc->mc_top] == mp &&
6327 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6328 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6331 mc->mc_db->md_entries--;
6332 mc->mc_flags |= C_DEL;
6335 /* otherwise fall thru and delete the sub-DB */
6338 if (leaf->mn_flags & F_SUBDATA) {
6339 /* add all the child DB's pages to the free list */
6340 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6341 if (rc == MDB_SUCCESS) {
6342 mc->mc_db->md_entries -=
6343 mc->mc_xcursor->mx_db.md_entries;
6348 return mdb_cursor_del0(mc, leaf);
6351 /** Allocate and initialize new pages for a database.
6352 * @param[in] mc a cursor on the database being added to.
6353 * @param[in] flags flags defining what type of page is being allocated.
6354 * @param[in] num the number of pages to allocate. This is usually 1,
6355 * unless allocating overflow pages for a large record.
6356 * @param[out] mp Address of a page, or NULL on failure.
6357 * @return 0 on success, non-zero on failure.
6360 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6365 if ((rc = mdb_page_alloc(mc, num, &np)))
6367 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6368 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6369 np->mp_flags = flags | P_DIRTY;
6370 np->mp_lower = PAGEHDRSZ;
6371 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6374 mc->mc_db->md_branch_pages++;
6375 else if (IS_LEAF(np))
6376 mc->mc_db->md_leaf_pages++;
6377 else if (IS_OVERFLOW(np)) {
6378 mc->mc_db->md_overflow_pages += num;
6386 /** Calculate the size of a leaf node.
6387 * The size depends on the environment's page size; if a data item
6388 * is too large it will be put onto an overflow page and the node
6389 * size will only include the key and not the data. Sizes are always
6390 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6391 * of the #MDB_node headers.
6392 * @param[in] env The environment handle.
6393 * @param[in] key The key for the node.
6394 * @param[in] data The data for the node.
6395 * @return The number of bytes needed to store the node.
6398 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6402 sz = LEAFSIZE(key, data);
6403 if (sz > env->me_nodemax) {
6404 /* put on overflow page */
6405 sz -= data->mv_size - sizeof(pgno_t);
6408 return EVEN(sz + sizeof(indx_t));
6411 /** Calculate the size of a branch node.
6412 * The size should depend on the environment's page size but since
6413 * we currently don't support spilling large keys onto overflow
6414 * pages, it's simply the size of the #MDB_node header plus the
6415 * size of the key. Sizes are always rounded up to an even number
6416 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6417 * @param[in] env The environment handle.
6418 * @param[in] key The key for the node.
6419 * @return The number of bytes needed to store the node.
6422 mdb_branch_size(MDB_env *env, MDB_val *key)
6427 if (sz > env->me_nodemax) {
6428 /* put on overflow page */
6429 /* not implemented */
6430 /* sz -= key->size - sizeof(pgno_t); */
6433 return sz + sizeof(indx_t);
6436 /** Add a node to the page pointed to by the cursor.
6437 * @param[in] mc The cursor for this operation.
6438 * @param[in] indx The index on the page where the new node should be added.
6439 * @param[in] key The key for the new node.
6440 * @param[in] data The data for the new node, if any.
6441 * @param[in] pgno The page number, if adding a branch node.
6442 * @param[in] flags Flags for the node.
6443 * @return 0 on success, non-zero on failure. Possible errors are:
6445 * <li>ENOMEM - failed to allocate overflow pages for the node.
6446 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6447 * should never happen since all callers already calculate the
6448 * page's free space before calling this function.
6452 mdb_node_add(MDB_cursor *mc, indx_t indx,
6453 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6456 size_t node_size = NODESIZE;
6460 MDB_page *mp = mc->mc_pg[mc->mc_top];
6461 MDB_page *ofp = NULL; /* overflow page */
6464 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6466 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6467 IS_LEAF(mp) ? "leaf" : "branch",
6468 IS_SUBP(mp) ? "sub-" : "",
6469 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6470 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6473 /* Move higher keys up one slot. */
6474 int ksize = mc->mc_db->md_pad, dif;
6475 char *ptr = LEAF2KEY(mp, indx, ksize);
6476 dif = NUMKEYS(mp) - indx;
6478 memmove(ptr+ksize, ptr, dif*ksize);
6479 /* insert new key */
6480 memcpy(ptr, key->mv_data, ksize);
6482 /* Just using these for counting */
6483 mp->mp_lower += sizeof(indx_t);
6484 mp->mp_upper -= ksize - sizeof(indx_t);
6488 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6490 node_size += key->mv_size;
6492 mdb_cassert(mc, data);
6493 if (F_ISSET(flags, F_BIGDATA)) {
6494 /* Data already on overflow page. */
6495 node_size += sizeof(pgno_t);
6496 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6497 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6499 /* Put data on overflow page. */
6500 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6501 data->mv_size, node_size+data->mv_size));
6502 node_size = EVEN(node_size + sizeof(pgno_t));
6503 if ((ssize_t)node_size > room)
6505 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6507 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6511 node_size += data->mv_size;
6514 node_size = EVEN(node_size);
6515 if ((ssize_t)node_size > room)
6519 /* Move higher pointers up one slot. */
6520 for (i = NUMKEYS(mp); i > indx; i--)
6521 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6523 /* Adjust free space offsets. */
6524 ofs = mp->mp_upper - node_size;
6525 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6526 mp->mp_ptrs[indx] = ofs;
6528 mp->mp_lower += sizeof(indx_t);
6530 /* Write the node data. */
6531 node = NODEPTR(mp, indx);
6532 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6533 node->mn_flags = flags;
6535 SETDSZ(node,data->mv_size);
6540 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6543 mdb_cassert(mc, key);
6545 if (F_ISSET(flags, F_BIGDATA))
6546 memcpy(node->mn_data + key->mv_size, data->mv_data,
6548 else if (F_ISSET(flags, MDB_RESERVE))
6549 data->mv_data = node->mn_data + key->mv_size;
6551 memcpy(node->mn_data + key->mv_size, data->mv_data,
6554 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6556 if (F_ISSET(flags, MDB_RESERVE))
6557 data->mv_data = METADATA(ofp);
6559 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6566 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6567 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6568 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6569 DPRINTF(("node size = %"Z"u", node_size));
6570 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6571 return MDB_PAGE_FULL;
6574 /** Delete the specified node from a page.
6575 * @param[in] mc Cursor pointing to the node to delete.
6576 * @param[in] ksize The size of a node. Only used if the page is
6577 * part of a #MDB_DUPFIXED database.
6580 mdb_node_del(MDB_cursor *mc, int ksize)
6582 MDB_page *mp = mc->mc_pg[mc->mc_top];
6583 indx_t indx = mc->mc_ki[mc->mc_top];
6585 indx_t i, j, numkeys, ptr;
6589 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6590 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6591 numkeys = NUMKEYS(mp);
6592 mdb_cassert(mc, indx < numkeys);
6595 int x = numkeys - 1 - indx;
6596 base = LEAF2KEY(mp, indx, ksize);
6598 memmove(base, base + ksize, x * ksize);
6599 mp->mp_lower -= sizeof(indx_t);
6600 mp->mp_upper += ksize - sizeof(indx_t);
6604 node = NODEPTR(mp, indx);
6605 sz = NODESIZE + node->mn_ksize;
6607 if (F_ISSET(node->mn_flags, F_BIGDATA))
6608 sz += sizeof(pgno_t);
6610 sz += NODEDSZ(node);
6614 ptr = mp->mp_ptrs[indx];
6615 for (i = j = 0; i < numkeys; i++) {
6617 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6618 if (mp->mp_ptrs[i] < ptr)
6619 mp->mp_ptrs[j] += sz;
6624 base = (char *)mp + mp->mp_upper;
6625 memmove(base + sz, base, ptr - mp->mp_upper);
6627 mp->mp_lower -= sizeof(indx_t);
6631 /** Compact the main page after deleting a node on a subpage.
6632 * @param[in] mp The main page to operate on.
6633 * @param[in] indx The index of the subpage on the main page.
6636 mdb_node_shrink(MDB_page *mp, indx_t indx)
6642 indx_t i, numkeys, ptr;
6644 node = NODEPTR(mp, indx);
6645 sp = (MDB_page *)NODEDATA(node);
6646 delta = SIZELEFT(sp);
6647 xp = (MDB_page *)((char *)sp + delta);
6649 /* shift subpage upward */
6651 nsize = NUMKEYS(sp) * sp->mp_pad;
6653 return; /* do not make the node uneven-sized */
6654 memmove(METADATA(xp), METADATA(sp), nsize);
6657 numkeys = NUMKEYS(sp);
6658 for (i=numkeys-1; i>=0; i--)
6659 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6661 xp->mp_upper = sp->mp_lower;
6662 xp->mp_lower = sp->mp_lower;
6663 xp->mp_flags = sp->mp_flags;
6664 xp->mp_pad = sp->mp_pad;
6665 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6667 nsize = NODEDSZ(node) - delta;
6668 SETDSZ(node, nsize);
6670 /* shift lower nodes upward */
6671 ptr = mp->mp_ptrs[indx];
6672 numkeys = NUMKEYS(mp);
6673 for (i = 0; i < numkeys; i++) {
6674 if (mp->mp_ptrs[i] <= ptr)
6675 mp->mp_ptrs[i] += delta;
6678 base = (char *)mp + mp->mp_upper;
6679 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6680 mp->mp_upper += delta;
6683 /** Initial setup of a sorted-dups cursor.
6684 * Sorted duplicates are implemented as a sub-database for the given key.
6685 * The duplicate data items are actually keys of the sub-database.
6686 * Operations on the duplicate data items are performed using a sub-cursor
6687 * initialized when the sub-database is first accessed. This function does
6688 * the preliminary setup of the sub-cursor, filling in the fields that
6689 * depend only on the parent DB.
6690 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6693 mdb_xcursor_init0(MDB_cursor *mc)
6695 MDB_xcursor *mx = mc->mc_xcursor;
6697 mx->mx_cursor.mc_xcursor = NULL;
6698 mx->mx_cursor.mc_txn = mc->mc_txn;
6699 mx->mx_cursor.mc_db = &mx->mx_db;
6700 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6701 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6702 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6703 mx->mx_cursor.mc_snum = 0;
6704 mx->mx_cursor.mc_top = 0;
6705 mx->mx_cursor.mc_flags = C_SUB;
6706 mx->mx_dbx.md_name.mv_size = 0;
6707 mx->mx_dbx.md_name.mv_data = NULL;
6708 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6709 mx->mx_dbx.md_dcmp = NULL;
6710 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6713 /** Final setup of a sorted-dups cursor.
6714 * Sets up the fields that depend on the data from the main cursor.
6715 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6716 * @param[in] node The data containing the #MDB_db record for the
6717 * sorted-dup database.
6720 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6722 MDB_xcursor *mx = mc->mc_xcursor;
6724 if (node->mn_flags & F_SUBDATA) {
6725 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6726 mx->mx_cursor.mc_pg[0] = 0;
6727 mx->mx_cursor.mc_snum = 0;
6728 mx->mx_cursor.mc_top = 0;
6729 mx->mx_cursor.mc_flags = C_SUB;
6731 MDB_page *fp = NODEDATA(node);
6732 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6733 mx->mx_db.md_flags = 0;
6734 mx->mx_db.md_depth = 1;
6735 mx->mx_db.md_branch_pages = 0;
6736 mx->mx_db.md_leaf_pages = 1;
6737 mx->mx_db.md_overflow_pages = 0;
6738 mx->mx_db.md_entries = NUMKEYS(fp);
6739 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6740 mx->mx_cursor.mc_snum = 1;
6741 mx->mx_cursor.mc_top = 0;
6742 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6743 mx->mx_cursor.mc_pg[0] = fp;
6744 mx->mx_cursor.mc_ki[0] = 0;
6745 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6746 mx->mx_db.md_flags = MDB_DUPFIXED;
6747 mx->mx_db.md_pad = fp->mp_pad;
6748 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6749 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6752 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6753 mx->mx_db.md_root));
6754 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6755 #if UINT_MAX < SIZE_MAX
6756 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6757 #ifdef MISALIGNED_OK
6758 mx->mx_dbx.md_cmp = mdb_cmp_long;
6760 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6765 /** Initialize a cursor for a given transaction and database. */
6767 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6770 mc->mc_backup = NULL;
6773 mc->mc_db = &txn->mt_dbs[dbi];
6774 mc->mc_dbx = &txn->mt_dbxs[dbi];
6775 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6780 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6781 mdb_tassert(txn, mx != NULL);
6782 mc->mc_xcursor = mx;
6783 mdb_xcursor_init0(mc);
6785 mc->mc_xcursor = NULL;
6787 if (*mc->mc_dbflag & DB_STALE) {
6788 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6793 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6796 size_t size = sizeof(MDB_cursor);
6798 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6801 if (txn->mt_flags & MDB_TXN_ERROR)
6804 /* Allow read access to the freelist */
6805 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6808 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6809 size += sizeof(MDB_xcursor);
6811 if ((mc = malloc(size)) != NULL) {
6812 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6813 if (txn->mt_cursors) {
6814 mc->mc_next = txn->mt_cursors[dbi];
6815 txn->mt_cursors[dbi] = mc;
6816 mc->mc_flags |= C_UNTRACK;
6828 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6830 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6833 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6836 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6840 /* Return the count of duplicate data items for the current key */
6842 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6846 if (mc == NULL || countp == NULL)
6849 if (mc->mc_xcursor == NULL)
6850 return MDB_INCOMPATIBLE;
6852 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6853 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6856 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6859 *countp = mc->mc_xcursor->mx_db.md_entries;
6865 mdb_cursor_close(MDB_cursor *mc)
6867 if (mc && !mc->mc_backup) {
6868 /* remove from txn, if tracked */
6869 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6870 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6871 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6873 *prev = mc->mc_next;
6880 mdb_cursor_txn(MDB_cursor *mc)
6882 if (!mc) return NULL;
6887 mdb_cursor_dbi(MDB_cursor *mc)
6892 /** Replace the key for a branch node with a new key.
6893 * @param[in] mc Cursor pointing to the node to operate on.
6894 * @param[in] key The new key to use.
6895 * @return 0 on success, non-zero on failure.
6898 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6904 int delta, ksize, oksize;
6905 indx_t ptr, i, numkeys, indx;
6908 indx = mc->mc_ki[mc->mc_top];
6909 mp = mc->mc_pg[mc->mc_top];
6910 node = NODEPTR(mp, indx);
6911 ptr = mp->mp_ptrs[indx];
6915 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6916 k2.mv_data = NODEKEY(node);
6917 k2.mv_size = node->mn_ksize;
6918 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6920 mdb_dkey(&k2, kbuf2),
6926 /* Sizes must be 2-byte aligned. */
6927 ksize = EVEN(key->mv_size);
6928 oksize = EVEN(node->mn_ksize);
6929 delta = ksize - oksize;
6931 /* Shift node contents if EVEN(key length) changed. */
6933 if (delta > 0 && SIZELEFT(mp) < delta) {
6935 /* not enough space left, do a delete and split */
6936 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6937 pgno = NODEPGNO(node);
6938 mdb_node_del(mc, 0);
6939 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6942 numkeys = NUMKEYS(mp);
6943 for (i = 0; i < numkeys; i++) {
6944 if (mp->mp_ptrs[i] <= ptr)
6945 mp->mp_ptrs[i] -= delta;
6948 base = (char *)mp + mp->mp_upper;
6949 len = ptr - mp->mp_upper + NODESIZE;
6950 memmove(base - delta, base, len);
6951 mp->mp_upper -= delta;
6953 node = NODEPTR(mp, indx);
6956 /* But even if no shift was needed, update ksize */
6957 if (node->mn_ksize != key->mv_size)
6958 node->mn_ksize = key->mv_size;
6961 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6967 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6969 /** Move a node from csrc to cdst.
6972 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6979 unsigned short flags;
6983 /* Mark src and dst as dirty. */
6984 if ((rc = mdb_page_touch(csrc)) ||
6985 (rc = mdb_page_touch(cdst)))
6988 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6989 key.mv_size = csrc->mc_db->md_pad;
6990 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6992 data.mv_data = NULL;
6996 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6997 mdb_cassert(csrc, !((size_t)srcnode & 1));
6998 srcpg = NODEPGNO(srcnode);
6999 flags = srcnode->mn_flags;
7000 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7001 unsigned int snum = csrc->mc_snum;
7003 /* must find the lowest key below src */
7004 mdb_page_search_lowest(csrc);
7005 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7006 key.mv_size = csrc->mc_db->md_pad;
7007 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7009 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7010 key.mv_size = NODEKSZ(s2);
7011 key.mv_data = NODEKEY(s2);
7013 csrc->mc_snum = snum--;
7014 csrc->mc_top = snum;
7016 key.mv_size = NODEKSZ(srcnode);
7017 key.mv_data = NODEKEY(srcnode);
7019 data.mv_size = NODEDSZ(srcnode);
7020 data.mv_data = NODEDATA(srcnode);
7022 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7023 unsigned int snum = cdst->mc_snum;
7026 /* must find the lowest key below dst */
7027 mdb_page_search_lowest(cdst);
7028 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
7029 bkey.mv_size = cdst->mc_db->md_pad;
7030 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
7032 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7033 bkey.mv_size = NODEKSZ(s2);
7034 bkey.mv_data = NODEKEY(s2);
7036 cdst->mc_snum = snum--;
7037 cdst->mc_top = snum;
7038 mdb_cursor_copy(cdst, &mn);
7040 rc = mdb_update_key(&mn, &bkey);
7045 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7046 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7047 csrc->mc_ki[csrc->mc_top],
7049 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7050 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7052 /* Add the node to the destination page.
7054 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7055 if (rc != MDB_SUCCESS)
7058 /* Delete the node from the source page.
7060 mdb_node_del(csrc, key.mv_size);
7063 /* Adjust other cursors pointing to mp */
7064 MDB_cursor *m2, *m3;
7065 MDB_dbi dbi = csrc->mc_dbi;
7066 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7068 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7069 if (csrc->mc_flags & C_SUB)
7070 m3 = &m2->mc_xcursor->mx_cursor;
7073 if (m3 == csrc) continue;
7074 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7075 csrc->mc_ki[csrc->mc_top]) {
7076 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7077 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7082 /* Update the parent separators.
7084 if (csrc->mc_ki[csrc->mc_top] == 0) {
7085 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7086 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7087 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7089 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7090 key.mv_size = NODEKSZ(srcnode);
7091 key.mv_data = NODEKEY(srcnode);
7093 DPRINTF(("update separator for source page %"Z"u to [%s]",
7094 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7095 mdb_cursor_copy(csrc, &mn);
7098 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7101 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7103 indx_t ix = csrc->mc_ki[csrc->mc_top];
7104 nullkey.mv_size = 0;
7105 csrc->mc_ki[csrc->mc_top] = 0;
7106 rc = mdb_update_key(csrc, &nullkey);
7107 csrc->mc_ki[csrc->mc_top] = ix;
7108 mdb_cassert(csrc, rc == MDB_SUCCESS);
7112 if (cdst->mc_ki[cdst->mc_top] == 0) {
7113 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7114 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7115 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7117 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7118 key.mv_size = NODEKSZ(srcnode);
7119 key.mv_data = NODEKEY(srcnode);
7121 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7122 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7123 mdb_cursor_copy(cdst, &mn);
7126 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7129 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7131 indx_t ix = cdst->mc_ki[cdst->mc_top];
7132 nullkey.mv_size = 0;
7133 cdst->mc_ki[cdst->mc_top] = 0;
7134 rc = mdb_update_key(cdst, &nullkey);
7135 cdst->mc_ki[cdst->mc_top] = ix;
7136 mdb_cassert(csrc, rc == MDB_SUCCESS);
7143 /** Merge one page into another.
7144 * The nodes from the page pointed to by \b csrc will
7145 * be copied to the page pointed to by \b cdst and then
7146 * the \b csrc page will be freed.
7147 * @param[in] csrc Cursor pointing to the source page.
7148 * @param[in] cdst Cursor pointing to the destination page.
7149 * @return 0 on success, non-zero on failure.
7152 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7160 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7161 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7163 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7164 mdb_cassert(csrc, cdst->mc_snum > 1);
7166 /* Mark dst as dirty. */
7167 if ((rc = mdb_page_touch(cdst)))
7170 /* Move all nodes from src to dst.
7172 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7173 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7174 key.mv_size = csrc->mc_db->md_pad;
7175 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7176 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7177 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7178 if (rc != MDB_SUCCESS)
7180 key.mv_data = (char *)key.mv_data + key.mv_size;
7183 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7184 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7185 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7186 unsigned int snum = csrc->mc_snum;
7188 /* must find the lowest key below src */
7189 mdb_page_search_lowest(csrc);
7190 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7191 key.mv_size = csrc->mc_db->md_pad;
7192 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7194 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7195 key.mv_size = NODEKSZ(s2);
7196 key.mv_data = NODEKEY(s2);
7198 csrc->mc_snum = snum--;
7199 csrc->mc_top = snum;
7201 key.mv_size = srcnode->mn_ksize;
7202 key.mv_data = NODEKEY(srcnode);
7205 data.mv_size = NODEDSZ(srcnode);
7206 data.mv_data = NODEDATA(srcnode);
7207 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7208 if (rc != MDB_SUCCESS)
7213 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7214 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7215 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7217 /* Unlink the src page from parent and add to free list.
7220 mdb_node_del(csrc, 0);
7221 if (csrc->mc_ki[csrc->mc_top] == 0) {
7223 rc = mdb_update_key(csrc, &key);
7231 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7232 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7235 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7236 csrc->mc_db->md_leaf_pages--;
7238 csrc->mc_db->md_branch_pages--;
7240 /* Adjust other cursors pointing to mp */
7241 MDB_cursor *m2, *m3;
7242 MDB_dbi dbi = csrc->mc_dbi;
7243 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7245 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7246 if (csrc->mc_flags & C_SUB)
7247 m3 = &m2->mc_xcursor->mx_cursor;
7250 if (m3 == csrc) continue;
7251 if (m3->mc_snum < csrc->mc_snum) continue;
7252 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7253 m3->mc_pg[csrc->mc_top] = mp;
7254 m3->mc_ki[csrc->mc_top] += nkeys;
7258 mdb_cursor_pop(csrc);
7260 return mdb_rebalance(csrc);
7263 /** Copy the contents of a cursor.
7264 * @param[in] csrc The cursor to copy from.
7265 * @param[out] cdst The cursor to copy to.
7268 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7272 cdst->mc_txn = csrc->mc_txn;
7273 cdst->mc_dbi = csrc->mc_dbi;
7274 cdst->mc_db = csrc->mc_db;
7275 cdst->mc_dbx = csrc->mc_dbx;
7276 cdst->mc_snum = csrc->mc_snum;
7277 cdst->mc_top = csrc->mc_top;
7278 cdst->mc_flags = csrc->mc_flags;
7280 for (i=0; i<csrc->mc_snum; i++) {
7281 cdst->mc_pg[i] = csrc->mc_pg[i];
7282 cdst->mc_ki[i] = csrc->mc_ki[i];
7286 /** Rebalance the tree after a delete operation.
7287 * @param[in] mc Cursor pointing to the page where rebalancing
7289 * @return 0 on success, non-zero on failure.
7292 mdb_rebalance(MDB_cursor *mc)
7296 unsigned int ptop, minkeys;
7299 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7300 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7301 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7302 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7303 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7305 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7306 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7307 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7308 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7312 if (mc->mc_snum < 2) {
7313 MDB_page *mp = mc->mc_pg[0];
7315 DPUTS("Can't rebalance a subpage, ignoring");
7318 if (NUMKEYS(mp) == 0) {
7319 DPUTS("tree is completely empty");
7320 mc->mc_db->md_root = P_INVALID;
7321 mc->mc_db->md_depth = 0;
7322 mc->mc_db->md_leaf_pages = 0;
7323 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7326 /* Adjust cursors pointing to mp */
7329 mc->mc_flags &= ~C_INITIALIZED;
7331 MDB_cursor *m2, *m3;
7332 MDB_dbi dbi = mc->mc_dbi;
7334 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7335 if (mc->mc_flags & C_SUB)
7336 m3 = &m2->mc_xcursor->mx_cursor;
7339 if (m3->mc_snum < mc->mc_snum) continue;
7340 if (m3->mc_pg[0] == mp) {
7343 m3->mc_flags &= ~C_INITIALIZED;
7347 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7348 DPUTS("collapsing root page!");
7349 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7352 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7353 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7356 mc->mc_db->md_depth--;
7357 mc->mc_db->md_branch_pages--;
7358 mc->mc_ki[0] = mc->mc_ki[1];
7360 /* Adjust other cursors pointing to mp */
7361 MDB_cursor *m2, *m3;
7362 MDB_dbi dbi = mc->mc_dbi;
7364 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7365 if (mc->mc_flags & C_SUB)
7366 m3 = &m2->mc_xcursor->mx_cursor;
7369 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7370 if (m3->mc_pg[0] == mp) {
7374 for (i=0; i<m3->mc_snum; i++) {
7375 m3->mc_pg[i] = m3->mc_pg[i+1];
7376 m3->mc_ki[i] = m3->mc_ki[i+1];
7382 DPUTS("root page doesn't need rebalancing");
7386 /* The parent (branch page) must have at least 2 pointers,
7387 * otherwise the tree is invalid.
7389 ptop = mc->mc_top-1;
7390 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7392 /* Leaf page fill factor is below the threshold.
7393 * Try to move keys from left or right neighbor, or
7394 * merge with a neighbor page.
7399 mdb_cursor_copy(mc, &mn);
7400 mn.mc_xcursor = NULL;
7402 if (mc->mc_ki[ptop] == 0) {
7403 /* We're the leftmost leaf in our parent.
7405 DPUTS("reading right neighbor");
7407 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7408 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7411 mn.mc_ki[mn.mc_top] = 0;
7412 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7414 /* There is at least one neighbor to the left.
7416 DPUTS("reading left neighbor");
7418 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7419 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7422 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7423 mc->mc_ki[mc->mc_top] = 0;
7426 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7427 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7428 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7430 /* If the neighbor page is above threshold and has enough keys,
7431 * move one key from it. Otherwise we should try to merge them.
7432 * (A branch page must never have less than 2 keys.)
7434 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7435 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7436 return mdb_node_move(&mn, mc);
7438 if (mc->mc_ki[ptop] == 0)
7439 rc = mdb_page_merge(&mn, mc);
7441 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7442 rc = mdb_page_merge(mc, &mn);
7443 mdb_cursor_copy(&mn, mc);
7445 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7450 /** Complete a delete operation started by #mdb_cursor_del(). */
7452 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7459 mp = mc->mc_pg[mc->mc_top];
7460 ki = mc->mc_ki[mc->mc_top];
7462 /* add overflow pages to free list */
7463 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7467 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7468 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7469 (rc = mdb_ovpage_free(mc, omp)))
7472 mdb_node_del(mc, mc->mc_db->md_pad);
7473 mc->mc_db->md_entries--;
7474 rc = mdb_rebalance(mc);
7475 if (rc != MDB_SUCCESS)
7476 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7478 MDB_cursor *m2, *m3;
7479 MDB_dbi dbi = mc->mc_dbi;
7481 mp = mc->mc_pg[mc->mc_top];
7482 nkeys = NUMKEYS(mp);
7484 /* if mc points past last node in page, find next sibling */
7485 if (mc->mc_ki[mc->mc_top] >= nkeys)
7486 mdb_cursor_sibling(mc, 1);
7488 /* Adjust other cursors pointing to mp */
7489 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7490 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7491 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7493 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7495 if (m3->mc_pg[mc->mc_top] == mp) {
7496 if (m3->mc_ki[mc->mc_top] >= ki) {
7497 m3->mc_flags |= C_DEL;
7498 if (m3->mc_ki[mc->mc_top] > ki)
7499 m3->mc_ki[mc->mc_top]--;
7501 if (m3->mc_ki[mc->mc_top] >= nkeys)
7502 mdb_cursor_sibling(m3, 1);
7505 mc->mc_flags |= C_DEL;
7512 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7513 MDB_val *key, MDB_val *data)
7518 MDB_val rdata, *xdata;
7525 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7527 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7530 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7531 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7533 mdb_cursor_init(&mc, txn, dbi, &mx);
7536 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7537 /* must ignore any data */
7548 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7550 /* let mdb_page_split know about this cursor if needed:
7551 * delete will trigger a rebalance; if it needs to move
7552 * a node from one page to another, it will have to
7553 * update the parent's separator key(s). If the new sepkey
7554 * is larger than the current one, the parent page may
7555 * run out of space, triggering a split. We need this
7556 * cursor to be consistent until the end of the rebalance.
7558 mc.mc_flags |= C_UNTRACK;
7559 mc.mc_next = txn->mt_cursors[dbi];
7560 txn->mt_cursors[dbi] = &mc;
7561 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7562 txn->mt_cursors[dbi] = mc.mc_next;
7567 /** Split a page and insert a new node.
7568 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7569 * The cursor will be updated to point to the actual page and index where
7570 * the node got inserted after the split.
7571 * @param[in] newkey The key for the newly inserted node.
7572 * @param[in] newdata The data for the newly inserted node.
7573 * @param[in] newpgno The page number, if the new node is a branch node.
7574 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7575 * @return 0 on success, non-zero on failure.
7578 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7579 unsigned int nflags)
7582 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7585 int i, j, split_indx, nkeys, pmax;
7586 MDB_env *env = mc->mc_txn->mt_env;
7588 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7589 MDB_page *copy = NULL;
7590 MDB_page *mp, *rp, *pp;
7595 mp = mc->mc_pg[mc->mc_top];
7596 newindx = mc->mc_ki[mc->mc_top];
7597 nkeys = NUMKEYS(mp);
7599 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7600 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7601 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7603 /* Create a right sibling. */
7604 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7606 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7608 if (mc->mc_snum < 2) {
7609 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7611 /* shift current top to make room for new parent */
7612 mc->mc_pg[1] = mc->mc_pg[0];
7613 mc->mc_ki[1] = mc->mc_ki[0];
7616 mc->mc_db->md_root = pp->mp_pgno;
7617 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7618 mc->mc_db->md_depth++;
7621 /* Add left (implicit) pointer. */
7622 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7623 /* undo the pre-push */
7624 mc->mc_pg[0] = mc->mc_pg[1];
7625 mc->mc_ki[0] = mc->mc_ki[1];
7626 mc->mc_db->md_root = mp->mp_pgno;
7627 mc->mc_db->md_depth--;
7634 ptop = mc->mc_top-1;
7635 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7638 mc->mc_flags |= C_SPLITTING;
7639 mdb_cursor_copy(mc, &mn);
7640 mn.mc_pg[mn.mc_top] = rp;
7641 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7643 if (nflags & MDB_APPEND) {
7644 mn.mc_ki[mn.mc_top] = 0;
7646 split_indx = newindx;
7650 split_indx = (nkeys+1) / 2;
7655 unsigned int lsize, rsize, ksize;
7656 /* Move half of the keys to the right sibling */
7658 x = mc->mc_ki[mc->mc_top] - split_indx;
7659 ksize = mc->mc_db->md_pad;
7660 split = LEAF2KEY(mp, split_indx, ksize);
7661 rsize = (nkeys - split_indx) * ksize;
7662 lsize = (nkeys - split_indx) * sizeof(indx_t);
7663 mp->mp_lower -= lsize;
7664 rp->mp_lower += lsize;
7665 mp->mp_upper += rsize - lsize;
7666 rp->mp_upper -= rsize - lsize;
7667 sepkey.mv_size = ksize;
7668 if (newindx == split_indx) {
7669 sepkey.mv_data = newkey->mv_data;
7671 sepkey.mv_data = split;
7674 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7675 memcpy(rp->mp_ptrs, split, rsize);
7676 sepkey.mv_data = rp->mp_ptrs;
7677 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7678 memcpy(ins, newkey->mv_data, ksize);
7679 mp->mp_lower += sizeof(indx_t);
7680 mp->mp_upper -= ksize - sizeof(indx_t);
7683 memcpy(rp->mp_ptrs, split, x * ksize);
7684 ins = LEAF2KEY(rp, x, ksize);
7685 memcpy(ins, newkey->mv_data, ksize);
7686 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7687 rp->mp_lower += sizeof(indx_t);
7688 rp->mp_upper -= ksize - sizeof(indx_t);
7689 mc->mc_ki[mc->mc_top] = x;
7690 mc->mc_pg[mc->mc_top] = rp;
7693 int psize, nsize, k;
7694 /* Maximum free space in an empty page */
7695 pmax = env->me_psize - PAGEHDRSZ;
7697 nsize = mdb_leaf_size(env, newkey, newdata);
7699 nsize = mdb_branch_size(env, newkey);
7700 nsize = EVEN(nsize);
7702 /* grab a page to hold a temporary copy */
7703 copy = mdb_page_malloc(mc->mc_txn, 1);
7706 copy->mp_pgno = mp->mp_pgno;
7707 copy->mp_flags = mp->mp_flags;
7708 copy->mp_lower = PAGEHDRSZ;
7709 copy->mp_upper = env->me_psize;
7711 /* prepare to insert */
7712 for (i=0, j=0; i<nkeys; i++) {
7714 copy->mp_ptrs[j++] = 0;
7716 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7719 /* When items are relatively large the split point needs
7720 * to be checked, because being off-by-one will make the
7721 * difference between success or failure in mdb_node_add.
7723 * It's also relevant if a page happens to be laid out
7724 * such that one half of its nodes are all "small" and
7725 * the other half of its nodes are "large." If the new
7726 * item is also "large" and falls on the half with
7727 * "large" nodes, it also may not fit.
7729 * As a final tweak, if the new item goes on the last
7730 * spot on the page (and thus, onto the new page), bias
7731 * the split so the new page is emptier than the old page.
7732 * This yields better packing during sequential inserts.
7734 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7735 /* Find split point */
7737 if (newindx <= split_indx || newindx >= nkeys) {
7739 k = newindx >= nkeys ? nkeys : split_indx+2;
7744 for (; i!=k; i+=j) {
7749 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7750 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7752 if (F_ISSET(node->mn_flags, F_BIGDATA))
7753 psize += sizeof(pgno_t);
7755 psize += NODEDSZ(node);
7757 psize = EVEN(psize);
7759 if (psize > pmax || i == k-j) {
7760 split_indx = i + (j<0);
7765 if (split_indx == newindx) {
7766 sepkey.mv_size = newkey->mv_size;
7767 sepkey.mv_data = newkey->mv_data;
7769 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7770 sepkey.mv_size = node->mn_ksize;
7771 sepkey.mv_data = NODEKEY(node);
7776 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7778 /* Copy separator key to the parent.
7780 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7784 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7787 if (mn.mc_snum == mc->mc_snum) {
7788 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7789 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7790 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7791 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7796 /* Right page might now have changed parent.
7797 * Check if left page also changed parent.
7799 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7800 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7801 for (i=0; i<ptop; i++) {
7802 mc->mc_pg[i] = mn.mc_pg[i];
7803 mc->mc_ki[i] = mn.mc_ki[i];
7805 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7806 if (mn.mc_ki[ptop]) {
7807 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7809 /* find right page's left sibling */
7810 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7811 mdb_cursor_sibling(mc, 0);
7816 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7819 mc->mc_flags ^= C_SPLITTING;
7820 if (rc != MDB_SUCCESS) {
7823 if (nflags & MDB_APPEND) {
7824 mc->mc_pg[mc->mc_top] = rp;
7825 mc->mc_ki[mc->mc_top] = 0;
7826 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7829 for (i=0; i<mc->mc_top; i++)
7830 mc->mc_ki[i] = mn.mc_ki[i];
7831 } else if (!IS_LEAF2(mp)) {
7833 mc->mc_pg[mc->mc_top] = rp;
7838 rkey.mv_data = newkey->mv_data;
7839 rkey.mv_size = newkey->mv_size;
7845 /* Update index for the new key. */
7846 mc->mc_ki[mc->mc_top] = j;
7848 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7849 rkey.mv_data = NODEKEY(node);
7850 rkey.mv_size = node->mn_ksize;
7852 xdata.mv_data = NODEDATA(node);
7853 xdata.mv_size = NODEDSZ(node);
7856 pgno = NODEPGNO(node);
7857 flags = node->mn_flags;
7860 if (!IS_LEAF(mp) && j == 0) {
7861 /* First branch index doesn't need key data. */
7865 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7867 /* return tmp page to freelist */
7868 mdb_page_free(env, copy);
7874 mc->mc_pg[mc->mc_top] = copy;
7879 } while (i != split_indx);
7881 nkeys = NUMKEYS(copy);
7882 for (i=0; i<nkeys; i++)
7883 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7884 mp->mp_lower = copy->mp_lower;
7885 mp->mp_upper = copy->mp_upper;
7886 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7887 env->me_psize - copy->mp_upper);
7889 /* reset back to original page */
7890 if (newindx < split_indx) {
7891 mc->mc_pg[mc->mc_top] = mp;
7892 if (nflags & MDB_RESERVE) {
7893 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7894 if (!(node->mn_flags & F_BIGDATA))
7895 newdata->mv_data = NODEDATA(node);
7898 mc->mc_pg[mc->mc_top] = rp;
7900 /* Make sure mc_ki is still valid.
7902 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7903 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7904 for (i=0; i<=ptop; i++) {
7905 mc->mc_pg[i] = mn.mc_pg[i];
7906 mc->mc_ki[i] = mn.mc_ki[i];
7910 /* return tmp page to freelist */
7911 mdb_page_free(env, copy);
7915 /* Adjust other cursors pointing to mp */
7916 MDB_cursor *m2, *m3;
7917 MDB_dbi dbi = mc->mc_dbi;
7918 int fixup = NUMKEYS(mp);
7920 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7921 if (mc->mc_flags & C_SUB)
7922 m3 = &m2->mc_xcursor->mx_cursor;
7927 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7929 if (m3->mc_flags & C_SPLITTING)
7934 for (k=m3->mc_top; k>=0; k--) {
7935 m3->mc_ki[k+1] = m3->mc_ki[k];
7936 m3->mc_pg[k+1] = m3->mc_pg[k];
7938 if (m3->mc_ki[0] >= split_indx) {
7943 m3->mc_pg[0] = mc->mc_pg[0];
7947 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7948 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7949 m3->mc_ki[mc->mc_top]++;
7950 if (m3->mc_ki[mc->mc_top] >= fixup) {
7951 m3->mc_pg[mc->mc_top] = rp;
7952 m3->mc_ki[mc->mc_top] -= fixup;
7953 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7955 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7956 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7961 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7966 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7967 MDB_val *key, MDB_val *data, unsigned int flags)
7972 if (key == NULL || data == NULL)
7975 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7978 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7981 mdb_cursor_init(&mc, txn, dbi, &mx);
7982 return mdb_cursor_put(&mc, key, data, flags);
7986 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7988 if ((flag & CHANGEABLE) != flag)
7991 env->me_flags |= flag;
7993 env->me_flags &= ~flag;
7998 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8003 *arg = env->me_flags;
8008 mdb_env_set_userctx(MDB_env *env, void *ctx)
8012 env->me_userctx = ctx;
8017 mdb_env_get_userctx(MDB_env *env)
8019 return env ? env->me_userctx : NULL;
8023 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8028 env->me_assert_func = func;
8034 mdb_env_get_path(MDB_env *env, const char **arg)
8039 *arg = env->me_path;
8044 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8053 /** Common code for #mdb_stat() and #mdb_env_stat().
8054 * @param[in] env the environment to operate in.
8055 * @param[in] db the #MDB_db record containing the stats to return.
8056 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8057 * @return 0, this function always succeeds.
8060 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8062 arg->ms_psize = env->me_psize;
8063 arg->ms_depth = db->md_depth;
8064 arg->ms_branch_pages = db->md_branch_pages;
8065 arg->ms_leaf_pages = db->md_leaf_pages;
8066 arg->ms_overflow_pages = db->md_overflow_pages;
8067 arg->ms_entries = db->md_entries;
8072 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8076 if (env == NULL || arg == NULL)
8079 toggle = mdb_env_pick_meta(env);
8081 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8085 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8089 if (env == NULL || arg == NULL)
8092 toggle = mdb_env_pick_meta(env);
8093 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
8094 arg->me_mapsize = env->me_mapsize;
8095 arg->me_maxreaders = env->me_maxreaders;
8097 /* me_numreaders may be zero if this process never used any readers. Use
8098 * the shared numreader count if it exists.
8100 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8102 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8103 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8107 /** Set the default comparison functions for a database.
8108 * Called immediately after a database is opened to set the defaults.
8109 * The user can then override them with #mdb_set_compare() or
8110 * #mdb_set_dupsort().
8111 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8112 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8115 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8117 uint16_t f = txn->mt_dbs[dbi].md_flags;
8119 txn->mt_dbxs[dbi].md_cmp =
8120 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8121 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8123 txn->mt_dbxs[dbi].md_dcmp =
8124 !(f & MDB_DUPSORT) ? 0 :
8125 ((f & MDB_INTEGERDUP)
8126 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8127 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8130 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8135 int rc, dbflag, exact;
8136 unsigned int unused = 0;
8139 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8140 mdb_default_cmp(txn, FREE_DBI);
8143 if ((flags & VALID_FLAGS) != flags)
8145 if (txn->mt_flags & MDB_TXN_ERROR)
8151 if (flags & PERSISTENT_FLAGS) {
8152 uint16_t f2 = flags & PERSISTENT_FLAGS;
8153 /* make sure flag changes get committed */
8154 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8155 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8156 txn->mt_flags |= MDB_TXN_DIRTY;
8159 mdb_default_cmp(txn, MAIN_DBI);
8163 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8164 mdb_default_cmp(txn, MAIN_DBI);
8167 /* Is the DB already open? */
8169 for (i=2; i<txn->mt_numdbs; i++) {
8170 if (!txn->mt_dbxs[i].md_name.mv_size) {
8171 /* Remember this free slot */
8172 if (!unused) unused = i;
8175 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8176 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8182 /* If no free slot and max hit, fail */
8183 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8184 return MDB_DBS_FULL;
8186 /* Cannot mix named databases with some mainDB flags */
8187 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8188 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8190 /* Find the DB info */
8191 dbflag = DB_NEW|DB_VALID;
8194 key.mv_data = (void *)name;
8195 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8196 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8197 if (rc == MDB_SUCCESS) {
8198 /* make sure this is actually a DB */
8199 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8200 if (!(node->mn_flags & F_SUBDATA))
8201 return MDB_INCOMPATIBLE;
8202 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8203 /* Create if requested */
8205 data.mv_size = sizeof(MDB_db);
8206 data.mv_data = &dummy;
8207 memset(&dummy, 0, sizeof(dummy));
8208 dummy.md_root = P_INVALID;
8209 dummy.md_flags = flags & PERSISTENT_FLAGS;
8210 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8214 /* OK, got info, add to table */
8215 if (rc == MDB_SUCCESS) {
8216 unsigned int slot = unused ? unused : txn->mt_numdbs;
8217 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8218 txn->mt_dbxs[slot].md_name.mv_size = len;
8219 txn->mt_dbxs[slot].md_rel = NULL;
8220 txn->mt_dbflags[slot] = dbflag;
8221 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8223 mdb_default_cmp(txn, slot);
8232 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8234 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8237 if (txn->mt_dbflags[dbi] & DB_STALE) {
8240 /* Stale, must read the DB's root. cursor_init does it for us. */
8241 mdb_cursor_init(&mc, txn, dbi, &mx);
8243 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8246 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8249 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8251 ptr = env->me_dbxs[dbi].md_name.mv_data;
8252 env->me_dbxs[dbi].md_name.mv_data = NULL;
8253 env->me_dbxs[dbi].md_name.mv_size = 0;
8254 env->me_dbflags[dbi] = 0;
8258 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8260 /* We could return the flags for the FREE_DBI too but what's the point? */
8261 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8263 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8267 /** Add all the DB's pages to the free list.
8268 * @param[in] mc Cursor on the DB to free.
8269 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8270 * @return 0 on success, non-zero on failure.
8273 mdb_drop0(MDB_cursor *mc, int subs)
8277 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8278 if (rc == MDB_SUCCESS) {
8279 MDB_txn *txn = mc->mc_txn;
8284 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8285 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8288 mdb_cursor_copy(mc, &mx);
8289 while (mc->mc_snum > 0) {
8290 MDB_page *mp = mc->mc_pg[mc->mc_top];
8291 unsigned n = NUMKEYS(mp);
8293 for (i=0; i<n; i++) {
8294 ni = NODEPTR(mp, i);
8295 if (ni->mn_flags & F_BIGDATA) {
8298 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8299 rc = mdb_page_get(txn, pg, &omp, NULL);
8302 mdb_cassert(mc, IS_OVERFLOW(omp));
8303 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8307 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8308 mdb_xcursor_init1(mc, ni);
8309 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8315 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8317 for (i=0; i<n; i++) {
8319 ni = NODEPTR(mp, i);
8322 mdb_midl_xappend(txn->mt_free_pgs, pg);
8327 mc->mc_ki[mc->mc_top] = i;
8328 rc = mdb_cursor_sibling(mc, 1);
8330 /* no more siblings, go back to beginning
8331 * of previous level.
8335 for (i=1; i<mc->mc_snum; i++) {
8337 mc->mc_pg[i] = mx.mc_pg[i];
8342 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8343 } else if (rc == MDB_NOTFOUND) {
8349 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8351 MDB_cursor *mc, *m2;
8354 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8357 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8360 rc = mdb_cursor_open(txn, dbi, &mc);
8364 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8365 /* Invalidate the dropped DB's cursors */
8366 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8367 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8371 /* Can't delete the main DB */
8372 if (del && dbi > MAIN_DBI) {
8373 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8375 txn->mt_dbflags[dbi] = DB_STALE;
8376 mdb_dbi_close(txn->mt_env, dbi);
8379 /* reset the DB record, mark it dirty */
8380 txn->mt_dbflags[dbi] |= DB_DIRTY;
8381 txn->mt_dbs[dbi].md_depth = 0;
8382 txn->mt_dbs[dbi].md_branch_pages = 0;
8383 txn->mt_dbs[dbi].md_leaf_pages = 0;
8384 txn->mt_dbs[dbi].md_overflow_pages = 0;
8385 txn->mt_dbs[dbi].md_entries = 0;
8386 txn->mt_dbs[dbi].md_root = P_INVALID;
8388 txn->mt_flags |= MDB_TXN_DIRTY;
8391 mdb_cursor_close(mc);
8395 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8397 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8400 txn->mt_dbxs[dbi].md_cmp = cmp;
8404 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8406 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8409 txn->mt_dbxs[dbi].md_dcmp = cmp;
8413 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8415 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8418 txn->mt_dbxs[dbi].md_rel = rel;
8422 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8424 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8427 txn->mt_dbxs[dbi].md_relctx = ctx;
8431 int mdb_env_get_maxkeysize(MDB_env *env)
8433 return ENV_MAXKEY(env);
8436 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8438 unsigned int i, rdrs;
8441 int rc = 0, first = 1;
8445 if (!env->me_txns) {
8446 return func("(no reader locks)\n", ctx);
8448 rdrs = env->me_txns->mti_numreaders;
8449 mr = env->me_txns->mti_readers;
8450 for (i=0; i<rdrs; i++) {
8452 txnid_t txnid = mr[i].mr_txnid;
8453 sprintf(buf, txnid == (txnid_t)-1 ?
8454 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
8455 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
8458 rc = func(" pid thread txnid\n", ctx);
8462 rc = func(buf, ctx);
8468 rc = func("(no active readers)\n", ctx);
8473 /** Insert pid into list if not already present.
8474 * return -1 if already present.
8476 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8478 /* binary search of pid in list */
8480 unsigned cursor = 1;
8482 unsigned n = ids[0];
8485 unsigned pivot = n >> 1;
8486 cursor = base + pivot + 1;
8487 val = pid - ids[cursor];
8492 } else if ( val > 0 ) {
8497 /* found, so it's a duplicate */
8506 for (n = ids[0]; n > cursor; n--)
8512 int mdb_reader_check(MDB_env *env, int *dead)
8514 unsigned int i, j, rdrs;
8516 MDB_PID_T *pids, pid;
8525 rdrs = env->me_txns->mti_numreaders;
8526 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8530 mr = env->me_txns->mti_readers;
8531 for (i=0; i<rdrs; i++) {
8532 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8534 if (mdb_pid_insert(pids, pid) == 0) {
8535 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8537 /* Recheck, a new process may have reused pid */
8538 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8539 for (j=i; j<rdrs; j++)
8540 if (mr[j].mr_pid == pid) {
8541 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8542 (unsigned) pid, mr[j].mr_txnid));
8547 UNLOCK_MUTEX_R(env);