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 /**< txn is unusable after an error */
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 /** Check \b txn and \b dbi arguments to a function */
1102 #define TXN_DBI_EXIST(txn, dbi) \
1103 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1105 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1106 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1107 static int mdb_page_touch(MDB_cursor *mc);
1109 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1110 static int mdb_page_search_root(MDB_cursor *mc,
1111 MDB_val *key, int modify);
1112 #define MDB_PS_MODIFY 1
1113 #define MDB_PS_ROOTONLY 2
1114 #define MDB_PS_FIRST 4
1115 #define MDB_PS_LAST 8
1116 static int mdb_page_search(MDB_cursor *mc,
1117 MDB_val *key, int flags);
1118 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1120 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1121 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1122 pgno_t newpgno, unsigned int nflags);
1124 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1125 static int mdb_env_pick_meta(const MDB_env *env);
1126 static int mdb_env_write_meta(MDB_txn *txn);
1127 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1128 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1130 static void mdb_env_close0(MDB_env *env, int excl);
1132 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1133 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1134 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1135 static void mdb_node_del(MDB_cursor *mc, int ksize);
1136 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1137 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1138 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1139 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1140 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1142 static int mdb_rebalance(MDB_cursor *mc);
1143 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1145 static void mdb_cursor_pop(MDB_cursor *mc);
1146 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1148 static int mdb_cursor_del0(MDB_cursor *mc);
1149 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1150 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1151 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1152 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1153 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1155 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1156 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1158 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1159 static void mdb_xcursor_init0(MDB_cursor *mc);
1160 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1162 static int mdb_drop0(MDB_cursor *mc, int subs);
1163 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1166 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1170 static SECURITY_DESCRIPTOR mdb_null_sd;
1171 static SECURITY_ATTRIBUTES mdb_all_sa;
1172 static int mdb_sec_inited;
1175 /** Return the library version info. */
1177 mdb_version(int *major, int *minor, int *patch)
1179 if (major) *major = MDB_VERSION_MAJOR;
1180 if (minor) *minor = MDB_VERSION_MINOR;
1181 if (patch) *patch = MDB_VERSION_PATCH;
1182 return MDB_VERSION_STRING;
1185 /** Table of descriptions for MDB @ref errors */
1186 static char *const mdb_errstr[] = {
1187 "MDB_KEYEXIST: Key/data pair already exists",
1188 "MDB_NOTFOUND: No matching key/data pair found",
1189 "MDB_PAGE_NOTFOUND: Requested page not found",
1190 "MDB_CORRUPTED: Located page was wrong type",
1191 "MDB_PANIC: Update of meta page failed",
1192 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1193 "MDB_INVALID: File is not an MDB file",
1194 "MDB_MAP_FULL: Environment mapsize limit reached",
1195 "MDB_DBS_FULL: Environment maxdbs limit reached",
1196 "MDB_READERS_FULL: Environment maxreaders limit reached",
1197 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1198 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1199 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1200 "MDB_PAGE_FULL: Internal error - page has no more space",
1201 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1202 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1203 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1204 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1205 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1209 mdb_strerror(int err)
1213 return ("Successful return: 0");
1215 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1216 i = err - MDB_KEYEXIST;
1217 return mdb_errstr[i];
1220 return strerror(err);
1223 /** assert(3) variant in cursor context */
1224 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1225 /** assert(3) variant in transaction context */
1226 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1227 /** assert(3) variant in environment context */
1228 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1231 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1232 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1235 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1236 const char *func, const char *file, int line)
1239 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1240 file, line, expr_txt, func);
1241 if (env->me_assert_func)
1242 env->me_assert_func(env, buf);
1243 fprintf(stderr, "%s\n", buf);
1247 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1251 /** Return the page number of \b mp which may be sub-page, for debug output */
1253 mdb_dbg_pgno(MDB_page *mp)
1256 COPY_PGNO(ret, mp->mp_pgno);
1260 /** Display a key in hexadecimal and return the address of the result.
1261 * @param[in] key the key to display
1262 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1263 * @return The key in hexadecimal form.
1266 mdb_dkey(MDB_val *key, char *buf)
1269 unsigned char *c = key->mv_data;
1275 if (key->mv_size > DKBUF_MAXKEYSIZE)
1276 return "MDB_MAXKEYSIZE";
1277 /* may want to make this a dynamic check: if the key is mostly
1278 * printable characters, print it as-is instead of converting to hex.
1282 for (i=0; i<key->mv_size; i++)
1283 ptr += sprintf(ptr, "%02x", *c++);
1285 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1291 mdb_leafnode_type(MDB_node *n)
1293 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1294 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1295 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1298 /** Display all the keys in the page. */
1300 mdb_page_list(MDB_page *mp)
1302 pgno_t pgno = mdb_dbg_pgno(mp);
1303 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1305 unsigned int i, nkeys, nsize, total = 0;
1309 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1310 case P_BRANCH: type = "Branch page"; break;
1311 case P_LEAF: type = "Leaf page"; break;
1312 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1313 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1314 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1316 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1317 pgno, mp->mp_pages, state);
1320 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1321 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1324 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1328 nkeys = NUMKEYS(mp);
1329 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1331 for (i=0; i<nkeys; i++) {
1332 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1333 key.mv_size = nsize = mp->mp_pad;
1334 key.mv_data = LEAF2KEY(mp, i, nsize);
1336 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1339 node = NODEPTR(mp, i);
1340 key.mv_size = node->mn_ksize;
1341 key.mv_data = node->mn_data;
1342 nsize = NODESIZE + key.mv_size;
1343 if (IS_BRANCH(mp)) {
1344 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1348 if (F_ISSET(node->mn_flags, F_BIGDATA))
1349 nsize += sizeof(pgno_t);
1351 nsize += NODEDSZ(node);
1353 nsize += sizeof(indx_t);
1354 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1355 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1357 total = EVEN(total);
1359 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1360 IS_LEAF2(mp) ? PAGEHDRSZ : mp->mp_lower, total, SIZELEFT(mp));
1364 mdb_cursor_chk(MDB_cursor *mc)
1370 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1371 for (i=0; i<mc->mc_top; i++) {
1373 node = NODEPTR(mp, mc->mc_ki[i]);
1374 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1377 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1383 /** Count all the pages in each DB and in the freelist
1384 * and make sure it matches the actual number of pages
1386 * All named DBs must be open for a correct count.
1388 static void mdb_audit(MDB_txn *txn)
1392 MDB_ID freecount, count;
1397 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1398 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1399 freecount += *(MDB_ID *)data.mv_data;
1400 mdb_tassert(txn, rc == MDB_NOTFOUND);
1403 for (i = 0; i<txn->mt_numdbs; i++) {
1405 if (!(txn->mt_dbflags[i] & DB_VALID))
1407 mdb_cursor_init(&mc, txn, i, &mx);
1408 if (txn->mt_dbs[i].md_root == P_INVALID)
1410 count += txn->mt_dbs[i].md_branch_pages +
1411 txn->mt_dbs[i].md_leaf_pages +
1412 txn->mt_dbs[i].md_overflow_pages;
1413 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1414 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1415 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1418 mp = mc.mc_pg[mc.mc_top];
1419 for (j=0; j<NUMKEYS(mp); j++) {
1420 MDB_node *leaf = NODEPTR(mp, j);
1421 if (leaf->mn_flags & F_SUBDATA) {
1423 memcpy(&db, NODEDATA(leaf), sizeof(db));
1424 count += db.md_branch_pages + db.md_leaf_pages +
1425 db.md_overflow_pages;
1429 mdb_tassert(txn, rc == MDB_NOTFOUND);
1432 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1433 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1434 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1440 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1442 return txn->mt_dbxs[dbi].md_cmp(a, b);
1446 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1448 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1451 /** Allocate memory for a page.
1452 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1455 mdb_page_malloc(MDB_txn *txn, unsigned num)
1457 MDB_env *env = txn->mt_env;
1458 MDB_page *ret = env->me_dpages;
1459 size_t psize = env->me_psize, sz = psize, off;
1460 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1461 * For a single page alloc, we init everything after the page header.
1462 * For multi-page, we init the final page; if the caller needed that
1463 * many pages they will be filling in at least up to the last page.
1467 VGMEMP_ALLOC(env, ret, sz);
1468 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1469 env->me_dpages = ret->mp_next;
1472 psize -= off = PAGEHDRSZ;
1477 if ((ret = malloc(sz)) != NULL) {
1478 VGMEMP_ALLOC(env, ret, sz);
1479 if (!(env->me_flags & MDB_NOMEMINIT)) {
1480 memset((char *)ret + off, 0, psize);
1484 txn->mt_flags |= MDB_TXN_ERROR;
1489 /** Free a single page.
1490 * Saves single pages to a list, for future reuse.
1491 * (This is not used for multi-page overflow pages.)
1494 mdb_page_free(MDB_env *env, MDB_page *mp)
1496 mp->mp_next = env->me_dpages;
1497 VGMEMP_FREE(env, mp);
1498 env->me_dpages = mp;
1501 /** Free a dirty page */
1503 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1505 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1506 mdb_page_free(env, dp);
1508 /* large pages just get freed directly */
1509 VGMEMP_FREE(env, dp);
1514 /** Return all dirty pages to dpage list */
1516 mdb_dlist_free(MDB_txn *txn)
1518 MDB_env *env = txn->mt_env;
1519 MDB_ID2L dl = txn->mt_u.dirty_list;
1520 unsigned i, n = dl[0].mid;
1522 for (i = 1; i <= n; i++) {
1523 mdb_dpage_free(env, dl[i].mptr);
1528 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1529 * @param[in] mc A cursor handle for the current operation.
1530 * @param[in] pflags Flags of the pages to update:
1531 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1532 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1533 * @return 0 on success, non-zero on failure.
1536 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1538 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1539 MDB_txn *txn = mc->mc_txn;
1545 int rc = MDB_SUCCESS, level;
1547 /* Mark pages seen by cursors */
1548 if (mc->mc_flags & C_UNTRACK)
1549 mc = NULL; /* will find mc in mt_cursors */
1550 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1551 for (; mc; mc=mc->mc_next) {
1552 if (!(mc->mc_flags & C_INITIALIZED))
1554 for (m3 = mc;; m3 = &mx->mx_cursor) {
1556 for (j=0; j<m3->mc_snum; j++) {
1558 if ((mp->mp_flags & Mask) == pflags)
1559 mp->mp_flags ^= P_KEEP;
1561 mx = m3->mc_xcursor;
1562 /* Proceed to mx if it is at a sub-database */
1563 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1565 if (! (mp && (mp->mp_flags & P_LEAF)))
1567 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1568 if (!(leaf->mn_flags & F_SUBDATA))
1577 /* Mark dirty root pages */
1578 for (i=0; i<txn->mt_numdbs; i++) {
1579 if (txn->mt_dbflags[i] & DB_DIRTY) {
1580 pgno_t pgno = txn->mt_dbs[i].md_root;
1581 if (pgno == P_INVALID)
1583 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1585 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1586 dp->mp_flags ^= P_KEEP;
1594 static int mdb_page_flush(MDB_txn *txn, int keep);
1596 /** Spill pages from the dirty list back to disk.
1597 * This is intended to prevent running into #MDB_TXN_FULL situations,
1598 * but note that they may still occur in a few cases:
1599 * 1) our estimate of the txn size could be too small. Currently this
1600 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1601 * 2) child txns may run out of space if their parents dirtied a
1602 * lot of pages and never spilled them. TODO: we probably should do
1603 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1604 * the parent's dirty_room is below a given threshold.
1606 * Otherwise, if not using nested txns, it is expected that apps will
1607 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1608 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1609 * If the txn never references them again, they can be left alone.
1610 * If the txn only reads them, they can be used without any fuss.
1611 * If the txn writes them again, they can be dirtied immediately without
1612 * going thru all of the work of #mdb_page_touch(). Such references are
1613 * handled by #mdb_page_unspill().
1615 * Also note, we never spill DB root pages, nor pages of active cursors,
1616 * because we'll need these back again soon anyway. And in nested txns,
1617 * we can't spill a page in a child txn if it was already spilled in a
1618 * parent txn. That would alter the parent txns' data even though
1619 * the child hasn't committed yet, and we'd have no way to undo it if
1620 * the child aborted.
1622 * @param[in] m0 cursor A cursor handle identifying the transaction and
1623 * database for which we are checking space.
1624 * @param[in] key For a put operation, the key being stored.
1625 * @param[in] data For a put operation, the data being stored.
1626 * @return 0 on success, non-zero on failure.
1629 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1631 MDB_txn *txn = m0->mc_txn;
1633 MDB_ID2L dl = txn->mt_u.dirty_list;
1634 unsigned int i, j, need;
1637 if (m0->mc_flags & C_SUB)
1640 /* Estimate how much space this op will take */
1641 i = m0->mc_db->md_depth;
1642 /* Named DBs also dirty the main DB */
1643 if (m0->mc_dbi > MAIN_DBI)
1644 i += txn->mt_dbs[MAIN_DBI].md_depth;
1645 /* For puts, roughly factor in the key+data size */
1647 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1648 i += i; /* double it for good measure */
1651 if (txn->mt_dirty_room > i)
1654 if (!txn->mt_spill_pgs) {
1655 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1656 if (!txn->mt_spill_pgs)
1659 /* purge deleted slots */
1660 MDB_IDL sl = txn->mt_spill_pgs;
1661 unsigned int num = sl[0];
1663 for (i=1; i<=num; i++) {
1670 /* Preserve pages which may soon be dirtied again */
1671 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1674 /* Less aggressive spill - we originally spilled the entire dirty list,
1675 * with a few exceptions for cursor pages and DB root pages. But this
1676 * turns out to be a lot of wasted effort because in a large txn many
1677 * of those pages will need to be used again. So now we spill only 1/8th
1678 * of the dirty pages. Testing revealed this to be a good tradeoff,
1679 * better than 1/2, 1/4, or 1/10.
1681 if (need < MDB_IDL_UM_MAX / 8)
1682 need = MDB_IDL_UM_MAX / 8;
1684 /* Save the page IDs of all the pages we're flushing */
1685 /* flush from the tail forward, this saves a lot of shifting later on. */
1686 for (i=dl[0].mid; i && need; i--) {
1687 MDB_ID pn = dl[i].mid << 1;
1689 if (dp->mp_flags & P_KEEP)
1691 /* Can't spill twice, make sure it's not already in a parent's
1694 if (txn->mt_parent) {
1696 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1697 if (tx2->mt_spill_pgs) {
1698 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1699 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1700 dp->mp_flags |= P_KEEP;
1708 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1712 mdb_midl_sort(txn->mt_spill_pgs);
1714 /* Flush the spilled part of dirty list */
1715 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1718 /* Reset any dirty pages we kept that page_flush didn't see */
1719 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1722 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1726 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1728 mdb_find_oldest(MDB_txn *txn)
1731 txnid_t mr, oldest = txn->mt_txnid - 1;
1732 if (txn->mt_env->me_txns) {
1733 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1734 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1745 /** Add a page to the txn's dirty list */
1747 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1750 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1752 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1753 insert = mdb_mid2l_append;
1755 insert = mdb_mid2l_insert;
1757 mid.mid = mp->mp_pgno;
1759 rc = insert(txn->mt_u.dirty_list, &mid);
1760 mdb_tassert(txn, rc == 0);
1761 txn->mt_dirty_room--;
1764 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1765 * me_pghead and mt_next_pgno.
1767 * If there are free pages available from older transactions, they
1768 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1769 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1770 * and move me_pglast to say which records were consumed. Only this
1771 * function can create me_pghead and move me_pglast/mt_next_pgno.
1772 * @param[in] mc cursor A cursor handle identifying the transaction and
1773 * database for which we are allocating.
1774 * @param[in] num the number of pages to allocate.
1775 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1776 * will always be satisfied by a single contiguous chunk of memory.
1777 * @return 0 on success, non-zero on failure.
1780 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1782 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1783 /* Get at most <Max_retries> more freeDB records once me_pghead
1784 * has enough pages. If not enough, use new pages from the map.
1785 * If <Paranoid> and mc is updating the freeDB, only get new
1786 * records if me_pghead is empty. Then the freelist cannot play
1787 * catch-up with itself by growing while trying to save it.
1789 enum { Paranoid = 1, Max_retries = 500 };
1791 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1793 int rc, retry = num * 20;
1794 MDB_txn *txn = mc->mc_txn;
1795 MDB_env *env = txn->mt_env;
1796 pgno_t pgno, *mop = env->me_pghead;
1797 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1799 txnid_t oldest = 0, last;
1805 /* If our dirty list is already full, we can't do anything */
1806 if (txn->mt_dirty_room == 0) {
1811 for (op = MDB_FIRST;; op = MDB_NEXT) {
1814 pgno_t *idl, old_id, new_id;
1816 /* Seek a big enough contiguous page range. Prefer
1817 * pages at the tail, just truncating the list.
1823 if (mop[i-n2] == pgno+n2)
1830 if (op == MDB_FIRST) { /* 1st iteration */
1831 /* Prepare to fetch more and coalesce */
1832 oldest = mdb_find_oldest(txn);
1833 last = env->me_pglast;
1834 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1837 key.mv_data = &last; /* will look up last+1 */
1838 key.mv_size = sizeof(last);
1840 if (Paranoid && mc->mc_dbi == FREE_DBI)
1843 if (Paranoid && retry < 0 && mop_len)
1847 /* Do not fetch more if the record will be too recent */
1850 rc = mdb_cursor_get(&m2, &key, NULL, op);
1852 if (rc == MDB_NOTFOUND)
1856 last = *(txnid_t*)key.mv_data;
1859 np = m2.mc_pg[m2.mc_top];
1860 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1861 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1864 idl = (MDB_ID *) data.mv_data;
1867 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1872 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1874 mop = env->me_pghead;
1876 env->me_pglast = last;
1878 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1879 last, txn->mt_dbs[FREE_DBI].md_root, i));
1881 DPRINTF(("IDL %"Z"u", idl[k]));
1883 /* Merge in descending sorted order */
1886 mop[0] = (pgno_t)-1;
1890 for (; old_id < new_id; old_id = mop[--j])
1897 /* Use new pages from the map when nothing suitable in the freeDB */
1899 pgno = txn->mt_next_pgno;
1900 if (pgno + num >= env->me_maxpg) {
1901 DPUTS("DB size maxed out");
1907 if (env->me_flags & MDB_WRITEMAP) {
1908 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1910 if (!(np = mdb_page_malloc(txn, num))) {
1916 mop[0] = mop_len -= num;
1917 /* Move any stragglers down */
1918 for (j = i-num; j < mop_len; )
1919 mop[++j] = mop[++i];
1921 txn->mt_next_pgno = pgno + num;
1924 mdb_page_dirty(txn, np);
1930 txn->mt_flags |= MDB_TXN_ERROR;
1934 /** Copy the used portions of a non-overflow page.
1935 * @param[in] dst page to copy into
1936 * @param[in] src page to copy from
1937 * @param[in] psize size of a page
1940 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1942 enum { Align = sizeof(pgno_t) };
1943 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1945 /* If page isn't full, just copy the used portion. Adjust
1946 * alignment so memcpy may copy words instead of bytes.
1948 if ((unused &= -Align) && !IS_LEAF2(src)) {
1950 memcpy(dst, src, (lower + (Align-1)) & -Align);
1951 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1954 memcpy(dst, src, psize - unused);
1958 /** Pull a page off the txn's spill list, if present.
1959 * If a page being referenced was spilled to disk in this txn, bring
1960 * it back and make it dirty/writable again.
1961 * @param[in] txn the transaction handle.
1962 * @param[in] mp the page being referenced. It must not be dirty.
1963 * @param[out] ret the writable page, if any. ret is unchanged if
1964 * mp wasn't spilled.
1967 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1969 MDB_env *env = txn->mt_env;
1972 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1974 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1975 if (!tx2->mt_spill_pgs)
1977 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1978 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1981 if (txn->mt_dirty_room == 0)
1982 return MDB_TXN_FULL;
1983 if (IS_OVERFLOW(mp))
1987 if (env->me_flags & MDB_WRITEMAP) {
1990 np = mdb_page_malloc(txn, num);
1994 memcpy(np, mp, num * env->me_psize);
1996 mdb_page_copy(np, mp, env->me_psize);
1999 /* If in current txn, this page is no longer spilled.
2000 * If it happens to be the last page, truncate the spill list.
2001 * Otherwise mark it as deleted by setting the LSB.
2003 if (x == txn->mt_spill_pgs[0])
2004 txn->mt_spill_pgs[0]--;
2006 txn->mt_spill_pgs[x] |= 1;
2007 } /* otherwise, if belonging to a parent txn, the
2008 * page remains spilled until child commits
2011 mdb_page_dirty(txn, np);
2012 np->mp_flags |= P_DIRTY;
2020 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2021 * @param[in] mc cursor pointing to the page to be touched
2022 * @return 0 on success, non-zero on failure.
2025 mdb_page_touch(MDB_cursor *mc)
2027 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2028 MDB_txn *txn = mc->mc_txn;
2029 MDB_cursor *m2, *m3;
2033 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2034 if (txn->mt_flags & MDB_TXN_SPILLS) {
2036 rc = mdb_page_unspill(txn, mp, &np);
2042 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2043 (rc = mdb_page_alloc(mc, 1, &np)))
2046 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2047 mp->mp_pgno, pgno));
2048 mdb_cassert(mc, mp->mp_pgno != pgno);
2049 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2050 /* Update the parent page, if any, to point to the new page */
2052 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2053 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2054 SETPGNO(node, pgno);
2056 mc->mc_db->md_root = pgno;
2058 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2059 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2061 /* If txn has a parent, make sure the page is in our
2065 unsigned x = mdb_mid2l_search(dl, pgno);
2066 if (x <= dl[0].mid && dl[x].mid == pgno) {
2067 if (mp != dl[x].mptr) { /* bad cursor? */
2068 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2069 txn->mt_flags |= MDB_TXN_ERROR;
2070 return MDB_CORRUPTED;
2075 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2077 np = mdb_page_malloc(txn, 1);
2082 rc = mdb_mid2l_insert(dl, &mid);
2083 mdb_cassert(mc, rc == 0);
2088 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2090 np->mp_flags |= P_DIRTY;
2093 /* Adjust cursors pointing to mp */
2094 mc->mc_pg[mc->mc_top] = np;
2095 m2 = txn->mt_cursors[mc->mc_dbi];
2096 if (mc->mc_flags & C_SUB) {
2097 for (; m2; m2=m2->mc_next) {
2098 m3 = &m2->mc_xcursor->mx_cursor;
2099 if (m3->mc_snum < mc->mc_snum) continue;
2100 if (m3->mc_pg[mc->mc_top] == mp)
2101 m3->mc_pg[mc->mc_top] = np;
2104 for (; m2; m2=m2->mc_next) {
2105 if (m2->mc_snum < mc->mc_snum) continue;
2106 if (m2->mc_pg[mc->mc_top] == mp) {
2107 m2->mc_pg[mc->mc_top] = np;
2108 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2110 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2112 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2113 if (!(leaf->mn_flags & F_SUBDATA))
2114 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2122 txn->mt_flags |= MDB_TXN_ERROR;
2127 mdb_env_sync(MDB_env *env, int force)
2130 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2131 if (env->me_flags & MDB_WRITEMAP) {
2132 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2133 ? MS_ASYNC : MS_SYNC;
2134 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2137 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2141 if (MDB_FDATASYNC(env->me_fd))
2148 /** Back up parent txn's cursors, then grab the originals for tracking */
2150 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2152 MDB_cursor *mc, *bk;
2157 for (i = src->mt_numdbs; --i >= 0; ) {
2158 if ((mc = src->mt_cursors[i]) != NULL) {
2159 size = sizeof(MDB_cursor);
2161 size += sizeof(MDB_xcursor);
2162 for (; mc; mc = bk->mc_next) {
2168 mc->mc_db = &dst->mt_dbs[i];
2169 /* Kill pointers into src - and dst to reduce abuse: The
2170 * user may not use mc until dst ends. Otherwise we'd...
2172 mc->mc_txn = NULL; /* ...set this to dst */
2173 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2174 if ((mx = mc->mc_xcursor) != NULL) {
2175 *(MDB_xcursor *)(bk+1) = *mx;
2176 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2178 mc->mc_next = dst->mt_cursors[i];
2179 dst->mt_cursors[i] = mc;
2186 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2187 * @param[in] txn the transaction handle.
2188 * @param[in] merge true to keep changes to parent cursors, false to revert.
2189 * @return 0 on success, non-zero on failure.
2192 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2194 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2198 for (i = txn->mt_numdbs; --i >= 0; ) {
2199 for (mc = cursors[i]; mc; mc = next) {
2201 if ((bk = mc->mc_backup) != NULL) {
2203 /* Commit changes to parent txn */
2204 mc->mc_next = bk->mc_next;
2205 mc->mc_backup = bk->mc_backup;
2206 mc->mc_txn = bk->mc_txn;
2207 mc->mc_db = bk->mc_db;
2208 mc->mc_dbflag = bk->mc_dbflag;
2209 if ((mx = mc->mc_xcursor) != NULL)
2210 mx->mx_cursor.mc_txn = bk->mc_txn;
2212 /* Abort nested txn */
2214 if ((mx = mc->mc_xcursor) != NULL)
2215 *mx = *(MDB_xcursor *)(bk+1);
2219 /* Only malloced cursors are permanently tracked. */
2227 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2230 mdb_txn_reset0(MDB_txn *txn, const char *act);
2232 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2238 Pidset = F_SETLK, Pidcheck = F_GETLK
2242 /** Set or check a pid lock. Set returns 0 on success.
2243 * Check returns 0 if the process is certainly dead, nonzero if it may
2244 * be alive (the lock exists or an error happened so we do not know).
2246 * On Windows Pidset is a no-op, we merely check for the existence
2247 * of the process with the given pid. On POSIX we use a single byte
2248 * lock on the lockfile, set at an offset equal to the pid.
2251 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2253 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2256 if (op == Pidcheck) {
2257 h = OpenProcess(env->me_pidquery, FALSE, pid);
2258 /* No documented "no such process" code, but other program use this: */
2260 return ErrCode() != ERROR_INVALID_PARAMETER;
2261 /* A process exists until all handles to it close. Has it exited? */
2262 ret = WaitForSingleObject(h, 0) != 0;
2269 struct flock lock_info;
2270 memset(&lock_info, 0, sizeof(lock_info));
2271 lock_info.l_type = F_WRLCK;
2272 lock_info.l_whence = SEEK_SET;
2273 lock_info.l_start = pid;
2274 lock_info.l_len = 1;
2275 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2276 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2278 } else if ((rc = ErrCode()) == EINTR) {
2286 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2287 * @param[in] txn the transaction handle to initialize
2288 * @return 0 on success, non-zero on failure.
2291 mdb_txn_renew0(MDB_txn *txn)
2293 MDB_env *env = txn->mt_env;
2294 MDB_txninfo *ti = env->me_txns;
2298 int rc, new_notls = 0;
2301 txn->mt_numdbs = env->me_numdbs;
2302 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2304 if (txn->mt_flags & MDB_TXN_RDONLY) {
2306 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2307 txn->mt_txnid = meta->mm_txnid;
2308 txn->mt_u.reader = NULL;
2310 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2311 pthread_getspecific(env->me_txkey);
2313 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2314 return MDB_BAD_RSLOT;
2316 MDB_PID_T pid = env->me_pid;
2317 pthread_t tid = pthread_self();
2319 if (!env->me_live_reader) {
2320 rc = mdb_reader_pid(env, Pidset, pid);
2323 env->me_live_reader = 1;
2327 nr = ti->mti_numreaders;
2328 for (i=0; i<nr; i++)
2329 if (ti->mti_readers[i].mr_pid == 0)
2331 if (i == env->me_maxreaders) {
2332 UNLOCK_MUTEX_R(env);
2333 return MDB_READERS_FULL;
2335 ti->mti_readers[i].mr_pid = pid;
2336 ti->mti_readers[i].mr_tid = tid;
2338 ti->mti_numreaders = ++nr;
2339 /* Save numreaders for un-mutexed mdb_env_close() */
2340 env->me_numreaders = nr;
2341 UNLOCK_MUTEX_R(env);
2343 r = &ti->mti_readers[i];
2344 new_notls = (env->me_flags & MDB_NOTLS);
2345 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2350 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2351 txn->mt_u.reader = r;
2352 meta = env->me_metas[txn->mt_txnid & 1];
2358 txn->mt_txnid = ti->mti_txnid;
2359 meta = env->me_metas[txn->mt_txnid & 1];
2361 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2362 txn->mt_txnid = meta->mm_txnid;
2366 if (txn->mt_txnid == mdb_debug_start)
2369 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2370 txn->mt_u.dirty_list = env->me_dirty_list;
2371 txn->mt_u.dirty_list[0].mid = 0;
2372 txn->mt_free_pgs = env->me_free_pgs;
2373 txn->mt_free_pgs[0] = 0;
2374 txn->mt_spill_pgs = NULL;
2378 /* Copy the DB info and flags */
2379 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2381 /* Moved to here to avoid a data race in read TXNs */
2382 txn->mt_next_pgno = meta->mm_last_pg+1;
2384 for (i=2; i<txn->mt_numdbs; i++) {
2385 x = env->me_dbflags[i];
2386 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2387 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2389 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2391 if (env->me_maxpg < txn->mt_next_pgno) {
2392 mdb_txn_reset0(txn, "renew0-mapfail");
2394 txn->mt_u.reader->mr_pid = 0;
2395 txn->mt_u.reader = NULL;
2397 return MDB_MAP_RESIZED;
2404 mdb_txn_renew(MDB_txn *txn)
2408 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2411 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2412 DPUTS("environment had fatal error, must shutdown!");
2416 rc = mdb_txn_renew0(txn);
2417 if (rc == MDB_SUCCESS) {
2418 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2419 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2420 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2426 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2430 int rc, size, tsize = sizeof(MDB_txn);
2432 if (env->me_flags & MDB_FATAL_ERROR) {
2433 DPUTS("environment had fatal error, must shutdown!");
2436 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2439 /* Nested transactions: Max 1 child, write txns only, no writemap */
2440 if (parent->mt_child ||
2441 (flags & MDB_RDONLY) ||
2442 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2443 (env->me_flags & MDB_WRITEMAP))
2445 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2447 tsize = sizeof(MDB_ntxn);
2449 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2450 if (!(flags & MDB_RDONLY))
2451 size += env->me_maxdbs * sizeof(MDB_cursor *);
2453 if ((txn = calloc(1, size)) == NULL) {
2454 DPRINTF(("calloc: %s", strerror(ErrCode())));
2457 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2458 if (flags & MDB_RDONLY) {
2459 txn->mt_flags |= MDB_TXN_RDONLY;
2460 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2462 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2463 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2469 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2470 if (!txn->mt_u.dirty_list ||
2471 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2473 free(txn->mt_u.dirty_list);
2477 txn->mt_txnid = parent->mt_txnid;
2478 txn->mt_dirty_room = parent->mt_dirty_room;
2479 txn->mt_u.dirty_list[0].mid = 0;
2480 txn->mt_spill_pgs = NULL;
2481 txn->mt_next_pgno = parent->mt_next_pgno;
2482 parent->mt_child = txn;
2483 txn->mt_parent = parent;
2484 txn->mt_numdbs = parent->mt_numdbs;
2485 txn->mt_flags = parent->mt_flags;
2486 txn->mt_dbxs = parent->mt_dbxs;
2487 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2488 /* Copy parent's mt_dbflags, but clear DB_NEW */
2489 for (i=0; i<txn->mt_numdbs; i++)
2490 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2492 ntxn = (MDB_ntxn *)txn;
2493 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2494 if (env->me_pghead) {
2495 size = MDB_IDL_SIZEOF(env->me_pghead);
2496 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2498 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2503 rc = mdb_cursor_shadow(parent, txn);
2505 mdb_txn_reset0(txn, "beginchild-fail");
2507 rc = mdb_txn_renew0(txn);
2513 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2514 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2515 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2522 mdb_txn_env(MDB_txn *txn)
2524 if(!txn) return NULL;
2528 /** Export or close DBI handles opened in this txn. */
2530 mdb_dbis_update(MDB_txn *txn, int keep)
2533 MDB_dbi n = txn->mt_numdbs;
2534 MDB_env *env = txn->mt_env;
2535 unsigned char *tdbflags = txn->mt_dbflags;
2537 for (i = n; --i >= 2;) {
2538 if (tdbflags[i] & DB_NEW) {
2540 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2542 char *ptr = env->me_dbxs[i].md_name.mv_data;
2543 env->me_dbxs[i].md_name.mv_data = NULL;
2544 env->me_dbxs[i].md_name.mv_size = 0;
2545 env->me_dbflags[i] = 0;
2550 if (keep && env->me_numdbs < n)
2554 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2555 * May be called twice for readonly txns: First reset it, then abort.
2556 * @param[in] txn the transaction handle to reset
2557 * @param[in] act why the transaction is being reset
2560 mdb_txn_reset0(MDB_txn *txn, const char *act)
2562 MDB_env *env = txn->mt_env;
2564 /* Close any DBI handles opened in this txn */
2565 mdb_dbis_update(txn, 0);
2567 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2568 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2569 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2571 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2572 if (txn->mt_u.reader) {
2573 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2574 if (!(env->me_flags & MDB_NOTLS))
2575 txn->mt_u.reader = NULL; /* txn does not own reader */
2577 txn->mt_numdbs = 0; /* close nothing if called again */
2578 txn->mt_dbxs = NULL; /* mark txn as reset */
2580 mdb_cursors_close(txn, 0);
2582 if (!(env->me_flags & MDB_WRITEMAP)) {
2583 mdb_dlist_free(txn);
2585 mdb_midl_free(env->me_pghead);
2587 if (txn->mt_parent) {
2588 txn->mt_parent->mt_child = NULL;
2589 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2590 mdb_midl_free(txn->mt_free_pgs);
2591 mdb_midl_free(txn->mt_spill_pgs);
2592 free(txn->mt_u.dirty_list);
2596 if (mdb_midl_shrink(&txn->mt_free_pgs))
2597 env->me_free_pgs = txn->mt_free_pgs;
2598 env->me_pghead = NULL;
2602 /* The writer mutex was locked in mdb_txn_begin. */
2604 UNLOCK_MUTEX_W(env);
2609 mdb_txn_reset(MDB_txn *txn)
2614 /* This call is only valid for read-only txns */
2615 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2618 mdb_txn_reset0(txn, "reset");
2622 mdb_txn_abort(MDB_txn *txn)
2628 mdb_txn_abort(txn->mt_child);
2630 mdb_txn_reset0(txn, "abort");
2631 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2632 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2633 txn->mt_u.reader->mr_pid = 0;
2638 /** Save the freelist as of this transaction to the freeDB.
2639 * This changes the freelist. Keep trying until it stabilizes.
2642 mdb_freelist_save(MDB_txn *txn)
2644 /* env->me_pghead[] can grow and shrink during this call.
2645 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2646 * Page numbers cannot disappear from txn->mt_free_pgs[].
2649 MDB_env *env = txn->mt_env;
2650 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2651 txnid_t pglast = 0, head_id = 0;
2652 pgno_t freecnt = 0, *free_pgs, *mop;
2653 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2655 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2657 if (env->me_pghead) {
2658 /* Make sure first page of freeDB is touched and on freelist */
2659 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2660 if (rc && rc != MDB_NOTFOUND)
2664 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2665 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2666 ? SSIZE_MAX : maxfree_1pg;
2669 /* Come back here after each Put() in case freelist changed */
2674 /* If using records from freeDB which we have not yet
2675 * deleted, delete them and any we reserved for me_pghead.
2677 while (pglast < env->me_pglast) {
2678 rc = mdb_cursor_first(&mc, &key, NULL);
2681 pglast = head_id = *(txnid_t *)key.mv_data;
2682 total_room = head_room = 0;
2683 mdb_tassert(txn, pglast <= env->me_pglast);
2684 rc = mdb_cursor_del(&mc, 0);
2689 /* Save the IDL of pages freed by this txn, to a single record */
2690 if (freecnt < txn->mt_free_pgs[0]) {
2692 /* Make sure last page of freeDB is touched and on freelist */
2693 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2694 if (rc && rc != MDB_NOTFOUND)
2697 free_pgs = txn->mt_free_pgs;
2698 /* Write to last page of freeDB */
2699 key.mv_size = sizeof(txn->mt_txnid);
2700 key.mv_data = &txn->mt_txnid;
2702 freecnt = free_pgs[0];
2703 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2704 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2707 /* Retry if mt_free_pgs[] grew during the Put() */
2708 free_pgs = txn->mt_free_pgs;
2709 } while (freecnt < free_pgs[0]);
2710 mdb_midl_sort(free_pgs);
2711 memcpy(data.mv_data, free_pgs, data.mv_size);
2714 unsigned int i = free_pgs[0];
2715 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2716 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2718 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2724 mop = env->me_pghead;
2725 mop_len = mop ? mop[0] : 0;
2727 /* Reserve records for me_pghead[]. Split it if multi-page,
2728 * to avoid searching freeDB for a page range. Use keys in
2729 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2731 if (total_room >= mop_len) {
2732 if (total_room == mop_len || --more < 0)
2734 } else if (head_room >= maxfree_1pg && head_id > 1) {
2735 /* Keep current record (overflow page), add a new one */
2739 /* (Re)write {key = head_id, IDL length = head_room} */
2740 total_room -= head_room;
2741 head_room = mop_len - total_room;
2742 if (head_room > maxfree_1pg && head_id > 1) {
2743 /* Overflow multi-page for part of me_pghead */
2744 head_room /= head_id; /* amortize page sizes */
2745 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2746 } else if (head_room < 0) {
2747 /* Rare case, not bothering to delete this record */
2750 key.mv_size = sizeof(head_id);
2751 key.mv_data = &head_id;
2752 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2753 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2756 /* IDL is initially empty, zero out at least the length */
2757 pgs = (pgno_t *)data.mv_data;
2758 j = head_room > clean_limit ? head_room : 0;
2762 total_room += head_room;
2765 /* Fill in the reserved me_pghead records */
2771 rc = mdb_cursor_first(&mc, &key, &data);
2772 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2773 txnid_t id = *(txnid_t *)key.mv_data;
2774 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2777 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2779 if (len > mop_len) {
2781 data.mv_size = (len + 1) * sizeof(MDB_ID);
2783 data.mv_data = mop -= len;
2786 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2788 if (rc || !(mop_len -= len))
2795 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2796 * @param[in] txn the transaction that's being committed
2797 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2798 * @return 0 on success, non-zero on failure.
2801 mdb_page_flush(MDB_txn *txn, int keep)
2803 MDB_env *env = txn->mt_env;
2804 MDB_ID2L dl = txn->mt_u.dirty_list;
2805 unsigned psize = env->me_psize, j;
2806 int i, pagecount = dl[0].mid, rc;
2807 size_t size = 0, pos = 0;
2809 MDB_page *dp = NULL;
2813 struct iovec iov[MDB_COMMIT_PAGES];
2814 ssize_t wpos = 0, wsize = 0, wres;
2815 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2821 if (env->me_flags & MDB_WRITEMAP) {
2822 /* Clear dirty flags */
2823 while (++i <= pagecount) {
2825 /* Don't flush this page yet */
2826 if (dp->mp_flags & P_KEEP) {
2827 dp->mp_flags ^= P_KEEP;
2831 dp->mp_flags &= ~P_DIRTY;
2836 /* Write the pages */
2838 if (++i <= pagecount) {
2840 /* Don't flush this page yet */
2841 if (dp->mp_flags & P_KEEP) {
2842 dp->mp_flags ^= P_KEEP;
2847 /* clear dirty flag */
2848 dp->mp_flags &= ~P_DIRTY;
2851 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2856 /* Windows actually supports scatter/gather I/O, but only on
2857 * unbuffered file handles. Since we're relying on the OS page
2858 * cache for all our data, that's self-defeating. So we just
2859 * write pages one at a time. We use the ov structure to set
2860 * the write offset, to at least save the overhead of a Seek
2863 DPRINTF(("committing page %"Z"u", pgno));
2864 memset(&ov, 0, sizeof(ov));
2865 ov.Offset = pos & 0xffffffff;
2866 ov.OffsetHigh = pos >> 16 >> 16;
2867 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2869 DPRINTF(("WriteFile: %d", rc));
2873 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2874 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2876 /* Write previous page(s) */
2877 #ifdef MDB_USE_PWRITEV
2878 wres = pwritev(env->me_fd, iov, n, wpos);
2881 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2883 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2885 DPRINTF(("lseek: %s", strerror(rc)));
2888 wres = writev(env->me_fd, iov, n);
2891 if (wres != wsize) {
2894 DPRINTF(("Write error: %s", strerror(rc)));
2896 rc = EIO; /* TODO: Use which error code? */
2897 DPUTS("short write, filesystem full?");
2908 DPRINTF(("committing page %"Z"u", pgno));
2909 next_pos = pos + size;
2910 iov[n].iov_len = size;
2911 iov[n].iov_base = (char *)dp;
2917 for (i = keep; ++i <= pagecount; ) {
2919 /* This is a page we skipped above */
2922 dl[j].mid = dp->mp_pgno;
2925 mdb_dpage_free(env, dp);
2930 txn->mt_dirty_room += i - j;
2936 mdb_txn_commit(MDB_txn *txn)
2942 if (txn == NULL || txn->mt_env == NULL)
2945 if (txn->mt_child) {
2946 rc = mdb_txn_commit(txn->mt_child);
2947 txn->mt_child = NULL;
2954 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2955 mdb_dbis_update(txn, 1);
2956 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2961 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2962 DPUTS("error flag is set, can't commit");
2964 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2969 if (txn->mt_parent) {
2970 MDB_txn *parent = txn->mt_parent;
2973 unsigned x, y, len, ps_len;
2975 /* Append our free list to parent's */
2976 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2979 mdb_midl_free(txn->mt_free_pgs);
2980 /* Failures after this must either undo the changes
2981 * to the parent or set MDB_TXN_ERROR in the parent.
2984 parent->mt_next_pgno = txn->mt_next_pgno;
2985 parent->mt_flags = txn->mt_flags;
2987 /* Merge our cursors into parent's and close them */
2988 mdb_cursors_close(txn, 1);
2990 /* Update parent's DB table. */
2991 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2992 parent->mt_numdbs = txn->mt_numdbs;
2993 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2994 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2995 for (i=2; i<txn->mt_numdbs; i++) {
2996 /* preserve parent's DB_NEW status */
2997 x = parent->mt_dbflags[i] & DB_NEW;
2998 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3001 dst = parent->mt_u.dirty_list;
3002 src = txn->mt_u.dirty_list;
3003 /* Remove anything in our dirty list from parent's spill list */
3004 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3006 pspill[0] = (pgno_t)-1;
3007 /* Mark our dirty pages as deleted in parent spill list */
3008 for (i=0, len=src[0].mid; ++i <= len; ) {
3009 MDB_ID pn = src[i].mid << 1;
3010 while (pn > pspill[x])
3012 if (pn == pspill[x]) {
3017 /* Squash deleted pagenums if we deleted any */
3018 for (x=y; ++x <= ps_len; )
3019 if (!(pspill[x] & 1))
3020 pspill[++y] = pspill[x];
3024 /* Find len = length of merging our dirty list with parent's */
3026 dst[0].mid = 0; /* simplify loops */
3027 if (parent->mt_parent) {
3028 len = x + src[0].mid;
3029 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3030 for (i = x; y && i; y--) {
3031 pgno_t yp = src[y].mid;
3032 while (yp < dst[i].mid)
3034 if (yp == dst[i].mid) {
3039 } else { /* Simplify the above for single-ancestor case */
3040 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3042 /* Merge our dirty list with parent's */
3044 for (i = len; y; dst[i--] = src[y--]) {
3045 pgno_t yp = src[y].mid;
3046 while (yp < dst[x].mid)
3047 dst[i--] = dst[x--];
3048 if (yp == dst[x].mid)
3049 free(dst[x--].mptr);
3051 mdb_tassert(txn, i == x);
3053 free(txn->mt_u.dirty_list);
3054 parent->mt_dirty_room = txn->mt_dirty_room;
3055 if (txn->mt_spill_pgs) {
3056 if (parent->mt_spill_pgs) {
3057 /* TODO: Prevent failure here, so parent does not fail */
3058 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3060 parent->mt_flags |= MDB_TXN_ERROR;
3061 mdb_midl_free(txn->mt_spill_pgs);
3062 mdb_midl_sort(parent->mt_spill_pgs);
3064 parent->mt_spill_pgs = txn->mt_spill_pgs;
3068 parent->mt_child = NULL;
3069 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3074 if (txn != env->me_txn) {
3075 DPUTS("attempt to commit unknown transaction");
3080 mdb_cursors_close(txn, 0);
3082 if (!txn->mt_u.dirty_list[0].mid &&
3083 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3086 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3087 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3089 /* Update DB root pointers */
3090 if (txn->mt_numdbs > 2) {
3094 data.mv_size = sizeof(MDB_db);
3096 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3097 for (i = 2; i < txn->mt_numdbs; i++) {
3098 if (txn->mt_dbflags[i] & DB_DIRTY) {
3099 data.mv_data = &txn->mt_dbs[i];
3100 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3107 rc = mdb_freelist_save(txn);
3111 mdb_midl_free(env->me_pghead);
3112 env->me_pghead = NULL;
3113 if (mdb_midl_shrink(&txn->mt_free_pgs))
3114 env->me_free_pgs = txn->mt_free_pgs;
3120 if ((rc = mdb_page_flush(txn, 0)) ||
3121 (rc = mdb_env_sync(env, 0)) ||
3122 (rc = mdb_env_write_meta(txn)))
3128 mdb_dbis_update(txn, 1);
3131 UNLOCK_MUTEX_W(env);
3141 /** Read the environment parameters of a DB environment before
3142 * mapping it into memory.
3143 * @param[in] env the environment handle
3144 * @param[out] meta address of where to store the meta information
3145 * @return 0 on success, non-zero on failure.
3148 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3154 enum { Size = sizeof(pbuf) };
3156 /* We don't know the page size yet, so use a minimum value.
3157 * Read both meta pages so we can use the latest one.
3160 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3164 memset(&ov, 0, sizeof(ov));
3166 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3167 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3170 rc = pread(env->me_fd, &pbuf, Size, off);
3173 if (rc == 0 && off == 0)
3175 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3176 DPRINTF(("read: %s", mdb_strerror(rc)));
3180 p = (MDB_page *)&pbuf;
3182 if (!F_ISSET(p->mp_flags, P_META)) {
3183 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3188 if (m->mm_magic != MDB_MAGIC) {
3189 DPUTS("meta has invalid magic");
3193 if (m->mm_version != MDB_DATA_VERSION) {
3194 DPRINTF(("database is version %u, expected version %u",
3195 m->mm_version, MDB_DATA_VERSION));
3196 return MDB_VERSION_MISMATCH;
3199 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3205 /** Write the environment parameters of a freshly created DB environment.
3206 * @param[in] env the environment handle
3207 * @param[out] meta address of where to store the meta information
3208 * @return 0 on success, non-zero on failure.
3211 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3219 memset(&ov, 0, sizeof(ov));
3220 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3222 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3225 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3226 len = pwrite(fd, ptr, size, pos); \
3227 rc = (len >= 0); } while(0)
3230 DPUTS("writing new meta page");
3232 psize = env->me_psize;
3234 meta->mm_magic = MDB_MAGIC;
3235 meta->mm_version = MDB_DATA_VERSION;
3236 meta->mm_mapsize = env->me_mapsize;
3237 meta->mm_psize = psize;
3238 meta->mm_last_pg = 1;
3239 meta->mm_flags = env->me_flags & 0xffff;
3240 meta->mm_flags |= MDB_INTEGERKEY;
3241 meta->mm_dbs[0].md_root = P_INVALID;
3242 meta->mm_dbs[1].md_root = P_INVALID;
3244 p = calloc(2, psize);
3246 p->mp_flags = P_META;
3247 *(MDB_meta *)METADATA(p) = *meta;
3249 q = (MDB_page *)((char *)p + psize);
3251 q->mp_flags = P_META;
3252 *(MDB_meta *)METADATA(q) = *meta;
3254 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3257 else if ((unsigned) len == psize * 2)
3265 /** Update the environment info to commit a transaction.
3266 * @param[in] txn the transaction that's being committed
3267 * @return 0 on success, non-zero on failure.
3270 mdb_env_write_meta(MDB_txn *txn)
3273 MDB_meta meta, metab, *mp;
3275 int rc, len, toggle;
3284 toggle = txn->mt_txnid & 1;
3285 DPRINTF(("writing meta page %d for root page %"Z"u",
3286 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3289 mp = env->me_metas[toggle];
3291 if (env->me_flags & MDB_WRITEMAP) {
3292 /* Persist any increases of mapsize config */
3293 if (env->me_mapsize > mp->mm_mapsize)
3294 mp->mm_mapsize = env->me_mapsize;
3295 mp->mm_dbs[0] = txn->mt_dbs[0];
3296 mp->mm_dbs[1] = txn->mt_dbs[1];
3297 mp->mm_last_pg = txn->mt_next_pgno - 1;
3298 mp->mm_txnid = txn->mt_txnid;
3299 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3300 unsigned meta_size = env->me_psize;
3301 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3304 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3305 if (meta_size < env->me_os_psize)
3306 meta_size += meta_size;
3311 if (MDB_MSYNC(ptr, meta_size, rc)) {
3318 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3319 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3321 ptr = (char *)&meta;
3322 if (env->me_mapsize > mp->mm_mapsize) {
3323 /* Persist any increases of mapsize config */
3324 meta.mm_mapsize = env->me_mapsize;
3325 off = offsetof(MDB_meta, mm_mapsize);
3327 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3329 len = sizeof(MDB_meta) - off;
3332 meta.mm_dbs[0] = txn->mt_dbs[0];
3333 meta.mm_dbs[1] = txn->mt_dbs[1];
3334 meta.mm_last_pg = txn->mt_next_pgno - 1;
3335 meta.mm_txnid = txn->mt_txnid;
3338 off += env->me_psize;
3341 /* Write to the SYNC fd */
3342 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3343 env->me_fd : env->me_mfd;
3346 memset(&ov, 0, sizeof(ov));
3348 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3352 rc = pwrite(mfd, ptr, len, off);
3355 rc = rc < 0 ? ErrCode() : EIO;
3356 DPUTS("write failed, disk error?");
3357 /* On a failure, the pagecache still contains the new data.
3358 * Write some old data back, to prevent it from being used.
3359 * Use the non-SYNC fd; we know it will fail anyway.
3361 meta.mm_last_pg = metab.mm_last_pg;
3362 meta.mm_txnid = metab.mm_txnid;
3364 memset(&ov, 0, sizeof(ov));
3366 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3368 r2 = pwrite(env->me_fd, ptr, len, off);
3369 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3372 env->me_flags |= MDB_FATAL_ERROR;
3376 /* Memory ordering issues are irrelevant; since the entire writer
3377 * is wrapped by wmutex, all of these changes will become visible
3378 * after the wmutex is unlocked. Since the DB is multi-version,
3379 * readers will get consistent data regardless of how fresh or
3380 * how stale their view of these values is.
3383 env->me_txns->mti_txnid = txn->mt_txnid;
3388 /** Check both meta pages to see which one is newer.
3389 * @param[in] env the environment handle
3390 * @return meta toggle (0 or 1).
3393 mdb_env_pick_meta(const MDB_env *env)
3395 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3399 mdb_env_create(MDB_env **env)
3403 e = calloc(1, sizeof(MDB_env));
3407 e->me_maxreaders = DEFAULT_READERS;
3408 e->me_maxdbs = e->me_numdbs = 2;
3409 e->me_fd = INVALID_HANDLE_VALUE;
3410 e->me_lfd = INVALID_HANDLE_VALUE;
3411 e->me_mfd = INVALID_HANDLE_VALUE;
3412 #ifdef MDB_USE_POSIX_SEM
3413 e->me_rmutex = SEM_FAILED;
3414 e->me_wmutex = SEM_FAILED;
3416 e->me_pid = getpid();
3417 GET_PAGESIZE(e->me_os_psize);
3418 VGMEMP_CREATE(e,0,0);
3424 mdb_env_map(MDB_env *env, void *addr, int newsize)
3427 unsigned int flags = env->me_flags;
3431 LONG sizelo, sizehi;
3432 sizelo = env->me_mapsize & 0xffffffff;
3433 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3435 /* Windows won't create mappings for zero length files.
3436 * Just allocate the maxsize right now.
3439 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3440 || !SetEndOfFile(env->me_fd)
3441 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3444 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3445 PAGE_READWRITE : PAGE_READONLY,
3446 sizehi, sizelo, NULL);
3449 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3450 FILE_MAP_WRITE : FILE_MAP_READ,
3451 0, 0, env->me_mapsize, addr);
3452 rc = env->me_map ? 0 : ErrCode();
3457 int prot = PROT_READ;
3458 if (flags & MDB_WRITEMAP) {
3460 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3463 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3465 if (env->me_map == MAP_FAILED) {
3470 if (flags & MDB_NORDAHEAD) {
3471 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3473 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3475 #ifdef POSIX_MADV_RANDOM
3476 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3477 #endif /* POSIX_MADV_RANDOM */
3478 #endif /* MADV_RANDOM */
3482 /* Can happen because the address argument to mmap() is just a
3483 * hint. mmap() can pick another, e.g. if the range is in use.
3484 * The MAP_FIXED flag would prevent that, but then mmap could
3485 * instead unmap existing pages to make room for the new map.
3487 if (addr && env->me_map != addr)
3488 return EBUSY; /* TODO: Make a new MDB_* error code? */
3490 p = (MDB_page *)env->me_map;
3491 env->me_metas[0] = METADATA(p);
3492 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3498 mdb_env_set_mapsize(MDB_env *env, size_t size)
3500 /* If env is already open, caller is responsible for making
3501 * sure there are no active txns.
3509 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3510 else if (size < env->me_mapsize) {
3511 /* If the configured size is smaller, make sure it's
3512 * still big enough. Silently round up to minimum if not.
3514 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3518 munmap(env->me_map, env->me_mapsize);
3519 env->me_mapsize = size;
3520 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3521 rc = mdb_env_map(env, old, 1);
3525 env->me_mapsize = size;
3527 env->me_maxpg = env->me_mapsize / env->me_psize;
3532 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3536 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3541 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3543 if (env->me_map || readers < 1)
3545 env->me_maxreaders = readers;
3550 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3552 if (!env || !readers)
3554 *readers = env->me_maxreaders;
3558 /** Further setup required for opening an MDB environment
3561 mdb_env_open2(MDB_env *env)
3563 unsigned int flags = env->me_flags;
3564 int i, newenv = 0, rc;
3568 /* See if we should use QueryLimited */
3570 if ((rc & 0xff) > 5)
3571 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3573 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3576 memset(&meta, 0, sizeof(meta));
3578 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3581 DPUTS("new mdbenv");
3583 env->me_psize = env->me_os_psize;
3584 if (env->me_psize > MAX_PAGESIZE)
3585 env->me_psize = MAX_PAGESIZE;
3587 env->me_psize = meta.mm_psize;
3590 /* Was a mapsize configured? */
3591 if (!env->me_mapsize) {
3592 /* If this is a new environment, take the default,
3593 * else use the size recorded in the existing env.
3595 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3596 } else if (env->me_mapsize < meta.mm_mapsize) {
3597 /* If the configured size is smaller, make sure it's
3598 * still big enough. Silently round up to minimum if not.
3600 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3601 if (env->me_mapsize < minsize)
3602 env->me_mapsize = minsize;
3605 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3610 if (flags & MDB_FIXEDMAP)
3611 meta.mm_address = env->me_map;
3612 i = mdb_env_init_meta(env, &meta);
3613 if (i != MDB_SUCCESS) {
3618 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3619 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3621 #if !(MDB_MAXKEYSIZE)
3622 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3624 env->me_maxpg = env->me_mapsize / env->me_psize;
3628 int toggle = mdb_env_pick_meta(env);
3629 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3631 DPRINTF(("opened database version %u, pagesize %u",
3632 env->me_metas[0]->mm_version, env->me_psize));
3633 DPRINTF(("using meta page %d", toggle));
3634 DPRINTF(("depth: %u", db->md_depth));
3635 DPRINTF(("entries: %"Z"u", db->md_entries));
3636 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3637 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3638 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3639 DPRINTF(("root: %"Z"u", db->md_root));
3647 /** Release a reader thread's slot in the reader lock table.
3648 * This function is called automatically when a thread exits.
3649 * @param[in] ptr This points to the slot in the reader lock table.
3652 mdb_env_reader_dest(void *ptr)
3654 MDB_reader *reader = ptr;
3660 /** Junk for arranging thread-specific callbacks on Windows. This is
3661 * necessarily platform and compiler-specific. Windows supports up
3662 * to 1088 keys. Let's assume nobody opens more than 64 environments
3663 * in a single process, for now. They can override this if needed.
3665 #ifndef MAX_TLS_KEYS
3666 #define MAX_TLS_KEYS 64
3668 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3669 static int mdb_tls_nkeys;
3671 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3675 case DLL_PROCESS_ATTACH: break;
3676 case DLL_THREAD_ATTACH: break;
3677 case DLL_THREAD_DETACH:
3678 for (i=0; i<mdb_tls_nkeys; i++) {
3679 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3681 mdb_env_reader_dest(r);
3685 case DLL_PROCESS_DETACH: break;
3690 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3692 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3696 /* Force some symbol references.
3697 * _tls_used forces the linker to create the TLS directory if not already done
3698 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3700 #pragma comment(linker, "/INCLUDE:_tls_used")
3701 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3702 #pragma const_seg(".CRT$XLB")
3703 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3704 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3707 #pragma comment(linker, "/INCLUDE:__tls_used")
3708 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3709 #pragma data_seg(".CRT$XLB")
3710 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3712 #endif /* WIN 32/64 */
3713 #endif /* !__GNUC__ */
3716 /** Downgrade the exclusive lock on the region back to shared */
3718 mdb_env_share_locks(MDB_env *env, int *excl)
3720 int rc = 0, toggle = mdb_env_pick_meta(env);
3722 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3727 /* First acquire a shared lock. The Unlock will
3728 * then release the existing exclusive lock.
3730 memset(&ov, 0, sizeof(ov));
3731 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3734 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3740 struct flock lock_info;
3741 /* The shared lock replaces the existing lock */
3742 memset((void *)&lock_info, 0, sizeof(lock_info));
3743 lock_info.l_type = F_RDLCK;
3744 lock_info.l_whence = SEEK_SET;
3745 lock_info.l_start = 0;
3746 lock_info.l_len = 1;
3747 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3748 (rc = ErrCode()) == EINTR) ;
3749 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3756 /** Try to get exlusive lock, otherwise shared.
3757 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3760 mdb_env_excl_lock(MDB_env *env, int *excl)
3764 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3768 memset(&ov, 0, sizeof(ov));
3769 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3776 struct flock lock_info;
3777 memset((void *)&lock_info, 0, sizeof(lock_info));
3778 lock_info.l_type = F_WRLCK;
3779 lock_info.l_whence = SEEK_SET;
3780 lock_info.l_start = 0;
3781 lock_info.l_len = 1;
3782 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3783 (rc = ErrCode()) == EINTR) ;
3787 # ifdef MDB_USE_POSIX_SEM
3788 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3791 lock_info.l_type = F_RDLCK;
3792 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3793 (rc = ErrCode()) == EINTR) ;
3803 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3805 * @(#) $Revision: 5.1 $
3806 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3807 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3809 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3813 * Please do not copyright this code. This code is in the public domain.
3815 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3816 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3817 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3818 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3819 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3820 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3821 * PERFORMANCE OF THIS SOFTWARE.
3824 * chongo <Landon Curt Noll> /\oo/\
3825 * http://www.isthe.com/chongo/
3827 * Share and Enjoy! :-)
3830 typedef unsigned long long mdb_hash_t;
3831 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3833 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3834 * @param[in] val value to hash
3835 * @param[in] hval initial value for hash
3836 * @return 64 bit hash
3838 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3839 * hval arg on the first call.
3842 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3844 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3845 unsigned char *end = s + val->mv_size;
3847 * FNV-1a hash each octet of the string
3850 /* xor the bottom with the current octet */
3851 hval ^= (mdb_hash_t)*s++;
3853 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3854 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3855 (hval << 7) + (hval << 8) + (hval << 40);
3857 /* return our new hash value */
3861 /** Hash the string and output the encoded hash.
3862 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3863 * very short name limits. We don't care about the encoding being reversible,
3864 * we just want to preserve as many bits of the input as possible in a
3865 * small printable string.
3866 * @param[in] str string to hash
3867 * @param[out] encbuf an array of 11 chars to hold the hash
3869 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3872 mdb_pack85(unsigned long l, char *out)
3876 for (i=0; i<5; i++) {
3877 *out++ = mdb_a85[l % 85];
3883 mdb_hash_enc(MDB_val *val, char *encbuf)
3885 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3887 mdb_pack85(h, encbuf);
3888 mdb_pack85(h>>32, encbuf+5);
3893 /** Open and/or initialize the lock region for the environment.
3894 * @param[in] env The MDB environment.
3895 * @param[in] lpath The pathname of the file used for the lock region.
3896 * @param[in] mode The Unix permissions for the file, if we create it.
3897 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3898 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3899 * @return 0 on success, non-zero on failure.
3902 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3905 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3907 # define MDB_ERRCODE_ROFS EROFS
3908 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3909 # define MDB_CLOEXEC O_CLOEXEC
3912 # define MDB_CLOEXEC 0
3919 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3920 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3921 FILE_ATTRIBUTE_NORMAL, NULL);
3923 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3925 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3927 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3932 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3933 /* Lose record locks when exec*() */
3934 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3935 fcntl(env->me_lfd, F_SETFD, fdflags);
3938 if (!(env->me_flags & MDB_NOTLS)) {
3939 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3942 env->me_flags |= MDB_ENV_TXKEY;
3944 /* Windows TLS callbacks need help finding their TLS info. */
3945 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3949 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3953 /* Try to get exclusive lock. If we succeed, then
3954 * nobody is using the lock region and we should initialize it.
3956 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3959 size = GetFileSize(env->me_lfd, NULL);
3961 size = lseek(env->me_lfd, 0, SEEK_END);
3962 if (size == -1) goto fail_errno;
3964 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3965 if (size < rsize && *excl > 0) {
3967 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3968 || !SetEndOfFile(env->me_lfd))
3971 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3975 size = rsize - sizeof(MDB_txninfo);
3976 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3981 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3983 if (!mh) goto fail_errno;
3984 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3986 if (!env->me_txns) goto fail_errno;
3988 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3990 if (m == MAP_FAILED) goto fail_errno;
3996 BY_HANDLE_FILE_INFORMATION stbuf;
4005 if (!mdb_sec_inited) {
4006 InitializeSecurityDescriptor(&mdb_null_sd,
4007 SECURITY_DESCRIPTOR_REVISION);
4008 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4009 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4010 mdb_all_sa.bInheritHandle = FALSE;
4011 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4014 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4015 idbuf.volume = stbuf.dwVolumeSerialNumber;
4016 idbuf.nhigh = stbuf.nFileIndexHigh;
4017 idbuf.nlow = stbuf.nFileIndexLow;
4018 val.mv_data = &idbuf;
4019 val.mv_size = sizeof(idbuf);
4020 mdb_hash_enc(&val, encbuf);
4021 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4022 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4023 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4024 if (!env->me_rmutex) goto fail_errno;
4025 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4026 if (!env->me_wmutex) goto fail_errno;
4027 #elif defined(MDB_USE_POSIX_SEM)
4036 #if defined(__NetBSD__)
4037 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4039 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4040 idbuf.dev = stbuf.st_dev;
4041 idbuf.ino = stbuf.st_ino;
4042 val.mv_data = &idbuf;
4043 val.mv_size = sizeof(idbuf);
4044 mdb_hash_enc(&val, encbuf);
4045 #ifdef MDB_SHORT_SEMNAMES
4046 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4048 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4049 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4050 /* Clean up after a previous run, if needed: Try to
4051 * remove both semaphores before doing anything else.
4053 sem_unlink(env->me_txns->mti_rmname);
4054 sem_unlink(env->me_txns->mti_wmname);
4055 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4056 O_CREAT|O_EXCL, mode, 1);
4057 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4058 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4059 O_CREAT|O_EXCL, mode, 1);
4060 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4061 #else /* MDB_USE_POSIX_SEM */
4062 pthread_mutexattr_t mattr;
4064 if ((rc = pthread_mutexattr_init(&mattr))
4065 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4066 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4067 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4069 pthread_mutexattr_destroy(&mattr);
4070 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4072 env->me_txns->mti_magic = MDB_MAGIC;
4073 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4074 env->me_txns->mti_txnid = 0;
4075 env->me_txns->mti_numreaders = 0;
4078 if (env->me_txns->mti_magic != MDB_MAGIC) {
4079 DPUTS("lock region has invalid magic");
4083 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4084 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4085 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4086 rc = MDB_VERSION_MISMATCH;
4090 if (rc && rc != EACCES && rc != EAGAIN) {
4094 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4095 if (!env->me_rmutex) goto fail_errno;
4096 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4097 if (!env->me_wmutex) goto fail_errno;
4098 #elif defined(MDB_USE_POSIX_SEM)
4099 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4100 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4101 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4102 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4113 /** The name of the lock file in the DB environment */
4114 #define LOCKNAME "/lock.mdb"
4115 /** The name of the data file in the DB environment */
4116 #define DATANAME "/data.mdb"
4117 /** The suffix of the lock file when no subdir is used */
4118 #define LOCKSUFF "-lock"
4119 /** Only a subset of the @ref mdb_env flags can be changed
4120 * at runtime. Changing other flags requires closing the
4121 * environment and re-opening it with the new flags.
4123 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4124 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4125 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4127 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4128 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4132 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4134 int oflags, rc, len, excl = -1;
4135 char *lpath, *dpath;
4137 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4141 if (flags & MDB_NOSUBDIR) {
4142 rc = len + sizeof(LOCKSUFF) + len + 1;
4144 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4149 if (flags & MDB_NOSUBDIR) {
4150 dpath = lpath + len + sizeof(LOCKSUFF);
4151 sprintf(lpath, "%s" LOCKSUFF, path);
4152 strcpy(dpath, path);
4154 dpath = lpath + len + sizeof(LOCKNAME);
4155 sprintf(lpath, "%s" LOCKNAME, path);
4156 sprintf(dpath, "%s" DATANAME, path);
4160 flags |= env->me_flags;
4161 if (flags & MDB_RDONLY) {
4162 /* silently ignore WRITEMAP when we're only getting read access */
4163 flags &= ~MDB_WRITEMAP;
4165 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4166 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4169 env->me_flags = flags |= MDB_ENV_ACTIVE;
4173 env->me_path = strdup(path);
4174 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4175 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4176 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4181 /* For RDONLY, get lockfile after we know datafile exists */
4182 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4183 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4189 if (F_ISSET(flags, MDB_RDONLY)) {
4190 oflags = GENERIC_READ;
4191 len = OPEN_EXISTING;
4193 oflags = GENERIC_READ|GENERIC_WRITE;
4196 mode = FILE_ATTRIBUTE_NORMAL;
4197 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4198 NULL, len, mode, NULL);
4200 if (F_ISSET(flags, MDB_RDONLY))
4203 oflags = O_RDWR | O_CREAT;
4205 env->me_fd = open(dpath, oflags, mode);
4207 if (env->me_fd == INVALID_HANDLE_VALUE) {
4212 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4213 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4218 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4219 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4220 env->me_mfd = env->me_fd;
4222 /* Synchronous fd for meta writes. Needed even with
4223 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4226 len = OPEN_EXISTING;
4227 env->me_mfd = CreateFile(dpath, oflags,
4228 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4229 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4232 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4234 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4239 DPRINTF(("opened dbenv %p", (void *) env));
4241 rc = mdb_env_share_locks(env, &excl);
4245 if (!((flags & MDB_RDONLY) ||
4246 (env->me_pbuf = calloc(1, env->me_psize))))
4252 mdb_env_close0(env, excl);
4258 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4260 mdb_env_close0(MDB_env *env, int excl)
4264 if (!(env->me_flags & MDB_ENV_ACTIVE))
4267 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4268 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4269 free(env->me_dbxs[i].md_name.mv_data);
4272 free(env->me_dbflags);
4275 free(env->me_dirty_list);
4276 mdb_midl_free(env->me_free_pgs);
4278 if (env->me_flags & MDB_ENV_TXKEY) {
4279 pthread_key_delete(env->me_txkey);
4281 /* Delete our key from the global list */
4282 for (i=0; i<mdb_tls_nkeys; i++)
4283 if (mdb_tls_keys[i] == env->me_txkey) {
4284 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4292 munmap(env->me_map, env->me_mapsize);
4294 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4295 (void) close(env->me_mfd);
4296 if (env->me_fd != INVALID_HANDLE_VALUE)
4297 (void) close(env->me_fd);
4299 MDB_PID_T pid = env->me_pid;
4300 /* Clearing readers is done in this function because
4301 * me_txkey with its destructor must be disabled first.
4303 for (i = env->me_numreaders; --i >= 0; )
4304 if (env->me_txns->mti_readers[i].mr_pid == pid)
4305 env->me_txns->mti_readers[i].mr_pid = 0;
4307 if (env->me_rmutex) {
4308 CloseHandle(env->me_rmutex);
4309 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4311 /* Windows automatically destroys the mutexes when
4312 * the last handle closes.
4314 #elif defined(MDB_USE_POSIX_SEM)
4315 if (env->me_rmutex != SEM_FAILED) {
4316 sem_close(env->me_rmutex);
4317 if (env->me_wmutex != SEM_FAILED)
4318 sem_close(env->me_wmutex);
4319 /* If we have the filelock: If we are the
4320 * only remaining user, clean up semaphores.
4323 mdb_env_excl_lock(env, &excl);
4325 sem_unlink(env->me_txns->mti_rmname);
4326 sem_unlink(env->me_txns->mti_wmname);
4330 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4332 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4335 /* Unlock the lockfile. Windows would have unlocked it
4336 * after closing anyway, but not necessarily at once.
4338 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4341 (void) close(env->me_lfd);
4344 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4348 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4350 MDB_txn *txn = NULL;
4356 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4360 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4363 /* Do the lock/unlock of the reader mutex before starting the
4364 * write txn. Otherwise other read txns could block writers.
4366 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4371 /* We must start the actual read txn after blocking writers */
4372 mdb_txn_reset0(txn, "reset-stage1");
4374 /* Temporarily block writers until we snapshot the meta pages */
4377 rc = mdb_txn_renew0(txn);
4379 UNLOCK_MUTEX_W(env);
4384 wsize = env->me_psize * 2;
4388 DO_WRITE(rc, fd, ptr, w2, len);
4392 } else if (len > 0) {
4398 /* Non-blocking or async handles are not supported */
4404 UNLOCK_MUTEX_W(env);
4409 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4411 if (wsize > MAX_WRITE)
4415 DO_WRITE(rc, fd, ptr, w2, len);
4419 } else if (len > 0) {
4436 mdb_env_copy(MDB_env *env, const char *path)
4440 HANDLE newfd = INVALID_HANDLE_VALUE;
4442 if (env->me_flags & MDB_NOSUBDIR) {
4443 lpath = (char *)path;
4446 len += sizeof(DATANAME);
4447 lpath = malloc(len);
4450 sprintf(lpath, "%s" DATANAME, path);
4453 /* The destination path must exist, but the destination file must not.
4454 * We don't want the OS to cache the writes, since the source data is
4455 * already in the OS cache.
4458 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4459 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4461 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4463 if (newfd == INVALID_HANDLE_VALUE) {
4469 /* Set O_DIRECT if the file system supports it */
4470 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4471 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4473 #ifdef F_NOCACHE /* __APPLE__ */
4474 rc = fcntl(newfd, F_NOCACHE, 1);
4481 rc = mdb_env_copyfd(env, newfd);
4484 if (!(env->me_flags & MDB_NOSUBDIR))
4486 if (newfd != INVALID_HANDLE_VALUE)
4487 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4494 mdb_env_close(MDB_env *env)
4501 VGMEMP_DESTROY(env);
4502 while ((dp = env->me_dpages) != NULL) {
4503 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4504 env->me_dpages = dp->mp_next;
4508 mdb_env_close0(env, 0);
4512 /** Compare two items pointing at aligned size_t's */
4514 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4516 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4517 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4520 /** Compare two items pointing at aligned unsigned int's */
4522 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4524 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4525 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4528 /** Compare two items pointing at unsigned ints of unknown alignment.
4529 * Nodes and keys are guaranteed to be 2-byte aligned.
4532 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4534 #if BYTE_ORDER == LITTLE_ENDIAN
4535 unsigned short *u, *c;
4538 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4539 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4542 } while(!x && u > (unsigned short *)a->mv_data);
4545 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4549 /** Compare two items pointing at size_t's of unknown alignment. */
4550 #ifdef MISALIGNED_OK
4551 # define mdb_cmp_clong mdb_cmp_long
4553 # define mdb_cmp_clong mdb_cmp_cint
4556 /** Compare two items lexically */
4558 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4565 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4571 diff = memcmp(a->mv_data, b->mv_data, len);
4572 return diff ? diff : len_diff<0 ? -1 : len_diff;
4575 /** Compare two items in reverse byte order */
4577 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4579 const unsigned char *p1, *p2, *p1_lim;
4583 p1_lim = (const unsigned char *)a->mv_data;
4584 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4585 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4587 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4593 while (p1 > p1_lim) {
4594 diff = *--p1 - *--p2;
4598 return len_diff<0 ? -1 : len_diff;
4601 /** Search for key within a page, using binary search.
4602 * Returns the smallest entry larger or equal to the key.
4603 * If exactp is non-null, stores whether the found entry was an exact match
4604 * in *exactp (1 or 0).
4605 * Updates the cursor index with the index of the found entry.
4606 * If no entry larger or equal to the key is found, returns NULL.
4609 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4611 unsigned int i = 0, nkeys;
4614 MDB_page *mp = mc->mc_pg[mc->mc_top];
4615 MDB_node *node = NULL;
4620 nkeys = NUMKEYS(mp);
4622 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4623 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4626 low = IS_LEAF(mp) ? 0 : 1;
4628 cmp = mc->mc_dbx->md_cmp;
4630 /* Branch pages have no data, so if using integer keys,
4631 * alignment is guaranteed. Use faster mdb_cmp_int.
4633 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4634 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4641 nodekey.mv_size = mc->mc_db->md_pad;
4642 node = NODEPTR(mp, 0); /* fake */
4643 while (low <= high) {
4644 i = (low + high) >> 1;
4645 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4646 rc = cmp(key, &nodekey);
4647 DPRINTF(("found leaf index %u [%s], rc = %i",
4648 i, DKEY(&nodekey), rc));
4657 while (low <= high) {
4658 i = (low + high) >> 1;
4660 node = NODEPTR(mp, i);
4661 nodekey.mv_size = NODEKSZ(node);
4662 nodekey.mv_data = NODEKEY(node);
4664 rc = cmp(key, &nodekey);
4667 DPRINTF(("found leaf index %u [%s], rc = %i",
4668 i, DKEY(&nodekey), rc));
4670 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4671 i, DKEY(&nodekey), NODEPGNO(node), rc));
4682 if (rc > 0) { /* Found entry is less than the key. */
4683 i++; /* Skip to get the smallest entry larger than key. */
4685 node = NODEPTR(mp, i);
4688 *exactp = (rc == 0 && nkeys > 0);
4689 /* store the key index */
4690 mc->mc_ki[mc->mc_top] = i;
4692 /* There is no entry larger or equal to the key. */
4695 /* nodeptr is fake for LEAF2 */
4701 mdb_cursor_adjust(MDB_cursor *mc, func)
4705 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4706 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4713 /** Pop a page off the top of the cursor's stack. */
4715 mdb_cursor_pop(MDB_cursor *mc)
4719 MDB_page *top = mc->mc_pg[mc->mc_top];
4725 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4726 DDBI(mc), (void *) mc));
4730 /** Push a page onto the top of the cursor's stack. */
4732 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4734 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4735 DDBI(mc), (void *) mc));
4737 if (mc->mc_snum >= CURSOR_STACK) {
4738 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4739 return MDB_CURSOR_FULL;
4742 mc->mc_top = mc->mc_snum++;
4743 mc->mc_pg[mc->mc_top] = mp;
4744 mc->mc_ki[mc->mc_top] = 0;
4749 /** Find the address of the page corresponding to a given page number.
4750 * @param[in] txn the transaction for this access.
4751 * @param[in] pgno the page number for the page to retrieve.
4752 * @param[out] ret address of a pointer where the page's address will be stored.
4753 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4754 * @return 0 on success, non-zero on failure.
4757 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4759 MDB_env *env = txn->mt_env;
4763 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4767 MDB_ID2L dl = tx2->mt_u.dirty_list;
4769 /* Spilled pages were dirtied in this txn and flushed
4770 * because the dirty list got full. Bring this page
4771 * back in from the map (but don't unspill it here,
4772 * leave that unless page_touch happens again).
4774 if (tx2->mt_spill_pgs) {
4775 MDB_ID pn = pgno << 1;
4776 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4777 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4778 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4783 unsigned x = mdb_mid2l_search(dl, pgno);
4784 if (x <= dl[0].mid && dl[x].mid == pgno) {
4790 } while ((tx2 = tx2->mt_parent) != NULL);
4793 if (pgno < txn->mt_next_pgno) {
4795 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4797 DPRINTF(("page %"Z"u not found", pgno));
4798 txn->mt_flags |= MDB_TXN_ERROR;
4799 return MDB_PAGE_NOTFOUND;
4809 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4810 * The cursor is at the root page, set up the rest of it.
4813 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4815 MDB_page *mp = mc->mc_pg[mc->mc_top];
4819 while (IS_BRANCH(mp)) {
4823 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4824 mdb_cassert(mc, NUMKEYS(mp) > 1);
4825 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4827 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4829 if (flags & MDB_PS_LAST)
4830 i = NUMKEYS(mp) - 1;
4833 node = mdb_node_search(mc, key, &exact);
4835 i = NUMKEYS(mp) - 1;
4837 i = mc->mc_ki[mc->mc_top];
4839 mdb_cassert(mc, i > 0);
4843 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4846 mdb_cassert(mc, i < NUMKEYS(mp));
4847 node = NODEPTR(mp, i);
4849 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4852 mc->mc_ki[mc->mc_top] = i;
4853 if ((rc = mdb_cursor_push(mc, mp)))
4856 if (flags & MDB_PS_MODIFY) {
4857 if ((rc = mdb_page_touch(mc)) != 0)
4859 mp = mc->mc_pg[mc->mc_top];
4864 DPRINTF(("internal error, index points to a %02X page!?",
4866 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4867 return MDB_CORRUPTED;
4870 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4871 key ? DKEY(key) : "null"));
4872 mc->mc_flags |= C_INITIALIZED;
4873 mc->mc_flags &= ~C_EOF;
4878 /** Search for the lowest key under the current branch page.
4879 * This just bypasses a NUMKEYS check in the current page
4880 * before calling mdb_page_search_root(), because the callers
4881 * are all in situations where the current page is known to
4885 mdb_page_search_lowest(MDB_cursor *mc)
4887 MDB_page *mp = mc->mc_pg[mc->mc_top];
4888 MDB_node *node = NODEPTR(mp, 0);
4891 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4894 mc->mc_ki[mc->mc_top] = 0;
4895 if ((rc = mdb_cursor_push(mc, mp)))
4897 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4900 /** Search for the page a given key should be in.
4901 * Push it and its parent pages on the cursor stack.
4902 * @param[in,out] mc the cursor for this operation.
4903 * @param[in] key the key to search for, or NULL for first/last page.
4904 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4905 * are touched (updated with new page numbers).
4906 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4907 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4908 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4909 * @return 0 on success, non-zero on failure.
4912 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4917 /* Make sure the txn is still viable, then find the root from
4918 * the txn's db table and set it as the root of the cursor's stack.
4920 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4921 DPUTS("transaction has failed, must abort");
4924 /* Make sure we're using an up-to-date root */
4925 if (*mc->mc_dbflag & DB_STALE) {
4927 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4928 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4935 MDB_node *leaf = mdb_node_search(&mc2,
4936 &mc->mc_dbx->md_name, &exact);
4938 return MDB_NOTFOUND;
4939 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4942 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4944 /* The txn may not know this DBI, or another process may
4945 * have dropped and recreated the DB with other flags.
4947 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4948 return MDB_INCOMPATIBLE;
4949 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4951 *mc->mc_dbflag &= ~DB_STALE;
4953 root = mc->mc_db->md_root;
4955 if (root == P_INVALID) { /* Tree is empty. */
4956 DPUTS("tree is empty");
4957 return MDB_NOTFOUND;
4961 mdb_cassert(mc, root > 1);
4962 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4963 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4969 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4970 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4972 if (flags & MDB_PS_MODIFY) {
4973 if ((rc = mdb_page_touch(mc)))
4977 if (flags & MDB_PS_ROOTONLY)
4980 return mdb_page_search_root(mc, key, flags);
4984 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4986 MDB_txn *txn = mc->mc_txn;
4987 pgno_t pg = mp->mp_pgno;
4988 unsigned x = 0, ovpages = mp->mp_pages;
4989 MDB_env *env = txn->mt_env;
4990 MDB_IDL sl = txn->mt_spill_pgs;
4991 MDB_ID pn = pg << 1;
4994 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4995 /* If the page is dirty or on the spill list we just acquired it,
4996 * so we should give it back to our current free list, if any.
4997 * Otherwise put it onto the list of pages we freed in this txn.
4999 * Won't create me_pghead: me_pglast must be inited along with it.
5000 * Unsupported in nested txns: They would need to hide the page
5001 * range in ancestor txns' dirty and spilled lists.
5003 if (env->me_pghead &&
5005 ((mp->mp_flags & P_DIRTY) ||
5006 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5010 MDB_ID2 *dl, ix, iy;
5011 rc = mdb_midl_need(&env->me_pghead, ovpages);
5014 if (!(mp->mp_flags & P_DIRTY)) {
5015 /* This page is no longer spilled */
5022 /* Remove from dirty list */
5023 dl = txn->mt_u.dirty_list;
5025 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5031 mdb_cassert(mc, x > 1);
5033 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5034 txn->mt_flags |= MDB_TXN_ERROR;
5035 return MDB_CORRUPTED;
5038 if (!(env->me_flags & MDB_WRITEMAP))
5039 mdb_dpage_free(env, mp);
5041 /* Insert in me_pghead */
5042 mop = env->me_pghead;
5043 j = mop[0] + ovpages;
5044 for (i = mop[0]; i && mop[i] < pg; i--)
5050 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5054 mc->mc_db->md_overflow_pages -= ovpages;
5058 /** Return the data associated with a given node.
5059 * @param[in] txn The transaction for this operation.
5060 * @param[in] leaf The node being read.
5061 * @param[out] data Updated to point to the node's data.
5062 * @return 0 on success, non-zero on failure.
5065 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5067 MDB_page *omp; /* overflow page */
5071 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5072 data->mv_size = NODEDSZ(leaf);
5073 data->mv_data = NODEDATA(leaf);
5077 /* Read overflow data.
5079 data->mv_size = NODEDSZ(leaf);
5080 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5081 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5082 DPRINTF(("read overflow page %"Z"u failed", pgno));
5085 data->mv_data = METADATA(omp);
5091 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5092 MDB_val *key, MDB_val *data)
5099 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5101 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5104 if (txn->mt_flags & MDB_TXN_ERROR)
5107 mdb_cursor_init(&mc, txn, dbi, &mx);
5108 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5111 /** Find a sibling for a page.
5112 * Replaces the page at the top of the cursor's stack with the
5113 * specified sibling, if one exists.
5114 * @param[in] mc The cursor for this operation.
5115 * @param[in] move_right Non-zero if the right sibling is requested,
5116 * otherwise the left sibling.
5117 * @return 0 on success, non-zero on failure.
5120 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5126 if (mc->mc_snum < 2) {
5127 return MDB_NOTFOUND; /* root has no siblings */
5131 DPRINTF(("parent page is page %"Z"u, index %u",
5132 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5134 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5135 : (mc->mc_ki[mc->mc_top] == 0)) {
5136 DPRINTF(("no more keys left, moving to %s sibling",
5137 move_right ? "right" : "left"));
5138 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5139 /* undo cursor_pop before returning */
5146 mc->mc_ki[mc->mc_top]++;
5148 mc->mc_ki[mc->mc_top]--;
5149 DPRINTF(("just moving to %s index key %u",
5150 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5152 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5154 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5155 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5156 /* mc will be inconsistent if caller does mc_snum++ as above */
5157 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5161 mdb_cursor_push(mc, mp);
5163 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5168 /** Move the cursor to the next data item. */
5170 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5176 if (mc->mc_flags & C_EOF) {
5177 return MDB_NOTFOUND;
5180 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5182 mp = mc->mc_pg[mc->mc_top];
5184 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5185 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5186 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5187 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5188 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5189 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5190 if (rc == MDB_SUCCESS)
5191 MDB_GET_KEY(leaf, key);
5196 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5197 if (op == MDB_NEXT_DUP)
5198 return MDB_NOTFOUND;
5202 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5203 mdb_dbg_pgno(mp), (void *) mc));
5204 if (mc->mc_flags & C_DEL)
5207 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5208 DPUTS("=====> move to next sibling page");
5209 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5210 mc->mc_flags |= C_EOF;
5213 mp = mc->mc_pg[mc->mc_top];
5214 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5216 mc->mc_ki[mc->mc_top]++;
5219 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5220 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5223 key->mv_size = mc->mc_db->md_pad;
5224 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5228 mdb_cassert(mc, IS_LEAF(mp));
5229 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5231 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5232 mdb_xcursor_init1(mc, leaf);
5235 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5238 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5239 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5240 if (rc != MDB_SUCCESS)
5245 MDB_GET_KEY(leaf, key);
5249 /** Move the cursor to the previous data item. */
5251 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5257 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5259 mp = mc->mc_pg[mc->mc_top];
5261 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5262 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5263 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5264 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5265 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5266 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5267 if (rc == MDB_SUCCESS)
5268 MDB_GET_KEY(leaf, key);
5272 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5273 if (op == MDB_PREV_DUP)
5274 return MDB_NOTFOUND;
5279 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5280 mdb_dbg_pgno(mp), (void *) mc));
5282 if (mc->mc_ki[mc->mc_top] == 0) {
5283 DPUTS("=====> move to prev sibling page");
5284 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5287 mp = mc->mc_pg[mc->mc_top];
5288 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5289 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5291 mc->mc_ki[mc->mc_top]--;
5293 mc->mc_flags &= ~C_EOF;
5295 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5296 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5299 key->mv_size = mc->mc_db->md_pad;
5300 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5304 mdb_cassert(mc, IS_LEAF(mp));
5305 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5307 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5308 mdb_xcursor_init1(mc, leaf);
5311 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5314 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5315 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5316 if (rc != MDB_SUCCESS)
5321 MDB_GET_KEY(leaf, key);
5325 /** Set the cursor on a specific data item. */
5327 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5328 MDB_cursor_op op, int *exactp)
5332 MDB_node *leaf = NULL;
5335 if (key->mv_size == 0)
5336 return MDB_BAD_VALSIZE;
5339 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5341 /* See if we're already on the right page */
5342 if (mc->mc_flags & C_INITIALIZED) {
5345 mp = mc->mc_pg[mc->mc_top];
5347 mc->mc_ki[mc->mc_top] = 0;
5348 return MDB_NOTFOUND;
5350 if (mp->mp_flags & P_LEAF2) {
5351 nodekey.mv_size = mc->mc_db->md_pad;
5352 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5354 leaf = NODEPTR(mp, 0);
5355 MDB_GET_KEY2(leaf, nodekey);
5357 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5359 /* Probably happens rarely, but first node on the page
5360 * was the one we wanted.
5362 mc->mc_ki[mc->mc_top] = 0;
5369 unsigned int nkeys = NUMKEYS(mp);
5371 if (mp->mp_flags & P_LEAF2) {
5372 nodekey.mv_data = LEAF2KEY(mp,
5373 nkeys-1, nodekey.mv_size);
5375 leaf = NODEPTR(mp, nkeys-1);
5376 MDB_GET_KEY2(leaf, nodekey);
5378 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5380 /* last node was the one we wanted */
5381 mc->mc_ki[mc->mc_top] = nkeys-1;
5387 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5388 /* This is definitely the right page, skip search_page */
5389 if (mp->mp_flags & P_LEAF2) {
5390 nodekey.mv_data = LEAF2KEY(mp,
5391 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5393 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5394 MDB_GET_KEY2(leaf, nodekey);
5396 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5398 /* current node was the one we wanted */
5408 /* If any parents have right-sibs, search.
5409 * Otherwise, there's nothing further.
5411 for (i=0; i<mc->mc_top; i++)
5413 NUMKEYS(mc->mc_pg[i])-1)
5415 if (i == mc->mc_top) {
5416 /* There are no other pages */
5417 mc->mc_ki[mc->mc_top] = nkeys;
5418 return MDB_NOTFOUND;
5422 /* There are no other pages */
5423 mc->mc_ki[mc->mc_top] = 0;
5424 if (op == MDB_SET_RANGE && !exactp) {
5428 return MDB_NOTFOUND;
5432 rc = mdb_page_search(mc, key, 0);
5433 if (rc != MDB_SUCCESS)
5436 mp = mc->mc_pg[mc->mc_top];
5437 mdb_cassert(mc, IS_LEAF(mp));
5440 leaf = mdb_node_search(mc, key, exactp);
5441 if (exactp != NULL && !*exactp) {
5442 /* MDB_SET specified and not an exact match. */
5443 return MDB_NOTFOUND;
5447 DPUTS("===> inexact leaf not found, goto sibling");
5448 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5449 return rc; /* no entries matched */
5450 mp = mc->mc_pg[mc->mc_top];
5451 mdb_cassert(mc, IS_LEAF(mp));
5452 leaf = NODEPTR(mp, 0);
5456 mc->mc_flags |= C_INITIALIZED;
5457 mc->mc_flags &= ~C_EOF;
5460 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5461 key->mv_size = mc->mc_db->md_pad;
5462 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5467 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5468 mdb_xcursor_init1(mc, leaf);
5471 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5472 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5473 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5476 if (op == MDB_GET_BOTH) {
5482 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5483 if (rc != MDB_SUCCESS)
5486 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5488 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5490 rc = mc->mc_dbx->md_dcmp(data, &d2);
5492 if (op == MDB_GET_BOTH || rc > 0)
5493 return MDB_NOTFOUND;
5500 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5501 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5506 /* The key already matches in all other cases */
5507 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5508 MDB_GET_KEY(leaf, key);
5509 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5514 /** Move the cursor to the first item in the database. */
5516 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5522 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5524 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5525 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5526 if (rc != MDB_SUCCESS)
5529 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5531 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5532 mc->mc_flags |= C_INITIALIZED;
5533 mc->mc_flags &= ~C_EOF;
5535 mc->mc_ki[mc->mc_top] = 0;
5537 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5538 key->mv_size = mc->mc_db->md_pad;
5539 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5544 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5545 mdb_xcursor_init1(mc, leaf);
5546 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5550 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5554 MDB_GET_KEY(leaf, key);
5558 /** Move the cursor to the last item in the database. */
5560 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5566 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5568 if (!(mc->mc_flags & C_EOF)) {
5570 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5571 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5572 if (rc != MDB_SUCCESS)
5575 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5578 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5579 mc->mc_flags |= C_INITIALIZED|C_EOF;
5580 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5582 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5583 key->mv_size = mc->mc_db->md_pad;
5584 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5589 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5590 mdb_xcursor_init1(mc, leaf);
5591 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5595 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5600 MDB_GET_KEY(leaf, key);
5605 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5610 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5615 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5619 case MDB_GET_CURRENT:
5620 if (!(mc->mc_flags & C_INITIALIZED)) {
5623 MDB_page *mp = mc->mc_pg[mc->mc_top];
5624 int nkeys = NUMKEYS(mp);
5625 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5626 mc->mc_ki[mc->mc_top] = nkeys;
5632 key->mv_size = mc->mc_db->md_pad;
5633 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5635 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5636 MDB_GET_KEY(leaf, key);
5638 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5639 if (mc->mc_flags & C_DEL)
5640 mdb_xcursor_init1(mc, leaf);
5641 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5643 rc = mdb_node_read(mc->mc_txn, leaf, data);
5650 case MDB_GET_BOTH_RANGE:
5655 if (mc->mc_xcursor == NULL) {
5656 rc = MDB_INCOMPATIBLE;
5666 rc = mdb_cursor_set(mc, key, data, op,
5667 op == MDB_SET_RANGE ? NULL : &exact);
5670 case MDB_GET_MULTIPLE:
5671 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5675 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5676 rc = MDB_INCOMPATIBLE;
5680 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5681 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5684 case MDB_NEXT_MULTIPLE:
5689 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5690 rc = MDB_INCOMPATIBLE;
5693 if (!(mc->mc_flags & C_INITIALIZED))
5694 rc = mdb_cursor_first(mc, key, data);
5696 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5697 if (rc == MDB_SUCCESS) {
5698 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5701 mx = &mc->mc_xcursor->mx_cursor;
5702 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5704 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5705 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5713 case MDB_NEXT_NODUP:
5714 if (!(mc->mc_flags & C_INITIALIZED))
5715 rc = mdb_cursor_first(mc, key, data);
5717 rc = mdb_cursor_next(mc, key, data, op);
5721 case MDB_PREV_NODUP:
5722 if (!(mc->mc_flags & C_INITIALIZED)) {
5723 rc = mdb_cursor_last(mc, key, data);
5726 mc->mc_flags |= C_INITIALIZED;
5727 mc->mc_ki[mc->mc_top]++;
5729 rc = mdb_cursor_prev(mc, key, data, op);
5732 rc = mdb_cursor_first(mc, key, data);
5735 mfunc = mdb_cursor_first;
5737 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5741 if (mc->mc_xcursor == NULL) {
5742 rc = MDB_INCOMPATIBLE;
5745 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5749 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5752 rc = mdb_cursor_last(mc, key, data);
5755 mfunc = mdb_cursor_last;
5758 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5763 if (mc->mc_flags & C_DEL)
5764 mc->mc_flags ^= C_DEL;
5769 /** Touch all the pages in the cursor stack. Set mc_top.
5770 * Makes sure all the pages are writable, before attempting a write operation.
5771 * @param[in] mc The cursor to operate on.
5774 mdb_cursor_touch(MDB_cursor *mc)
5776 int rc = MDB_SUCCESS;
5778 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5781 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5782 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5785 *mc->mc_dbflag |= DB_DIRTY;
5790 rc = mdb_page_touch(mc);
5791 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5792 mc->mc_top = mc->mc_snum-1;
5797 /** Do not spill pages to disk if txn is getting full, may fail instead */
5798 #define MDB_NOSPILL 0x8000
5801 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5804 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5806 MDB_node *leaf = NULL;
5809 MDB_val xdata, *rdata, dkey, olddata;
5811 int do_sub = 0, insert_key, insert_data;
5812 unsigned int mcount = 0, dcount = 0, nospill;
5815 unsigned int nflags;
5818 if (mc == NULL || key == NULL)
5821 env = mc->mc_txn->mt_env;
5823 /* Check this first so counter will always be zero on any
5826 if (flags & MDB_MULTIPLE) {
5827 dcount = data[1].mv_size;
5828 data[1].mv_size = 0;
5829 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5830 return MDB_INCOMPATIBLE;
5833 nospill = flags & MDB_NOSPILL;
5834 flags &= ~MDB_NOSPILL;
5836 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5837 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5839 if (key->mv_size-1 >= ENV_MAXKEY(env))
5840 return MDB_BAD_VALSIZE;
5842 #if SIZE_MAX > MAXDATASIZE
5843 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5844 return MDB_BAD_VALSIZE;
5846 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5847 return MDB_BAD_VALSIZE;
5850 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5851 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5855 if (flags == MDB_CURRENT) {
5856 if (!(mc->mc_flags & C_INITIALIZED))
5859 } else if (mc->mc_db->md_root == P_INVALID) {
5860 /* new database, cursor has nothing to point to */
5863 mc->mc_flags &= ~C_INITIALIZED;
5868 if (flags & MDB_APPEND) {
5870 rc = mdb_cursor_last(mc, &k2, &d2);
5872 rc = mc->mc_dbx->md_cmp(key, &k2);
5875 mc->mc_ki[mc->mc_top]++;
5877 /* new key is <= last key */
5882 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5884 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5885 DPRINTF(("duplicate key [%s]", DKEY(key)));
5887 return MDB_KEYEXIST;
5889 if (rc && rc != MDB_NOTFOUND)
5893 if (mc->mc_flags & C_DEL)
5894 mc->mc_flags ^= C_DEL;
5896 /* Cursor is positioned, check for room in the dirty list */
5898 if (flags & MDB_MULTIPLE) {
5900 xdata.mv_size = data->mv_size * dcount;
5904 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5908 if (rc == MDB_NO_ROOT) {
5910 /* new database, write a root leaf page */
5911 DPUTS("allocating new root leaf page");
5912 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5915 mdb_cursor_push(mc, np);
5916 mc->mc_db->md_root = np->mp_pgno;
5917 mc->mc_db->md_depth++;
5918 *mc->mc_dbflag |= DB_DIRTY;
5919 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5921 np->mp_flags |= P_LEAF2;
5922 mc->mc_flags |= C_INITIALIZED;
5924 /* make sure all cursor pages are writable */
5925 rc2 = mdb_cursor_touch(mc);
5930 insert_key = insert_data = rc;
5932 /* The key does not exist */
5933 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5934 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5935 LEAFSIZE(key, data) > env->me_nodemax)
5937 /* Too big for a node, insert in sub-DB */
5938 fp_flags = P_LEAF|P_DIRTY;
5940 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5941 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5945 /* there's only a key anyway, so this is a no-op */
5946 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5948 unsigned int ksize = mc->mc_db->md_pad;
5949 if (key->mv_size != ksize)
5950 return MDB_BAD_VALSIZE;
5951 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5952 memcpy(ptr, key->mv_data, ksize);
5957 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5958 olddata.mv_size = NODEDSZ(leaf);
5959 olddata.mv_data = NODEDATA(leaf);
5962 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5963 /* Prepare (sub-)page/sub-DB to accept the new item,
5964 * if needed. fp: old sub-page or a header faking
5965 * it. mp: new (sub-)page. offset: growth in page
5966 * size. xdata: node data with new page or DB.
5968 unsigned i, offset = 0;
5969 mp = fp = xdata.mv_data = env->me_pbuf;
5970 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5972 /* Was a single item before, must convert now */
5973 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5974 /* Just overwrite the current item */
5975 if (flags == MDB_CURRENT)
5978 #if UINT_MAX < SIZE_MAX
5979 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5980 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
5982 /* does data match? */
5983 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5984 if (flags & MDB_NODUPDATA)
5985 return MDB_KEYEXIST;
5990 /* Back up original data item */
5991 dkey.mv_size = olddata.mv_size;
5992 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5994 /* Make sub-page header for the dup items, with dummy body */
5995 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5996 fp->mp_lower = PAGEHDRSZ;
5997 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5998 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5999 fp->mp_flags |= P_LEAF2;
6000 fp->mp_pad = data->mv_size;
6001 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6003 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6004 (dkey.mv_size & 1) + (data->mv_size & 1);
6006 fp->mp_upper = xdata.mv_size;
6007 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
6008 } else if (leaf->mn_flags & F_SUBDATA) {
6009 /* Data is on sub-DB, just store it */
6010 flags |= F_DUPDATA|F_SUBDATA;
6013 /* Data is on sub-page */
6014 fp = olddata.mv_data;
6017 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6018 offset = EVEN(NODESIZE + sizeof(indx_t) +
6022 offset = fp->mp_pad;
6023 if (SIZELEFT(fp) < offset) {
6024 offset *= 4; /* space for 4 more */
6027 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6029 fp->mp_flags |= P_DIRTY;
6030 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6031 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6035 xdata.mv_size = olddata.mv_size + offset;
6038 fp_flags = fp->mp_flags;
6039 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6040 /* Too big for a sub-page, convert to sub-DB */
6041 fp_flags &= ~P_SUBP;
6043 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6044 fp_flags |= P_LEAF2;
6045 dummy.md_pad = fp->mp_pad;
6046 dummy.md_flags = MDB_DUPFIXED;
6047 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6048 dummy.md_flags |= MDB_INTEGERKEY;
6054 dummy.md_branch_pages = 0;
6055 dummy.md_leaf_pages = 1;
6056 dummy.md_overflow_pages = 0;
6057 dummy.md_entries = NUMKEYS(fp);
6058 xdata.mv_size = sizeof(MDB_db);
6059 xdata.mv_data = &dummy;
6060 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6062 offset = env->me_psize - olddata.mv_size;
6063 flags |= F_DUPDATA|F_SUBDATA;
6064 dummy.md_root = mp->mp_pgno;
6067 mp->mp_flags = fp_flags | P_DIRTY;
6068 mp->mp_pad = fp->mp_pad;
6069 mp->mp_lower = fp->mp_lower;
6070 mp->mp_upper = fp->mp_upper + offset;
6071 if (fp_flags & P_LEAF2) {
6072 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6074 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6075 olddata.mv_size - fp->mp_upper);
6076 for (i=0; i<NUMKEYS(fp); i++)
6077 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6085 mdb_node_del(mc, 0);
6089 /* overflow page overwrites need special handling */
6090 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6093 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6095 memcpy(&pg, olddata.mv_data, sizeof(pg));
6096 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6098 ovpages = omp->mp_pages;
6100 /* Is the ov page large enough? */
6101 if (ovpages >= dpages) {
6102 if (!(omp->mp_flags & P_DIRTY) &&
6103 (level || (env->me_flags & MDB_WRITEMAP)))
6105 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6108 level = 0; /* dirty in this txn or clean */
6111 if (omp->mp_flags & P_DIRTY) {
6112 /* yes, overwrite it. Note in this case we don't
6113 * bother to try shrinking the page if the new data
6114 * is smaller than the overflow threshold.
6117 /* It is writable only in a parent txn */
6118 size_t sz = (size_t) env->me_psize * ovpages, off;
6119 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6125 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6126 mdb_cassert(mc, rc2 == 0);
6127 if (!(flags & MDB_RESERVE)) {
6128 /* Copy end of page, adjusting alignment so
6129 * compiler may copy words instead of bytes.
6131 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6132 memcpy((size_t *)((char *)np + off),
6133 (size_t *)((char *)omp + off), sz - off);
6136 memcpy(np, omp, sz); /* Copy beginning of page */
6139 SETDSZ(leaf, data->mv_size);
6140 if (F_ISSET(flags, MDB_RESERVE))
6141 data->mv_data = METADATA(omp);
6143 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6147 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6149 } else if (data->mv_size == olddata.mv_size) {
6150 /* same size, just replace it. Note that we could
6151 * also reuse this node if the new data is smaller,
6152 * but instead we opt to shrink the node in that case.
6154 if (F_ISSET(flags, MDB_RESERVE))
6155 data->mv_data = olddata.mv_data;
6156 else if (!(mc->mc_flags & C_SUB))
6157 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6159 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6162 mdb_node_del(mc, 0);
6168 nflags = flags & NODE_ADD_FLAGS;
6169 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6170 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6171 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6172 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6174 nflags |= MDB_SPLIT_REPLACE;
6175 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6177 /* There is room already in this leaf page. */
6178 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6179 if (rc == 0 && insert_key) {
6180 /* Adjust other cursors pointing to mp */
6181 MDB_cursor *m2, *m3;
6182 MDB_dbi dbi = mc->mc_dbi;
6183 unsigned i = mc->mc_top;
6184 MDB_page *mp = mc->mc_pg[i];
6186 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6187 if (mc->mc_flags & C_SUB)
6188 m3 = &m2->mc_xcursor->mx_cursor;
6191 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6192 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6199 if (rc == MDB_SUCCESS) {
6200 /* Now store the actual data in the child DB. Note that we're
6201 * storing the user data in the keys field, so there are strict
6202 * size limits on dupdata. The actual data fields of the child
6203 * DB are all zero size.
6211 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6212 if (flags & MDB_CURRENT) {
6213 xflags = MDB_CURRENT|MDB_NOSPILL;
6215 mdb_xcursor_init1(mc, leaf);
6216 xflags = (flags & MDB_NODUPDATA) ?
6217 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6219 /* converted, write the original data first */
6221 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6225 /* Adjust other cursors pointing to mp */
6227 unsigned i = mc->mc_top;
6228 MDB_page *mp = mc->mc_pg[i];
6230 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6231 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6232 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6233 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6234 mdb_xcursor_init1(m2, leaf);
6238 /* we've done our job */
6241 ecount = mc->mc_xcursor->mx_db.md_entries;
6242 if (flags & MDB_APPENDDUP)
6243 xflags |= MDB_APPEND;
6244 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6245 if (flags & F_SUBDATA) {
6246 void *db = NODEDATA(leaf);
6247 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6249 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6251 /* Increment count unless we just replaced an existing item. */
6253 mc->mc_db->md_entries++;
6255 /* Invalidate txn if we created an empty sub-DB */
6258 /* If we succeeded and the key didn't exist before,
6259 * make sure the cursor is marked valid.
6261 mc->mc_flags |= C_INITIALIZED;
6263 if (flags & MDB_MULTIPLE) {
6266 /* let caller know how many succeeded, if any */
6267 data[1].mv_size = mcount;
6268 if (mcount < dcount) {
6269 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6270 insert_key = insert_data = 0;
6277 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6280 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6285 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6291 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6292 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6294 if (!(mc->mc_flags & C_INITIALIZED))
6297 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6298 return MDB_NOTFOUND;
6300 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6303 rc = mdb_cursor_touch(mc);
6307 mp = mc->mc_pg[mc->mc_top];
6310 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6312 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6313 if (flags & MDB_NODUPDATA) {
6314 /* mdb_cursor_del0() will subtract the final entry */
6315 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6317 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6318 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6320 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6323 /* If sub-DB still has entries, we're done */
6324 if (mc->mc_xcursor->mx_db.md_entries) {
6325 if (leaf->mn_flags & F_SUBDATA) {
6326 /* update subDB info */
6327 void *db = NODEDATA(leaf);
6328 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6331 /* shrink fake page */
6332 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6333 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6334 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6335 /* fix other sub-DB cursors pointed at this fake page */
6336 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6337 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6338 if (m2->mc_pg[mc->mc_top] == mp &&
6339 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6340 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6343 mc->mc_db->md_entries--;
6344 mc->mc_flags |= C_DEL;
6347 /* otherwise fall thru and delete the sub-DB */
6350 if (leaf->mn_flags & F_SUBDATA) {
6351 /* add all the child DB's pages to the free list */
6352 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6358 /* add overflow pages to free list */
6359 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6363 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6364 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6365 (rc = mdb_ovpage_free(mc, omp)))
6370 return mdb_cursor_del0(mc);
6373 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6377 /** Allocate and initialize new pages for a database.
6378 * @param[in] mc a cursor on the database being added to.
6379 * @param[in] flags flags defining what type of page is being allocated.
6380 * @param[in] num the number of pages to allocate. This is usually 1,
6381 * unless allocating overflow pages for a large record.
6382 * @param[out] mp Address of a page, or NULL on failure.
6383 * @return 0 on success, non-zero on failure.
6386 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6391 if ((rc = mdb_page_alloc(mc, num, &np)))
6393 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6394 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6395 np->mp_flags = flags | P_DIRTY;
6396 np->mp_lower = PAGEHDRSZ;
6397 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6400 mc->mc_db->md_branch_pages++;
6401 else if (IS_LEAF(np))
6402 mc->mc_db->md_leaf_pages++;
6403 else if (IS_OVERFLOW(np)) {
6404 mc->mc_db->md_overflow_pages += num;
6412 /** Calculate the size of a leaf node.
6413 * The size depends on the environment's page size; if a data item
6414 * is too large it will be put onto an overflow page and the node
6415 * size will only include the key and not the data. Sizes are always
6416 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6417 * of the #MDB_node headers.
6418 * @param[in] env The environment handle.
6419 * @param[in] key The key for the node.
6420 * @param[in] data The data for the node.
6421 * @return The number of bytes needed to store the node.
6424 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6428 sz = LEAFSIZE(key, data);
6429 if (sz > env->me_nodemax) {
6430 /* put on overflow page */
6431 sz -= data->mv_size - sizeof(pgno_t);
6434 return EVEN(sz + sizeof(indx_t));
6437 /** Calculate the size of a branch node.
6438 * The size should depend on the environment's page size but since
6439 * we currently don't support spilling large keys onto overflow
6440 * pages, it's simply the size of the #MDB_node header plus the
6441 * size of the key. Sizes are always rounded up to an even number
6442 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6443 * @param[in] env The environment handle.
6444 * @param[in] key The key for the node.
6445 * @return The number of bytes needed to store the node.
6448 mdb_branch_size(MDB_env *env, MDB_val *key)
6453 if (sz > env->me_nodemax) {
6454 /* put on overflow page */
6455 /* not implemented */
6456 /* sz -= key->size - sizeof(pgno_t); */
6459 return sz + sizeof(indx_t);
6462 /** Add a node to the page pointed to by the cursor.
6463 * @param[in] mc The cursor for this operation.
6464 * @param[in] indx The index on the page where the new node should be added.
6465 * @param[in] key The key for the new node.
6466 * @param[in] data The data for the new node, if any.
6467 * @param[in] pgno The page number, if adding a branch node.
6468 * @param[in] flags Flags for the node.
6469 * @return 0 on success, non-zero on failure. Possible errors are:
6471 * <li>ENOMEM - failed to allocate overflow pages for the node.
6472 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6473 * should never happen since all callers already calculate the
6474 * page's free space before calling this function.
6478 mdb_node_add(MDB_cursor *mc, indx_t indx,
6479 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6482 size_t node_size = NODESIZE;
6486 MDB_page *mp = mc->mc_pg[mc->mc_top];
6487 MDB_page *ofp = NULL; /* overflow page */
6490 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6492 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6493 IS_LEAF(mp) ? "leaf" : "branch",
6494 IS_SUBP(mp) ? "sub-" : "",
6495 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6496 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6499 /* Move higher keys up one slot. */
6500 int ksize = mc->mc_db->md_pad, dif;
6501 char *ptr = LEAF2KEY(mp, indx, ksize);
6502 dif = NUMKEYS(mp) - indx;
6504 memmove(ptr+ksize, ptr, dif*ksize);
6505 /* insert new key */
6506 memcpy(ptr, key->mv_data, ksize);
6508 /* Just using these for counting */
6509 mp->mp_lower += sizeof(indx_t);
6510 mp->mp_upper -= ksize - sizeof(indx_t);
6514 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6516 node_size += key->mv_size;
6518 mdb_cassert(mc, data);
6519 if (F_ISSET(flags, F_BIGDATA)) {
6520 /* Data already on overflow page. */
6521 node_size += sizeof(pgno_t);
6522 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6523 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6525 /* Put data on overflow page. */
6526 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6527 data->mv_size, node_size+data->mv_size));
6528 node_size = EVEN(node_size + sizeof(pgno_t));
6529 if ((ssize_t)node_size > room)
6531 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6533 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6537 node_size += data->mv_size;
6540 node_size = EVEN(node_size);
6541 if ((ssize_t)node_size > room)
6545 /* Move higher pointers up one slot. */
6546 for (i = NUMKEYS(mp); i > indx; i--)
6547 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6549 /* Adjust free space offsets. */
6550 ofs = mp->mp_upper - node_size;
6551 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6552 mp->mp_ptrs[indx] = ofs;
6554 mp->mp_lower += sizeof(indx_t);
6556 /* Write the node data. */
6557 node = NODEPTR(mp, indx);
6558 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6559 node->mn_flags = flags;
6561 SETDSZ(node,data->mv_size);
6566 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6569 mdb_cassert(mc, key);
6571 if (F_ISSET(flags, F_BIGDATA))
6572 memcpy(node->mn_data + key->mv_size, data->mv_data,
6574 else if (F_ISSET(flags, MDB_RESERVE))
6575 data->mv_data = node->mn_data + key->mv_size;
6577 memcpy(node->mn_data + key->mv_size, data->mv_data,
6580 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6582 if (F_ISSET(flags, MDB_RESERVE))
6583 data->mv_data = METADATA(ofp);
6585 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6592 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6593 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6594 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6595 DPRINTF(("node size = %"Z"u", node_size));
6596 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6597 return MDB_PAGE_FULL;
6600 /** Delete the specified node from a page.
6601 * @param[in] mc Cursor pointing to the node to delete.
6602 * @param[in] ksize The size of a node. Only used if the page is
6603 * part of a #MDB_DUPFIXED database.
6606 mdb_node_del(MDB_cursor *mc, int ksize)
6608 MDB_page *mp = mc->mc_pg[mc->mc_top];
6609 indx_t indx = mc->mc_ki[mc->mc_top];
6611 indx_t i, j, numkeys, ptr;
6615 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6616 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6617 numkeys = NUMKEYS(mp);
6618 mdb_cassert(mc, indx < numkeys);
6621 int x = numkeys - 1 - indx;
6622 base = LEAF2KEY(mp, indx, ksize);
6624 memmove(base, base + ksize, x * ksize);
6625 mp->mp_lower -= sizeof(indx_t);
6626 mp->mp_upper += ksize - sizeof(indx_t);
6630 node = NODEPTR(mp, indx);
6631 sz = NODESIZE + node->mn_ksize;
6633 if (F_ISSET(node->mn_flags, F_BIGDATA))
6634 sz += sizeof(pgno_t);
6636 sz += NODEDSZ(node);
6640 ptr = mp->mp_ptrs[indx];
6641 for (i = j = 0; i < numkeys; i++) {
6643 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6644 if (mp->mp_ptrs[i] < ptr)
6645 mp->mp_ptrs[j] += sz;
6650 base = (char *)mp + mp->mp_upper;
6651 memmove(base + sz, base, ptr - mp->mp_upper);
6653 mp->mp_lower -= sizeof(indx_t);
6657 /** Compact the main page after deleting a node on a subpage.
6658 * @param[in] mp The main page to operate on.
6659 * @param[in] indx The index of the subpage on the main page.
6662 mdb_node_shrink(MDB_page *mp, indx_t indx)
6668 indx_t i, numkeys, ptr;
6670 node = NODEPTR(mp, indx);
6671 sp = (MDB_page *)NODEDATA(node);
6672 delta = SIZELEFT(sp);
6673 xp = (MDB_page *)((char *)sp + delta);
6675 /* shift subpage upward */
6677 nsize = NUMKEYS(sp) * sp->mp_pad;
6679 return; /* do not make the node uneven-sized */
6680 memmove(METADATA(xp), METADATA(sp), nsize);
6683 numkeys = NUMKEYS(sp);
6684 for (i=numkeys-1; i>=0; i--)
6685 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6687 xp->mp_upper = sp->mp_lower;
6688 xp->mp_lower = sp->mp_lower;
6689 xp->mp_flags = sp->mp_flags;
6690 xp->mp_pad = sp->mp_pad;
6691 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6693 nsize = NODEDSZ(node) - delta;
6694 SETDSZ(node, nsize);
6696 /* shift lower nodes upward */
6697 ptr = mp->mp_ptrs[indx];
6698 numkeys = NUMKEYS(mp);
6699 for (i = 0; i < numkeys; i++) {
6700 if (mp->mp_ptrs[i] <= ptr)
6701 mp->mp_ptrs[i] += delta;
6704 base = (char *)mp + mp->mp_upper;
6705 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6706 mp->mp_upper += delta;
6709 /** Initial setup of a sorted-dups cursor.
6710 * Sorted duplicates are implemented as a sub-database for the given key.
6711 * The duplicate data items are actually keys of the sub-database.
6712 * Operations on the duplicate data items are performed using a sub-cursor
6713 * initialized when the sub-database is first accessed. This function does
6714 * the preliminary setup of the sub-cursor, filling in the fields that
6715 * depend only on the parent DB.
6716 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6719 mdb_xcursor_init0(MDB_cursor *mc)
6721 MDB_xcursor *mx = mc->mc_xcursor;
6723 mx->mx_cursor.mc_xcursor = NULL;
6724 mx->mx_cursor.mc_txn = mc->mc_txn;
6725 mx->mx_cursor.mc_db = &mx->mx_db;
6726 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6727 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6728 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6729 mx->mx_cursor.mc_snum = 0;
6730 mx->mx_cursor.mc_top = 0;
6731 mx->mx_cursor.mc_flags = C_SUB;
6732 mx->mx_dbx.md_name.mv_size = 0;
6733 mx->mx_dbx.md_name.mv_data = NULL;
6734 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6735 mx->mx_dbx.md_dcmp = NULL;
6736 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6739 /** Final setup of a sorted-dups cursor.
6740 * Sets up the fields that depend on the data from the main cursor.
6741 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6742 * @param[in] node The data containing the #MDB_db record for the
6743 * sorted-dup database.
6746 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6748 MDB_xcursor *mx = mc->mc_xcursor;
6750 if (node->mn_flags & F_SUBDATA) {
6751 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6752 mx->mx_cursor.mc_pg[0] = 0;
6753 mx->mx_cursor.mc_snum = 0;
6754 mx->mx_cursor.mc_top = 0;
6755 mx->mx_cursor.mc_flags = C_SUB;
6757 MDB_page *fp = NODEDATA(node);
6758 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6759 mx->mx_db.md_flags = 0;
6760 mx->mx_db.md_depth = 1;
6761 mx->mx_db.md_branch_pages = 0;
6762 mx->mx_db.md_leaf_pages = 1;
6763 mx->mx_db.md_overflow_pages = 0;
6764 mx->mx_db.md_entries = NUMKEYS(fp);
6765 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6766 mx->mx_cursor.mc_snum = 1;
6767 mx->mx_cursor.mc_top = 0;
6768 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6769 mx->mx_cursor.mc_pg[0] = fp;
6770 mx->mx_cursor.mc_ki[0] = 0;
6771 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6772 mx->mx_db.md_flags = MDB_DUPFIXED;
6773 mx->mx_db.md_pad = fp->mp_pad;
6774 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6775 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6778 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6779 mx->mx_db.md_root));
6780 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6781 #if UINT_MAX < SIZE_MAX
6782 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6783 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6787 /** Initialize a cursor for a given transaction and database. */
6789 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6792 mc->mc_backup = NULL;
6795 mc->mc_db = &txn->mt_dbs[dbi];
6796 mc->mc_dbx = &txn->mt_dbxs[dbi];
6797 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6802 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6803 mdb_tassert(txn, mx != NULL);
6804 mc->mc_xcursor = mx;
6805 mdb_xcursor_init0(mc);
6807 mc->mc_xcursor = NULL;
6809 if (*mc->mc_dbflag & DB_STALE) {
6810 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6815 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6818 size_t size = sizeof(MDB_cursor);
6820 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6823 if (txn->mt_flags & MDB_TXN_ERROR)
6826 /* Allow read access to the freelist */
6827 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6830 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6831 size += sizeof(MDB_xcursor);
6833 if ((mc = malloc(size)) != NULL) {
6834 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6835 if (txn->mt_cursors) {
6836 mc->mc_next = txn->mt_cursors[dbi];
6837 txn->mt_cursors[dbi] = mc;
6838 mc->mc_flags |= C_UNTRACK;
6850 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6852 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6855 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6858 if (txn->mt_flags & MDB_TXN_ERROR)
6861 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6865 /* Return the count of duplicate data items for the current key */
6867 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6871 if (mc == NULL || countp == NULL)
6874 if (mc->mc_xcursor == NULL)
6875 return MDB_INCOMPATIBLE;
6877 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6880 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6881 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6884 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6887 *countp = mc->mc_xcursor->mx_db.md_entries;
6893 mdb_cursor_close(MDB_cursor *mc)
6895 if (mc && !mc->mc_backup) {
6896 /* remove from txn, if tracked */
6897 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6898 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6899 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6901 *prev = mc->mc_next;
6908 mdb_cursor_txn(MDB_cursor *mc)
6910 if (!mc) return NULL;
6915 mdb_cursor_dbi(MDB_cursor *mc)
6920 /** Replace the key for a branch node with a new key.
6921 * @param[in] mc Cursor pointing to the node to operate on.
6922 * @param[in] key The new key to use.
6923 * @return 0 on success, non-zero on failure.
6926 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6932 int delta, ksize, oksize;
6933 indx_t ptr, i, numkeys, indx;
6936 indx = mc->mc_ki[mc->mc_top];
6937 mp = mc->mc_pg[mc->mc_top];
6938 node = NODEPTR(mp, indx);
6939 ptr = mp->mp_ptrs[indx];
6943 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6944 k2.mv_data = NODEKEY(node);
6945 k2.mv_size = node->mn_ksize;
6946 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6948 mdb_dkey(&k2, kbuf2),
6954 /* Sizes must be 2-byte aligned. */
6955 ksize = EVEN(key->mv_size);
6956 oksize = EVEN(node->mn_ksize);
6957 delta = ksize - oksize;
6959 /* Shift node contents if EVEN(key length) changed. */
6961 if (delta > 0 && SIZELEFT(mp) < delta) {
6963 /* not enough space left, do a delete and split */
6964 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6965 pgno = NODEPGNO(node);
6966 mdb_node_del(mc, 0);
6967 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6970 numkeys = NUMKEYS(mp);
6971 for (i = 0; i < numkeys; i++) {
6972 if (mp->mp_ptrs[i] <= ptr)
6973 mp->mp_ptrs[i] -= delta;
6976 base = (char *)mp + mp->mp_upper;
6977 len = ptr - mp->mp_upper + NODESIZE;
6978 memmove(base - delta, base, len);
6979 mp->mp_upper -= delta;
6981 node = NODEPTR(mp, indx);
6984 /* But even if no shift was needed, update ksize */
6985 if (node->mn_ksize != key->mv_size)
6986 node->mn_ksize = key->mv_size;
6989 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6995 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6997 /** Move a node from csrc to cdst.
7000 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7007 unsigned short flags;
7011 /* Mark src and dst as dirty. */
7012 if ((rc = mdb_page_touch(csrc)) ||
7013 (rc = mdb_page_touch(cdst)))
7016 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7017 key.mv_size = csrc->mc_db->md_pad;
7018 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7020 data.mv_data = NULL;
7024 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7025 mdb_cassert(csrc, !((size_t)srcnode & 1));
7026 srcpg = NODEPGNO(srcnode);
7027 flags = srcnode->mn_flags;
7028 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7029 unsigned int snum = csrc->mc_snum;
7031 /* must find the lowest key below src */
7032 rc = mdb_page_search_lowest(csrc);
7035 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7036 key.mv_size = csrc->mc_db->md_pad;
7037 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7039 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7040 key.mv_size = NODEKSZ(s2);
7041 key.mv_data = NODEKEY(s2);
7043 csrc->mc_snum = snum--;
7044 csrc->mc_top = snum;
7046 key.mv_size = NODEKSZ(srcnode);
7047 key.mv_data = NODEKEY(srcnode);
7049 data.mv_size = NODEDSZ(srcnode);
7050 data.mv_data = NODEDATA(srcnode);
7052 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7053 unsigned int snum = cdst->mc_snum;
7056 /* must find the lowest key below dst */
7057 rc = mdb_page_search_lowest(cdst);
7060 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
7061 bkey.mv_size = cdst->mc_db->md_pad;
7062 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
7064 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7065 bkey.mv_size = NODEKSZ(s2);
7066 bkey.mv_data = NODEKEY(s2);
7068 cdst->mc_snum = snum--;
7069 cdst->mc_top = snum;
7070 mdb_cursor_copy(cdst, &mn);
7072 rc = mdb_update_key(&mn, &bkey);
7077 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7078 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7079 csrc->mc_ki[csrc->mc_top],
7081 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7082 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7084 /* Add the node to the destination page.
7086 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7087 if (rc != MDB_SUCCESS)
7090 /* Delete the node from the source page.
7092 mdb_node_del(csrc, key.mv_size);
7095 /* Adjust other cursors pointing to mp */
7096 MDB_cursor *m2, *m3;
7097 MDB_dbi dbi = csrc->mc_dbi;
7098 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7100 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7101 if (csrc->mc_flags & C_SUB)
7102 m3 = &m2->mc_xcursor->mx_cursor;
7105 if (m3 == csrc) continue;
7106 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7107 csrc->mc_ki[csrc->mc_top]) {
7108 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7109 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7114 /* Update the parent separators.
7116 if (csrc->mc_ki[csrc->mc_top] == 0) {
7117 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7118 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7119 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7121 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7122 key.mv_size = NODEKSZ(srcnode);
7123 key.mv_data = NODEKEY(srcnode);
7125 DPRINTF(("update separator for source page %"Z"u to [%s]",
7126 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7127 mdb_cursor_copy(csrc, &mn);
7130 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7133 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7135 indx_t ix = csrc->mc_ki[csrc->mc_top];
7136 nullkey.mv_size = 0;
7137 csrc->mc_ki[csrc->mc_top] = 0;
7138 rc = mdb_update_key(csrc, &nullkey);
7139 csrc->mc_ki[csrc->mc_top] = ix;
7140 mdb_cassert(csrc, rc == MDB_SUCCESS);
7144 if (cdst->mc_ki[cdst->mc_top] == 0) {
7145 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7146 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7147 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7149 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7150 key.mv_size = NODEKSZ(srcnode);
7151 key.mv_data = NODEKEY(srcnode);
7153 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7154 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7155 mdb_cursor_copy(cdst, &mn);
7158 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7161 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7163 indx_t ix = cdst->mc_ki[cdst->mc_top];
7164 nullkey.mv_size = 0;
7165 cdst->mc_ki[cdst->mc_top] = 0;
7166 rc = mdb_update_key(cdst, &nullkey);
7167 cdst->mc_ki[cdst->mc_top] = ix;
7168 mdb_cassert(csrc, rc == MDB_SUCCESS);
7175 /** Merge one page into another.
7176 * The nodes from the page pointed to by \b csrc will
7177 * be copied to the page pointed to by \b cdst and then
7178 * the \b csrc page will be freed.
7179 * @param[in] csrc Cursor pointing to the source page.
7180 * @param[in] cdst Cursor pointing to the destination page.
7181 * @return 0 on success, non-zero on failure.
7184 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7192 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7193 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7195 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7196 mdb_cassert(csrc, cdst->mc_snum > 1);
7198 /* Mark dst as dirty. */
7199 if ((rc = mdb_page_touch(cdst)))
7202 /* Move all nodes from src to dst.
7204 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7205 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7206 key.mv_size = csrc->mc_db->md_pad;
7207 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7208 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7209 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7210 if (rc != MDB_SUCCESS)
7212 key.mv_data = (char *)key.mv_data + key.mv_size;
7215 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7216 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7217 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7218 unsigned int snum = csrc->mc_snum;
7220 /* must find the lowest key below src */
7221 rc = mdb_page_search_lowest(csrc);
7224 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7225 key.mv_size = csrc->mc_db->md_pad;
7226 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7228 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7229 key.mv_size = NODEKSZ(s2);
7230 key.mv_data = NODEKEY(s2);
7232 csrc->mc_snum = snum--;
7233 csrc->mc_top = snum;
7235 key.mv_size = srcnode->mn_ksize;
7236 key.mv_data = NODEKEY(srcnode);
7239 data.mv_size = NODEDSZ(srcnode);
7240 data.mv_data = NODEDATA(srcnode);
7241 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7242 if (rc != MDB_SUCCESS)
7247 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7248 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7249 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7251 /* Unlink the src page from parent and add to free list.
7254 mdb_node_del(csrc, 0);
7255 if (csrc->mc_ki[csrc->mc_top] == 0) {
7257 rc = mdb_update_key(csrc, &key);
7265 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7266 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7269 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7270 csrc->mc_db->md_leaf_pages--;
7272 csrc->mc_db->md_branch_pages--;
7274 /* Adjust other cursors pointing to mp */
7275 MDB_cursor *m2, *m3;
7276 MDB_dbi dbi = csrc->mc_dbi;
7277 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7279 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7280 if (csrc->mc_flags & C_SUB)
7281 m3 = &m2->mc_xcursor->mx_cursor;
7284 if (m3 == csrc) continue;
7285 if (m3->mc_snum < csrc->mc_snum) continue;
7286 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7287 m3->mc_pg[csrc->mc_top] = mp;
7288 m3->mc_ki[csrc->mc_top] += nkeys;
7293 unsigned int snum = cdst->mc_snum;
7294 uint16_t depth = cdst->mc_db->md_depth;
7295 mdb_cursor_pop(cdst);
7296 rc = mdb_rebalance(cdst);
7297 /* Did the tree shrink? */
7298 if (depth > cdst->mc_db->md_depth)
7300 cdst->mc_snum = snum;
7301 cdst->mc_top = snum-1;
7306 /** Copy the contents of a cursor.
7307 * @param[in] csrc The cursor to copy from.
7308 * @param[out] cdst The cursor to copy to.
7311 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7315 cdst->mc_txn = csrc->mc_txn;
7316 cdst->mc_dbi = csrc->mc_dbi;
7317 cdst->mc_db = csrc->mc_db;
7318 cdst->mc_dbx = csrc->mc_dbx;
7319 cdst->mc_snum = csrc->mc_snum;
7320 cdst->mc_top = csrc->mc_top;
7321 cdst->mc_flags = csrc->mc_flags;
7323 for (i=0; i<csrc->mc_snum; i++) {
7324 cdst->mc_pg[i] = csrc->mc_pg[i];
7325 cdst->mc_ki[i] = csrc->mc_ki[i];
7329 /** Rebalance the tree after a delete operation.
7330 * @param[in] mc Cursor pointing to the page where rebalancing
7332 * @return 0 on success, non-zero on failure.
7335 mdb_rebalance(MDB_cursor *mc)
7339 unsigned int ptop, minkeys;
7343 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7344 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7345 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7346 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7347 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7349 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7350 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7351 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7352 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7356 if (mc->mc_snum < 2) {
7357 MDB_page *mp = mc->mc_pg[0];
7359 DPUTS("Can't rebalance a subpage, ignoring");
7362 if (NUMKEYS(mp) == 0) {
7363 DPUTS("tree is completely empty");
7364 mc->mc_db->md_root = P_INVALID;
7365 mc->mc_db->md_depth = 0;
7366 mc->mc_db->md_leaf_pages = 0;
7367 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7370 /* Adjust cursors pointing to mp */
7373 mc->mc_flags &= ~C_INITIALIZED;
7375 MDB_cursor *m2, *m3;
7376 MDB_dbi dbi = mc->mc_dbi;
7378 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7379 if (mc->mc_flags & C_SUB)
7380 m3 = &m2->mc_xcursor->mx_cursor;
7383 if (m3->mc_snum < mc->mc_snum) continue;
7384 if (m3->mc_pg[0] == mp) {
7387 m3->mc_flags &= ~C_INITIALIZED;
7391 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7393 DPUTS("collapsing root page!");
7394 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7397 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7398 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7401 mc->mc_db->md_depth--;
7402 mc->mc_db->md_branch_pages--;
7403 mc->mc_ki[0] = mc->mc_ki[1];
7404 for (i = 1; i<mc->mc_db->md_depth; i++) {
7405 mc->mc_pg[i] = mc->mc_pg[i+1];
7406 mc->mc_ki[i] = mc->mc_ki[i+1];
7409 /* Adjust other cursors pointing to mp */
7410 MDB_cursor *m2, *m3;
7411 MDB_dbi dbi = mc->mc_dbi;
7413 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7414 if (mc->mc_flags & C_SUB)
7415 m3 = &m2->mc_xcursor->mx_cursor;
7418 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7419 if (m3->mc_pg[0] == mp) {
7422 for (i=0; i<m3->mc_snum; i++) {
7423 m3->mc_pg[i] = m3->mc_pg[i+1];
7424 m3->mc_ki[i] = m3->mc_ki[i+1];
7430 DPUTS("root page doesn't need rebalancing");
7434 /* The parent (branch page) must have at least 2 pointers,
7435 * otherwise the tree is invalid.
7437 ptop = mc->mc_top-1;
7438 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7440 /* Leaf page fill factor is below the threshold.
7441 * Try to move keys from left or right neighbor, or
7442 * merge with a neighbor page.
7447 mdb_cursor_copy(mc, &mn);
7448 mn.mc_xcursor = NULL;
7450 oldki = mc->mc_ki[mc->mc_top];
7451 if (mc->mc_ki[ptop] == 0) {
7452 /* We're the leftmost leaf in our parent.
7454 DPUTS("reading right neighbor");
7456 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7457 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7460 mn.mc_ki[mn.mc_top] = 0;
7461 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7463 /* There is at least one neighbor to the left.
7465 DPUTS("reading left neighbor");
7467 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7468 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7471 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7472 mc->mc_ki[mc->mc_top] = 0;
7475 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7476 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7477 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7479 /* If the neighbor page is above threshold and has enough keys,
7480 * move one key from it. Otherwise we should try to merge them.
7481 * (A branch page must never have less than 2 keys.)
7483 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7484 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7485 rc = mdb_node_move(&mn, mc);
7486 if (mc->mc_ki[ptop]) {
7490 if (mc->mc_ki[ptop] == 0) {
7491 rc = mdb_page_merge(&mn, mc);
7493 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7494 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7495 rc = mdb_page_merge(mc, &mn);
7496 mdb_cursor_copy(&mn, mc);
7498 mc->mc_flags &= ~C_EOF;
7500 mc->mc_ki[mc->mc_top] = oldki;
7504 /** Complete a delete operation started by #mdb_cursor_del(). */
7506 mdb_cursor_del0(MDB_cursor *mc)
7513 ki = mc->mc_ki[mc->mc_top];
7514 mdb_node_del(mc, mc->mc_db->md_pad);
7515 mc->mc_db->md_entries--;
7516 rc = mdb_rebalance(mc);
7518 if (rc == MDB_SUCCESS) {
7519 MDB_cursor *m2, *m3;
7520 MDB_dbi dbi = mc->mc_dbi;
7522 mp = mc->mc_pg[mc->mc_top];
7523 nkeys = NUMKEYS(mp);
7525 /* if mc points past last node in page, find next sibling */
7526 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7527 rc = mdb_cursor_sibling(mc, 1);
7528 if (rc == MDB_NOTFOUND)
7532 /* Adjust other cursors pointing to mp */
7533 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7534 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7535 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7537 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7539 if (m3->mc_pg[mc->mc_top] == mp) {
7540 if (m3->mc_ki[mc->mc_top] >= ki) {
7541 m3->mc_flags |= C_DEL;
7542 if (m3->mc_ki[mc->mc_top] > ki)
7543 m3->mc_ki[mc->mc_top]--;
7545 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7546 rc = mdb_cursor_sibling(m3, 1);
7547 if (rc == MDB_NOTFOUND)
7552 mc->mc_flags |= C_DEL;
7556 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7561 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7562 MDB_val *key, MDB_val *data)
7564 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7567 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7568 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7570 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7571 /* must ignore any data */
7575 return mdb_del0(txn, dbi, key, data, 0);
7579 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7580 MDB_val *key, MDB_val *data, unsigned flags)
7585 MDB_val rdata, *xdata;
7589 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7591 mdb_cursor_init(&mc, txn, dbi, &mx);
7600 flags |= MDB_NODUPDATA;
7602 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7604 /* let mdb_page_split know about this cursor if needed:
7605 * delete will trigger a rebalance; if it needs to move
7606 * a node from one page to another, it will have to
7607 * update the parent's separator key(s). If the new sepkey
7608 * is larger than the current one, the parent page may
7609 * run out of space, triggering a split. We need this
7610 * cursor to be consistent until the end of the rebalance.
7612 mc.mc_flags |= C_UNTRACK;
7613 mc.mc_next = txn->mt_cursors[dbi];
7614 txn->mt_cursors[dbi] = &mc;
7615 rc = mdb_cursor_del(&mc, flags);
7616 txn->mt_cursors[dbi] = mc.mc_next;
7621 /** Split a page and insert a new node.
7622 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7623 * The cursor will be updated to point to the actual page and index where
7624 * the node got inserted after the split.
7625 * @param[in] newkey The key for the newly inserted node.
7626 * @param[in] newdata The data for the newly inserted node.
7627 * @param[in] newpgno The page number, if the new node is a branch node.
7628 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7629 * @return 0 on success, non-zero on failure.
7632 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7633 unsigned int nflags)
7636 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7639 int i, j, split_indx, nkeys, pmax;
7640 MDB_env *env = mc->mc_txn->mt_env;
7642 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7643 MDB_page *copy = NULL;
7644 MDB_page *mp, *rp, *pp;
7649 mp = mc->mc_pg[mc->mc_top];
7650 newindx = mc->mc_ki[mc->mc_top];
7651 nkeys = NUMKEYS(mp);
7653 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7654 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7655 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7657 /* Create a right sibling. */
7658 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7660 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7662 if (mc->mc_snum < 2) {
7663 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7665 /* shift current top to make room for new parent */
7666 mc->mc_pg[1] = mc->mc_pg[0];
7667 mc->mc_ki[1] = mc->mc_ki[0];
7670 mc->mc_db->md_root = pp->mp_pgno;
7671 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7672 mc->mc_db->md_depth++;
7675 /* Add left (implicit) pointer. */
7676 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7677 /* undo the pre-push */
7678 mc->mc_pg[0] = mc->mc_pg[1];
7679 mc->mc_ki[0] = mc->mc_ki[1];
7680 mc->mc_db->md_root = mp->mp_pgno;
7681 mc->mc_db->md_depth--;
7688 ptop = mc->mc_top-1;
7689 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7692 mc->mc_flags |= C_SPLITTING;
7693 mdb_cursor_copy(mc, &mn);
7694 mn.mc_pg[mn.mc_top] = rp;
7695 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7697 if (nflags & MDB_APPEND) {
7698 mn.mc_ki[mn.mc_top] = 0;
7700 split_indx = newindx;
7704 split_indx = (nkeys+1) / 2;
7709 unsigned int lsize, rsize, ksize;
7710 /* Move half of the keys to the right sibling */
7711 x = mc->mc_ki[mc->mc_top] - split_indx;
7712 ksize = mc->mc_db->md_pad;
7713 split = LEAF2KEY(mp, split_indx, ksize);
7714 rsize = (nkeys - split_indx) * ksize;
7715 lsize = (nkeys - split_indx) * sizeof(indx_t);
7716 mp->mp_lower -= lsize;
7717 rp->mp_lower += lsize;
7718 mp->mp_upper += rsize - lsize;
7719 rp->mp_upper -= rsize - lsize;
7720 sepkey.mv_size = ksize;
7721 if (newindx == split_indx) {
7722 sepkey.mv_data = newkey->mv_data;
7724 sepkey.mv_data = split;
7727 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7728 memcpy(rp->mp_ptrs, split, rsize);
7729 sepkey.mv_data = rp->mp_ptrs;
7730 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7731 memcpy(ins, newkey->mv_data, ksize);
7732 mp->mp_lower += sizeof(indx_t);
7733 mp->mp_upper -= ksize - sizeof(indx_t);
7736 memcpy(rp->mp_ptrs, split, x * ksize);
7737 ins = LEAF2KEY(rp, x, ksize);
7738 memcpy(ins, newkey->mv_data, ksize);
7739 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7740 rp->mp_lower += sizeof(indx_t);
7741 rp->mp_upper -= ksize - sizeof(indx_t);
7742 mc->mc_ki[mc->mc_top] = x;
7743 mc->mc_pg[mc->mc_top] = rp;
7746 int psize, nsize, k;
7747 /* Maximum free space in an empty page */
7748 pmax = env->me_psize - PAGEHDRSZ;
7750 nsize = mdb_leaf_size(env, newkey, newdata);
7752 nsize = mdb_branch_size(env, newkey);
7753 nsize = EVEN(nsize);
7755 /* grab a page to hold a temporary copy */
7756 copy = mdb_page_malloc(mc->mc_txn, 1);
7761 copy->mp_pgno = mp->mp_pgno;
7762 copy->mp_flags = mp->mp_flags;
7763 copy->mp_lower = PAGEHDRSZ;
7764 copy->mp_upper = env->me_psize;
7766 /* prepare to insert */
7767 for (i=0, j=0; i<nkeys; i++) {
7769 copy->mp_ptrs[j++] = 0;
7771 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7774 /* When items are relatively large the split point needs
7775 * to be checked, because being off-by-one will make the
7776 * difference between success or failure in mdb_node_add.
7778 * It's also relevant if a page happens to be laid out
7779 * such that one half of its nodes are all "small" and
7780 * the other half of its nodes are "large." If the new
7781 * item is also "large" and falls on the half with
7782 * "large" nodes, it also may not fit.
7784 * As a final tweak, if the new item goes on the last
7785 * spot on the page (and thus, onto the new page), bias
7786 * the split so the new page is emptier than the old page.
7787 * This yields better packing during sequential inserts.
7789 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7790 /* Find split point */
7792 if (newindx <= split_indx || newindx >= nkeys) {
7794 k = newindx >= nkeys ? nkeys : split_indx+2;
7799 for (; i!=k; i+=j) {
7804 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7805 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7807 if (F_ISSET(node->mn_flags, F_BIGDATA))
7808 psize += sizeof(pgno_t);
7810 psize += NODEDSZ(node);
7812 psize = EVEN(psize);
7814 if (psize > pmax || i == k-j) {
7815 split_indx = i + (j<0);
7820 if (split_indx == newindx) {
7821 sepkey.mv_size = newkey->mv_size;
7822 sepkey.mv_data = newkey->mv_data;
7824 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7825 sepkey.mv_size = node->mn_ksize;
7826 sepkey.mv_data = NODEKEY(node);
7831 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7833 /* Copy separator key to the parent.
7835 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7839 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7844 if (mn.mc_snum == mc->mc_snum) {
7845 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7846 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7847 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7848 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7853 /* Right page might now have changed parent.
7854 * Check if left page also changed parent.
7856 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7857 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7858 for (i=0; i<ptop; i++) {
7859 mc->mc_pg[i] = mn.mc_pg[i];
7860 mc->mc_ki[i] = mn.mc_ki[i];
7862 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7863 if (mn.mc_ki[ptop]) {
7864 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7866 /* find right page's left sibling */
7867 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7868 mdb_cursor_sibling(mc, 0);
7873 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7876 mc->mc_flags ^= C_SPLITTING;
7877 if (rc != MDB_SUCCESS) {
7880 if (nflags & MDB_APPEND) {
7881 mc->mc_pg[mc->mc_top] = rp;
7882 mc->mc_ki[mc->mc_top] = 0;
7883 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7886 for (i=0; i<mc->mc_top; i++)
7887 mc->mc_ki[i] = mn.mc_ki[i];
7888 } else if (!IS_LEAF2(mp)) {
7890 mc->mc_pg[mc->mc_top] = rp;
7895 rkey.mv_data = newkey->mv_data;
7896 rkey.mv_size = newkey->mv_size;
7902 /* Update index for the new key. */
7903 mc->mc_ki[mc->mc_top] = j;
7905 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7906 rkey.mv_data = NODEKEY(node);
7907 rkey.mv_size = node->mn_ksize;
7909 xdata.mv_data = NODEDATA(node);
7910 xdata.mv_size = NODEDSZ(node);
7913 pgno = NODEPGNO(node);
7914 flags = node->mn_flags;
7917 if (!IS_LEAF(mp) && j == 0) {
7918 /* First branch index doesn't need key data. */
7922 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7928 mc->mc_pg[mc->mc_top] = copy;
7933 } while (i != split_indx);
7935 nkeys = NUMKEYS(copy);
7936 for (i=0; i<nkeys; i++)
7937 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7938 mp->mp_lower = copy->mp_lower;
7939 mp->mp_upper = copy->mp_upper;
7940 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7941 env->me_psize - copy->mp_upper);
7943 /* reset back to original page */
7944 if (newindx < split_indx) {
7945 mc->mc_pg[mc->mc_top] = mp;
7946 if (nflags & MDB_RESERVE) {
7947 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7948 if (!(node->mn_flags & F_BIGDATA))
7949 newdata->mv_data = NODEDATA(node);
7952 mc->mc_pg[mc->mc_top] = rp;
7954 /* Make sure mc_ki is still valid.
7956 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7957 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7958 for (i=0; i<=ptop; i++) {
7959 mc->mc_pg[i] = mn.mc_pg[i];
7960 mc->mc_ki[i] = mn.mc_ki[i];
7967 /* Adjust other cursors pointing to mp */
7968 MDB_cursor *m2, *m3;
7969 MDB_dbi dbi = mc->mc_dbi;
7970 int fixup = NUMKEYS(mp);
7972 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7973 if (mc->mc_flags & C_SUB)
7974 m3 = &m2->mc_xcursor->mx_cursor;
7979 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7981 if (m3->mc_flags & C_SPLITTING)
7986 for (k=m3->mc_top; k>=0; k--) {
7987 m3->mc_ki[k+1] = m3->mc_ki[k];
7988 m3->mc_pg[k+1] = m3->mc_pg[k];
7990 if (m3->mc_ki[0] >= split_indx) {
7995 m3->mc_pg[0] = mc->mc_pg[0];
7999 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8000 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8001 m3->mc_ki[mc->mc_top]++;
8002 if (m3->mc_ki[mc->mc_top] >= fixup) {
8003 m3->mc_pg[mc->mc_top] = rp;
8004 m3->mc_ki[mc->mc_top] -= fixup;
8005 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8007 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8008 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8013 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8016 if (copy) /* tmp page */
8017 mdb_page_free(env, copy);
8019 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8024 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8025 MDB_val *key, MDB_val *data, unsigned int flags)
8030 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8033 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8036 mdb_cursor_init(&mc, txn, dbi, &mx);
8037 return mdb_cursor_put(&mc, key, data, flags);
8041 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8043 if ((flag & CHANGEABLE) != flag)
8046 env->me_flags |= flag;
8048 env->me_flags &= ~flag;
8053 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8058 *arg = env->me_flags;
8063 mdb_env_set_userctx(MDB_env *env, void *ctx)
8067 env->me_userctx = ctx;
8072 mdb_env_get_userctx(MDB_env *env)
8074 return env ? env->me_userctx : NULL;
8078 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8083 env->me_assert_func = func;
8089 mdb_env_get_path(MDB_env *env, const char **arg)
8094 *arg = env->me_path;
8099 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8108 /** Common code for #mdb_stat() and #mdb_env_stat().
8109 * @param[in] env the environment to operate in.
8110 * @param[in] db the #MDB_db record containing the stats to return.
8111 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8112 * @return 0, this function always succeeds.
8115 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8117 arg->ms_psize = env->me_psize;
8118 arg->ms_depth = db->md_depth;
8119 arg->ms_branch_pages = db->md_branch_pages;
8120 arg->ms_leaf_pages = db->md_leaf_pages;
8121 arg->ms_overflow_pages = db->md_overflow_pages;
8122 arg->ms_entries = db->md_entries;
8127 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8131 if (env == NULL || arg == NULL)
8134 toggle = mdb_env_pick_meta(env);
8136 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8140 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8144 if (env == NULL || arg == NULL)
8147 toggle = mdb_env_pick_meta(env);
8148 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
8149 arg->me_mapsize = env->me_mapsize;
8150 arg->me_maxreaders = env->me_maxreaders;
8152 /* me_numreaders may be zero if this process never used any readers. Use
8153 * the shared numreader count if it exists.
8155 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8157 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8158 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8162 /** Set the default comparison functions for a database.
8163 * Called immediately after a database is opened to set the defaults.
8164 * The user can then override them with #mdb_set_compare() or
8165 * #mdb_set_dupsort().
8166 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8167 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8170 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8172 uint16_t f = txn->mt_dbs[dbi].md_flags;
8174 txn->mt_dbxs[dbi].md_cmp =
8175 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8176 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8178 txn->mt_dbxs[dbi].md_dcmp =
8179 !(f & MDB_DUPSORT) ? 0 :
8180 ((f & MDB_INTEGERDUP)
8181 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8182 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8185 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8190 int rc, dbflag, exact;
8191 unsigned int unused = 0;
8194 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8195 mdb_default_cmp(txn, FREE_DBI);
8198 if ((flags & VALID_FLAGS) != flags)
8200 if (txn->mt_flags & MDB_TXN_ERROR)
8206 if (flags & PERSISTENT_FLAGS) {
8207 uint16_t f2 = flags & PERSISTENT_FLAGS;
8208 /* make sure flag changes get committed */
8209 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8210 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8211 txn->mt_flags |= MDB_TXN_DIRTY;
8214 mdb_default_cmp(txn, MAIN_DBI);
8218 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8219 mdb_default_cmp(txn, MAIN_DBI);
8222 /* Is the DB already open? */
8224 for (i=2; i<txn->mt_numdbs; i++) {
8225 if (!txn->mt_dbxs[i].md_name.mv_size) {
8226 /* Remember this free slot */
8227 if (!unused) unused = i;
8230 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8231 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8237 /* If no free slot and max hit, fail */
8238 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8239 return MDB_DBS_FULL;
8241 /* Cannot mix named databases with some mainDB flags */
8242 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8243 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8245 /* Find the DB info */
8246 dbflag = DB_NEW|DB_VALID;
8249 key.mv_data = (void *)name;
8250 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8251 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8252 if (rc == MDB_SUCCESS) {
8253 /* make sure this is actually a DB */
8254 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8255 if (!(node->mn_flags & F_SUBDATA))
8256 return MDB_INCOMPATIBLE;
8257 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8258 /* Create if requested */
8260 data.mv_size = sizeof(MDB_db);
8261 data.mv_data = &dummy;
8262 memset(&dummy, 0, sizeof(dummy));
8263 dummy.md_root = P_INVALID;
8264 dummy.md_flags = flags & PERSISTENT_FLAGS;
8265 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8269 /* OK, got info, add to table */
8270 if (rc == MDB_SUCCESS) {
8271 unsigned int slot = unused ? unused : txn->mt_numdbs;
8272 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8273 txn->mt_dbxs[slot].md_name.mv_size = len;
8274 txn->mt_dbxs[slot].md_rel = NULL;
8275 txn->mt_dbflags[slot] = dbflag;
8276 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8278 mdb_default_cmp(txn, slot);
8287 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8289 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8292 if (txn->mt_flags & MDB_TXN_ERROR)
8295 if (txn->mt_dbflags[dbi] & DB_STALE) {
8298 /* Stale, must read the DB's root. cursor_init does it for us. */
8299 mdb_cursor_init(&mc, txn, dbi, &mx);
8301 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8304 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8307 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8309 ptr = env->me_dbxs[dbi].md_name.mv_data;
8310 env->me_dbxs[dbi].md_name.mv_data = NULL;
8311 env->me_dbxs[dbi].md_name.mv_size = 0;
8312 env->me_dbflags[dbi] = 0;
8316 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8318 /* We could return the flags for the FREE_DBI too but what's the point? */
8319 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8321 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8325 /** Add all the DB's pages to the free list.
8326 * @param[in] mc Cursor on the DB to free.
8327 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8328 * @return 0 on success, non-zero on failure.
8331 mdb_drop0(MDB_cursor *mc, int subs)
8335 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8336 if (rc == MDB_SUCCESS) {
8337 MDB_txn *txn = mc->mc_txn;
8342 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8343 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8346 mdb_cursor_copy(mc, &mx);
8347 while (mc->mc_snum > 0) {
8348 MDB_page *mp = mc->mc_pg[mc->mc_top];
8349 unsigned n = NUMKEYS(mp);
8351 for (i=0; i<n; i++) {
8352 ni = NODEPTR(mp, i);
8353 if (ni->mn_flags & F_BIGDATA) {
8356 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8357 rc = mdb_page_get(txn, pg, &omp, NULL);
8360 mdb_cassert(mc, IS_OVERFLOW(omp));
8361 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8365 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8366 mdb_xcursor_init1(mc, ni);
8367 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8373 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8375 for (i=0; i<n; i++) {
8377 ni = NODEPTR(mp, i);
8380 mdb_midl_xappend(txn->mt_free_pgs, pg);
8385 mc->mc_ki[mc->mc_top] = i;
8386 rc = mdb_cursor_sibling(mc, 1);
8388 if (rc != MDB_NOTFOUND)
8390 /* no more siblings, go back to beginning
8391 * of previous level.
8395 for (i=1; i<mc->mc_snum; i++) {
8397 mc->mc_pg[i] = mx.mc_pg[i];
8402 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8405 txn->mt_flags |= MDB_TXN_ERROR;
8406 } else if (rc == MDB_NOTFOUND) {
8412 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8414 MDB_cursor *mc, *m2;
8417 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8420 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8423 rc = mdb_cursor_open(txn, dbi, &mc);
8427 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8428 /* Invalidate the dropped DB's cursors */
8429 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8430 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8434 /* Can't delete the main DB */
8435 if (del && dbi > MAIN_DBI) {
8436 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
8438 txn->mt_dbflags[dbi] = DB_STALE;
8439 mdb_dbi_close(txn->mt_env, dbi);
8441 txn->mt_flags |= MDB_TXN_ERROR;
8444 /* reset the DB record, mark it dirty */
8445 txn->mt_dbflags[dbi] |= DB_DIRTY;
8446 txn->mt_dbs[dbi].md_depth = 0;
8447 txn->mt_dbs[dbi].md_branch_pages = 0;
8448 txn->mt_dbs[dbi].md_leaf_pages = 0;
8449 txn->mt_dbs[dbi].md_overflow_pages = 0;
8450 txn->mt_dbs[dbi].md_entries = 0;
8451 txn->mt_dbs[dbi].md_root = P_INVALID;
8453 txn->mt_flags |= MDB_TXN_DIRTY;
8456 mdb_cursor_close(mc);
8460 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8462 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8465 txn->mt_dbxs[dbi].md_cmp = cmp;
8469 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8471 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8474 txn->mt_dbxs[dbi].md_dcmp = cmp;
8478 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8480 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8483 txn->mt_dbxs[dbi].md_rel = rel;
8487 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8489 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8492 txn->mt_dbxs[dbi].md_relctx = ctx;
8496 int mdb_env_get_maxkeysize(MDB_env *env)
8498 return ENV_MAXKEY(env);
8501 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8503 unsigned int i, rdrs;
8506 int rc = 0, first = 1;
8510 if (!env->me_txns) {
8511 return func("(no reader locks)\n", ctx);
8513 rdrs = env->me_txns->mti_numreaders;
8514 mr = env->me_txns->mti_readers;
8515 for (i=0; i<rdrs; i++) {
8517 txnid_t txnid = mr[i].mr_txnid;
8518 sprintf(buf, txnid == (txnid_t)-1 ?
8519 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
8520 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
8523 rc = func(" pid thread txnid\n", ctx);
8527 rc = func(buf, ctx);
8533 rc = func("(no active readers)\n", ctx);
8538 /** Insert pid into list if not already present.
8539 * return -1 if already present.
8541 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8543 /* binary search of pid in list */
8545 unsigned cursor = 1;
8547 unsigned n = ids[0];
8550 unsigned pivot = n >> 1;
8551 cursor = base + pivot + 1;
8552 val = pid - ids[cursor];
8557 } else if ( val > 0 ) {
8562 /* found, so it's a duplicate */
8571 for (n = ids[0]; n > cursor; n--)
8577 int mdb_reader_check(MDB_env *env, int *dead)
8579 unsigned int i, j, rdrs;
8581 MDB_PID_T *pids, pid;
8590 rdrs = env->me_txns->mti_numreaders;
8591 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8595 mr = env->me_txns->mti_readers;
8596 for (i=0; i<rdrs; i++) {
8597 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8599 if (mdb_pid_insert(pids, pid) == 0) {
8600 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8602 /* Recheck, a new process may have reused pid */
8603 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8604 for (j=i; j<rdrs; j++)
8605 if (mr[j].mr_pid == pid) {
8606 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8607 (unsigned) pid, mr[j].mr_txnid));
8612 UNLOCK_MUTEX_R(env);